Split (2)

train · 1.33k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

filename stringlengths 17 38 ⌀	gt_das stringlengths 41 7.52k ⌀	genshare_das stringlengths 10 46.3k ⌀	md_content stringlengths 0 359k	healing_error stringclasses 4 values
10.1371_journal.pcbi.1011775.pdf	Data Availability Statement: The software “ImportRisk-v1.0.0” to compute the import risk is available under the Zenodo repository https://doi. org/10.5281/zenodo.7852476.	The software 'ImportRisk-v1.0.0' to compute the import risk is available under the Zenodo repository https://doi. org/10.5281/zenodo.7852476 .	RESEARCH ARTICLE Inferring country-specific import risk of diseases from the world air transportation network Pascal P. KlamserID Clara Jongen1,2, Frank Schlosser1,2, Dirk BrockmannID 1,2, Adrian ZachariaeID 1,2, Benjamin F. MaierID 1,2,7* 1,2,3,4, Olga Baranov5,6, a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Klamser PP, Zachariae A, Maier BF, Baranov O, Jongen C, Schlosser F, et al. (2024) Inferring country-specific import risk of diseases from the world air transportation network. PLoS Comput Biol 20(1): e1011775. https://doi.org/ 10.1371/journal.pcbi.1011775 Editor: Yamir Moreno, University of Zaragoza: Universidad de Zaragoza, SPAIN Received: May 3, 2023 Accepted: December 21, 2023 Published: January 24, 2024 Peer Review History: PLOS recognizes the benefits of transparency in the peer review process; therefore, we enable the publication of all of the content of peer review and author responses alongside final, published articles. The editorial history of this article is available here: https://doi.org/10.1371/journal.pcbi.1011775 Copyright: © 2024 Klamser et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: The software “ImportRisk-v1.0.0” to compute the import risk is available under the Zenodo repository https://doi. org/10.5281/zenodo.7852476. 1 Department of Biology, Institute for Theoretical Biology, Humboldt-Universita¨t zu Berlin, Berlin, Germany, 2 Robert Koch Institute, Berlin, Germany, 3 DTU Compute, Technical University of Denmark, Kongens Lyngby, Denmark, 4 Copenhagen Center for Social Data Science, University of Copenhagen, Copenhagen, Denmark, 5 Division of Infectious Diseases and Tropical Medicine, University Hospital, LMU Munich, Munich, Germany, 6 German Center for Infection Research (DZIF), Partner Site Munich, Munich, Germany, 7 Center Synergy of Systems (SynoSys), Center for Interdisciplinary Digital Sciences, Technische Universita¨t Dresden, Dresden, Germany * [email protected] Abstract Disease propagation between countries strongly depends on their effective distance, a mea- sure derived from the world air transportation network (WAN). It reduces the complex spreading patterns of a pandemic to a wave-like propagation from the outbreak country, establishing a linear relationship to the arrival time of the unmitigated spread of a disease. However, in the early stages of an outbreak, what concerns decision-makers in countries is understanding the relative risk of active cases arriving in their country—essentially, the likeli- hood that an active case boarding an airplane at the outbreak location will reach them. While there are data-fitted models available to estimate these risks, accurate mechanistic, parameter-free models are still lacking. Therefore, we introduce the ‘import risk’ model in this study, which defines import probabilities using the effective-distance framework. The model assumes that airline passengers are distributed along the shortest path tree that starts at the outbreak’s origin. In combination with a random walk, we account for all possible paths, thus inferring predominant connecting flights. Our model outperforms other mobility models, such as the radiation and gravity model with varying distance types, and it improves further if additional geographic information is included. The import risk model’s precision increases for countries with stronger connections within the WAN, and it reveals a geo- graphic distance dependence that implies a pull- rather than a push-dynamic in the distribu- tion process. Author summary For the spread of a contagious disease, human mobility puts distant places in proximity and geographically closer targets may be effectively much further away. The worldwide flight network is crucial for long distance travels and the previously proposed ‘effective PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 1 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN Funding: B.F.M received funding through Grant CF20-0044, HOPE: How Democracies Cope with Covid-19, from the Carlsberg Foundation and was supported as an Add-On Fellow for Interdisciplinary Life Science by the Joachim Herz Stiftung. P.P.K, A.Z, F.S received funding through Grant D81870, COVID-19 Lockdown-Monitor, from Germany’s Federal Ministry of Health. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors declare no competing interests. distance’ translates this mobility into a distance measure that correlates with the disease arrival time. We use the effective distance to generate a bottom-up and thus parameter- free distribution process of passengers on the flight network, which takes into account all possible flight routes. This allows us to determine the import probability of a disease. Our ‘import risk’ model outperforms or matches established mobility models, some of which require calibration with scarce or costly data. In contrast, our approach relies on minimal flight network data, that is the number of planes between airports and their passenger capacities, but not on passenger data. Its bottom-up approach enables future studies on country-specific measures for controlling and containing infected passengers, a challenge with existing models. Thus, the ‘import risk’ model’s strength lies in its data simplicity, this relevance to pandemics, and parameter-free design. Introduction The recent decades have seen a considerable increase in mobility: The worldwide number of passenger cars in use increased by an average of about 4% each year between 2006 and 2015, reaching approximately 1 billion in 2015 [1]. This growth is comparable to the yearly increase in the number of sea containers shipped [2], and the global scheduled air passenger count also experienced an annual growth of about 6% between 2004 and 2019 [3] In essence, the world is becoming increasingly interconnected in terms of passenger mobility, both on a small scale (cars) and a large scale (air traffic), as well as in the import and export of goods. This height- ened connectivity facilitates the distribution of goods and people, as demonstrated by the dis- tribution of over 400 invasive species through agricultural imports, which is best predicted by the global trade network [4]. A prime example of unwanted side effects of well-connected regions is the potential for pandemics, accompanied by death, economic damage and the potential stigmatization of survivors, migrants and minorities [5–7]. Already the first plague pandemic that started AD 541 in the Nile Delta of Egypt spread in 8 years across the territories (Mediterranean, Northern Europe and Near East) of 2 affected empires because of the intense commerce in the Roman Empire [6]. Nowadays, the intensified exchange reduces the time until a pandemic reaches all parts of the world to months as for the 2009 H1N1 virus that spread from Mexico in 5 months to all continents [8, 9] or the recent COVID-19 pandemic whose variants spread within a few months across the globe [10–13]. The connection strength between world regions is only partly explained by their geographic proximity. Instead, due to historic geopolitical relations [14, 15] pandemics spread rather along an effective distance that is derived from the world air transportation network (WAN) [16–19], or, if applied on a smaller scale, also from other means of transportation [16, 20]. According to the effective distance, region B is closest to region A if the passenger flow from A to B is greater than to other destinations. An intriguing extension is the multipath effective dis- tance, which enhances the prediction of disease arrival times by considering all paths taken by a random walker on the WAN [17]. The effective distance is regularly used to analyze the impact of mobility on the spread of diseases, as for example for MERS [21], Ebola [22], Zika [23] and most recently COVID-19 [20, 24–26]. While it enables a qualitative estimation of dis- ease arrival times, its applicability is severely restricted when it comes to describing the impor- tation of infected passengers from a specific source to a target. However, these import events are highly relevant for political decision-makers and to enable modeling predictions. In this work, we describe these import events via the “import probability” p(B\|A), which is equivalent to the origin-destination (OD) matrix whose element TBA represents the number of PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 2 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN trips from A to B, with the difference that the probability is normalized by all trips starting in A, i.e. p(B\|A) = TBA/TA. There are mobility models that fit the OD matrix, requiring a reference OD matrix as seen in the gravity model [27–31]. Additionally, some models integrate OD matrix-fitted models on a smaller scale with the OD matrix of the global air transportation net- work, creating a multiscale mobility network to represent all modes of transportation [32, 33]. Note that the multiscale mobility model has been successfully employed to analyze past pan- demics [34–36]. Yet, it can be extremely difficult to obtain the OD matrix and most often it is estimated by small surveys [37] or alongside a census [38]. Even for the air transportation net- work derived from a booking system, the OD is only an approximation since passengers increasingly book directly at the airlines (in 2015 30% of all Lufthansa flights were booked directly which increased to 52% in 2018 [39]) and not via the big GDS (global distribution sys- tems) from which most OD-estimates are derived [40, 41]. This means that to exactly compute the air transportation OD matrix, bookings of all GDSs and about 900 airlines must be pur- chased/estimated and combined. Thus, models that do not rely on an existing reference OD matrix are important and those either assume an underlying decision process without integrat- ing traffic information as the radiation model [42, 43] or they apply a maximum entropy approach to distribute the unknown OD trips along possible routes of a known traffic network [30, 44, 45]. However, none of the above approaches use the effective distance with its qualita- tive link to disease propagation and none is based on a mechanistic distribution process on a traffic network. To our understanding, a mechanistic process mimics the detailed movement behavior of the passengers on the traffic network, and neither uses only quantities of and between the locations (gravity and radiation model) nor relies on principles of system in ther- modynamic equilibrium (maximum entropy model), in other words it is a bottom-up approach. This approach grants us a mechanistic understanding of the observed patterns, enabling us to investigate how modifications impact passenger distribution. For instance, we can analyze how containment interventions along distribution routes reduce the import prob- ability of infected passengers. In this work, we introduce the import risk model, based on a distribution process following the shortest path tree of the WAN based on effective distance. This process is combined with a random walker that explores all potential paths within the WAN. We are using WAN data from the year 2014 and compare it to the Global Transnational Mobility Dataset from 2014 [40], as a ground truth baseline. Additionally, we investigate the discrepancy to the import risk and alternative mobility models as the gravity [27, 31] and radiation model [43] through multi- ple comparison measures. We find that the import risk model outperforms the alternative models and improves only slightly when it includes not only WAN information but also the geodetic distance between airports. Lastly, we evaluate the quality of import probability estima- tion for specific countries and assess if and how the geodesic distance is encoded in the import risk estimate. Results Relating the WAN, OD-probability and the effective distance In this work, we introduce the import risk, which estimates the probability of a passenger departing from airport A to conclude their journey at any airport worldwide, even those not directly connected to the origin airport. The estimation is based on the traffic flow of airplanes and the respective maximal passenger capacity between airports, a.k.a. the world air transpor- tation network (WAN), provided by the Official Airline Guide (OAG) [46]. This inference- problem is intriguing because it is much easier to monitor the origin and destination of air- planes, than of passengers with possibly multiple connecting flights until their final PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 3 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN destination. In our study, we use the WAN from 2014 (Fig 1A) and compare the derived import probabilities to a reference dataset. The reference import probability is based on the Global Transnational Mobility Dataset (GTN) from 2014 [40, 47], which combines an origin- final-destination dataset from a major global distribution system (GDS) with a tourism dataset from the World Tourism Organization (Fig 1B, see Material and methods for more details on the data). Before introducing the import risk model, we contrast the two datasets, introduce the effective distance [16] and quantify its potential as the base metric for our proposed model. By comparing the world air transportation network (WAN) with the country-specific refer- ence import probability from the GTN (compare Fig 1A and 1B), we see that the airports con- nected via direct links belong to countries that also have a high import probability. Nevertheless, due to physical constraints and logistical optimization, not all countries with non-zero import probabilities are directly connected to airports in the source country; instead, they are reached via connecting flights. In the context of import probability, estimates based on geodesic distance and the population of the target country are useful but exhibit limitations in certain scenarios. For instance, the import probability for Italy is approximately 1.4 times greater than that for Germany, even though Germany is geographically closer to Canada and Fig 1. The relation between WAN, OD-probability, SPT and effective distance. A: The world air transportation network (WAN) represents the direct flight connections and maximal seat capacities between airports in 2014, here shown for flights starting from five selected countries. It is based on flight-schedule-data. The lines are bundled and do not represent the specific flight route, but illustrate the links to airports abroad. B: The reference import probability from Canada to all countries, based on the OD matrix (Origin-Destination) of the Global Transnational Mobility Data set [40, 47] in 2014. It combines origin and final-destination trips between countries from the SABRE and the World Tourism Organization (UNWTO). The lines illustrate the connection to the common source country. C: Based on the effective distance deff = d0 − ln(p) a shortest path tree (SPT) is constructed with the largest Canadian airport as source (YYZ: Toronto Pearson International Airport). The link color and thickness shows the hop distance, i.e. number of connecting flights. D: exponential decay of the reference import probability (as in B but for all countries as source) with the effective distance deff (derived from the SPT (C) of the WAN (A)). Each dot represents a country-country link, the lines are medians including either all source countries or only from a specific continent. Maps are created with geopandas [48]. https://doi.org/10.1371/journal.pcbi.1011775.g001 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 4 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN has a larger population. The effective distance is an alternative network-based distance mea- sure that does not rely solely on direct connections and geographic information [16–19]. Instead, it is based on the passenger flow Fij from j to i and its relationship to the outflow Fj through the transition probability Pij = Fij/Fj. Together with a constant distance offset d0, the effective distance between directly connected airports is deffðij jÞ ¼ d0 (cid:0) lnðPijÞ : ð1Þ The effective distance between airports without direct connection is the cumulative distance along the shortest path tree (SPT) derived from deff, as illustrated for the largest Canadian air- port (Toronto Pearson Airport, YYZ) in Fig 1C. Note that a distance offset of d0 = 0 would make two routes indistinguishable as long as the product of the transition probabilities along each route is the same, but with d0 > 0 the one route with fewer connecting flights is effectively shorter. Previous studies have demonstrated that the arrival time of diseases in countries exhibits a linear dependence on their effective distance [16–19]. We show that the import probability also correlates with deff (Fig 1D), whereby the correlation is higher than for other distance measures (see Fig A in S1 Text). In fact, the import probability decays exponentially with effective distance (linear decay on a semi-log scale in Fig 1D) which can be reproduced in a simplified model for a passenger that travels at a constant effective speed and has a constant exit rate. Therefore, the effective distance seems to be a good representation of the underlying distribution process, and is a promising candidate for the base of our proposed import risk model, to directly estimate the import probability. Import risk model The idea behind the import risk model is a combination of two elements: (i) a random walk with an exit probability of the walker to finish its travel at the current node and (ii) a distribu- tion mechanism derived from the deff SPT (Fig 2). The use of a random walk is motivated by Iannelli et al. [17] who could improve the arrival-order prediction of deff by including all possi- ble paths. The exit probability enables us to combine the random walk with a distribution mechanism that assigns the likelihood of each node being the final destination, as explained in detail in the second step. In the first step, we use the transition network representation of the WAN and let a random walker start at source n0 and after each step it either exits at the current node i with exit probability qi or continues to walk. Let us define the walker’s probability to continue walking to node n given it was at node n − 1 before and originally started in n0 by Sn;n(cid:0) 1ðn0Þ ¼ Pn;n(cid:0) 1ð1 (cid:0) qn(cid:0) 1ðn0ÞÞ ; ð2Þ with Pn,n−1 as the transition probability from n − 1 to n. Now the probability to walk along a path Γ starting at n0 and exiting at n is the probability to continue walking Si,j along each link (i, j) that is part of the path times the exit probability of the final node pðGÞ ¼ qn Y ði;jÞ2G Si;j ; ð3Þ where we omitted the explicit dependence on the source n0. Our goal is to describe all possible paths the walker can take from n0 to n. We will use the matrix S, whose elements are the proba- bilities to continue walking Si,j. The element (i, j) of the product of the matrix with itself S � S = S2 sums over all paths of length l = 2 that end at i and start at j. Next, we can define the probability of a walker to exit at n after traversing all paths of length l as plðnjn0Þ ¼ qnðSlÞn;n0 : PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 ð4Þ 5 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN Fig 2. Import risk scheme. Starting from the transition network (left) the shortest path tree is computed based on the effective distance (center bottom). Based on the shortest path tree, the exit probabilities q� = q(�\|?) are computed. In the formula, the geometric symbols represent the estimated population of the respective node, which can also be distance-weighted (depending on the exact model). A random walk-process with exit probability is defined (top): at each step, the walker either exits the node with prob. q� = q(�\|?), or continues walking with prob. (1 − q�). The import risk p1(�\|?) (right) is the probability of a walker to exit at node � given it started at node ? under consideration of all possible paths. https://doi.org/10.1371/journal.pcbi.1011775.g002 Finally, the import risk is the probability to exit at n given all paths of all lengths p1ðnjn0Þ ¼ qn ! X1 Sl l¼1 n;n0 ¼ qnððI (cid:0) SÞ(cid:0) 1 (cid:0) IÞn;n0 ; ð5Þ where we used the convergence of the geometric series with identity matrix I. In the second step, we approximate the exit probability qi(n0) that we used above, but did not specify yet. Thereby, we assume that passengers start at source airport n0, travel along the SPT and exit at node i with an exit-probability qiðn0Þ ¼ NðiÞ NðiÞ þ NðOðijn0ÞÞ ð6Þ with N(i) as the population at airport i and O(i\|n) as the set of all offspring nodes downstream of i on the SPT centered at source n0. Hence, the exit probability at node i is determined by the ratio of the population at node i to the combined populations of all downstream nodes of i on the SPT, inclusive of node i. We estimate the population at airport i using its outflow on the WAN, denoted as N(i) = Fi. To aggregate the import probabilities at the country level, we sum the targets and apply a weighted average to the source airports, with population serving as the weighting factor. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 6 / 26 PLOS COMPUTATIONAL BIOLOGY Infer import risk from the WAN To elucidate how additional information about the geographic distance between nodes influences p1, we explore two variations of the import risk model: In the variation with “geo- desic distance weighted” exit probability the populations in Eq 6 are substituted with ^N ðijn0Þ ¼ NðiÞ=di;n0 is the geodesic distance between i and n0. To control for increasing model complexity, we study the “effective distance weighted” exit probability, where ^N ðijn0Þ ¼ NðiÞ=deffðijn0Þ, i.e. no geographic information is used, but the model struc- ture is equivalent. , where di;n0 Alternative models. Numerous alternative models estimate the OD-matrix, from which the import probability can be derived [30, 31, 42, 43, 49–52]. Among those, the gravity [27] and the intervening opportunity [42, 43] model are most widely used. A recent variant of the latter is the radiation model [43]. Although past studies have found that the gravity model out- performs the radiation model at small scale [38, 53, 54], especially the radiation model’s good performance at the large scale [38, 54] makes it an interesting model for mobility on the WAN. It was originally conceptualized for commuter flows [43] where the surrounding populations serve as a proxy for possible job opportunities. By estimating an airport’s population based on its outflow, we adjust the concept from job opportunities to tourism opportunities. Its deriva- tion from a mechanistic decision process makes it parameter free, and therefore similar and a good comparison to our model. However, it only requires information on the population den- sity and does not integrate flight data. We compare our model to the gravity model with an exponential and power-law distance dependence and the radiation model (see Material and methods for definitions). These models solely rely on the outflow data from the WAN to estimate the node’s population and the geo- graphic locations. To incorporate structural information of the WAN [55], the alternative mod- els are also implemented with the geodesic path distance (the geodesic distance along the SPT) and the effective distance, i.e. there are in total nine alternative models: the radiation model, the gravity model with exponential and with power-law distance decaying function, and each implemented with geodesic, geodesic path and effective distance. The exponents of the six grav- ity models are fitted to the reference import probability by assigning the best fitting exponent to each of the six comparison measures (Pearson correlation, root-mean-square error, common part of commuters, Kendalls rank correlation and the correlation and RMSE of the logarithmic measures, all defined in Material and methods) and taking their mean value (see Figs B and C in S1 Text). As comparison measures, we have chosen three measures that are related to the absolute error and three that are related to the relative error between estimate and reference. Symmetry by returning visitors. Each of the twelve models provides an estimate for the import probability p(i\|n0), which is used to compute the OD-matrix T through multiplication with the corresponding source population N(n0). By comparing the symmetry of T with the reference OD-matrix ^T, we find a much higher and qualitatively different symmetry in the ref- erence data (see Supplementary Note B, Fig D in S1 Text). The high symmetry is likely due to visitors (family, business, tourism, etc.) that dominate the international travel. They return to their home-location after a limited period [56] and only the minority of the travelers are migrants, i.e. stay permanently at the destination. Interestingly, the import risk model has the highest symmetry, but is still less symmetric than the reference data by a factor of 4. Therefore, before conducting a detailed comparison of the estimates, we rectify the import probability estimates by symmetrizing their OD-matrix (by extracting the symmetric part and recalculat- ing the import probability; for further details, refer to Material and methods and Supplemen- tary Note B in S1 Text). This correction can be seen as an alternative version of a doubly constrained model where normally the constraints on in- and out-flow are ensured by an itera- tive proportionate fitting [31]. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 7 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN Model comparison In the subsequent analysis, we evaluate the import probability estimates against the reference data through four approaches: (i) a direct comparison and assessment of their medians to identify potential systematic errors, (ii) the application of six distinct goodness-of-fit metrics to assess the individual model’s rank and relative performance, (iii) a classification task identi- fying countries with the highest import risk, particularly relevant in the context of a pandemic and (iv) a correlation study of the arrival time of 20 diseases and SARS-CoV-2 variants. Qualitative comparison. In Fig 3 the import probability estimate p(i\|n0) of each model is compared to the reference import probability ^pðijn0Þ. The gravity models exhibit the closest agreement with the reference data when the effective distance is employed, as indicated by the medians (Fig 3, first and second columns). In contrast, the median values of the radiation and import risk models are relatively stable and less influenced by variations in distance metrics or their associated weighting (third and fourth columns). All models overestimate the lowest median import probability (leftmost orange dot in Fig 3), since the estimated import probabil- ity is always nonzero, but a large proportion of the lowest reference import probabilities are zero due to the limited observation period and/or an insufficient number of departing passen- gers. The overestimation of the median import probability is observed up to p(i\|n0) � 10−4 for both the gravity and import risk models. However, this overestimation is notably absent in the case of the gravity model with an exponential distance decaying function and the effective Fig 3. Estimates of import probability by the gravity model with exponentially (1st column) and power law (2nd column) decaying distance function, the radiation model (3rd. column) and by the import risk model (4th column). The first three models (1st-3rd column) use as distance the geodesic (1st row), geodesic path (2nd row) and the effective (3rd row) distance. The import risk model is computed from the WAN with the geodesic distance (D) or the effective distance (L) as a weight for the exit probabilities or without weighting (H), i.e. in the last two cases (H, D) only WAN information is used. The orange line depicts the median and the gray line is y = x and illustrates perfect mapping. https://doi.org/10.1371/journal.pcbi.1011775.g003 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 8 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN distance metric (Fig 3I), where the median demonstrates the closest alignment with the refer- ence data. The radiation models (third column) systematically overestimates the highest import probabilities (p(i\|n0) >� 10−1) and consequently underestimates the lower import probabilities. Goodness of fit by multiple measures. We compared each model with the reference import probability via the Pearson correlation, the root-mean-square error (RMSE), and the common part of commuters. These measures are more sensitive to strong links, i.e. large import probabilities, which is important when the emphasis is placed on the countries that are most likely to import passengers. However, if the focus is to get a fair comparison including all links, logarithmic versions of the above measures or rank correlations are more appropriate. Thus, we also quantify the agreement by the correlation and the RMSE of the logarithm of the measures and by Kendall’s rank correlation. The three import risk model variations outper- form the other models in all but one measure, whereby the variation employing the geodesic Fig 4. Rank and relative performance of import risk estimation models. The different import probability models are compared via their rank (A) and relative performance (B), with the highest values representing the best approach. The rank and relative performance are shown for each (black dots) of the six comparison measures (corr, logcorr, RMSE, logRMSE, cpc, τKendall) the box illustrates the interquartile range, the horizontal line the median and the red triangle the mean. The colors of the boxes illustrate the different distance measures in use. The outlier measure of the import risk models (I.R.) is the logRMSE, where the gravity models with effective distance are performing best. See Material and methods for definitions of comparison measures and Figs E, F in S1 Text for absolute and detailed relative performance. https://doi.org/10.1371/journal.pcbi.1011775.g004 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 9 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN distance weighted exit probability performs best (Fig 4A). Following the import risk models, the two gravity models based on effective distance also exhibit strong rankings. In contrast, the remaining models lack consistent high rankings across all six measures and are more evenly distributed within the lower half. This model categorization also holds for the relative perfor- mance of the models (Fig 4B), with linear scaling of values in between (see Eq 22). In contrast to the rankings, the median relative performance shows a notable improvement when the gravity models incorporate effective distance. However, among the import risk models, the dif- ference in median relative performance remains marginal. The only measure where the import risk models are outperformed by the gravity models with effective distance is the logRMSE (Figs E, F in S1 Text). It is expected from the gravity models’ good agreement in median import probability with the reference data over wide ranges and the overestimation of low import probability by the import risk model. This overes- timation can be reduced by model-modifications that introduce parameters favoring the exit at nodes with large-populations (for details, see Supplementary Note C and Figs G, H in S1 Text). However, we refrain from adding complexity to the model, since its generic nature is its key aspect. Classification of ten top risk countries. In a pandemic context, it is of specific interest to identify the countries with the highest import probability. We analyzed how well the twelve proxy models can classify, if a country is among the ten countries with the highest import probability. Again, the import risk models outperform the other models and the one with geo- desic distance-weighted exit probabilities is the top predictor with a sensitivity of 71.1% (Fig 5D). All effective distance-based models have a high sensitivity (>� 65%), including the radia- tion model with 66.8% that had the lowest relative performance and second-lowest mean rank (Fig 5I–5K). For these high import probabilities, the import risk models now outperform the Fig 5. Classification of the 10 countries with the highest import probability by the gravity model with exponentially (1st column) and power law decaying (2nd column) distance function, the radiation model (3rd. column) and by the import risk model (4th column). A true or false positive (T. Pos. or F. Pos.) means that the country is or is not among the 10 countries with the highest reference import probability ^p. A false negative (F. Neg.) means that it belongs to the reference set but was not detected by the respective model. The pie chart illustrates the sensitivity of the models. https://doi.org/10.1371/journal.pcbi.1011775.g005 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 10 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN Fig 6. Correlation analysis: Disease arrival time vs. the effective model distance. Each model’s import probability is converted to an effective distance dM(i\|n0) = −ln(p(i\|n0)) with n0 as the outbreak country of the respective disease. The correlation results C(tA, dM) with the arrival time tA(i) of the disease in the target country i are grouped by model (A) and by the disease (B). As comparison distances, the correlation of the geodesic, geodesic path (on the effective shortest path tree) and the effective distance with tA are shown. Each dot represents a correlation result of the 21 considered outbreaks (H1N1 in 2009, COVID-19 in 2020 and the spread of 18 of its variants in the years 2020–2022). https://doi.org/10.1371/journal.pcbi.1011775.g006 other models also in terms of RMSE and logRMSE, i.e. the 10 countries at highest risk are not only classified best by the import risk model, but also quantitatively assessed best. Disease arrival time. In our final comparison, we evaluate the correlation between disease arrival times and the estimated import probability from the outbreak country of the disease. Note that the effective distance, which is the base of the import risk model, already has the clear relation to disease arrival times and the import risk model is developed to extend this qualitative relation to a quantitative number of passengers imported, as done in a recent study on the pandemic potential of SARS-CoV-2 variants [11]. However, a qualitative comparison to arrival time is of course possible via the negative logarithm of the import probability for each model, which we refer to as effective model distance, which linearly relates [16, 19] to the arrival time tA(i\|j) of a disease dMðijjÞ ¼ (cid:0) lnðpijÞ / tAðijjÞ ð7Þ with j as the disease outbreak country. The arrival time tA(i\|j) is the number of days between the disease outbreak and the day the first case is reported in the target country i. We evaluated the correlation C(tA, dM) for the H1N1 pandemic starting 2009 [8], the COVID-19 pandemic starting 2019 [57] and 18 of its variants. Additional to the import probability models, the cor- relations of the geodesic, geodesic path and effective distance with tA are included. Our analysis reveals that models employing the effective distance as the distance measure consistently out- perform those relying on the geodesic or geodesic path distance (Fig 6A). Interestingly, the gravity model with a power-law decaying distance function consistently performs well, regard- less of the specific distance measure employed. We do not observe a specific model that excels exclusively for certain diseases. Instead, we observe similar correlation values for the same dis- ease across models (Fig 6B), which suggests that there is considerable noise on the arrival time tA that varies between diseases. The noise could be related to the disease specific spreading speed: our assumption, that the outbreak country is the sole source, gets increasingly violated the slower the disease spreads, because other countries become secondary sources. A simple linear regression of the mean correlation hC(tA, dM)i and the mean arrival time htai supports this hypothesis (r = −0.44, p = 0.055, Fig K in S1 Text). PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 11 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN Import risk of countries and regions Having quantified the performance of the import risk model, we now focus on (i) country spe- cific differences in its prediction quality, (ii) possible limitations due to no concept of adminis- trative units (e.g. countries) whose airports are more interconnected and (iii) how the geodesic distance is encoded in the import risk model, i.e. how a distance dependence emerges from WAN information only. Country specific performance. In the import risk approach, we assume minimal knowl- edge of the system, i.e. only the WAN is known. Consequently, we differentiate countries only via their network properties, one of which is the degree of a node, or more precisely the node strength, since the WAN is a weighted network. It is the simplest metric that is also easily adjustable for the country-level perspective. At the country level, the node strength corre- sponds directly to the flow out of country C FC ¼ X X n2C m=2C Fmn : ð8Þ This country-specific characteristic signifies a country’s potential to influence the network’s structure, since flows from small-outflow countries are diluted by large-outflow countries. From an ecological point of view, the outflow is strongly correlated with the gross domestic product of a country (Fig N in S1 Text). The correlation (logcorr) between the logarithms of the import risk p1 and the reference import probability ^p1 improves with the outflow of the source country (Fig 7), as illustrated by Great Britain (GB) as the country with the largest out- flow in the WAN and Eritrea (ER) as one of the countries with the lowest outflow. The predic- tion improvement with the country’s outflow suggests that the WAN is dominated by large- outflow countries and therefore predictions worsen for countries with lower WAN outflow. However, the prediction improvement is also present in model alternatives that do not use WAN information at all (e.g. gravity with geodesic distance, Fig M in S1 Text). We rule the explanation out that the alternative models show this improvement due to preferential fitting of strong links—and therefore of large-outflow countries—since the models are fitted to the reference data by their import probabilities, which ensures equal weighting among countries. It rather suggests that the mobility behavior in low outflow regions is different, also supported by the sudden performance saturation for countries with a WAN outflow of FC >� 106 (Fig 7 and Fig M in S1 Text). Possibly, their passenger distribution is constrained by additional fac- tors and is limited to the regions in proximity. There are clear exceptions where the import risk estimation is worse compared to outbreak countries with a similar WAN outflow, as Australia (AU), Israel (IL) and Macao (MO). These countries are connected due to historical relations to specific regions that are either not in their direct neighborhood (European countries for AU and IL) or that are more important than the bare neighborhood would suggest, as Macao that is a special administrative region of China. For Macao the import risk to China is underestimated, which consequently overesti- mates the import to other countries, and for AU and IL Europe is underestimated which over- estimates other regions (Fig 7). AU, IL, and MO serve as examples illustrating that the WAN may not fully encapsulate all relevant information accessible to the import risk model. Another concept that is missing in our methodological approach is the idea of a country or another administrative unit. Instead, it treats airport pairs uniformly, disregarding their country affilia- tions. Since we know the international flights leaving a specific country from the WAN, we can run a self-consistency analysis, i.e. without the need of reference import probability data. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 12 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN Fig 7. Source countries’ prediction quality and WAN outflow. The correlation between the logarithm of the import risk and the reference import probability logcorr ¼ corrðlogðp1Þ; logð^pÞÞ improves with the outflow of the respective source country (top). Examples of source countries with particularly low (ER, Eritrea) and high (GB, Great Britain) outflow and log_corr are shown with their import risk and reference import risk to target countries (middle row). Countries with exceptionally low log_corr measures compared to source countries with a comparable outflow are either historically linked to specific regions as Australia (AU) and Israel (IL) to European countries (lower right panel) or politically as Macao (MO) as a special administrative region of China. https://doi.org/10.1371/journal.pcbi.1011775.g007 We can estimate the outflow leaving the country C by the import risk model by X X TC ¼ p1ðmjnÞNn : n2C m=2C ð9Þ If we compare it to FC the WAN flow out of country C (see Eq 8), it turns out that the import risk model systematically overestimates the flow out of a country (Fig I panel A in S1 Text). In fact, the relative error increases with the number of airports belonging to the country (Fig I panel B in S1 Text). Possible explanations for this overestimation include the absence of a country-specific concept within the import risk model and the unintentional inclusion of tran- sit passengers in the population count of airport catchment areas (since we use the outflow as a proxy for the population). However, we can easily correct for this overestimation on country- level analysis, by normalizing the airport population such that the WAN country outflow is recovered. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 13 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN Fig 8. Import risk aggregated on regional level “to target” vs. “from source” and its geodesic distance dependence. The geodesic distance between regions predicts the import risk p1 to a single target from all sources (A, B) better than from a single source to all targets (D, E) as can be seen by the p-values (C) of the power law fit p1(d) = c�d−α that is illustrated for each selected examples by a grey line (A, B, D, E). The fitted exponent α of the import risk to a single target decreases with the respective regional WAN flow out of the target region (F), i.e. the more connected a region, the weaker the import risk decays with distance. The dashed horizontal lines show the average import risk of a single target (A, B) or a single source (D, E). The color of the dots corresponds to the depicted world regions (right). Maps are created with geopandas [48]. https://doi.org/10.1371/journal.pcbi.1011775.g008 Geodesic distance dependence. The import risk model estimates import probabilities without explicit geodesic-distance information (excluding the variant with distance weighted exit probability). Since classical models have proven distance to be a good predictor for human mobility, we assume that it is encoded in the WAN structure and by consequence in the import risk estimate [58]. To enhance clarity, we aggregate the import risk data across twenty- two world regions. We observe that the import risks to individual targets decrease in a manner resembling a power-law as the geodesic distance to the sources increases (Fig 8A and 8B and Fig L in S1 Text). When we change our perspective and examine the distance-dependence from a single source to all target regions (Fig 8D and 8E), the observed dependence is less con- sistent with a power-law fit of the form p1 ¼ c � d(cid:0) a import risk is computed via a source-centric view (by computing the exit probability from the shortest path tree originating at each source), which suggests that the distance dependence should be best from one source to its possible targets. A possible explanation is that each target possesses its own attractiveness independent of the source region. This suggests that the distri- bution dynamics may resemble a pull mechanism rather than a push mechanism. Indeed, we find that the fitted exponent α from the power-law fit decreases as the WAN flow out of the target region increases, which can serve as a proxy for the attractiveness of a region (Fig 8F). In other words, the more attractive a region, the larger the import risks from more distant source regions. The fitted exponent c has a high rank correlation with α (τKendall = 0.89), i.e. also the coefficient is dependent on the attractiveness of the region. (Fig 8C). This is surprising, since the ij PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 14 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN Discussion and conclusion Motivated by the import probability’s strong dependence on the effective distance, we imple- mented the import risk model based on the effective distance shortest path tree’s exit probabil- ity in combination with a random walk on the WAN. As a result, we can infer the passenger trip distribution within the traffic network of their transport vehicle (WAN). When we com- pare our parameter-free model to variations of established mobility models, we observe that it surpasses the alternatives in most comparison measures. The only exception is where the two parameter-fitted gravity models with effective distance perform the best. The import risk model is the most accurate in determining countries with the highest import probability and is one of the models that correlate best with the time of arrival of 20 diseases, showcasing its importance for epidemic-related problems. However, it systematically overestimates low import probabilities and its performance worsens for countries with a passenger outflow below a million per year. Despite the lack of any explicit geodesic distance information, the import risk model recovers a geodesic distance dependence. This distinction is more promi- nent when considering all sources to a single target compared to the reverse scenario. We attri- bute this phenomenon to a target’s specific attractiveness, which we estimate using its node strength, i.e. the target’s passenger outflow. The only measure where the gravity models with effective distance outperform the import risk models is the logRMSE. This is likely due to their good agreement over wide ranges of the import probability (Fig 3I and 3J). The import risk model performs poorly with respect to the logRMSE due to its systematic overestimation of low import probabilities. Note, that the sec- ond parameter free model, the radiation model, systematically underestimates low import probabilities in the same way as the import risk model does. This is expected, since deviation from the assumptions cannot be corrected by any parameter adjustment. We identified several ways to reduce the import risk’s overestimation of low import probabilities by introducing an additional parameter that scales the population of the respective airport, changes the exit prob- ability along the shortest path tree or only the exit probability of specific nodes (for details, see Supplementary Note C and Figs G, H in S1 Text). In conclusion, we find that introducing modifications that enhance the probability of exiting at airports or nodes with large popula- tions mitigates the issue of overestimation. However, we leave this as a possible extension of our model and highlight that it outperformed the other models in all correlation measures, illustrating its high potential. The radiation model’s poor performance can likely be attributed to its initial design, which focused on small-scale commuter flows driven by work opportunities [43], which shows that bottom-up approaches are often limited to their specific use case but can be adapted, such as the extended radiation model [59], which is no longer parameter-free and has similar perfor- mance to the gravity model [54]. Interestingly, the radiation model is the only one that does not improve with inclusion of flight network information via the geodesic path or the effective distance (Fig 4). The radiation model’s insensitivity to network information can be attributed to the fact that it only extracts rank information from the distance data, resulting in a signifi- cant loss of information. The rank representation has the problem that airports that directly follow in their rank with respect to a source airport could be separated by a mountain range or ocean, i.e. the rank difference is minimal but the actual distance immense. This argument holds for any distance information. We corrected the import probability by the symmetrization of the respective OD-matrices which corresponds to a specific form of a doubly-constrained model. Normally, the constraints only ensure that the out- and inflow of each location corresponds to the observations [31, 52, 54], in contrast, we assume that both equal each other because of returning visitors. We PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 15 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN repeated the model comparison without the correction: it reduced the agreement with the ref- erence data for all but five of the seventy-two model-measure combinations (Fig F in S1 Text), which is in agreement with previous studies that report a better performance of doubly con- strained models [54]. Importantly, the import risk model still outperforms the other models if the import probability estimates are not corrected (compare Fig 4 with Fig J in S1 Text). It’s crucial to note that the assumption of returning visitors is applicable when visitors and tourists dominate while migrants can be disregarded. However, this assumption may not hold for links between low- and high-income countries or conflict regions. In the disease arrival time analysis, all models that use the effective distance perform simi- larly well, including all gravity models with power-law distance decay. The disease arrival time tA correlates with the logarithm of the estimated import probabilities, i.e. the results should be in agreement with the logcorr goodness of fit results. The models with effective distance vary only by maximal 0.07 in their logcorr measures and these are based on 183 countries as poten- tial source (Fig J in S1 Text). However, the 20 diseases in the arrival time analysis have only 10 unique outbreak countries. Additionally, due to factors like varying testing rates between countries, the uncertainty in arrival times, and other factors, the sample size is likely insuffi- cient to recover the logcorr results. In order to decrease the noise on tA, we repeated the analy- sis by extrapolating the arrival time via a logarithmic fit on the early cases, i.e. assuming an initial exponential growth (see Supplementary Note D in S1 Text). As a result of this proce- dure, some countries with insufficient data for extrapolation had to be excluded, which in turn led to the exclusion of more diseases. Nevertheless, the results are consistent with the tA esti- mation by 1st count (compare Fig 6 and Fig P in S1 Text). We found that without providing any geodesic distance information to the import risk model, a distance dependence is recovered that is stronger for import probabilities to a single target, than from a single source, even if the import probability is computed from a source- centric view. Since the WAN is spatially embedded and has a network dimension of three [58], its connections reflect up to a certain degree the characteristics of the embedding space. This explains the import risk model’s ability to capture distance dependence in general. That dis- tance is a better predictor in the target-centric view aligns well with a previous study in which a target-specific human-mobility model collapses mobility data to multiple targets by assigning each target a specific attractiveness that is proportional to the target’s population [51]. The import risk model predictions worsen for countries with a small outflow on the WAN, and since the country’s WAN outflow is proportional to its gross domestic product, the model performs less good for countries with a lower GDP, i.e. small population and/or low to middle income countries. This is unfortunate, as our model derives Origin-Destination (OD) infor- mation (costly to directly monitor) from cost-effective traffic flow monitoring, making it par- ticularly valuable for regions with limited resources. However, we find that the model alternatives (gravity, radiation) also perform poorly for low-outflow countries and that the pas- senger distribution of the latter is most likely constrained by the GDP and thus limited to the target-regions in effective proximity. To circumvent this problem, one could aggregate neigh- boring low-outflow countries until the conglomerate crosses the outflow threshold of FC = 106 above which we observe a performance saturation (Fig 7 and Fig M in S1 Text). Of course, this compromise comes with a lower spatial resolution and we emphasize the need for future research in this direction. While we have assessed the model’s performance on the world air transportation network, its applicability extends to other modes of transportation such as subway systems, cars, buses, and trains. Future research will explore the specific conditions under which this model can be effectively applied. Furthermore, there is room for improvement in the basic estimation of the traveling population within an airport’s catchment area based solely on its outflow. This PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 16 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN estimation does not currently account for the significant role of hubs and the missing informa- tion about transit passengers. The simple framework that only relies on the traffic network is appealing, but in certain scenarios its prediction can be refined by using information about the GDP, Gini-coefficient or population density. Our comparison focused on the parameter-free radiation model and the fitted gravity model, but we acknowledge the existence of promising variations and alternative models that were not included in this study [30, 31, 54, 59]. However, the gravity model is widely applied and has been shown to perform equally well [59] or better than alternatives [54]. There are exceptions, e.g. an iterative computation of a gravity-like model outperforms the common gravity model in cases where the complete mobility network is not available [29]. Additionally, the radiation model outperforms the gravity model for long-distance connections [38, 54]. Still, the simplicity of the gravity model and its adaptability by parameter adjustment make it a strong counterpart. The model alternatives make use of the WAN-structure information by using the effective distance as done in e.g. Ren et al. [60] where the radiation model with time- distance was better than the travel-distance on the road network to predict the traffic on each link. Similarly, we observed that the effective distance, which is related to the arrival time of diseases, outperforms geodesic path-distance in predicting import probabilities. The import risk model is fundamentally different from classic approaches that estimate OD trips from traffic data, because the latter find the OD trips that best reproduce the traffic data [28, 30, 44, 45], while our model runs a distribution process on the traffic data network. Thus, our model is a mechanistic bottom-up approach, while the classic approaches either fit and require the knowledge of the reference trip data [28, 30] or are based on the assumption that the trip distribution across the links follows the maximum entropy principle, i.e. the OD trips are considered as most likely that can be realized by the largest number of microstates [44, 45]. Note that maximum entropy approaches require an estimation of routes and their alternatives between each OD pair, while we allow all routes to be taken by the random walker. To the best of our knowledge, our model stands as unique in its mechanistic nature, enabling the study of modifications to its underlying distribution process. This includes strategies for containment aimed at slowing or restricting a pandemic, for instance. A straight forward implementation could be the testing of a fraction of passengers Ci � 1 at every transit airport i, which corre- sponds to reducing the probability to continue walking of an infected passenger (Eq 2) to ~Sn;n(cid:0) 1ðn0; CÞ ¼ ð1 (cid:0) Cn(cid:0) 1Þ � Pn;n(cid:0) 1ð1 (cid:0) qn(cid:0) 1ðn0ÞÞ : With C = [C1, C2, . . .] one could allow for a varying testing capacity between the airports. Material and methods Data sources The WAN provided by OAG (Official Airline Guide) [46] contains the number of flights and the respective maximum seat capacity Fi,j between airports i and j aggregated for the year 2014. The reference import probability ^pðmjnÞ ¼ ^T mn= ^T n is based on the “Global Transnational Mobility Dataset” [40, 47] that assigns the number of trips in 2014 ^T mn from country n to m worldwide by combining the world air transportation origin-final-destination data set from the company SABRE, and cross-boarder visits with an overnight stay from the UNWTO (World Tourism Organization). Thus, ^pðmjnÞ represents not only the mobility via air travel but also via other means (sea, road, rail). However, air travel dominates long distance trips which makes it a fair reference set of the air transportation origin-final-destination matrix. For details on how the data sets were combined, see Supplementary Note A in S1 Text. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 17 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN Alternative models The gravity model states that the number of trips between regions n and m increase with their population sizes (Nn and Nm) and decrease with distance dnm Tmn ¼ On Nn Nm f ðdnmÞ ; ð10Þ with f(d) as a function that grows monotonically with distance d, most often chosen as either a power-law f(d) = dγ or an exponential f(dnm) = eγd. In the radiation model, the trips from n to m depend on their respective population sizes Nn, Nm (or other measures as job opportunities) and on the number of people smn that are in a circle with radius rmn centered around location n including Nn and Nm: Tmn ¼ On Nn Nm ðsmn (cid:0) NmÞsmn : ð11Þ The import probability of both models is computed by normalizing the trips with respect to the source-region pðmjnÞ ¼ TmnP jTjn ¼ Tmn Tn : ð12Þ Trip-symmetrization We correct the import probability via symmetrizing the OD-matrix by (i) compute the esti- mated OD-matrix m;n ¼ pð0ÞðmjnÞNn Tð0Þ from the import probability estimate, (ii) correct it by computing its symmetric part S ¼ ðT þ T>Þ=2 and (iii) compute the corresponding corrected import probability via pð1ÞðAjBÞ ¼ SAB=SB : ð13Þ ð14Þ ð15Þ By going through these steps, the asymmetry is reduced heavily but still persists. Thus, we repeat steps (i) till (iii) until p(3)(A\|B), which returns for all models a comparable asymmetry in mean and median to the reference data (see Supplementary Note B in S1 Text for details). Comparison measures We compare the import probability models with the reference data via the Pearson correlation corrðx; yÞ ¼ E½ðx (cid:0) �xÞðy (cid:0) �yÞ� sxsy ; with E½x� � �x as average, the root-mean-square error RMSEðx; yÞ ¼ q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi E½ðx (cid:0) yÞ2� ; PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 ð16Þ ð17Þ 18 / 26 PLOS COMPUTATIONAL BIOLOGYthe common part of commuters [59] cpcðx; yÞ ¼ P 2 P ij minðxij; yijÞ P ijxij þ ijyij Infer import risk from the WAN ; ð18Þ which is 1 if all links are identical and 0 if none of them agrees. All the above measures are more sensitive to strong links, i.e. large import probabilities. However, if the focus is to get a fair comparison including all links, we are more interested in logarithmic versions of the above measures or rank correlations. Thus, we compare the logarithm of the import probabili- ties via correlation logcorrðx; yÞ ¼ corrðlogðxÞ; logðyÞÞ ; root-mean-square error logRMSEðx; yÞ ¼ RMSEðlogðxÞ; logðyÞÞ ; and use the Kendall rank correlation coefficient tKendall ¼ q C (cid:0) D ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðC þ D þ TxÞðC þ D þ TyÞ ; ð19Þ ð20Þ ð21Þ with C and D as the number of concordant and discordant pairs and Tx and Ty as ties only in x and y, respectively. To simplify and generalize the comparison we combine the six above defined measures by computing the mean rank of each model, i.e. the best correlating model has the highest (12) and the worst the lowest (0) rank and the mean rank of one model is the average of all six ranks. To quantify the mean difference between the models we define the relative performance of one model M as rel:perf:ðf ðxM; yÞÞ ¼ f ðxMÞ (cid:0) worstðf ðxkÞ; kÞ bestðf ðxkÞ; kÞ (cid:0) worstðf ðxkÞ; kÞ ; ð22Þ with f(xM) = f(xM, y) as the specific comparison function and best( f (xk), k) and worst( f (xk), k) as the best and worst performing value of all models using this comparison function. Note, that best(. . .) = max(. . .) apart for the rmse-measures, where it is min(. . .) (analog for worst (. . .)). Disease arrival times The disease arrival time tA(i) in country i is estimated by the date of the first reported case for H1N1 and SARS-CoV-2. For the SARS-CoV-2 variants we use the first sequenced sample in this country. However, for certain variants some sequenced samples appear in the statistics month before the outbreak date declared by the WHO [61], which we treat as misclassifica- tions, discard them and use instead the first sample after the WHO listed outbreak for the respective country (see Supplementary Note D for details and Fig O in S1 Text). For each of the diseases/variants we used the WAN that we have access to and that is closest to the respec- tive outbreak date (see Table B in S1 Text) and as outbreak country we used the one listed by the WHO as first country with first sequenced sample of the respective variant [61]. For the H1N1 outbreak in 2009 we used the case data provided by FluNet [62, 63] (the column AH1N12009), for the COVID-19 cases we use the WHO COVID-19 dashboard [64] accessed PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 19 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN through ourworldindata.org, the number of sequenced samples was accessed through GISAID [65–67] using the file gisaid_variants_statistics.json. Supporting information S1 Text. Supplementary Note A: Origin-destination data (“Global Transnational Mobility Dataset”). Supplementary Note B: Symmetrized flows. Supplementary Note C: On the overes- timation of low import probabilities. Supplementary Note D: Disease arrival time analysis. Table A. Filtering criteria for the log-cases fit to extrapolate the arrival times tA. A country is excluded if (C0:) the detection is too sparse before peak-0 (less than 6 weeks of data), (C1:) the number of cases at peak-0 is below 30 (otherwise the signal is too noisy), (C2:) the extrapolated arrival time is before the WHO-outbreak date. N is the number of countries for which case data could be generated. NC0 and NC1 are the countries that pass criteria C0 and C1. NC0 & C1 and NC0 & C1 & C2 are the numbers of countries that pass multiple criteria. Table B. Disease and SARS-CoV-2 Variant outbreak information and WAN date. For each disease/variant the outbreak country and the date of the WAN used to compute the import probability esti- mates with the different models is displayed. Note that we only have the WAN from the years 2014 and 2019 in a yearly resolution and from 2020–2022 in monthly resolution. We repeated the analysis for COVID with the WAN from the month 2020–01-01, instead of using the yearly WAN from 2019, which gave comparable results. Fig A. Import probability dependence on the geographic distance (A), the effective distance (B) and the geographic path distance (C). The orange line represents the median and C(x, y) is the correlation between the two measures either log-transformed or not. The geographic distance between countries is averaged over all airport pairs. The geographic path distance is the geographic distance along the shortest path derived from the WAN using deff, i.e. it is a combination of geographic and network informa- tion. The axis scale corresponds to the one with the highest correlation, i.e. log-log for distance and path distance (A, C) and y-log for the effective distance (B). Fig B. Gravity model scans. Parameter dependence of measures that compare the model estimated import probability with the reference import risk ^pðijn0Þ. Thereby is “corr” the correlation, “cpc” the common part of commuters, “log_corr” the correlation on log-scale, “rmse” the root mean squared error and “kendalltau” the rank correlation via Kendalls tau. Two versions of the gravity model are shown with an exponentially decaying distance function f(d) = e−γd (left column: A, C, E), and a power law decaying distance function f(d) = d−β (right column: B, D, F). As distance the geo- desic distance (first row: A, B), the geodesic path distance (second row: C, D) and the effective distance (third row: E, F) are used. The dotted horizontal lines show the comparison measure with the import risk as model and have the same respective color. Fig C. Mean optimal parameters for gravity models. For each gravity model with exponentially and power law decaying distance function and with one of the three different distance measures (geodesic dis- tance, geodesic path distance and effective distance), the exponent γ or β that results in the best fit to the reference import risk is shown. The comparison is quantified via the correlation (corr), correlation between the log-transformed import risks (log_corr), root mean square error (rmse), root mean square error of the log-transformed import risks (log_rmse), Kendall rank correlation (kendalltau) and the common part of commuters (cpc). The mean optimal parameter for each model is marked by a horizontal line and their values are γ = [6.71, 6.41] * 10−4 for geographic and geo. path distance and γ = 0.84 for the effective distance, and β = [1.90, 1, 95, 5.10] for geo., geo. path, and effective distance, respectively. Fig D. Symmetry check for OD-matrix. Each dot represents the number of passengers that travel between 2 countries and back. The OD-matrix is computed by the radiation model (1st. column), gravity model with exponentially (2nd column) and power law decaying (3rd column) distance PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 20 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN function and by the import risk model (4th column). The OD-matrix of the models is com- puted by multiplying the import probability with the source-outflow. The reference trips and return trips have the highest symmetry (5th column, M). The orange line depicts the median and the gray line is y = x and illustrates perfect symmetry. The mean (AVG(asym)) and median (MED(asym)) asymmetry of the flows, computed according to Eq. C in S1 Text., are shown in each panel. The reference trips (M) show the lowest asymmetry, especially for large passenger flows. Fig E. Relative comparison measures for the import probability estimates. The rank (A) and the relative performance (B) for the different import probability estimation models. The model that agrees best (worst) with the reference import risk according a specific measure has the highest (lowest) rank and a relative performance of one (zero). The relative perfor- mance is then a linear interpolation between the best and worst model. The comparison mea- sures are the correlation (corr), correlation between the log-transformed import risks (log_corr), Root-mean-square error (rmse), Root-mean-square error of the log-transformed import risks (log_rmse), Kendall rank correlation (kendalltau) and the common part of com- muters (cpc). As exponents of the gravity models the mean optimal parameter is used (hori- zontal lines in Fig C in S1 Text.). Fig F. Absolute comparison measures for the import probability estimates. The comparison measures are the correlation (corr), correlation between the log-transformed import risks (log_corr), Root-mean-square error (rmse), Root- mean-square error of the log-transformed import risks (log_rmse), Kendall rank correlation (kendalltau) and the common part of commuters (cpc). As exponents of the gravity models the mean optimal parameter is used (horizontal lines in Fig C in S1 Text.). The colors depict the 4 different models. The solid circles are the models with corrected import probability by symmetrizing their OD-matrix, and the transparent squares are the non-corrected import probabilities of the respective model. Fig G. Import risk comparison and its deviation from a linear relation. Scatter plot (left) and only median and IQR with an exponential fit (right). Fig H. Variations of the import risk model to investigate how additional parameters influ- ence the relation between the import risk and the reference import risk. A: the flow scaling exponent ν that estimates the travelling population N(i) of the airport i depending on its WAN outflow Fi via NðiÞ ¼ Fn i (default: ν = 1). B: the effective distance offset d0 that penalizes larger hop-distances in the effective distance deff(i\|n0) = d0 − ln(Pij) when creating the shortest path tree (default: d0 = 1). C: the descendant fraction introduced in the shortest path exit probabil- ity, where 0.5 is the default value and values larger than 0.5 mean that the exiting at the descen- dant (or offspring) nodes compared to the current node becomes more likely. D: different weight options introduced for the shortest path tree exit probability. Per default, the node pop- ulations are not weighted. The weight is the inverse of either the geodesic or the effective dis- tance. E: manually set shortest path exit probability of leaf nodes (dead-end nodes). Per default, the exit probability is 1. A decrease to 0.9 or 0.8 does not visually change the median. Fig I. Country outflow reconstruction by import risk. The flow in the WAN leaving a coun- try FC is estimated by the import risk model by TC = ∑n2C ∑m=2C p1(m\|n)Nn. Both measures are directly compared (A) and the relative error is computed depending on the number of airports in the respective country Narpts (B). The import risk model does not include the concept of a country which partly explains the overestimation for larger airports. Another explanation is the overestimation of the respective airport population Nn = Fn by the WAN outflow for the import risk model (the true population is smaller because of the transit passengers that need to be excluded). Note that the WAN is used here, i.e. we check for self-consistency of the model and no reference data is included. Fig J. Uncorrected models: rank and relative performance. Same analysis as in the main text in Fig 4), however, here the uncorrected model predictions are used, i.e. without symmetrizing the OD-matrix. Fig K. Mean correlation between arrival time and effective model distance vs. the speed of the disease estimated by the mean arrival PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 21 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN time htA(C)iC, averaged over all countries C. The correlation C(tA, dM) between arrival time tA and effective model distance dM is averaged over all models. The size of the datapoints illus- trates the number of countries that were reached by the disease. Fig L. Import risk between world regions to a specific target region. In contrast to its derivation the import risk is dis- played in a target-centric view, i.e. each panel displays the import probability to a single target region from all source regions. The distance between world regions is the mean distance between their airport locations. The grey line represents a power-law fit p1 = c � d−α. The mean import risk is marked for each world region by a horizontal dashed line. The 22 target- world-regions are sorted according to their mean import risk. Maps are created with geopan- das [48]. Fig M. Source countries prediction quality and WAN outflow for two gravity models. Same model-result representation as in Fig 7 but here instead of the import risk model, the gravity model with power-law distance decaying function using the geodesic dgeo (left) or effective deff (right) distance is applied. Also for these models the logcorr between import probability estimates p(i\|n0) and the reference data ^pðijn0Þ improves for countries with a larger outflow in the WAN. Fig N. WAN flow out of countries vs. population and GDP The WAN flow out of a country is best mapped by its gross domestic product (GDP, C) com- pared to its population (A) or per capita GDP (B). The linear double-logarithmic regression results are shown in the lower part of each panel (r- and p-value). The size of each country cor- responds to its population (A) and the color codes its continent. GDP is taken from the World Bank Dataset for the year 2014 [69]. Fig O. Variant outbreak detection and fraction of sequenced samples for each of the considered variants. To illustrate the spread of the variant and how often it occurs worlwide the fraction of the variant in all sequenced probes is plotted, i.e. if it reaches 1, all sequenced probes are the respective variant. The official WHO outbreak date [61] is highlighted as red dotted vertical line. We estimated an outbreak date by 45 days before the fraction of sequenced samples reached 2.5% of its world-wide peak. The orange ver- tical lines (lower row of lines) show for each country the arrival of the variant, estimated by the first sequenced probe (“count1”). The black vertical lines (upper row of lines) show the arrival times after the outbreak which are used in the main text. Fig P. Correlation analysis with log- cases estimated arrival time. Each model’s import probability is converted to an effective dis- tance dM(i\|n0) = −ln(p(i\|n0)) with n0 as the outbreak country of the respective disease. The cor- relation results C(tA, dM) with the arrival time tA(i) of the disease in the target country i are grouped by model (A) and by the disease (B). As comparison distances, the correlation of the geodesic, geodesic path (on the effective shortest path tree) and the effective distance with tA are shown. Each dot represents a correlation result of the 10 considered outbreaks (H1N1 in 2009, COVID-19 in 2020 and the spread of 8 of its variants in the years 2020–2022). For the analysis only those diseases/variants were used with more than 10 datapoints (see Table A in S1 Text.). Fig Q. New case numbers of the Alpha variant for countries that passed the selec- tion criteria for the log-cases fit to extrapolate the arrival time tA in the attempt to reduce noise. The vertical dashed line marks the outbreak as listed by the WHO [61], the yellow star is the extrapolated arrival time from the log-cases fit that is illustrated by a yellow line. To deter- mine the peak-0 (marked by a vertical line) we used a difference analysis on the smoothed new-cases data. Fig R. New case numbers of the Alpha variant for countries that failed the selection criteria for the log-cases fit to extrapolate the arrival time tA in the attempt to reduce noise. The vertical dashed line marks the outbreak as listed by the WHO [61]. Those countries that passed the criteria C0 and C1 (see Table A in S1 Text. for details) show the log-cases fit. Note that the latter have an extrapolated tA before the outbreak date listed by the WHO. To determine the peak-0 (marked by a vertical line) we used a difference analysis on the smoothed new-cases data. (PDF) PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 22 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN Acknowledgments We acknowledge Marc Wiedermann for insightful comments. Author Contributions Conceptualization: Pascal P. Klamser, Dirk Brockmann. Data curation: Pascal P. Klamser, Adrian Zachariae, Benjamin F. Maier, Olga Baranov, Clara Jongen, Frank Schlosser. Formal analysis: Pascal P. Klamser, Adrian Zachariae, Benjamin F. Maier. Funding acquisition: Dirk Brockmann. Methodology: Pascal P. Klamser, Benjamin F. Maier, Frank Schlosser, Dirk Brockmann. Software: Pascal P. Klamser, Adrian Zachariae, Benjamin F. Maier. Visualization: Pascal P. Klamser. Writing – original draft: Pascal P. Klamser. Writing – review & editing: Pascal P. Klamser, Adrian Zachariae, Olga Baranov, Clara Jon- gen, Dirk Brockmann. References 1. Carlier M. Number of passenger cars and commercial vehicles in use worldwide from 2006 to 2015; 2021. Available from: https://www.statista.com/statistics/281134/number-of-vehicles-in-use-worldwide/. 2. OECD. Container transport (indicator); 2023. Available from: https://data.oecd.org/transport/container- transport.htm. 3. Statista Research Department. Global air traffic—scheduled passengers 2004-2022; 2023. Available from: https://www.statista.com/statistics/564717/airline-industry-passenger-traffic-globally/. 4. Chapman D, Purse BV, Roy HE, Bullock JM. Global trade networks determine the distribution of inva- sive non-native species. Global Ecology and Biogeography. 2017; 26(8):907–917. https://doi.org/10. 1111/geb.12599 5. Yashadhana A, Derbas A, Biles J, Grant J. Pandemic-related racial discrimination and its health impact among non-Indigenous racially minoritized peoples in high-income contexts: a systematic review. Health Promotion International. 2021; 37(2). 6. Hays JN. Epidemics and pandemics: their impacts on human history. ABC-CLIO; 2005. 7. Daftary A, Frick M, Venkatesan N, Pai M. Fighting TB stigma: we need to apply lessons learnt from HIV activism. BMJ Global Health. 2017; 2(4):e000515. https://doi.org/10.1136/bmjgh-2017-000515 PMID: 29225954 8. 9. 10. Fineberg HV. Pandemic Preparedness and Response — Lessons from the H1N1 Influenza of 2009. New England Journal of Medicine. 2014; 370(14):1335–1342. https://doi.org/10.1056/NEJMra1208802 PMID: 24693893 Fraser C, Donnelly CA, Cauchemez S, Hanage WP, Van Kerkhove MD, Hollingsworth TD, et al. Pan- demic Potential of a Strain of Influenza A (H1N1): Early Findings. Science. 2009; 324(5934):1557– 1561. https://doi.org/10.1126/science.1176062 PMID: 19433588 Jia JS, Lu X, Yuan Y, Xu G, Jia J, Christakis NA. Population flow drives spatio-temporal distribution of COVID-19 in China. Nature. 2020; 582(7812):389–394. https://doi.org/10.1038/s41586-020-2284-y PMID: 32349120 11. Klamser PP, D’Andrea V, Di Lauro F, Zachariae A, Bontorin S, Di Nardo A, et al. Enhancing global pre- paredness during an ongoing pandemic from partial and noisy data. PNAS Nexus. 2023; 2(6). https:// doi.org/10.1093/pnasnexus/pgad192 PMID: 37351112 12. Hadfield J, Megill C, Bell SM, Huddleston J, Potter B, Callender C, et al. Nextstrain: real-time tracking of pathogen evolution. Bioinformatics. 2018; 34(23):4121–4123. https://doi.org/10.1093/bioinformatics/ bty407 PMID: 29790939 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 23 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN 13. Tegally H, Wilkinson E, Martin D, Moir M, Brito A, Giovanetti M, et al. Global Expansion of SARS-CoV-2 Variants of Concern: Dispersal Patterns and Influence of Air Travel. medRxiv. 2022. https://doi.org/10. 1101/2022.11.22.22282629 PMID: 36451885 14. Sacco PL, Arenas A, De Domenico M. The Resilience of the Multirelational Structure of Geopolitical Treaties is Critically Linked to Past Colonial World Order and Offshore Fiscal Havens. Complexity. 2023; 2023:1–9. https://doi.org/10.1155/2023/5280604 15. Kissinger H. World Order. Penguin Books; 2015. 16. Brockmann D, Helbing D. The Hidden Geometry of Complex, Network-Driven Contagion Phenomena. Science. 2013; 342(6164):1337–1342. https://doi.org/10.1126/science.1245200 PMID: 24337289 17. Iannelli F, Koher A, Brockmann D, Ho¨vel P, Sokolov IM. Effective distances for epidemics spreading on complex networks. Physical Review E. 2017; 95(1):012313. https://doi.org/10.1103/PhysRevE.95. 012313 PMID: 28208446 18. Gautreau A, Barrat A, Barthe´lemy M. Arrival time statistics in global disease spread. Journal of Statisti- cal Mechanics: Theory and Experiment. 2007; 2007(09):L09001–L09001. 19. Gautreau A, Barrat A, Barthe´lemy M. Global disease spread: Statistics and estimation of arrival times. Journal of Theoretical Biology. 2008; 251(3):509–522. https://doi.org/10.1016/j.jtbi.2007.12.001 PMID: 18222486 20. Nohara Y, Manabe T. Impact of human mobility and networking on spread of COVID-19 at the time of the 1st and 2nd epidemic waves in Japan: An effective distance approach. PLOS ONE. 2022; 17(8): e0272996. https://doi.org/10.1371/journal.pone.0272996 PMID: 35951674 21. Nah K, Otsuki S, Chowell G, Nishiura H. Predicting the international spread of Middle East respiratory syndrome (MERS). BMC Infectious Diseases. 2016; 16(1):356. https://doi.org/10.1186/s12879-016- 1675-z PMID: 27449387 22. Otsuki S, Nishiura H. Reduced Risk of Importing Ebola Virus Disease because of Travel Restrictions in 2014: A Retrospective Epidemiological Modeling Study. PLOS ONE. 2016; 11(9):e0163418. https://doi. org/10.1371/journal.pone.0163418 PMID: 27657544 23. Nah K, Mizumoto K, Miyamatsu Y, Yasuda Y, Kinoshita R, Nishiura H. Estimating risks of importation and local transmission of Zika virus infection. PeerJ. 2016; 4:e1904. https://doi.org/10.7717/peerj.1904 PMID: 27069825 24. Edsberg Møllgaard P, Lehmann S, Alessandretti L. Understanding components of mobility during the COVID-19 pandemic. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences. 2022; 380 (2214). 25. Coelho FC, Lana RM, Cruz OG, Villela DAM, Bastos LS, Pastore y Piontti A, et al. Assessing the spread of COVID-19 in Brazil: Mobility, morbidity and social vulnerability. PLOS ONE. 2020; 15(9):e0238214. https://doi.org/10.1371/journal.pone.0238214 PMID: 32946442 26. Adiga A, Venkatramanan S, Schlitt J, Peddireddy A, Dickerman A, Bura A, et al. Evaluating the impact of international airline suspensions on the early global spread of COVID-19. medRxiv. 2020. https://doi. org/10.1101/2020.02.20.20025882 PMID: 32511466 27. Zipf GK. The P 1 P 2 D Hypothesis: On the Intercity Movement of Persons. American Sociological Review. 1946; 11(6):677. https://doi.org/10.2307/2087063 28. Cascetta E, Nguyen S. A unified framework for estimating or updating origin/destination matrices from traffic counts. Transportation Research Part B: Methodological. 1988; 22(6):437–455. https://doi.org/ 10.1016/0191-2615(88)90024-0 29. Lenormand M, Huet S, Gargiulo F, Deffuant G. A Universal Model of Commuting Networks. PLoS ONE. 2012; 7(10):e45985. https://doi.org/10.1371/journal.pone.0045985 PMID: 23049691 30. Abrahamsson T. Estimation of Origin-Destination Matrices Using Traffic Counts—A Literature Survey. Laxenburg, Austria: IIASA; 1998. Available from: https://pure.iiasa.ac.at/id/eprint/5627/. 31. Barbosa H, Barthelemy M, Ghoshal G, James CR, Lenormand M, Louail T, et al. Human mobility: Mod- els and applications. Physics Reports. 2018; 734:1–74. https://doi.org/10.1016/j.physrep.2018.01.001 32. Balcan D, Colizza V, Gonc¸alves B, Hu H, Ramasco JJ, Vespignani A. Multiscale mobility networks and the spatial spreading of infectious diseases. Proceedings of the National Academy of Sciences. 2009; 106(51):21484–21489. https://doi.org/10.1073/pnas.0906910106 PMID: 20018697 33. Balcan D, Gonc¸alves B, Hu H, Ramasco JJ, Colizza V, Vespignani A. Modeling the spatial spread of infectious diseases: The GLobal Epidemic and Mobility computational model. Journal of Computational Science. 2010; 1(3):132–145. https://doi.org/10.1016/j.jocs.2010.07.002 PMID: 21415939 34. Tizzoni M, Bajardi P, Poletto C, Ramasco JJ, Balcan D, Gonc¸alves B, et al. Real-time numerical fore- cast of global epidemic spreading: case study of 2009 A/H1N1pdm. BMC Medicine. 2012; 10(1):165. https://doi.org/10.1186/1741-7015-10-165 PMID: 23237460 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 24 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN 35. Poletto C, Gomes MF, Pastore y Piontti A, Rossi L, Bioglio L, Chao DL, et al. Assessing the impact of travel restrictions on international spread of the 2014 West African Ebola epidemic. Eurosurveillance. 2014; 19(42). https://doi.org/10.2807/1560-7917.es2014.19.42.20936 PMID: 25358040 36. Poletto C, Pelat C, Le´vy-Bruhl D, Yazdanpanah Y, Boe¨lle PY, Colizza V. Assessment of the Middle East respiratory syndrome coronavirus (MERS-CoV) epidemic in the Middle East and risk of interna- tional spread using a novel maximum likelihood analysis approach. Eurosurveillance. 2014; 19(23). https://doi.org/10.2807/1560-7917.ES2014.19.23.20824 PMID: 24957746 37. Go´mez-Gardeñes J, Soriano-Paños D, Arenas A. Critical regimes driven by recurrent mobility patterns of reaction–diffusion processes in networks. Nature Physics. 2018; 14(4):391–395. https://doi.org/10. 1038/s41567-017-0022-7 38. Masucci AP, Serras J, Johansson A, Batty M. Gravity versus radiation models: On the importance of scale and heterogeneity in commuting flows. Physical Review E. 2013; 88(2):022812. https://doi.org/ 10.1103/PhysRevE.88.022812 PMID: 24032888 39. O’Neill S. Lufthansa Now Drives More Than Half Its Bookings Directly; 2019. Available from: https:// skift.com/2019/03/14/lufthansa-now-drives-more-than-half-its-bookings-directly/. 40. Recchi E, Deutschmann E, Vespe M. Estimating Transnational Human Mobility on a Global Scale. SSRN Electronic Journal. 2019. https://doi.org/10.2139/ssrn.3384000 41. Christidis P, Christodoulou A. The Predictive Capacity of Air Travel Patterns during the Global Spread of the COVID-19 Pandemic: Risk, Uncertainty and Randomness. International Journal of Environmental Research and Public Health. 2020; 17(10):3356. https://doi.org/10.3390/ijerph17103356 PMID: 32408602 42. Stouffer SA. Intervening Opportunities: A Theory Relating Mobility and Distance. American Sociological Review. 1940; 5(6):845. https://doi.org/10.2307/2084520 43. Simini F, Gonza´lez MC, Maritan A, Baraba´si AL. A universal model for mobility and migration patterns. Nature. 2012; 484(7392):96–100. https://doi.org/10.1038/nature10856 PMID: 22367540 44. de Grange L, Gonza´ lez F, Bekhor S. Path Flow and Trip Matrix Estimation Using Link Flow Density. Networks and Spatial Economics. 2017; 17(1):173–195. https://doi.org/10.1007/s11067-016-9322-1 45. Englezou Y, Timotheou S, Panayiotou CG. Estimating the Origin-Destination Matrix using link count observations from Unmanned Aerial Vehicles. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). IEEE; 2021. p. 3539–3544. Available from: https://ieeexplore.ieee.org/ document/9564959/. 46. Official Airline Guide. OAG Global Airline Schedule Data; 2014. Available from: https://www.oag.com/ airline-schedules-data. 47. Recchi E, Deutschmann E, Vespe M. Global Transnational Mobility Dataset; 2019. Available from: https://doi.org/10.5281/zenodo.3911054. 48. Jordahl K, den Bossche JV, Fleischmann M, Wasserman J, McBride J, Gerard J, et al. geopandas/geo- pandas: v0.8.1; 2020. Available from: https://doi.org/10.5281/zenodo.3946761. 49. Song C, Koren T, Wang P, Baraba´ si AL. Modelling the scaling properties of human mobility. Nature Physics. 2010; 6(10):818–823. https://doi.org/10.1038/nphys1760 50. Brockmann D, Hufnagel L, Geisel T. The scaling laws of human travel. Nature. 2006; 439(7075):462– 465. https://doi.org/10.1038/nature04292 PMID: 16437114 51. Schla¨ pfer M, Dong L, O’Keeffe K, Santi P, Szell M, Salat H, et al. The universal visitation law of human mobility. Nature. 2021; 593(7860):522–527. https://doi.org/10.1038/s41586-021-03480-9 PMID: 34040209 52. Noulas A, Scellato S, Lambiotte R, Pontil M, Mascolo C. A Tale of Many Cities: Universal Patterns in Human Urban Mobility. PLoS ONE. 2012; 7(5):e37027. https://doi.org/10.1371/journal.pone.0037027 PMID: 22666339 53. 54. Liang X, Zhao J, Dong L, Xu K. Unraveling the origin of exponential law in intra-urban human mobility. Scientific Reports. 2013; 3(1):2983. https://doi.org/10.1038/srep02983 PMID: 24136012 Lenormand M, Bassolas A, Ramasco JJ. Systematic comparison of trip distribution laws and models. Journal of Transport Geography. 2016; 51:158–169. https://doi.org/10.1016/j.jtrangeo.2015.12.008 55. Pastore y Piontti A, Gomes MFDC, Samay N, Perra N, Vespignani A. The infection tree of global epi- demics. Network Science. 2014; 2(1):132–137. https://doi.org/10.1017/nws.2014.5 56. Belik V, Geisel T, Brockmann D. Natural Human Mobility Patterns and Spatial Spread of Infectious Dis- eases. Physical Review X. 2011; 1(1):011001. https://doi.org/10.1103/PhysRevX.1.011001 57. Ciotti M, Ciccozzi M, Terrinoni A, Jiang WC, Wang CB, Bernardini S. The COVID-19 pandemic. Critical Reviews in Clinical Laboratory Sciences. 2020; 57(6):365–388. https://doi.org/10.1080/10408363. 2020.1783198 PMID: 32645276 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 25 / 26 PLOS COMPUTATIONAL BIOLOGYInfer import risk from the WAN 58. Daqing L, Kosmidis K, Bunde A, Havlin S. Dimension of spatially embedded networks. Nature Physics. 2011; 7(6):481–484. https://doi.org/10.1038/nphys1932 59. Yang Y, Herrera C, Eagle N, Gonza´lez MC. Limits of predictability in commuting flows in the absence of data for calibration. Scientific Reports. 2014; 4(1):5662. https://doi.org/10.1038/srep05662 PMID: 25012599 60. Ren Y, Ercsey-Ravasz M, Wang P, Gonza´lez MC, Toroczkai Z. Predicting commuter flows in spatial networks using a radiation model based on temporal ranges. Nature Communications. 2014; 5(1):5347. https://doi.org/10.1038/ncomms6347 PMID: 25373437 61. 62. Technical Advisory Group on SARS-CoV-2 Virus Evolution. Historical working definitions and primary actions for SARS-CoV-2 variants; 2023. Available from: https://www.who.int/publications/m/item/ historical-working-definitions-and-primary-actions-for-sars-cov-2-variants. Flahault A, Dias-Ferrao V, Chaberty P, Esteves K, Valleron AJ, Lavanchy D. FluNet as a tool for global monitoring of influenza on the Web. Jama. 1998; 280(15):1330–1332. https://doi.org/10.1001/jama. 280.15.1330 PMID: 9794312 63. Geneva: World Health Organization. WHO: Global Influenza Programme—FluNet; 1997. Available from: https://www.who.int/tools/flunet. 64. Geneva: World Health Organization. WHO COVID-19 Dashboard; 2020. Available from: https:// covid19.who.int/. 65. Elbe S, Buckland-Merrett G. Data, disease and diplomacy: GISAID’s innovative contribution to global health. Global challenges. 2017; 1(1):33–46. https://doi.org/10.1002/gch2.1018 PMID: 31565258 66. Shu Y, McCauley J. GISAID: Global initiative on sharing all influenza data–from vision to reality. Euro- surveillance. 2017; 22(13):30494. https://doi.org/10.2807/1560-7917.ES.2017.22.13.30494 PMID: 28382917 67. Khare S, Gurry C, Freitas L, Schultz MB, Bach G, Diallo A, et al. GISAID’s role in pandemic response. China CDC Weekly. 2021; 3(49):1049. https://doi.org/10.46234/ccdcw2021.255 PMID: 34934514 68. Maier BF, Klamser PP, Zachariae A, Schlosser F, Brockmann D. ImportRisk-v1.0.0; 2023. Available from: https://doi.org/10.5281/zenodo.7852477. 69. World Bank. The World Bank: GDP per capita; 2023. Available from: https://data.worldbank.org/ indicator/NY.GDP.PCAP.CD. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1011775 January 24, 2024 26 / 26 PLOS COMPUTATIONAL BIOLOGY	null
10.1167_jov.23.6.7.pdf	The configurations and codes for this experiment are available in: Github github.com/Znasif/ColorCalibration, Google Colab colab.research.google.com/drive/ 1eHDXJPRn3HaeAqcdGbOKHVJSiNBXBCMv.	null	Journal of Vision (2023) 23(6):7, 1–20 1 Calibration of head mounted displays for vision research with virtual reality Nasif Zaman Prithul Sarker Alireza Tavakkoli Department of Computer Science and Engineering, University of Nevada, Reno, NV, USA Department of Computer Science and Engineering, University of Nevada, Reno, NV, USA Department of Computer Science and Engineering, University of Nevada, Reno, NV, USA Immersion in virtual environments is an important analog for scientists. Situations that cannot be safely organized in the real world are being simulated virtually to observe, evaluate, and train aspects of human behavior for psychology, therapy, and assessment. However, creating an immersive environment using traditional graphics practices may create conﬂict with a researcher’s goal of evaluating user response to well-deﬁned visual stimuli. Standard computer monitors may display color-accurate stimuli, but it is generally viewed from a seating position, where the participant can see real-world visual context. In this article, we propose a novel means to allow vision scientists to exert ﬁner control over the participants visual stimuli and context. We propose and verify a device-agnostic approach to color calibration by analyzing display properties such as luminance, spectral distribution, and chromaticity. We evaluated ﬁve diﬀerent head-mounted displays from diﬀerent manufacturers and showed how our approach produces conforming visual outputs. Introduction Recent advances in virtual reality (VR) head- mounted displays (HMD) have enabled wide adoption of the technology in diverse research areas. High-acuity screen resolution (Varjo XR-3, 2022), a wider field of view (Pimax 8k, 2022), higly affordable wireless VR (Oculus Rift, 2022), all-inclusive AR, VR, and eye-tracking capabilities (Vive Pro Eye, 2020) showcase how different commercial products have varied usefulness for research on fields such as vision science (Li et al., 2022), psychology (Ventura, Baños, Botella, & Mohamudally, 2018), therapeutics (Emmelkamp & Meyerbröker, 2021; Hilty et al., 2020), training (Kaplan et al., 2021; Lee, Park, & Park, 2019) and simulation (Stock, Erler, & Stork, 2018). Increasingly, researchers are adopting game engines such as Unreal Engine and unity to design and present immersive environments and stimuli to their subjects. Some of these applications require precise color specifications and display. However, no standard procedure exists that help researchers calibrate these HMDs and specify a color in their chosen color space for visualization. Increasingly, VR technology is replacing traditional displays for more immersive experiments and assessments. For example, virtual simulation of ocular pathologies such as color vision deficiency (Cwierz, Díaz-Barrancas, Llinás, & Pardo, 2021), cataracts (Krösl et al., 2019; Krösl et al., 2020), and macular degeneration (Zaman, Tavakkoli, & Zuckerbrod, 2020) are helping researchers to quantify the effects of these diseases on quality of life. Additionally, VR-based visual assessments are being used to diagnose glaucoma (Skalicky & Kong, 2019), age-related macular degeneration (Zaman et al., 2020). Binocular compensation available in AR displays are being used to correct neuronal loss in experimental settings. Traditional optics may soon be replaced digital spectacles that manipulate the camera feed for recovery of visual function loss. However, such solutions in the fields of assessment, simulation, and rehabilitation would entail a module that can be easily calibrated for accurate color representation that is common for most medical usage of color displays. Several lines of work exist that characterize the chromatic properties of different VR headsets and compare the perceptual performance with more traditional displays (Toscani et al., 2019), physical objective tests (Díaz-Barrancas, Cwierz, Pardo, Pérez, & Suero, 2020; Gil Rodríguez et al., 2022; Cwierz et al., 2021), and so on. In Diaz-Barrancas, Cwierz, Pardo, Perez, and Suero (2020), Díaz-Barrancas et al. (2020), Citation: Zaman, N., Sarker, P., & Tavakkoli, A. (2023). Calibration of head mounted displays for vision research with virtual reality. Journal of Vision, 23(6):7, 1–20, https://doi.org/10.1167/jov.23.6.7. https://doi.org/10.1167/jov.23.6.7 Received December 13, 2022; published June 14, 2023 ISSN 1534-7362 Copyright 2023 The Authors This work is licensed under a Creative Commons Attribution 4.0 International License. Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 2 and; Cwierz et al. (2021), the authors have implemented a way to reconstruct hyperspectral representation of physical scenes from multispectral images using CS-2000 tele-spectroradiometer. The reconstructed virtual scenes were Color-checker box, Ishihara test and Farnsworth-Munsell 100 Hue test. Although such a line of work is focused on creating an accurate virtual representation of captured scene components, in a separate line of work Kim, Cheng, Beams, and Badano (2021) and Toscani et al. (2019) have identified complex behavior in VR renders with Unity and Unreal Engine 4 (UE4), respectively, that invalidate standard color calibration practices. In Kim et al. (2021), the authors experimented with different render configurations to find the most perceptually correct representation for medical applications. In Toscani et al. (2019), the authors have disabled this behavior and formulated how to accurately control color and luminance in HTC Vive Pro Eye with Unreal Engine. However, their solution involved disabling system-wide tonemapping that is set by default in Unreal. Although this procedure causes clipped linear gamma behavior per channel and normal luminance additivity, it changes the behavior of many built-in shaders and materials, compromising the realistic effects default to Unreal Engine levels. Therefore, if researchers require fine-grained control over the behaviors of specific shaders without altering the appearance of the rest of the scene, a modular approach is necessary. More detailed differences between this work and ours will be discussed in the next section. In their subsequent work (Gil Rodríguez et al., 2022), the researchers showed the application of such a system to a real-world scenario, establishing the color constancy of a virtual scene calibrated using their approach. However, disabling the postprocessing routine means that the scene appears brighter and the light sources are clipped, contrary to the more realistic postprocess-enabled pathway. In Murray, Patel, and Wiedenmann (2022), the authors use look-up tables and color grading built into Unity to calibrate the luminance of VR headsets. Their procedure does not include color correction. In this work, we present a framework for UE4 that calibrates and shows any viable color expressed in xyY, Luv, or Lab color spaces to the displays of a VR HMD in a modular manner. The main contribution of our proposed approach is the ability to present the scene such that parts of the view preserve default properties while specific stimuli behave according to specified chromaticity. Our packaged abstraction would allow researchers to create and render realistic stimuli without requiring extensive knowledge of the specific HMD device, spectrophotometer, or color representation in UE4. Furthermore, unlike Toscani et al. (2019), the proposed work would allow for a greater number of researchers to use VR technology for their workflow. Our contributions in this work include measuring and comparing spectral distributions of major commercial HMDs including HTC Vive Pro Eye, Oculus Rift, Pimax, Fove and Varjo XR-3 using the i1 Pro 2 spectrophotometer (i1Pro, 2022). We build a novel UE4 camera asset that, when placed in any map, allows the scene to be processed along two different graphic pipelines. One pathway allows default Unreal rendering behavior to persist, so that objects in the scene appear realistic and provide a sense of immersion. The second pathway objects in the scene to display color-correct properties that are imperative for verifiable and reproducible vision research. This is achieved by modifying the postprocess material so that every object in the scene would go through the first pathway or if given a certain custom-depth stencil value would be processed along the second pathway. Furthermore, it works alongside i1Pro 2 to calibrate and preserve parameters of conversion for a specific HMD with regards to CIE xy, CIE Luv, and RGB. Finally, we validate whether the predicted and measured values of random color space coordinates match up closely. Materials and methods Standard color calibration practices make some assumptions about the properties of the display. Similar to Toscani et al. (2019), we characterize the properties of the HMDs to apply the appropriate calibration protocol. Suppose, we define default postprocess tonemapping as a function of input emissive FLinear Color RGB values τ (RGB). FLinear Color is an Unreal Engine object that defines color within a range of [0,1]. This is distinct from the reflectance values considered in Toscani et al. (2019), where the scene objects were either illuminants or Default Lit shading materials and, therefore, depended on illumination and viewing conditions.Then, disabling tonemapping has the following effect on input emissive values: Lxyview = ˜τ (RGB), (1) where ˜τ is a composite postprocess that includes all other postprocess operations except for tonemapping τ . In our setup, we use self-emitting virtual surfaces and we do not disable system wide tonemapping, and instead, use a mapping function α(RGB) that converts stimuli RGB values so that, for the output chromaticity, Lxystimuli = ˜τ (τ (α(RGB))) (2) has the same effect as ˜τ , while the rest of the scene behaves as default: Lxyscene = ˜τ (τ (RGB)) (3) Scene objects are any part of the three-dimensional environment that is rendered using the Default Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 3 Lit shading model and, therefore, is influenced by external lighting, shadow, tonemapping, and other postprocesses. Stimuli are objects that have no inherent chromaticity and obtain their color value from prescribed Lxy values that are calibrated to display exact chromaticity output to the HMD screen. and the output luminance Y. These relations were established for three different scenarios: • Conventional postprocessing. • Postprocess tonemapping disabled. • Selective correction. Experimental setup Our experimental setups include a handheld spectrophotometer for recording HMD display outputs. The spectrophotometer is controlled using MATLAB script and the HMDs are controlled using Unreal Engine. We placed the spectrophotometer against the right display of the HMDs. The device was held in place so that it pointed to the center of the display. As we did not have access to wide field colorimeters such as the I29 used by Gil Rodríguez et al. (2022), this may have resulted in off-axis measurements. We make comparisons between the characteristics of the separate HMDs and between the left and right ocular displays to help establish reliable color and luminance calibration procedures. Calibration process For calibration, two sets of data were collected for each headset. The first set involved measuring luminance and chromaticities of ten linearly spaced points in the R = RGB(x, 0, 0), G = RGB(0, x, 0), and B = RGB(0, 0, x) channels as well as a combined channel (x, x, x) where x(cid:4){0, 1}. The latter set involved measuring spectral distributions (Sp), chromaticities (x, y), and luminance (Y) from red, green and blue primaries alongside the white point. Using the first set of measurements we modeled the gamma correction, relation between input RGB to the emissive property of the shaders in Unreal Engine With the second set of measurements we verified the established relation between the input RGB and the chromaticity of the output (x, y). Considerations before rendering in unreal engine Rendering in Unreal Engine entails several considerations that help to avoid unintended color effects. A scene in Unreal Engine consists of two types of objects: illumination sources and reflective objects. Unlit shaders behave the same way under any lighting condition and can be considered as a self-emissive surface material. Emissive attribute is the only component that produces chromaticity and luminosity in unlit shaders. Default-lit materials, in contrast, are affected by lighting conditions. Figure 1 demonstrates these differences. After a main shading pass, which applies lighting and specular properties to objects in default-lit mode and only emissive property in unlit mode, the scene is passed through a series of post processing steps that introduce effects like tonemapping, motion blur, flare, and bloom. Selecting unlit shader model does not have any impact on these postprocess steps because these attributes are not influenced by external lighting conditions, but by their own properties of self-emission. Therefore, to correct any tonemapping and color grading, a method can either be setup by placing an unbounded postprocess volume in the scene that impacts every scene object within render view (Toscani et al., 2019) or by changing the postprocess material of the scene camera. We chose Figure 1. Comparison between diﬀerent shading models and attributes. Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 4 Figure 2. Comparison between (Toscani et al., 2019) and our approach. to create a blueprint ready object (PostCamera) that inherits from the Camera object in Unreal Engine and is modified to disable all postprocessing steps solely through camera settings when placed within a scene and used as target view. Additionally, for rendering and measuring stimuli properties, we chose to disable all illumination sources except for the stimuli. The stimuli material has lambertian surface parameters enabled with emissive (RGB) values set to be visible to the scene camera. This material is applied to a square plane directly in front of the scene camera. Moreover, our approach allows users to visually represent two kinds of materials at once: i) material displaying conventional photorealistic properties and ii) material displaying color-correct properties. By assigning selected specific values to the custom depth pass for each object in the scene, photorealism-specific graphics routines are not applied to them. This means illumination, shadow, exposure, and so on, have no effect on the colorcorrect material rendered to the corresponding final HMD pixel output due to its separate graphics pipeline. In some classical approaches, the light source of the scene is manipulated to change the reflectance value of materials for the desired HMD output. Generally, this approach would be ineffectual in Unreal Engine rendering pipeline as the postprocess routines would still alter the chromaticity and luminance intensity of the scene objects. To make this classical approach work, Toscani et al. (2019) disabled the complete postprocess routines. However, this approach still has Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 5 Rendered ﬁeld of view (horizontal in degrees) HTC Vive Pro Eye Oculus Rift Fove 0 Pimax 5k Plus Varjo VR-3 107 88 90 160 115 Table 1. Speciﬁcations of VR headsets. Pixel density (pixel per degrees) 13.45 12.27 14.2 16 70 (central) 30 (peripheral) Display type AMOLED OLED OLED LCD uOLED (center) LCD (peripheral) See-through camera Native eye tracking Yes No No No Yes Yes No Yes No Yes some limitations. First, after calibration with a certain illumination, the scene illumination needs to remain same for all scenes that use the calibrated reflectance values. Instead, in our approach, we leverage the postprocess pathway in such a manner that objects with custom depth pass assigned to them will behave not as default lit shader materials, but as self-illuminants. By leveraging emissive property and the postprocess pathway, we ensure that the rendering results are final and independent of scene illumination. Thereby, calibration process does not need to be repeated if the scene illumination changes. Figure 2 summarizes the important differences between Toscani et al. (2019) and our work. Equipment For measuring display outputs we use the i1Pro 2 (i1Pro, 2022) spectrophotometer. We measure the HMDs mentioned in Table 1. The data in Table 1 are taken through independent measurements (Murphy, Maraj, & Hurter, 2018; Hmd geometry database, 2021) or manufacturer reports (Hmd comparison, 2022). Most VR devices have masked diagonal regions, decreasing the diagonal visible area from those reported by manufacturers. Therefore, the pixel densities were calculated by dividing horizontal resolution with horizontal field of view. The HTC Vive Pro Eye is widely used for its all-round functionalities which include a moderate resolution, field of view, eye tracking and video see-through camera (480 p). Oculus Rift has lower resolution (1,440 vs. 1,080 horizontal per eye) and field of view, but is a good choice for commercial therapeutics because of the affordability of the Oculus headsets. Pimax 5k Plus has a considerably wider field of view and resolution (2560 horizontal per eye). Fove 0 is a compact alternative to the HTC Vive Pro Eye. It has a limited field of view and resolution (1,280 horizontal per eye). However, its eye tracking api allows vision researchers to monitor external ocular properties in real time. The Varjo VR-3 is a considerable improvement over all the other HMDs discussed so far. It has the highest resolution displays (central 27 degrees 1,920 and peripheral 88 degrees 2,880 horizontal per eye), and cameras (1,080 p). It has an eye tracking api that gives functional access similar to Fove. However, its cost may make it unsuitable for some research and commercial assessment and therapeutics. These devices vary in display types (AMOLED: HTC Vive Pro, OLED: Fove 0, Oculus Rift, LCD: Pimax 5k and mixed: Varjo VR-3). The mixture of two different display types in Varjo VR-3 of uOLED (central 27°) and LCD (peripheral) requires additional consideration before application in color vision research. The applications were rendered in a computer with a 4.2-GHz Intel Core i7 processor and a Nvidia GeForce RTX 2080 graphics card. Rendering with conventional postprocessing This is the default behavior in UE4.27. In these set of experiments, all camera and postprocessing settings were left to default behavior. Relationship between input intensity and luminance In Unreal Engine, instances of light sources such as point light, and directional light have intensity values (cd and lux respectively) associated with them. However, physically based lighting may have unintended effects on stimuli color perception. Therefore, we chose to use the emissive values (RGB) of the stimuli material to reproduce the intended luminance (Y). This procedure allows stimuli chromaticityperception to be independent of scene illumination and context although stimuli perception will still be affected. Figure 3 shows the luminance (cd/m2) corresponding to the input R = RGB(x, 0, 0), G = RGB(0, x, 0) and B = RGB(0, 0, x) emissive for the HMDs. Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 6 Figure 3. Default conﬁguration. The X axis shows emissive values x in scene material: RGB(x, 0, 0) for red, RGB(0, x, 0) for green, RGB(0, 0, x) for blue, and RGB(x, x, x) for white. The Y axis shows luminance in cd/m2. HTC Vive Oculus Fove Pimax Varjo Red channel Green channel Blue channel mX 8.9 255.6 18.3 t 17.5 44.15 3.36 mX 54.6 319.8 41.3 t 22.7 53.1 6.75 mX 1.5 29.8 0.6 t 5.5 12.6 1.93 mX 7.8 73.4 18.7 t 5.4 16.5 2.9 mX 57.6 379.7 65.2 t 12.5 45.6 8.1 Table 2. Default conﬁguration: Slope and threshold for modeled relationship between luminance and emissive value. The figures demonstrate how for different HMDs the luminous intensities are different for the same emissive values. LED LCD displays can be brighter than variants of OLED displays. However, this is not evident in the results shown in Figure 3. As we shall see in the other variants of this experiments, Unreal Engine’s postprocess routines play a role in the diminished brightness of LED LCD displays. Furthermore, the default camera exposure lead to the intensity of emissive materials in the scene to saturate quickly. Here, our observations differ dramatically from the ones reported by Toscani et al. (2019) because they simply used default shaded materials and reflectance values, whereas we used emissive materials. Default emissive materials behave like light sources and reach a predefined intensity to emulate light exposure. For this reason, in subsequent methods, we disabled the autoexposure property. Further, in Figure 3, we see that green has higher luminance than white, which is also a result of this autoexposure property of self-emitting surface materials. All the displays now exhibit properties of a clipped linear function. The relationship is modeled as follows: (cid:2) x · mX (cid:4)[R,G,B], x · mX < t otherwise t, L = (4) where L is the luminance, mX is the slope corresponding with a specific channel, and t is the threshold beyond which the luminance does not vary with change in x. Table 2 shows the corresponding values of mX, t for each HMD. Table 2 shows that highest exposure effect is displayed by Varjo for green colors. Varjo is also the brightest in this default configuration, while Fove is the dimmest. Therefore, vision research that involves scotopic stimuli, with realistic rendering still enabled, can make use of Fove for their research. In contrast, for simulation of real-world performance, Varjo is ideal, because it exhibits the highest dynamic range (signified by the steep slope). However, HTC Vive can be a good alternative if affordability is a concern. All displays show a spike and saturation at the start, which is caused by the saturation due to camera exposure. Pimax and Varjo HMDs show a secondary spike. In Varjo, white is the brightest, whereas in all other displays, green is the brightest. The OLED displays (Oculus and Fove) show equal spread between the luminance of these colors. Fove and Pimax have considerably dimmer brightness compared to the other HMDs. As default gamma correction is enabled, displays show piece-wise linear behavior. Luminance additivity For standard displays, summation of R, G, and B channel intensities give grayscale intensities: L(RGB(x, x, x)) = L(RGB(x, 0, 0)) +L(RGB(0, x, 0)) + L(RGB(0, 0, x)) (5) demonstrating additive property. However, when postprocess tonemapping is enabled, as is default for UE4, such behavior is not visible (Figure 4). Instead, Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 7 Figure 4. Default conﬁguration. The X axis shows emissive values x and the Y axis shows the ratio (LR + LG + LB)/LW. HTC Vive Oculus Fove Pimax Varjo HTC Vive Oculus Fove Pimax Varjo (LR + LG + LB)/LW 2.74 2.27 2.07 1.78 1.39 Table 3. Default conﬁguration: Average luminance ratios for ten shades within x(cid:4)[0.1, 1.0]. cred cgreen cblue 0.001 0.02 0.04 0.002 0.007 0.006 0.002 0.003 0.002 0.002 0.002 0.005 0.005 0.017 0.023 Table 4. Channel constancy scaling error mse(c ∗ φX 0.1 , φX x ). subadditive property is observed for all HMDs. This shows considerably more subadditivity (approximately 200%) compared with the one observed by Toscani et al. (2019) because of our choice of self-emitting surfaces. Here, we can see that Fove is the only device that shows near constancy across x values. All other devices show an upfront increase in ratio (caused by exposure) and a dip at the tail (for x = 1, LW transitions from grey to white). For different devices, the ratio of predicted (sum of individual channels) and measured luminance are shown in Table 3. It further shows that Varjo is closest to displaying luminance additivity while HTC Vive is the farthest. It means, for HTC Vive, white is significantly less bright than the summation of the primaries. Channel constancy Channel constancy is maintained when scaling emissive values (RGB values) linearly also scales the spectral outputs by the same factor. Channel constancy means that scaling the channels independently would not change the chromaticity and only impact luminance. To determine whether color constancy is preserved for all HMDs, we carry out the following experiment. We measure the spectral output (φ(RGB)) for x = 0.1 in R = RGB(x, 0, 0), G = RGB(0, x, 0), and B = RGB(0, 0, x) channels separately. Next, we obtain spectral outputs sr·m2 ) for x(cid:4)[0.1, 1.0] for all the channels. If (Radiance W the spectral outputs are a linear scaling of x, such that: of display (Varjo with uOLED and LCD), the channels show multiple peaks for the primaries (Figure 6). The local peaks indicate multiple wavelengths are dominantly present in the composition of the primaries. For the rest of the HMDs the channels show a single peak. Moreover, the channels have the same scaling factor of 1 for all the devices except for Varjo. This means only Varjo violates channel constancy. Using the scaling factor (c) of the peak of each shade and 0.1 shade we calculate the mean error in Table 4. The error, e is calculated by taking the mean square error of all c ∗ φX channel constancy for all other devices. For Varjo the constancy error is still small, as we used the peak ratio x This negligible error demonstrates good 0.1 and φX However, this may simply be caused by auto exposure transforming intermediate brightness levels to the primary brightness level. Further violations of channel constancy is demonstrated in Figure 5. Only the Varjo HMD primaries fall inside the chromaticity coordinates while displaying a slight drift in the shades. For all the rest of the HMDs, the calibration process creates a calibration matrix that when applied to original primary RGB values in R = RGB(1, 0, 0), G = RGB(0, 1, 0), and B = RGB(0, 0, 1) output chromaticity values that fall outside the gamut. This means the calibration process based on autoexposure peak values is faulty We will further investigate the calibration process in the following subsection. φX x=[0.1,1.0] = c · φX 0.1; where X (cid:4)[R, G, B], (6) Calibration test then it can be said that the HMDs maintain channel constancy. For the only HMD with more than one type For standard color calibration to be applicable, luminance additivity and channel constancy must Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 8 Figure 5. Default conﬁguration. CIE 1931 Chromaticity diagram with ‘x’ denoting actual device output for diﬀerent shades (0.1 through 0.9) of the primaries. ‘o’ denotes the theoretical position of the actual primaries using the calibration matrix. be preserved (Brainard, 1989). However, as we have seen in the preceding sections, these properties are not demonstrated by any of the VR devices while Unreal Engine’s default postprocess tonemapping is enabled. This is further demonstrated when we try to use standard calibration practices to this system. We record the primaries and using the least square method find an M that minimizes the euclidean distance XYZ − M · RGBT. Original XYZ are obtained from the spectroradiometer measurements of the primaries by converting xyY to XYZ: X = x ∗ Y/y Y = Y/100 Z = (1 − x − y) ∗ Y. (8) (7) (9) To verify if the calibration was successful, we display eight different emissive values. These emissive values correspond to the eight corners of two cubes such that one corner of the bigger cube is located at RGB(0.2, 0.2, 0.2) and the other corner at RGB(0.8,0.8,0.8) and one corner of the smaller cube is located at RGB(0.4,0.4,0.4) and the other corner at RGB(0.6,0.6,0.6). Essentially, these combination of colors are selected simply because of their spread in both RGB and chromaticity space (Figure 7) which facilitates verification of the calibration process. Next, we measure the Lxy values with the spectroradiometer. Let us represent the corresponding CIE xy chromaticity values as Lxymeasured. Using the conversion matrix from the primaries, let us now obtain the calibrated position of the cube corners: Lxypredicted = Lxy(M · RGBT cube) (10) Figure 7 shows how the plotted Lxymeasured and Lxypredicted deviate, confirming that standard calibration is not effective. The theoretical coordinates of the cube points vary significantly between devices owing to the variation in the chromaticity of the primaries measured with the spectroradiometer. This is because we use those values to construct the conversion matrix M. Because the Varjo device had the primaries most closely situated with the theoretical position, the mapped cube points also are quite near to the actual device measurement. In the following sections we will demonstrate two methods to calibrate and display nominal values by 1) disabling tonemapping and 2) computationally correcting for tonemapping. Rendering with postprocess tonemapping disabled As we saw in the earlier section, the standard camera in Unreal Engine had auto exposure enabled by default which was responsible for the discrepancy with luminance additivity, and channel constancy properties needed for accurate color calibration. These effects were not present in Toscani et al. (2019) as their objects were reflective instead of self-illuminant. We therefore supplanted the standard camera with our PostCamera object. Our PostCamera object inherits from the camera actor object in Unreal Engine. It is setup to partially disable tonemapping at begin play. Disabling tonemapping entails overriding the autoexposure, bloom, motion blur, grain jitter, scene fringe, and graining. Additionally, we applied a postprocess material that altered the blendable location to Before Tonemapping. In Unreal Engine, blendables are sets of parameters such as base color, opacity, and so on that are passed on to graphics pipeline for rendering. Different stages of the rendering pipeline read and write to different blendables. When set to Before Tonemapping, PostProcessInput0 in the postprocess material editor provides access to scene color with all lighting HDR. Therefore, we use it to only counter the effects of tonemapping using a single pipeline for both stimuli and scene objects. Figure 8 shows the rendering pipeline and how the input emissive or chromaticity values are interpreted to render accurate pixels to the HMD. Now with these altered settings, we repeated the previous experiments. Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 9 Figure 6. Default conﬁguration. Per channel spectral distribution graph. The X axis denotes the wavelength in nanometers and the Y axis denotes the spectral output at that wavelength. The spectral distribution of diﬀerent shades of red, green, blue, and white (from left to right) are shown for each device (top to bottom). Relationship between input intensity and luminance Figure 9 shows the luminance (cd/m2) corre- sponding to the input R, G and B channels. The figures demonstrate how different HMDs react to disabling postprocess tonemapping. As the auto exposure is now disabled, the piece-wise-linear relation is not present. Instead, we can see that disabling the tonemapping has disabled gamma correction as well. This is unlike the relationship found by Toscani et al. (2019) where they found Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 10 Figure 7. Default conﬁguration. CIE 1931 Chromaticity diagram with ‘x’ denoting actual device output for diﬀerent points on the cubes discussed earlier.‘o’ denotes the theoretical position of the actual RGB values using the calibration matrix. Figure 8. Demonstration of rendering pipelines for scene object and stimuli object. Figure 9. Disabled tonemapping. The X axis shows emissive values x in RGB(x, 0, 0) for red, RGB(0, x, 0) for green, RGB(0, 0, x) for blue, and RGB(x, x, x) for white. The Y axis shows luminance in cd/m2. piecewise-linear relationship restored by the same setting. It is possible that their settings included additional routines to enable gamma correction. Otherwise, the luminance groups are conserved so that Fove and Pimax are still dimmer than the rest. Instead of the tail spike that was visible for Varjo with default settings, now the opposite can be seen. The primaries seem to be dimmer compared with a portion of the shades of the primaries. This also illustrates a significant limitation of the normal settings in Unreal Engine. If a stimulus required precise luminance control, that cannot be attained as auto exposure maps most intensities over a threshold to saturation. Using this disabled tonemapping state, intermediate luminance levels can be reached. Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 11 Figure 10. Disabled tonemapping. The X axis shows emissive values x, while the Y axis shows the ratio (LR + LG + LB)/LW. HTC Vive Oculus Fove Pimax Varjo HTC Vive Oculus Fove Pimax Varjo (LR + LG + LB)/LW 1.04 1.14 1.39 1.01 1.02 Table 5. Disabled tonemapping: Average luminance ratios for ten shades within x(cid:4)[0.1, 1.0]. cred cgreen cblue 0.017 0.019 0.012 0.028 0.018 0.013 0.069 0.058 0.056 0.01 0.017 0.007 0.078 0.047 0.052 Table 6. Disabled tonemapping: Channel constancy scaling error. Luminance additivity Figure 10 shows the luminance additivity property discussed earlier. The dashed lines represent predicted luminance (summation of channel luminance) and the circles represent actual grayscale luminance. As the predicted and measured luminance values align, we can say that the luminance additivity property is preserved. For different devices, the ratio of predicted (sum of individual channels) and measured luminance are shown in Table 5. The ratios are now much closer to 1.0. However, the OLED display (Oculus and Fove) ratios are slightly higher. Additionally, Fove HMD shows the same drop it exhibited in the normal settings. Channel constancy As with the earlier channel constancy experiment, here we determine whether spectral distributions at higher emissive values are some constant multiples of the φ(x = 0.1) distributions. Compared with single spectral distributions for single type displays, the new distributions have separate distributions for each shade of the primary colors (Figure 12). This is a direct result of turning off auto exposure, as now different shades have different brightness and therefore different peaks in their spectral distribution. Furthermore, the scaling factors c seem to denote multiplicative spectral profile and channel constancy. Using the same methods of the earlier section we calculate the mean squared error with the scaling factors shown in Table 6. Because auto exposure is now disabled, we can more easily visualize the difference in luminance for each shade. Preservation of channel constancy is also demonstrated in Figure 11. Fove shows significant drift in measured chromaticity for shades of the red, green, and blue colors. Other HMDs show negligible shift in chromaticity. This slight drift is caused solely by uncorrected gamma. Color gamut Figure 11. Disabled tonemapping. CIE 1931 Chromaticity diagram with ‘x’ denoting actual device output for diﬀerent shades (0.1 through 0.9) of the primaries. ‘o’ denotes the theoretical position of the actual primaries using the calibration matrix. Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 12 Figure 12. Disabled tonemapping. Per channel spectral distribution graph. The X axis denotes the wavelength in nanometers and the Y axis denotes the spectral output at that wavelength. The spectral distribution of diﬀerent shades of red, green, blue, and white (from left to right) are shown for each device (top to bottom). for Pimax and Varjo are also within the boundary of the coordinate system. Demonstrating that disabling tonemapping is a much better solution for visual stimuli presentation in VR-based vision science products. Color calibration test We record the new primaries and using the least square method find an M that minimizes the euclidean distance XYZ − M · RGBT. New XYZ are obtained Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 13 Figure 13. Disabled tonemapping. CIE 1931 Chromaticity diagram with ‘x’ denoting actual device output for diﬀerent points on the cubes discussed earlier.‘o’ denotes the theoretical position of the actual RGB values using the calibration matrix. The top row shows uncorrected thoretical positions of the cube RGBs, while the bottom row shows gamma corrected positions. from the spectroradiometer measurements of the primaries by converting xyY to XYZ. Figure 13 shows the positions of the cube corners in CIE XY space for predicted and measured xyY values with the spectroradiometer. Let us represent the corresponding CIE xy chromaticity values as Lxymeasured. Using the conversion matrix from the primaries, let us now obtain the calibrated position of the cube corners: Lxypredicted = Lxy(M · RGBT cube). (11) Figure 13 shows how the plotted Lxymeasured and Lxypredicted do not align perfectly, confirming that standard calibration is affected by the absence of gamma correction. We, therefore, designed the following scheme to apply gamma correction in the XYZ space using α(RGB) that converts emissive values to produce the same effect as disabling tonemapping. We apply a function ˜α(xyY ) on xyY after converting the predicted XYZ = Mcorrected · RGBT to xyY using the following equation: R = X + Y + Z (12) x = X/R (13) y = Y/R (14) Y = Y. (15) For an input xyYscene and white point xyYw, we use the function ˜α(xyY ) defined as: ˜α(xyscene) = xyw + xyscene − xyw (1 + B\|xyscene−xyw\|) ˜α(Yscene) = (cid:2) x · mX (cid:4)[R,G,B], x · mX < t otherwise t, (16) . (17) This correction is formulated by examining the relative positions of the predicted and measured chromaticities. The predicted positions are radially spread more outwards compared with the measured values. Our ˜α(xyY ) function revises the predicted values further inward radially. Figure 20 shows how the function now maps the predicted values very close to the measured ones for B = 0.01. The correction shown in Figure 20 is necessary because of the gamma correction being withheld by completely disabling tonemapping and postprocess settings. Essentially, our function is a gamma correction in the chromaticity space. Rendering with selective color correction We repeated the measures and analyses presented in the previous section after re-enabling the tonemapping. However, we still do it with our PostCamera substitution of standard camera. However, we disable all the overrides we made in the previous section, basically reinstating tonemapping. To correct all the effects of tonemapping, we rely solely on the postprocess material now. Instead of using the blendable mode that Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 14 Figure 14. Diﬀerence between the two rendering pipelines: i) tonemapping disabled and ii) tonemapping and LDR enabled, gamma corrected. before tonemapping, we now use after tonemapping so that the material output is now the final render phase. To selectively use custom processing on only the stimuli and leave the rest of the scene with Unreal Engine’s default realistic rendering, we make use of custom render depth pass. This property allows the postprocess material to selectively only apply counter-tonemapping to the stimuli target. In the postprocess material, we use scene texture: PostProcessInput2, which is the scene texture before the tonemapping pass but without gamma correction. We simply correct gamma and it has the intended corrective effect, as will be demonstrated in the following subsections. The differences with previous method is summarized in Figure 14. Relationship between input intensity and luminance Figure 15 shows the luminance (cd/m2) corresponding with the input R, G and B channels. The figures demonstrate how our postprocessing material alone alters different HMDs rendering, without changing other postprocess routines. All the displays now exhibit properties of a clipped linear function as the input emissive values. However, one distinction from the clipped linear function of Figure 3 is that now the saturation does not occur immediately. This is due to the proxy shutdown of auto exposure through post process materials. Now, the relationship between input Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 15 Figure 15. Tonemapping countered. The X axis shows emissive values x in : RGB(x, 0, 0) for red, RGB(0, x, 0) for green, RGB(0, 0, x) for blue, RGB(x, x, x) for white. The Y axis shows luminance in cd/m2. HTC Vive Oculus Fove Pimax Varjo HTC Vive Oculus Fove Pimax Varjo (LR + LG + LB)/LW 1.04 1.11 1.12 0.95 0.97 mX t mX t mX t mX t mX t Table 7. Tonemapping countered: Luminance ratios. intensity and luminance is similar to when the works of Toscani et al. (2019) disabled postprocessing. Vive, Oculus, and Pimax show complete linearity in the given emissive ranges in Figure 15. This means that vision experimenters can smoothly change the emissive values to increase the display brightness. Fove and Varjo show slight clipping, showing that the experiments would need to be mindful of the brightness beyond the threshold and either restrict the protocols within the cutoff brightness or account for the two different gradients for luminance variation. Interestingly, Varjo primaries are again a little dimmer compared to the 0.9 shade. Although Fove primaries are not dimmer, the rate of change is diminished. Luminance additivity When postprocess tonemapping is enabled and our α(RGB) function is applied to counteract tonemapping for stimuli, luminance additivity is reinstated (Figure 4). For different devices, the ratio of predicted (sum of individual channels) and measured luminance are shown in Table 7. Table 7 and Figure 16 demonstrate that all the displays are now showing almost perfect 15.5 15.5 24 Red channel Green channel 40 40.2 67.2 70.5 87.8 85 37 38.5 67 8 7 Blue channel 30 30 8.7 8.8 17.3 15.4 57 7 7.3 7.8 5 2.5 2.5 24 5 7 Table 8. Tonemapping countered: Slope and threshold for modeled relationship between luminance and emissive values. additivity, whereas OLED displays are very slightly off. Table 8 shows the corresponding values of mX, t for each HMD. Whereas Fove used to be the dimmest display in the previous settings, now Fove is one of the brightest displays. It is now seven times as bright as before. Brightness of all HMDs have increased in the current settings, while the dimmest point has remained similar. This results in the overall increase in dynamic range of the system. Transforming Fove from one of the worst devices to render HDR images to one of the best HMDs for that purpose. However, the difference would be more indicative of true potential with auto exposure disabled in the normal camera settings. Channel constancy By applying our corrective material, the new distributions have a higher peak corresponding Figure 16. Tonemapping countered. The X axis shows emissive values x while the Y axis shows the ratio (LR + LG + LB)/LW. Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 16 Figure 17. Tonemapping countered. Per channel spectral distribution graph. The X axis denotes the wavelength in nanometers and the Y axis denotes the spectral output at that wavelength. The spectral distribution of diﬀerent shades of red, green, blue, and white (from left to right) are shown for each device (top to bottom). with the overall increase in luminance (Figure 17). Moreover, constant scaling factors c across channels of a device represent multiplicative spectral profile and channel constancy. Table 9 shows the least squared error is consistent with our expectations of channel constancy. HTC Vive Oculus Fove Pimax Varjo cred cgreen cblue 0.012 0.013 0.009 0.028 0.026 0.019 0.052 0.046 0.041 0.021 0.021 0.024 0.028 0.012 0.026 Table 9. Tonemapping countered: Channel constancy scaling. Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 17 Figure 18. Tonemapping countered. CIE 1931 Chromaticity diagram with ‘x’ denoting actual device output for diﬀerent shades (0.1 through 0.9) of the primaries. ‘o’ denotes the theoretical position of the actual primaries using the calibration matrix. Figure 19. Demonstration of parallel rendering pipelines. (i) Default-lit and (ii) gamma-corrected and unlit. In Figure 18, we see that the drift in Fove HMD for Calibration test the chromaticity coordinates of intermediate shades of the whites and primaries is still present. All the primaries theoretical positions now reside within the chromaticity diagram and except for Fove, agree perfectly with measurements. With the postprocessing routines re-enabled and application of corrective material, the HMDs now conform to standard calibration procedure. Again, we tested its accuracy by rendering the cube corners. Figure 20. Tonemapping countered. CIE 1931 Chromaticity diagram with ‘x’ denoting actual device output for diﬀerent points on the cubes discussed earlier.‘o’ denotes the theoretical position of the actual RGB values using the calibration matrix. Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 18 We used the newly calibrated Mcorrected to compute the nominal values X Y Z = M · RGBT cubes and converting to Lxy space. In Figure 20, we see how the measured values and nominal values are very close, indicating that it is possible to control the color of the emitted light with our strategy. Figure 8 shows the rendering pipeline and how the input emissive or chromaticity values are interpreted to render accurate pixels to the HMD Discussion We have established that it is not necessary to disable level wide tonemapping and postprocess settings for color accuracy. Using our hybrid approach permits vision scientists to show color accurate stimuli in real-world settings. This ensures that components of default behavior of Unreal Engine such as auto exposure, high dynamic range, tonemapping etc can be retained for real-world objects while assigning objects of interest with separate graphics render path for color accuracy. This is further demonstrated in the Figure 21 where the primaries of all the HMDs align to the ideal red, green and blue. In the normal settings, the primaries appear more washed out and the whites appear darker for all HMDs. In the settings with tonemapping disabled, the colors appear closer to ideal but slightly off due to absence of gamma correction. HTC Vive Oculus Fove Pimax Varjo λred λgreen λblue 610 520 450 620 520 460 620 530 460 610–620 540 450 630 540 450 Table 10. Wavelength in nanometer of the peak spectral distribution for each primaries. The results are even more apparent when we look at the visualization of the cube RGBs in Figure 22 described earlier. The normal settings clearly appear washed out. Although Varjo still shows washed out effect in the tonemap disabled settings, the rest of the HMDs only have a slightly elevated brightness owing to uncorrected gamma. Table 10 shows how different the peaks of the spectral distributions are for each of the HMDs, which contribute to the overall change in the appearance and chromaticity coordinates of the primaries. Finally, in Figure 23, we can see the (cid:7)E (Brainard, 2003) perceived color difference for each of the discussed approaches for the cube colors. Our approach clearly shows significantly lower (cid:7)E. Values less than 1.0 is imperceptible to human eyes, and Vive shows the least perceptible difference in our approach, closely followed by Pimax. Values between 1 and 2 indicate differences perceptible through close observation between 2 and 10 are perceptible at a glance. This indicates that for most of the colors, close observation is required to perceive Figure 21. Approximate visualization of the primaries. Figure 22. (From left to right: Vive, Oculus, Fove, Pimax and Varjo.) Cube sRGB representation where the top rows are normal settings, middle rows are tonemapping disabled setting and the bottom rows are postprocess corrected settings. Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 19 Figure 23. Perceived error. (From left to right Standard postprocessing, disabled tonemapping, and our selective correction approach). The X axis shows index of the cube colors and the Y axis shows the corresponding (cid:7)E perceived error. differences when VR devices are calibrated using our approach. Indices 7 and 15 show white and gray points, which are brought closer together when postprocessing is disabled, compared with standard. Overall, the line graph shows that standard postprocessing has an easily perceptible color difference compared with input and the variability between devices is high. However, when the postprocessing routines are disabled, the interdevice variability reduces while the perceived errors converge. Conclusion Future studies that leverage different rendering pathways laid out in this current work should reveal how this impacts results of color vision research with VR-HMDs. Furthermore, this should help researchers pick the ideal HMD for their particular application. Their are still limitations to the brightness and color gamut achievable by any particular HMD, but with careful consideration of display properties and graphics rendering pipeline, most relevant stimuli can be generated with some of the cheaper HMDs. However, using the two different calibration processes laid out in this work, a wide range of virtual studies can be conducted without requiring the express knowledge of underlying display properties. Keywords: virtual reality, color calibration, color vision Acknowledgments Supported in part by National Institute of General Medical Sciences of the National Institutes of Health under grant numbers [P30 GM145646], the Department of Defense by grant number [FA9550-21-1-0207], and by the National Science Foundation by grant number [OAC-2201599]. The configurations and codes for this experiment are available in: Github github.com/Znasif/ColorCalibration, Google Colab colab.research.google.com/drive/ 1eHDXJPRn3HaeAqcdGbOKHVJSiNBXBCMv. Commercial relationships: none. Corresponding author: Nasif Zaman. Email: [email protected]. Address: 1664 North Virginia Street (M.S. 171) Reno, NV 89557, USA. References 8k. (2022). Pimax 8k. Available from https: //pimax.com/product/vision-8k-plus/. Brainard, D. H. (1989). Calibration of a computer controlled color monitor. Color Research & Application, 14(1), 23–34. Brainard, D. H. (2003). Color appearance and color difference specification. Science of Color, 2(191–216), 5. Cwierz, H., Díaz-Barrancas, F., Llinás, J. G., & Pardo, P. J. (2021). On the validity of virtual reality applications for professional use: A case study on color vision research and diagnosis. IEEE Access, 9, 138215–138224. Díaz-Barrancas, F., Cwierz, H., Pardo, P. J., Pérez, Á. L., & Suero, M. I. (2020). Spectral color management in virtual reality scenes. Sensors, 20(19), 5658. Diaz-Barrancas, F., Cwierz, H. C., Pardo, P. J., Perez, A. L., & Suero, M. I. (2020). Visual fidelity Journal of Vision (2023) 23(6):7, 1–20 Zaman, Sarker, & Tavakkoli 20 improvement in virtual reality through spectral textures applied to lighting simulations. Electronic Imaging, 2020(15), 259–1. chronic stroke patients: A systematic review and meta-analysis. BioMed Research International, 2019 (pp. 1–12). Emmelkamp, P. M., & Meyerbröker, K. (2021). Virtual reality therapy in mental health. Annual Review of Clinical Psychology, 17, 495–519. Eye. (2020). Htc vive pro. Available from https: //www.vive.com/us/product/vive-pro/. Gil Rodríguez, R., Bayer, F., Toscani, M., Guarnera, D., Guarnera, G. C., & Gegenfurtner, K. R. (2022). Colour calibration of a head mounted display for colour vision research using virtual reality. SN Computer Science, 3(1), 1–10. Hilty, D. M., Randhawa, K., Maheu, M. M., McKean, A. J., Pantera, R., Mishkind, M. C., . . . Rizzo, A. S. (2020). A review of telepresence, virtual reality, and augmented reality applied to clinical care. Journal of Technology in Behavioral Science, 5(2), 178–205. Hmd comparison. (2022). Available from https: //vr-compare.com/. Hmd geometry database. (2021). Available from https://github.com/risa2000/hmdq. i1Pro. (2022). i1 display pro. Available from https: //www.xrite.com/categories/calibration-profiling/ i1display-pro/. Kaplan, A. D., Cruit, J., Endsley, M., Beers, S. M., Sawyer, B. D., & Hancock, P. A. (2021). The effects of virtual reality, augmented reality, and mixed reality as training enhancement methods: A meta-analysis. Human Factors, 63(4), 706–726. Kim, A. S., Cheng, W.-C., Beams, R., & Badano, A. (2021). Color rendering in medical extended-reality applications. Journal of Digital Imaging, 34(1), 16–26. Krösl, K., Elvezio, C., Hürbe, M., Karst, S., Wimmer, M., & Feiner, S. (2019). Icthroughvr: Illuminating cataracts through virtual reality. In 2019 IEEE Conference on Virtual Reality and 3D User Interfaces (VR) (pp. 655–663). Osaka, Japan: IEEE. Krösl, K., Elvezio, C., Luidolt, L. R., Hürbe, M., Karst, S., & Feiner, S. et al. (2020). Cataract: Simulating cataracts in augmented reality. In 2020 IEEE International Symposium on Mixed and Augmented Reality (ISMAR) (pp. 682–693). Brazil: IEEE. Lee, H. S., Park, Y. J., & Park, S. W. (2019). The effects of virtual reality training on function in Li, Y., Kim, K., Erickson, A., Norouzi, N., Jules, J., & Bruder, G. et al. (2022). A scoping review of assistance and therapy with head-mounted displays for people who are visually impaired. ACM Transactions on Accessible Computing (TACCESS), 15(3), 1–28. Murphy, S., Maraj, C., & Hurter, J. (2018). Field of view: Evaluation of oculus rift and htc vive. Murray, R. F., Patel, K. Y., & Wiedenmann, E. S. (2022). Luminance calibration of virtual reality displays in unity. Journal of Vision, 22(13), 1–1. Rift. (2022). Oculus rift. Available from https://www. oculus.com/blog/first-look-at-the-rift-shipping-q1- 2016/. Skalicky, S. E., & Kong, G. Y. (2019). Novel means of clinical visual function testing among glaucoma patients, including virtual reality. Journal of Current Glaucoma Practice, 13(3), 83. Stock, S., Erler, C., & Stork, W. (2018). Realistic simulation of progressive vision diseases in virtual reality. In Proceedings of the 24th ACM Symposium on Virtual Reality Software and Technology (pp. 1–2). Tsukuba, Japan: ACM. Toscani, M., Gil, R., Guarnera, D., Guarnera, G., Kalouaz, A., & Gegenfurtner, K. R. (2019). Assessment of oled head mounted display for vision research with virtual reality. In 2019 15th International Conference on Signal-Image Technology & Internet-Based Systems (SITIS) (pp. 738–745). Sorrento, Italy: IEEE. Ventura, S., Baños, R. M., Botella, C., & Mohamudally, N. (2018). Virtual and augmented reality: New frontiers for clinical psychology. State of the art virtual reality and augmented reality knowhow, IntechOpen (pp. 99–118). London. XR-3. (2022). Varjo xr-3. Available from https: //varjo.com/products/xr-3/. Zaman, N., Tavakkoli, A., & Zuckerbrod, S. (2020). A mixed reality system for modeling perceptual deficit to correct neural errors and recover functional vision. In 2020 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW) (pp. 269–274). Atlanta, GA: IEEE.	null
10.1371_journal.pmed.1003998.pdf	null	Data will be available to successful applications for clearly specified research projects following /our-research/other-research- policy/data-sharing/ Discussion with the trial team is encouraged to determine whether the relevant data to support the application are available. .	RESEARCH ARTICLE Radiotherapy to the prostate for men with metastatic prostate cancer in the UK and Switzerland: Long-term results from the STAMPEDE randomised controlled trial 2, Noel W. Clarke3,4, 2, Silke GillessenID 1, Christopher D. BrawleyID 9,10, Alex Hoyle3,4, Rob J. JonesID 7, David P. Dearnaley1, Hassan Douis8, Duncan C. GilbertID Chris C. Parker1, Nicholas D. JamesID Adnan Ali3, Claire L. Amos2, Gerhardt Attard5, Simon Chowdhury6, Adrian CookID William CrossID Clare GilsonID Zafar I. Malik12, Malcolm D. Mason13, David MathesonID Mary RauchenbergerID Amit Bahl15, Alison BirtleID Joanna GaleID Anna LydonID Delia PudneyID Jacob TanguayID Trial Collaborative Group¶ 2,6, J Martin RussellID 16,17, Lisa Capaldi18, Omar Din19, Daniel Ford8, 23, 3,22, Mohammed KagziID 26, Omi ParikhID 12, Narayanan Nair SrihariID 20, Ann HenryID 24, Joe M. O’Sullivan25, Sangeeta A. PaiseyID 28, Vijay RamaniID 31, Mahesh K. B. ParmarID 2☯, Matthew R. SydesID 3,29, Peter RobsonID 21, Peter HoskinID 14, Robin Millman2, 2, Hannah RushID 27, 2, 11, Hannah Sweeney2, 30, 2, 11, Ruth E. LangleyID 2, 2☯, for the STAMPEDE 1 The Institute of Cancer Research and Royal Marsden NHS Foundation Trust, London, United Kingdom, 2 MRC Clinical Trials Unit at UCL, UCL, London, United Kingdom, 3 The Christie Hospital, Manchester, United Kingdom, 4 Salford Royal Hospitals, Manchester, United Kingdom, 5 UCL Cancer Institute, UCL, London, United Kingdom, 6 Guys and St Thomas’s NHS Foundation Trust, London, United Kingdom, 7 St James’s University Hospital, Leeds, United Kingdom, 8 University Hospitals Birmingham NHS Foundation Trust, Birmingham, United Kingdom, 9 Istituto Oncologico della Svizzera Italiana, EOC, Bellinzona, Switzerland, 10 Università della Svizzera Italiana, Lugano, Switzerland, 11 Institute of Cancer Sciences, University of Glasgow, Glasgow, United Kingdom, 12 Clatterbridge Cancer Centre, Liverpool, United Kingdom, 13 Cardiff University, Cardiff, United Kingdom, 14 University of Wolverhampton, Wolverhampton, United Kingdom, 15 University Hospitals Bristol NHS Trust, Bristol, United Kingdom, 16 Rosemere Cancer Centre, Lancs Teaching Hospitals, University of Manchester, Manchester, United Kingdom, 17 UCLan, Lanchashire, United Kingdom, 18 Worcestershire Acute Hospitals NHS Trust, Worcester, United Kingdom, 19 Weston Park Cancer Centre, Sheffield, United Kingdom, 20 Queen Alexandra Hospital, Portsmouth, United Kingdom, 21 University of Leeds, Leeds, United Kingdom, 22 Mount Vernon Cancer Centre, Northwood, United Kingdom, 23 The James Cook University Hospital, Middlesbrough, United Kingdom, 24 Torbay and South Devon NHS Trust, Torbay, United Kingdom, 25 Patrick G Johnston Centre for Cancer Research, Queen’s University Belfast, United Kingdom, 26 Hampshire Hospitals NHS Foundation Trust, Hampshire, United Kingdom, 27 Royal Preston Hospital, Preston, United Kingdom, 28 South West Wales Cancer Centre, Swansea, United Kingdom, 29 Manchester University Hospitals NHS Trust, Manchester, United Kingdom, 30 Shrewsbury & Telford Hospitals NHS Trust, Shrewsbury, United Kingdom, 31 Velindre Cancer Centre, Cardiff, United Kingdom ☯ These authors contributed equally to this work. ¶ Membership of the STAMPEDE Trial Collaborative Group is listed in S4 Text. * [email protected] (CCP); [email protected] (MRS) Abstract Background STAMPEDE has previously reported that radiotherapy (RT) to the prostate improved overall survival (OS) for patients with newly diagnosed prostate cancer with low metastatic burden, but not those with high-burden disease. In this final analysis, we report long-term findings on a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Parker CC, James ND, Brawley CD, Clarke NW, Ali A, Amos CL, et al. (2022) Radiotherapy to the prostate for men with metastatic prostate cancer in the UK and Switzerland: Long-term results from the STAMPEDE randomised controlled trial. PLoS Med 19(6): e1003998. https://doi.org/ 10.1371/journal.pmed.1003998 Academic Editor: James Derek Brenton, University of Cambridge, UNITED KINGDOM Received: January 4, 2022 Accepted: April 22, 2022 Published: June 7, 2022 Peer Review History: PLOS recognizes the benefits of transparency in the peer review process; therefore, we enable the publication of all of the content of peer review and author responses alongside final, published articles. The editorial history of this article is available here: https://doi.org/10.1371/journal.pmed.1003998 Copyright: © 2022 Parker et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: Data will be available to successful applications for clearly specified research projects following the MRC CTU at UCL PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 1 / 20 PLOS MEDICINEstandard data sharing processes: https://www. mrcctu.ucl.ac.uk/our-research/other-research- policy/data-sharing/ Discussion with the trial team is encouraged to determine whether the relevant data to support the application are available. Email to: [email protected]. Funding: Research support for this comparison and other comparisons in the STAMPEDE protocol was awarded by Cancer Research UK (CRUK_A12459) www.cancerresearchuk.org (for this comparison, co-authors CCP, DPD, MDM, MKBP, MR, MRS, NDJ; and additionally for other comparisons DG, DM, GA, REL, RM, WC); Medical Research Council (MRC_MC_UU_12023/25, MC_UU_00004/01 and MC_UU_00004/02) www. ukri.org/councils/mrc (to authors MKBP, MRS, REL); and Swiss Group for Clinical Cancer Research, www.sakk.ch (to co-author SG). Other research support for the STAMPEDE protocol was awarded by Astellas www.astellas.com, Clovis Oncology www.clovisoncology.com, Janssen www.janssen.com, Novartis www.novartis.com, Pfizer www.pfizer.com, Sanofi-Aventis www.sanofi. com. CCP, DPD and NDJ are supported by the National Institute for Health Research (NIHR) Biomedical Research Centre at The Royal Marsden NHS Foundation Trust and the Institute of Cancer Research, London. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: I have read the journal’s policy and the authors of this manuscript have the following competing interests: CCP reports personal fees from Bayer, personal fees from Janssen, personal fees from Clarity Pharmaceuticals, personal fees from Myovant, personal fees from ITM Oncologics, outside the submitted work NDJ received research funding to the institution from Astellas, Astra Zeneca &Janssen; receipt of honoraria/fees on the advisory board for Astra Zenenca, Clovis, Janssen, Merck, Novartis & Sanofi; received fees as a speaker for Bayer & Novartis NWC received honoraria from Astellas & Janssen; took a consulting/advisory role for Astellas, Janssen, Ferring, Bayer & Sanofi; was paid speakers fees b Janssen & Astellas; received funding for the institution from Astra Zeneca; received meeting and travel expenses from Janssen, Astellas, Sanofi, Astra Zeneca, Ferring & Ipsen GA reports personal fees from Sanofi Aventis, during the conduct of the study; personal fees and non-financial support from Astellas, personal fees and non-financial support from Medivation, personal fees from Novartis, personal fees from Millennium Pharmaceuticals, personal fees and non-financial Radiotherapy to the prostate for metastatic disease the primary outcome measure of OS and on the secondary outcome measures of symptom- atic local events, RT toxicity events, and quality of life (QoL). Methods and findings Patients were randomised at secondary care sites in the United Kingdom and Switzerland between January 2013 and September 2016, with 1:1 stratified allocation: 1,029 to standard of care (SOC) and 1,032 to SOC+RT. No masking of the treatment allocation was employed. A total of 1,939 had metastatic burden classifiable, with 42% low burden and 58% high burden, balanced by treatment allocation. Intention-to-treat (ITT) analyses used Cox regression and flexible parametric models (FPMs), adjusted for stratification factors age, nodal involvement, the World Health Organization (WHO) performance status, regular aspirin or nonsteroidal anti-inflammatory drug (NSAID) use, and planned docetaxel use. QoL in the first 2 years on trial was assessed using prospectively collected patient responses to QLQ-30 questionnaire. Patients were followed for a median of 61.3 months. Prostate RT improved OS in patients with low, but not high, metastatic burden (respectively: 202 deaths in SOC versus 156 in SOC+RT, hazard ratio (HR) = 0�64, 95% CI 0.52, 0.79, p < 0.001; 375 SOC versus 386 SOC+RT, HR = 1.11, 95% CI 0.96, 1.28, p = 0�164; interaction p < 0.001). No evidence of difference in time to symptomatic local events was found. There was no evidence of differ- ence in Global QoL or QLQ-30 Summary Score. Long-term urinary toxicity of grade 3 or worse was reported for 10 SOC and 10 SOC+RT; long-term bowel toxicity of grade 3 or worse was reported for 15 and 11, respectively. Conclusions Prostate RT improves OS, without detriment in QoL, in men with low-burden, newly diag- nosed, metastatic prostate cancer, indicating that it should be recommended as a SOC. Trial registration ClinicalTrials.gov NCT00268476, ISRCTN.com ISRCTN78818544. Author summary Why was this study done? • Prostate cancer is the most common cancer in males. • Radiotherapy (RT) to the prostate is widely used as a radical treatment for nonmeta- static prostate cancer. • A comparison was added to the STAMPEDE protocol to assess whether RT to the pros- tate would also be helpful for males with metastatic prostate cancer. A benefit in survival was targeted. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 2 / 20 PLOS MEDICINEsupport from Abbott Laboratories, personal fees and non-financial support from Essa Pharmaceuticals, personal fees and non-financial support from Bayer Healthcare Pharmaceuticals, personal fees from Takeda, grants from AstraZeneca, grants from Arno Therapeutics, grants from Innocrin Pharma, grants, personal fees and non-financial support from Janssen, personal fees from Veridex, personal fees and non-financial support from Roche/Ventana, personal fees and non-financial support from Pfizer, personal fees from The Institute of Cancer Research (ICR), outside the submitted work; and The Institute of Cancer Research (ICR) receives royalty income from abiraterone I receive a share of this income through the ICR’s Rewards to Discoverers Scheme SC received consulting fees from Telix, remedy & Huma; received payment for speaker fees and/or manuscript writing and/or educational events from Astra Zeneca, Novartis/AAA, Clovis, Janssen, Bayer, Pfizer, Beigene & Astellas; they were a member of the data safety monitoring/advisory board for Astra Zeneca, Novartis/AAA, Clovis, Janssen, Bayer, Pfizer, Beigene & Astellas DPD received payment to the institution from C33589/ A19727 Advances in Physics for Precision Radiotherapy; previous employer, The Institute of Cancer Research receives loyalty income from abiraterone, receives personal share of this income through ICR’s Rewards to Discoverer’s Scheme; honoraria for consultancy from Janssen; EP1933709B1 – Location and Stabilisation Device., European patent issued, Pending in Canada and India SG reports personal fees from Orion, personal fees from Janssen Cilag, personal fees from ProteoMedix, personal fees from Amgen, personal fees from MSD, other from Tolero Pharmaceuticals, other from Astellas Pharma, other from Janssen, other from MSD Merck Sharp&Dome, other from Bayer, other from Roche, other from Pfizer, other from Telixpharma, other from Amgen, other from Bristol-Myers Squibb, other from AAA International SA, other from Orion, other from Silvio Grasso Consulting, from Tolremo, outside the submitted work; In addition, Gillessen has a patent WO2009138392 issued and Menarini Silicon Biosystems (Advisory Board 2019) - not compensated Aranda (Advisory Board 2019) - not compensated RJJ received research funding to the institution from Bayer, Astellas & Pfizer; received honoraria on the advisory board for Janssen, Astellas, Bayer, Pfizer; received speaker fees from Janssen, Astellas, Bayer & Pfizer REL received an institutional grant from the MRC CG received research funding to the institution from Janssen, Clovis Oncology, Sanofi, Astellas, Medical Research Council & Cancer Radiotherapy to the prostate for metastatic disease • The trial previously reported a clinically relevant, statistically significant overall survival (OS) benefit for patients with a low metastatic burden but not for men with a high meta- static burden. • This long-term analysis assesses survival with substantially longer follow-up and more events and looked also at complications of local disease. What did the researchers do and find? • A randomised controlled trial of adding RT to the prostate to standard of care (SOC) was incorporated into the STAMPEDE protocol. • More than 2,000 patients joined the comparison between 2013 and 2016. • The data set was frozen in 2021 and analysed using standard methods. • There was a clear improvement in survival with prostate RT in the low metastatic bur- den group. • There was no improvement in survival with prostate RT in the high metastatic burden group. • Symptomatic local progression and the need for later local intervention were improved with RT in the low metastatic burden group. • In the low metastatic burden group, the improvement with RT was similar whether the RT was given with a daily schedule (over 4.5 weeks) or a weekly schedule (over 6 weeks). • The adverse effects of RT were manageable without any impact on long-term quality of life (QoL). What do these findings mean? • Prostate RT is a relatively cheap, widely accessible, and well-tolerated treatment. • Prostate RT is indicated in patients with newly diagnosed prostate cancer with a low metastatic burden. • RT to the prostate is not routinely indicated for patients with a high metastatic burden. Introduction Prostate radiotherapy (RT) is recommended for men with newly diagnosed, low-burden, met- astatic prostate cancer, but not for men with high-burden disease [1]. This recommendation is based largely on the initial results of the STAMPEDE trial, reported in 2018 [2]. In this rando- mised controlled trial of 2,061 men with newly diagnosed metastatic prostate cancer, prostate RT improved overall survival (OS) for men with low metastatic burden (hazard ratio [HR] PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 3 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease Research UK DF received speaker fees and/or manuscript writing and/or educational events from BMS, IPSEN, EUSA, Pfizer, ESAI; they received travel expenses from Janssen & IPSEN MDM is an advisory board member for Endocyte & Clovis AB received payment for lecture/presentation/speaker bureau/manuscript writing or educational event from Boston Scientific AJB received speaker fees and travel support from Janssen DF received payment for lectures for Janssen, Pfizer & BMS; support for attending conferences/meetings from Genisiscare & BMS AMH received research grants from CRUK and NIHR; support attending meetings from the European Association of Urologists; is a member of the European Association of Urologists & the Prostate Cancer Guidelines Group MK received travel, accommodation and conference fees as expenses from Bayer, travel and accommodation fees for Prostate cancer summits from Janssen AL received expenses for attending meetings and/or travel from Astellas, Bayer, BMS & MSD JMOS received speaker fees from AAA, Astellas, Bayer, Janssen, Novartis, Sanofi and participated as an advisory board member and/or member of the data safety monitoring board for AAA, Astellas, Bayer, Janssen, Novartis & Sanofi NNS received travel/meeting payments from Janssen JT received support for conference attendance from Janssen, Roche & Bayer; participated on the advisory board for Astra Zeneca, Astellas & Bayer MKBP received research funding to the Unit he directs from Acoria Pvt Ltd, Akagera, Amgen, Aspirin Foundation, Astellas, AstraZeneca, Baxter, Bayer, BMS US, Bri-Bio, Cepheid, Cipla, Clovis Inc, CSL Behring, Eli-Lilly, Emergent Biosolutions, Gilead Sciences, GlaxoSmithKline, Grifols, Janssen Products LP, Janssen-Cilag, Johnson & Johnson, Micronoma, Modus Theraputics, Mylan, Novartis, Pfizer, Sanofi, Serum Institute of India, Shionogi, Synteny Biotechnology, Takeda, Tibotec, Transgene, ViiV Healthcare, Virco and Xenothera MRS received research funding to the institution from Astellas, Clovis, Janssen, Novartis, Pfizer, Sanofi-Aventis; received speaker fees from Lilly Oncology & Janssen; independent member of data monitoring committees. All other authors have nothing to declare. Abbreviations: ADT, androgen deprivation therapy; AE, adverse event; CONSORT, Consolidated Standards of Reporting Trials; CTCAE, Common Terminology Criteria for Adverse Events; FPM, flexible parametric model; GnRH, gonadotrophin- releasing hormone; HR, hazard ratio; IQR, interquartile range; ITT, intention-to-treat; LIFS, local intervention–free survival; NSAID, nonsteroidal anti-inflammatory drug; OS, overall 0.68, 95% CI 0.52 to 0.90; p = 0.007), with no evidence of a meaningful effect on survival in men with high metastatic burden (HR 1.07, 95% CI 0.90 to 1.28; p = 0.420). That initial analy- sis, triggered by a preplanned number of events, was done after a median follow-up of 37 months and was based on 761 events. Here, we report the final analysis of OS, with an addi- tional 2 years follow-up. We hypothesised that prostate RT would reduce the complications of local disease progres- sion, such as urinary or bowel obstruction. If so, this could benefit men with metastatic disease, regardless of disease burden. Here, we report data on freedom from local interventions (e.g., urinary catheter, ureteric stents, nephrostomies, and colostomy). Any benefits of prostate RT need to be weighed against the risk of treatment-related adverse events (AEs). We report, for the first time, data from the trial on quality of life (QoL). The trial was stratified according to the choice of 1 of 2 RT dose-fractionation schedules, nominated prior to randomisation; 36 Gy in 6 fractions over 6 weeks, or 55 Gy in 20 fractions over 4 weeks. The 2 schedules were chosen in the expectation that they would be similarly effective. With the benefit of additional follow-up, and more events in the final analysis, we have tested for any differential impact on OS by choice of RT schedule. Methods Study participants Eligible patients had prostate cancer that was newly diagnosed, with no previous radical treat- ment, had metastatic disease confirmed on a bone scintigraphic scan and soft tissue imaging, and were within 12 weeks after starting androgen deprivation therapy (ADT). Patients were required to have no contraindications to RT and no clinically significant cardiovascular his- tory. Participants were recruited at secondary care sites in the UK and Switzerland. The trial was registered as NCT00268476 (ClinicalTrials.gov) and ISRCTN78818544 (ISRCTN.com). The trial was done in accordance with Good Clinical Practice guidelines and the Declaration of Helsinki and had relevant ethics (West Midlands–Edgbaston Research Eth- ics Committee) and regulatory approvals. All patients gave written informed consent. The rationale and design, including sample size calculations, have been described previously [2,3]. Full details are in the protocol at www.stampedetrial.org. Procedures All patients received lifelong hormone therapy as gonadotrophin-releasing hormones (GnRHs) agonists or antagonists or orchidectomy. In addition, docetaxel was permitted after it became available for this setting in the UK. Docetaxel, when used, was given as six 3 weekly cycles of 75mg/m2 with or without prednisolone 10 mg daily. External beam RT to the prostate was given as 1 of 2 schedules nominated prior to rando- misation: 36 Gy in 6 consecutive weekly fractions of 6 Gy or 55 Gy in 20 daily fractions of 2.75 Gy over 4 weeks. Treatment was given with the patient supine, with a full bladder and an empty rectum. The planning target volume consisted of the prostate only with an 8-mm mar- gin posteriorly and a 10-mm margin elsewhere. RT was to commence as soon as practicable after randomisation, and, if the patient was having docetaxel as part of standard of care (SOC), within 3 to 4 weeks after the last docetaxel dose. Patients were followed up 6 weekly until 6 months after randomisation, 12 weekly to 2 years, 6 monthly to 5 years, and then annually. Toxicities and symptoms were reported at regu- lar follow-up visits or when an AE was categorised as “serious.” These were graded with Com- mon Terminology Criteria for Adverse Events (CTCAE) v4�0. Separately, bowel and bladder adverse effects during RT and long-term possible RT effects were recorded using the RTOG PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 4 / 20 PLOS MEDICINEsurvival; PH, proportional hazard; QoL, quality of life; RMST, restricted mean event-free (“survival”) time; RT, radiotherapy; SLEFS, symptomatic local event-free survival; SOC, standard of care; WHO, World Health Organization. Radiotherapy to the prostate for metastatic disease scale [4]. Participants were asked to complete the EORTC QLQ-C30 at each scheduled follow- up appointment. Metastatic burden at randomisation was evaluated retrospectively through central imaging review of whole body scintigraphy and computerized tomography (CT) or MRI staging scans. Metastatic burden was classified according to the definition used in the CHAARTED trial [5] as either high (polymetastatic; �4 bone metastases with �1 outside the vertebral bodies or pel- vis and/or visceral metastases) or low (oligometastatic). Patients with only lymph node metas- tases, in the absence of bone or visceral disease, were therefore classified as low metastatic burden regardless of the number of nodal metastases. Randomisation and masking Patients were randomised centrally using a computerised algorithm, developed and main- tained by the trials unit. Minimisation with a random element of 20% was used (80% probably of allocation to a minimising treatment), stratifying for hospital, age at randomisation (<70 versus �70 years), nodal involvement (negative versus positive versus indeterminate), the World Health Organization (WHO) performance status (0 versus 1 or 2), planned form of ADT (orchidectomy versus LHRH (leuteinising hormone-releasing hormone) agonist versus LHRH antagonist versus dual androgen blockade), and regular aspirin or nonsteroidal anti- inflammatory drug (NSAID) use (yes or no). Planned docetaxel use was added as a stratifica- tion factor after use was permitted as part of SOC. Allocation was 1:1 to SOC only or SOC+RT. There was no blinding to treatment allocation. Primary and secondary outcomes The primary efficacy outcome measure was OS, defined as time from randomisation to death from any cause. Secondary outcomes for this long-term efficacy analysis included local inter- vention–free survival (LIFS)—consisting of time from randomisation to the first report on case report forms of TURP, ureteric stent, surgery for bowel obstruction, urinary catheter, nephrostomy, colostomy, death from prostate cancer—and symptomatic local event-free sur- vival (SLEFS), comprising any of these LIFS events or acute kidney injury, urinary tract infec- tion, or urinary tract obstruction. Cause of death was determined by the site investigator, with some cases reclassified as prostate cancer death according to predefined criteria which sug- gested this to be the likely cause. Patients without the event of interest were censored at the time last known to be event free. QoL analyses focused on Global QoL % and QLQ-30 Sum- mary Score %, as derived from patient reports at scheduled assessment time points in the first 2 years after randomisation (see S2 Text). Statistical analysis The primary outcome measure, OS, was assessed across all patients and separately within patient subgroups characterised by baseline metastatic burden (low versus high) and nomi- nated RT schedule (daily versus weekly). Standard survival analysis methods were used to analyse time-to-event data in Stata v16.1 (College Station, Texas, United States of America). A nonparametric stratified log-rank test was used to assess any difference in survival between treatment groups; this was stratified across the minimisation factors used at randomisation (except hospital and planned form of hormone therapy) plus protocol-specific time periods defined by other arms recruiting to STAMPEDE or changes to SOC which could affect the population being randomised. Cox proportional hazards (PHs) regression models adjusting for the same stratification factors and stratified by time period were used to estimate relative treatment effect; a HR less than 1�00 PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 5 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease favoured the research arm. Unadjusted estimates of treatment effect are also presented. Flexi- ble parametric models (FPMs) were fitted with degrees of freedom (5.5) and adjusted for strati- fication factors and time periods [6]. Medians and 5-year survival estimates are presented from the FPM fitted to the data. Kaplan–Meier curves, using the KMunicate format [7], show estimated survival over time. Following the fitting of Cox models, the PHs assumption was tested using a global Grambsch–Therneau test with log-transformed time; restricted mean event-free (“survival”) time (RMST) was emphasised in the presence of nonproportionality, using a t-star of 91 months as determined by the Royston and Parmar method [6]. Cox and Fine and Gray regression models [8] were used for cause-specific and competing risk analyses, respectively; competing risks were non-prostate cancer–related death for prostate cancer–spe- cific survival and death from any cause for SLEFS and LIFS. Evidence for different treatment effect across subgroups was assessed using the likelihood ratio p-value for an interaction term added to the relevant adjusted Cox/FPM model, or Wald test p-value from Fine and Gray model. Sensitivity analyses of local event outcomes examined the impact of excluding death from prostate cancer and a competing risks approach with death from any cause specified as a competing event. All tests are presented as 2 sided, with 95% CIs and the relevant p-value. Median follow-up was estimated using the Kaplan–Meier method with reverse censoring on death. All patients were included in the efficacy and QoL analyses according to allocated treatment on an intention-to-treat (ITT) basis; sensitivity analyses exclude patients who did not explicitly fulfill all of the eligibility criteria. AE data are shown for the safety population, in patients with at least 1 follow-up assessment and analysed according to whether RT was received within 1 year of randomisation (SOC+RT) or not (SOC). Analyses of the QoL outcomes included partly conditional and composite approaches, build- ing on the approaches previously used in the trial [9]. For the former, missing values were mul- tiply imputed using observed data using chained equations. Imputed values for assessments dating after a patient had died were restored to missing. Generalised estimating equations with an independence correlation matrix were used to estimate the expected value of the outcome for each treatment arm at each assessment time point. For the composite approach, observa- tions following the death of a patient were set to 0% (corresponding to the lowest possible QoL state). Mixed linear regression with random intercept and slope (with unstructured correlation specification) was used to model the outcome. Additional cross-sectional analyses estimated the difference in average QoL associated with treatment allocation in patients alive and with data available at a given assessment time point, controlling for baseline state. This trial is reported per the Consolidated Standards of Reporting Trials (CONSORT; see S3 Text). Results Patients Between January 22, 2013 and September 2, 2016, 2,061 patients were randomised from 117 hospitals in UK and Switzerland: 1,029 to SOC and 1,032 to SOC+RT. The data set was frozen on March 17, 2021 and included information up to November 30, 2020. Fig 1 shows the CON- SORT flow diagram for analyses presented in this paper. Table 1 shows baseline characteristics balanced across the allocated treatment groups. Table A in S1 Text shows baseline characteris- tics in 1,939 (94%) patients who were evaluable for disease burden, 819 (40%) with low- and 1,120 (54%) with high-burden disease. Median duration of follow-up was 61.3 months (interquartile range [IQR] = 53.8 to 73.1) and was similar in both treatment groups: SOC 61.0 (IQR = 53.8 to 72.6) and SOC+RT 61.6 (IQR = 53.8 to 73.1). PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 6 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease Fig 1. CONSORT diagram. AE, adverse event; CONSORT, Consolidated Standards of Reporting Trials; RT, radiotherapy to the prostate, SOC, standard of care. �Alive, no withdrawal of permission for continued data collection. https://doi.org/10.1371/journal.pmed.1003998.g001 OS by allocated treatment and metastatic burden A total of 1,183 deaths were reported, 609 in patients allocated to SOC and 574 in those allo- cated to SOC+RT (Fig 2, Table 2). In the low metastatic burden group, 358 had died: 202/409 SOC and 156/410 SOC+RT. Median survival was 63.6 months for SOC and 85.5 months for SOC+RT (5-year survival 53% versus 65%); adjusted HR = 0.64 (95% CI 0.52 to 0.79; p < 0.001 [p = 0.00004]) (Fig 3, Table 2). There was no evidence of non-PHs. In the high-burden disease group, 761 had died: 375/567 SOC and 386/553 SOC+RT. Median survival was 41.2 months in SOC and 38.8 months in SOC+RT (5-year survival 35% versus 30%): adjusted HR = 1.11 (95% CI 0.96 to 1.28; p = 0.164) (Fig 4, Table 2). There was no evidence of non-PHs. There was clear evidence of differential treatment effect according to metastatic burden: interaction test p < 0.001 [p = 0.00005]. Similar results were obtained from cause-specific and competing risk analyses (Table 2). A participant audit found 36 (<2%) patients with baseline data or documented protocol devia- tion inconsistent with the comparison’s full eligibility criteria. Sensitivity analyses found no impact from excluding these patients (Tables B and C in S1 Text). Analysis of time from ran- domisation to reported second-line treatments indicates no confounding of RT treatment effect on OS by postprogression abiraterone or enzalutamide therapy (S10 and S11 Figs). Exploration of OS by elected RT schedule In 980 patients nominated prior to randomisation for weekly RT, 575 had died: 282/482 SOC and 293/498 SOC+RT. Median survival was 52.2 months for SOC and 49.9 months for PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 7 / 20 PLOS MEDICINETable 1. Baseline characteristics of all patients in the comparison. Characteristic Age at randomisation (years) Median (IQR) SOC (n = 1,029) 68 (63 to 73) SOC+RT (n = 1,032) 68 (63 to 73) Radiotherapy to the prostate for metastatic disease WHO performance status Pain from prostate cancer Previous notable health issues T-category at randomisation N-category at randomisation Metastatic burden Sites of metastases Gleason sum score PSA pre-ADT (ng/ml) Time from diagnosis (days) Days from starting hormones Planned for SOC docetaxel Nominated RT schedule Range 0 1 to 2 Absent Present Missing Myocardial infarction Cerebrovascular disease Congestive heart failure Angina Hypertension T0 T1 T2 T3 T4 TX N0 N+ NX Low metastatic burden� High metastatic burden Not classified Bone Liver Lung Distant lymph nodes Other < = 7 8 to 10 Unknown Median (IQR) Range Median (IQR) Missing Median (IQR) Range Missing No Yes 36 Gy/6 f over 6 weeks 55 Gy/20 f over 4 weeks 37 to 86 732 (71%) 297 (29%) 826 (81%) 198 (19%) 5 67 (7%) 29 (3%) 5 (<1%) 46 (4%) 408 (40%) 0 (0%) 12 (1%) 84 (9%) 585 (62%) 260 (28%) 88 345 (36%) 620 (64%) 64 409 (42%) 567 (58%) 53 919 (89%) 23 (2%) 42 (4%) 295 (29%) 35 (3%) 173 (17%) 826 (83%) 30 98 (30 to 316) 1 to 20,590 73 (55 to 94) 1 53 (35 to 70) -3 to 84 17 845 (82%) 184 (18%) 482 (47%) 547 (53%) 45 to 87 734 (71%) 298 (29%) 855 (83%) 172 (17%) 5 58 (6%) 32 (3%) 8 (1%) 52 (5%) 444 (43%) 1 (<1%) 12 (1%) 89 (9%) 603 (63%) 247 (26%) 80 344 (36%) 620 (64%) 68 410 (43%) 553 (57%) 69 917 (89%) 19 (2%) 48 (5%) 304 (29%) 32 (3%) 175 (18%) 820 (82%) 37 97 (33 to 313) 1 to 11,156 73 (55 to 93) 2 55 (34 to 70) 0 to 86 13 849 (82%) 183 (18%) 498 (48%) 534 (52%) �Note: One patient classified with low-burden disease was subsequently restaged as nonmetastatic by the randomising site. They remain in the low metastatic burden subgroup for this analysis. ADT, androgen deprivation therapy; IQR, interquartile range; PSA, prostate specific antigen; RT, radiotherapy to the prostate; SOC, standard of care; WHO, World Health Organization. https://doi.org/10.1371/journal.pmed.1003998.t001 PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 8 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease Fig 2. OS in all patients. Adjusted HR = 0.90 (95% CI 0.81 to 1.01; p = 0.081). HR, hazard ratio; OS, overall survival; RT, radiotherapy to the prostate; SOC, standard of care. https://doi.org/10.1371/journal.pmed.1003998.g002 SOC+RT (5-year survival: 44% versus 42%); adjusted HR = 1.00 (95% CI 0.85 to 1.18); p = 0.974 (Fig 5, Table 2). In 1,081 patients nominated prior to randomisation for daily RT, 608 died: 327/547 SOC and 281/534 SOC+RT. Median survival was 47.8 months in SOC and 55.5 months in SOC+RT (5-year survival 41% versus 47%); adjusted HR = 0.83 (95% CI 0.71 to 0.97; p = 0.022) (Fig 6, Table 2). There was no good evidence of interaction in the treatment effect by RT schedule: interaction p = 0.088. Given that RT improved OS in the low metastatic burden patients, RT schedule was further explored in this subgroup. In 360 patients nominated for weekly RT, 162 had died: 94/190 SOC and 68/170 SOC+RT; adjusted HR = 0.67 (95% CI 0.49 to 0.93; p = 0.015 [p = 0.0155]). In 459 patients nominated for daily RT, 196 had died: 108/219 SOC and 88/240 SOC+RT; adjusted HR = 0.62 (95% CI 0.47–0.83; p = 0.001 [p = 0.00112]). There was no good evidence of interaction in the treatment effect by RT schedule: interaction p = 0.732. SLEFS by allocated treatment A total of 1,209 (59%) patients were reported as experiencing at least 1 symptomatic local event: 608 SOC and 601 SOC+RT. In 789 cases (400 SOC and 389 SOC+RT), death from pros- tate cancer was the only event recorded. Table 3 summarises the reported incidence of each type of event. There was no evidence of a difference in time to first reported event by treatment arm: adjusted HR = 1.00 (95% CI 0.90 to 1.13; p = 0.931); median symptomatic local event– free survival 43.8 months SOC, 43.3 months SOC+RT (5-year SLEFS survival 39% versus 40%) (S1 Fig, Table 2). A total of 1,086 (53%) patients had 1 or more local intervention events reported, 556 SOC and 530 SOC+RT, of which death from prostate cancer was the only event in 78% and 81% of cases. Median local intervention event–free survival was 51.1 months in SOC and 53.6 months PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 9 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease Table 2. Summary of estimated treatment effect for main outcome measures: all patients and metastatic burden subgroups. Outcome measure Patient group Adjusted HR~ Unadjusted HR^ OS All patients 0.90 (0.81 to 1.01) 0.90 (0.81 to 1.01) Low metastatic burden 0.64 (0.52 to 0.79) 0.66 (0.54 to 0.82) High metastatic burden 1.11 (0.96 to 1.28) 1.08 (0.94 to 1.25) Weekly RT (36 Gy/6 f) 1.00 (0.85 to 1.18) 1.01 (0.86 to 1.19) (Low metastatic burden) 0.67 (0.49 to 0.93) 0.71 (0.52 to 0.97) (High metastatic burden) 1.22 (0.99 to 1.50) 1.19 (0.97 to 1.46) Daily RT (55 Gy/20 f) 0.83 (0.71 to 0.97) 0.81 (0.69 to 0.95) (Low metastatic burden) 0.62 (0.47 to 0.83) 0.63 (0.48 to 0.84) (High metastatic burden) 1.02 (0.83 to 1.25) 0.99 (0.81 to 1.21) Prostate cancer–specific survival � All patients 0.92 (0.81 to 1.04) 0.92 (0.81 to 1.04) SLEFS# LIFS# Low metastatic burden 0.62 (0.49 to 0.79) 0.64 (0.50 to 0.81) High metastatic burden 1.12 (0.96 to 1.31) 1.10 (0.94 to 1.28) All patients 1.00 (0.90 to 1.13) 1.00 (0.90 to 1.12) Low metastatic burden 0.72 (0.59 to 0.88) 0.73 (0.60 to 0.90) High metastatic burden 1.23 (1.06 to 1.42) 1.21 (1.05 to 1.40) All patients 0.94 (0.83 to 1.06) 0.93 (0.83 to 1.05) Low metastatic burden 0.62 (0.49 to 0.77) 0.63 (0.50 to 0.78) High metastatic burden 1.18 (1.01 to 1.37) 1.16 (1.00 to 1.34) Event free at 5 years+ SOC SOC +RT RMST+ SOC SOC+RT Difference 42% 53% 35% 44% 54% 37% 41% 52% 33% 49% 62% 41% 39% 46% 33% 44% 54% 38% 45% 65% 30% 42% 64% 29% 47% 66% 32% 51% 72% 35% 40% 58% 26% 47% 67% 32% 52.9 60.6 47.7 53.9 61.3 48.9 52.2 59.9 46.8 57.6 65.7 51.8 49.2 54.5 45.1 53.5 59.7 49.0 55.5 69.0 45.5 53.6 68.2 44.5 57.2 69.5 46.6 59.5 73.7 49.0 48.9 61.8 39.4 55.1 69.1 44.7 2.5 (−0.2 to 5.2) 8.4 (4.5 to 12.2) −2.2 (−5.7 to 1.2) −0.3 (−3.4 to 2.8) 6.9 (0.6 to 13.2) −4.3 (−9.6 to 0.9) 5.0 (1.1 to 8.9) 9.6 (4.0 to 15.2) −0.2 (−4.5 to 4.0) 1.9 (−1.1 to 5.0) 8.0 (4.0 to 12.0) −2.8 (−6.6 to 1.0) −0.3 (−3.5 to 2.8) 7.2 (2.5 to 11.9) −5.8 (−9.7 to −1.9) 1.6 (−1.5 to 4.7) 9.5 (5.2 to 13.8) −4.4 (−8.4 to −0.4) Note: HR and RMST difference are for SOC+RT relative to SOC. �Cause-specific treatment × metastatic burden interaction test p < 0.001 [p = 0.0000977]. Competing risks analysis: overall adjusted sub-HR = 0.93 (95% CI 0.82 to 1.05; p = 0.260); low-burden adjusted sub-HR = 0.66 (95% CI 0.52 to 0.83; p = 0.001); high-burden adjusted sub-HR = 1.11 (95% CI 0.95 to 1.29; p = 0.189); treatment × metastatic burden interaction test p < 0.001 [p = 0.000350]. #SLEFS: treatment × metastatic burden interaction test p < 0.001 [p = 0.0000314]. LIFS interaction p < 0.001 [p = 2.53 × 10−6]. ~Estimates from Cox models adjusting for age, nodal involvement, WHO performance status, regular aspirin or NSAID use, and planned SOC docetaxel at randomisation, stratified by randomisation time period. ^Estimates from unadjusted, unstratified Cox models. +Survival probabilities and RMST estimates are taken from FPMs with t-star = 91 months. HR, hazard ratio; LIFS, local intervention–free survival; NSAID, nonsteroidal anti-inflammatory drug; RMST, restricted mean event-free (“survival”) time; RT, radiotherapy to the prostate; SLEFS, symptomatic local event–free survival; SOC, standard of care; WHO, World Health Organization. https://doi.org/10.1371/journal.pmed.1003998.t002 in SOC+RT (5-year survival 44% versus 47%); adjusted HR = 0.94 (95% CI 0.83 to 1.06; p = 0.286) (S2 Fig, Table 2, Table D in S1 Text). Table E in S1 Text presents the results of sen- sitivity analyses. AEs by allocated treatment Urinary-related late AEs of grade 3 were reported for 20 (2%) patients who received RT within 1 year after randomisation; 10 (2%) were planned for weekly and 10 (2%) for daily treatment; no grade 4 or 5 urinary-related events were reported. Bowel-related late AEs of grade 3 or 4 were reported for 26 (3%) patients, 15 (3%) planned for weekly and 11 (2%) daily treatment (Table 4, Table F in S1 Text). For 610 patients with data available at 2 years, grade 3 urinary AEs were reported for 3 (0.5%) and grade 3 bowel AEs for 6 (1%) (Table G in S1 Text). At 4 years, 2/467 (0.4%) patients had grade 3 or 4 bowel toxicity (Table H in S1 Text). PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 10 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease Fig 3. OS in patients in the low-burden metastatic disease group. Adjusted HR = 0.64 (95% CI 0.52 to 0.79; p < 0.001 [p = 0.00004]). HR, hazard ratio; OS, overall survival; RT, radiotherapy to the prostate; SOC, standard of care. https://doi.org/10.1371/journal.pmed.1003998.g003 Fig 4. OS in patients in the high-burden metastatic disease group. Adjusted HR = 1.11 (95% CI 0.96 to 1.28; p = 0.164). HR, hazard ratio; OS, overall survival; RT, radiotherapy to the prostate; SOC, standard of care. https://doi.org/10.1371/journal.pmed.1003998.g004 PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 11 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease Fig 5. OS in patients nominated for weekly RT (36 Gy/6 f) prior to randomisation. Adjusted HR = 1.00 (95% CI 0.85 to 1.18; p = 0.974). HR, hazard ratio; OS, overall survival; RT, radiotherapy to the prostate; SOC, standard of care. https://doi.org/10.1371/journal.pmed.1003998.g005 Fig 6. OS in patients nominated for daily RT (55 Gy/20 f) prior to randomisation. Adjusted HR = 0.83 (95% CI 0.71 to 0.97; p = 0.022). HR, hazard ratio; OS, overall survival; RT, radiotherapy to the prostate; SOC, standard of care. https://doi.org/10.1371/journal.pmed.1003998.g006 PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 12 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease Table 3. First symptomatic local event reported (patients with event reported). Type of event Urinary tract infection Urinary catheter Acute kidney injury TURP Urinary tract obstruction Ureteric stent Nephrostomy Colostomy Surgery for bowel obstruction PCa death SOC (n = 608) 57 (9%) SOC+RT (n = 601) 80 (13%) 52 (9%) 33 (5%) 24 (4%) 15 (2%) 19 (3%) 5 (1%) 3 (<1%) 0 (0%) 400 (66%) 44 (7%) 34 (6%) 24 (4%) 15 (3%) 8 (1%) 2 (<1%) 3 (1%) 2 (<1%) 389 (65%) PCa, prostate cancer; RT, radiotherapy to the prostate; SOC, standard of care; TURP, transurethral resection of the prostate. https://doi.org/10.1371/journal.pmed.1003998.t003 Over the entire reported follow-up period, at least 1 grade 3 to 5 AE was reported for 458 (44%) of SOC and 451 (45%) SOC+RT patients. Areas of focus for this long-term analysis were endocrine disorders: 160/1,052 (15%) SOC versus 155/992 (16%) SOC+RT; musculoskel- etal disorders: 112/1,052 (11%) SOC, 104/992 (10%) SOC+RT; blood and bone marrow disor- ders: 56/1,052 (5%) SOC, 49/992 (5%) SOC+RT; cardiovascular disorders: 46/1,052 (4%) SOC, 56/992 (6%) SOC+RT; renal disorders: 50/1,052 (5%) SOC, 52/992 (5%) SOC+RT; general dis- orders: 57/1,052 (5%) SOC, 43/992 (4%) SOC+RT; gastrointestinal disorders: 47/1,052 (4%) SOC, 52/992 (5%) SOC+RT; lab abnormalities: 49/1,052 (5%) SOC, 48/992 (5%) SOC+RT (Table I in S1 Text, S7 Fig). At 2 years, of 715 patients with data available, a grade 3 to 5 AE was reported for 52/320 (16%) SOC and 54/395 (14%) SOC+RT (Table J in S1 Text, S8 Fig). At 4 years, based on 358 patients, this was 12/133 (9%) SOC versus 29/225 (13%) SOC+RT (Table K in S1 Text, S9 Fig). Table 4. Patients with grade 3/4 worst late RT toxicity score reported over entire time on trial. Toxicity area SOC+RT Urinary Hematuria Urethral stricture Cystitis Bowel Proctitis Diarrhea Rectal–anal stricture Rectal ulcer Bowel obstruction Weekly, 36 Gy/6 f (n = 473) 10 (2%) 4 (1%) 3 (1%) 3 (1%) 15 (3%) 9 (2%) 6 (1%) 0 (0%) 0 (0%) 1 (<1%) Daily, 55 Gy/20 f (n = 517) 10 (2%) 4 (1%) 4 (1%) 4 (1%) 11 (2%) 5 (1%) 6 (1%) 0 (0%) 1 (<1%) 1 (<1%) Note: SOC+RT in safety population (RTOG scale; patients with RT started within 1 year of randomisation). There were no reported grade 5 late RT toxicity events. RT, radiotherapy to the prostate; SOC, standard of care. https://doi.org/10.1371/journal.pmed.1003998.t004 PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 13 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease Table 5. Summary of QoL analyses. Outcome measure Analysis Average over first 2 years on trial Difference (95% CI) Global QoL (%) QLQ-30 Summary Score (%) Partly conditional Composite outcome Cross-sectional: 12 weeks Cross-sectional: 24 weeks Cross-sectional: 60 weeks Cross-sectional: 104 weeks Partly conditional Composite outcome Cross-sectional: 12 weeks Cross-sectional: 24 weeks Cross-sectional: 60 weeks Cross-sectional: 104 weeks SOC 73.2% 60.3% n/a n/a n/a n/a 85.4% 70.6% n/a n/a n/a n/a SOC+RT 72.4% 61.6% n/a n/a n/a n/a 84.2% 71.7% n/a n/a n/a n/a −0.8% (−2.5% to 0.9%) 1.3% (−1.1% to 3.8%) −2.9% (−4.8% to −1.0%) −0.9% (−3.1% to 1.3%) −1.4% (−4.1% to 1.3%) 1.8% (−2.4% to 6.0%) −1.2% (−2.4% to 0.0%) 1.2% (−1.3% to 3.6%) −2.0% (−3.2% to −0.8%) −1.0% (−2.3% to 0.4%) −1.0% (−2.8% to 0.7%) 0.9% (−1.8% to 3.6%) Note: Partly conditional estimates are based on observed and multiply imputed data from patients alive at scheduled assessments within the first 2 years since randomisation. Composite outcome estimates are based on observed data and implied imputation of missing data from scheduled assessments when a patient was alive, and the assumption of a patient’s Global QoL/QLQ-30 Summary Score being 0% at all scheduled assessments after they have died. Cross-sectional analyses estimate the difference in average Global Qol/QLQ-30 Summary Score between SOC+RT and SOC treatment groups at the specified scheduled assessment, controlling for response at baseline, in complete cases only (i.e., in patients with outcome data provided at baseline and who have survived and for who outcome data is available at the specified scheduled assessment). QoL, quality of life; RT, radiotherapy to the prostate; SOC, standard of care. https://doi.org/10.1371/journal.pmed.1003998.t005 QoL There was no evidence of a difference in QoL scores over time between the allocated treatment groups. Average Global QoL in the first 2 years after randomisation across all patients was 73.2% SOC and 72.4% SOC+RT; absolute difference −0.8% (95% CI −2.5% to 0.9%), p = 0.349 (partly conditional analysis) (Table 5, S3 Fig). When including patients who had died prior to an assessment as having a Global QoL score of 0% at that assessment, average Global QoL was 60.3% SOC versus 61.6% SOC+RT; absolute difference 1.3% (95% CI -1.1% to 3.8%), p = 0.287 (composite outcome analysis) (Table 5, S4 Fig). Average QLQ-30 Summary Score in the first 2 years across all patients was 85.4% SOC and 84.2% SOC+RT; absolute difference −1.2% (95% CI −2.4% to 0.0%), p = 0.050 (partly condi- tional analysis) (Table 5, S5 Fig). When assuming a value of 0% for assessments after a patient had died, average Summary Score was 70.6% SOC and 71.7% SOC+RT; absolute difference 1.2% (95% CI −1.3% to 3.6%), p = 0.365 (composite outcome analysis) (Table 5, S6 Fig). Cross-sectional analyses of both Global QoL and QLQ-30 Summary Score indicated evi- dence of poorer QoL at week 12 after randomisation for patients allocated to SOC+RT— Global QoL absolute difference −2.9% (95% CI −4.8% to −1.0%, p = 0.003); Summary Score absolute difference −2.0% (95% CI −3.2% to −0.8%, p = 0.001)—but not at other assessments (Table 5). Discussion This final analysis has confirmed that prostate RT improves OS in men with newly diagnosed, low-burden metastatic prostate cancer, but not in men with high-burden disease. The magni- tude of the survival benefit is substantial and clinically relevant, particularly given that prostate RT is a relatively cheap, widely accessible, and well-tolerated treatment. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 14 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease These results of the final analysis confirm the findings from the initial analysis. The addi- tional 2 years of follow-up, and the subsequent increase in the number of events for analysis, has reduced the CIs around the point estimate of the HR of the OS benefit for prostate RT. However, the point estimate itself has changed very little, improving from 0.68 to 0.64 for men in the low metastatic disease risk group. This result is consistent with that from the smaller HORRAD trial [10]. Our new data strongly support those guidelines already recommending the use of prostate RT in men with low-burden metastatic disease. We have not found any ben- efit for prostate RT in men with high-burden disease, either in OS or in preventing interven- tions for local disease progression. We found no compelling evidence of a difference in efficacy or toxicity between the 2 RT dose-schedules tested. The weekly schedule of 36 Gy in 6 fractions over 6 weeks has an obvious practical advantage in terms of convenience and may be preferred for that reason. A daily schedule might be preferred if pelvic nodal RT were to be used in addition or if RT dose escala- tion was thought to be appropriate. Prostate RT did not have any long-term impact on QoL either in this trial, or in the HORRAD trial [11]. The risk of toxicity from prostate RT, although low, could be further reduced by the use of more contemporary intensity modulated techniques [12]. The criteria used in the trial to classify cases as low or high burden were taken from those used in the CHAARTED trial [5]. These criteria are based on the presence or absence of vis- ceral disease on CT scan, together with the number and the location of bone metastases on bone scan. Patients with only lymph node metastases have low-burden disease, regardless of the extent of nodal disease. There is no good reason to think that these criteria are optimal for identifying those patients with metastatic disease who stand to benefit from prostate RT. The initial analysis of STAMPEDE suggested that the survival benefit from prostate RT gradually decreased in magnitude as the number of bone metastases visible on a baseline bone scan increased [13]. One could decide to identify patients suitable for prostate RT based solely on the number of bone metastases visible on baseline bone scan, regardless of location. A count- of-metastases approach would be simpler to use in the clinic than the CHAARTED definition and would likely increase the number of men considered suitable for prostate RT. The trial has several strengths, including the randomised design, the large number of events for analysis, and recruitment from over 100 centres, which adds to the generalisability of the results. The main limitations of the study are the changes in clinical practice since the trial started, particularly with regard to imaging techniques and systemic treatment. The trial recruited between 2013 and 2016 and, while this has the benefit of long follow-up, it also means that newer imaging techniques, such as PSMA (Prostate-specific membrane antigen) PET and whole body MRI, were unavailable. It is important to note that low-burden disease in the trial was defined according to bone scan and CT scan. There is no agreed definition of met- astatic disease burden based solely on PSMA PET or on whole body MRI. In patients without visceral disease but who have more than 4 bone metastases on PET or MRI, a bone scan may be required in addition, in order to determine suitability for prostate RT. If this is not practica- ble, and there remains uncertainty as to whether a patient has high- or low-burden disease, there is a strong argument for using prostate RT. The systemic treatment of metastatic prostate cancer has changed since the trial recruited. Standard treatment for men with low-burden metastatic disease now includes one of the newer hormone agents (abiraterone or apalutamide or enzalutamide) in addition to ADT. The effect of these agents on the survival benefit of prostate RT is unknown. Similarly, the effect of prostate RT on the survival benefit of the newer hormonal agents is also unknown. Based on current evidence, it is reasonable to assume that both prostate RT and one of the newer hor- monal agents should be considered SOC for low-burden metastatic disease, in addition to PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 15 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease ADT. The PEACE-1 Trial is testing the use of prostate RT in men receiving ADT + abiraterone. In summary, this final analysis confirms that prostate RT improves OS in men with low- burden, newly diagnosed, metastatic prostate cancer, indicating that it should be recom- mended as a SOC. Supporting information S1 Fig. SLEFS in all patients. Adjusted HR = 1.00 (95% CI 0.90 to 1.13; p = 0.931). HR, hazard ratio; RT, radiotherapy to the prostate; SLE, symptomatic local event; SOC, standard of care. (TIF) S2 Fig. LIFS in all patients. Adjusted HR = 0.94 (95% CI 0.83 to 1.06; p = 0.286). HR, hazard ratio; LI, local intervention; RT, radiotherapy to the prostate; SOC, standard of care. (TIF) S3 Fig. Model-estimated Global QoL (partly conditional analysis in all patients). Difference in weighted average: −0.8% (95% CI −2.5% to 0.9%; p = 0.349). QoL, quality of life; RT, radio- therapy to the prostate; SOC, standard of care. (TIF) S4 Fig. Model-estimated Global QoL (composite outcome analysis in all patients). Differ- ence in weighted average: 1.3% (95% CI −1.1% to 3.8%; p = 0.287). QoL, quality of life; RT, radiotherapy to the prostate; SOC, standard of care. (TIF) S5 Fig. Model-estimated QLQ-30 Summary Score (partly conditional analysis in all patients). Difference in weighted average: −1.2% (95% CI −2.4% to 0.0%; p = 0.050). RT, radiotherapy to the prostate; SOC, standard of care. (TIF) S6 Fig. Model-estimated QLQ-30 Summary Score (composite outcome analysis in all patients). Difference in weighted average: 1.2% (95% CI −1.3% to 3.6%; p = 0.365). RT, radio- therapy to the prostate; SOC, standard of care. (TIF) S7 Fig. Highest grade AE reported over entire time on trial (CTCAE v4.0, all patients). AE, adverse event; CTCAE, Common Terminology Criteria for Adverse Events; RT, radiotherapy to the prostate; SOC, standard of care. (TIF) S8 Fig. Highest grade AE reported at 2 years in patients prior to disease progression (CTCAE v4.0). AE, adverse event; CTCAE, Common Terminology Criteria for Adverse Events; RT, radiotherapy to the prostate; SOC, standard of care. (TIF) S9 Fig. Highest Grade AE reported at 4 years in patients prior to disease progression (CTCAE v4.0). AE, adverse event; CTCAE, Common Terminology Criteria for Adverse Events; RT, radiotherapy to the prostate; SOC, standard of care. (TIF) S10 Fig. Time to reported initiation of abiraterone or enzalutamide from randomisation. RT, radiotherapy to the prostate; SOC, standard of care. (TIF) PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 16 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease S11 Fig. Time to reported initiation of abiraterone or enzalutamide from FFS event. FFS, failure-free survival; RT, radiotherapy to the prostate; SOC, standard of care. (TIF) S1 Text. Table A in S1 Text. Baseline characteristics for metastatic volume analyses. ADT, androgen deprivation therapy; IQR, interquartile range; PSA, prostate specific antigen; RT, radio- therapy to the prostate; SOC, standard of care; WHO, World Health Organization. Table B in S1 Text. Eligibility status following participant audit. RT, radiotherapy to the prostate; SOC, stan- dard of care. Table C in S1 Text. Sensitivity analyses on OS based on explicit eligibility. ITT, intention-to-treat; OS, overall survival; RT, radiotherapy to the prostate; SOC, standard of care. Table D in S1 Text. First local intervention event reported (patients with event reported). PCa, prostate cancer; RT, radiotherapy to the prostate; SOC, standard of care; TURP, transure- thral resection of the prostate. Table E in S1 Text. Summary of analyses of time to local event outcomes. �Subdistribution HR for competing risks models. ^Cox model, adjusting for age, nodal involvement, WHO performance status, regular aspirin or NSAID use and planned SOC doce- taxel at randomisation, stratified by randomisation time period. +Fine and Gray model with out- come excluding PCa death and death from any cause as competing risk. NSAID, nonsteroidal anti-inflammatory drug; PCa, prostate cancer; SOC, standard of care; WHO, World Health Orga- nization. Table F in S1 Text. Grade 3 to 5 late RT toxicities reported over entire time on trial (RTOG). Note: Treatment arms correspond to safety population; patients with �1 Follow-Up CRF returned. RT, radiotherapy to the prostate; RTOG, Radiation Therapy Oncology Group; SOC, standard of care. Table G in S1 Text. Grade 3 to 5 late RT toxicities reported at 2 years (RTOG). Note: Treatment arms correspond to safety population; patients with �1 Follow-Up CRF returned and no reported progression at 2 years. RT, radiotherapy to the prostate; RTOG, Radiation Therapy Oncology Group; SOC, standard of care. Table H in S1 Text. Grade 3 to 5 late RT toxicities reported at 4 years (RTOG). Note: Treatment arms correspond to safety popu- lation; patients with �1 Follow-Up CRF returned and no reported progression at 4 years. RT, radiotherapy to the prostate; RTOG, Radiation Therapy Oncology Group; SOC, standard of care. Table I in S1 Text. Grade 3 to 5 AEs reported over entire time on trial, overall and for selected body systems (CTCAE). Note: Treatment arms correspond to safety population; patients with �1 Follow-Up/SAE CRF returned. AE, adverse event; RT, radiotherapy to the prostate; RTOG, Radiation Therapy Oncology Group; SOC, standard of care. Table J in S1 Text. Grade 3 to 5 AEs reported at 2 years, overall and for selected body systems (CTCAE). Note: Treatment arms correspond to safety population; patients with �1 Follow-Up/SAE CRF returned and no reported progression at 2 years. AE, adverse event; CTCAE, Common Terminology Criteria for Adverse Events; RT, radiotherapy to the prostate; SOC, standard of care. Table K in S1 Text. Grade 3 to 5 AEs reported at 4 years, overall and for selected body systems (CTCAE). Note: Treatment arms correspond to safety population; patients with �1 Follow-Up/SAE CRF returned and no reported progression at 4 years. AE, adverse event; CTCAE, Common Terminology Crite- ria for Adverse Events; RT, radiotherapy to the prostate; SOC, standard of care. (DOCX) S2 Text. Statistical analysis plan. (PDF) S3 Text. CONSORT checklist. CONSORT, Consolidated Standards of Reporting Trials. (PDF) S4 Text. List of investigators, oversight committees, and contributors. (PDF) PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 17 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease Acknowledgments Large-scale trials do not happen without huge collaborations. Thanks to all central and site staff who have made the STAMPEDE trial happen. See S4 Text for full list of investigators, oversight committees, and contributors. In particular, thanks to all the people who have cho- sen to participate in STAMPEDE and their families and friends who have supported them. The views expressed are those of the authors and not necessarily those of the NIHR or the Department of Health and Social Care. Investigators and collaborators See Credit List included as S4 Text and on the STAMPEDE trial website: http://www. stampedetrial.org/media-section/presentation-repository/trial-recognition. Author Contributions Conceptualization: Chris C. Parker, Nicholas D. James, Noel W. Clarke, Peter Robson, Mahesh K. B. Parmar, Matthew R. Sydes. Data curation: Chris C. Parker, Nicholas D. James, Christopher D. Brawley, Noel W. Clarke, Adnan Ali, Claire L. Amos, Gerhardt Attard, Simon Chowdhury, Adrian Cook, William Cross, David P. Dearnaley, Duncan C. Gilbert, Clare Gilson, Silke Gillessen, Alex Hoyle, Rob J. Jones, Ruth E. Langley, Zafar I. Malik, Malcolm D. Mason, David Matheson, Robin Millman, Mary Rauchenberger, Hannah Rush, J Martin Russell, Hannah Sweeney, Amit Bahl, Alison Birtle, Lisa Capaldi, Omar Din, Daniel Ford, Joanna Gale, Ann Henry, Peter Hoskin, Mohammed Kagzi, Anna Lydon, Joe M. O’Sullivan, Sangeeta A. Paisey, Omi Parikh, Delia Pudney, Vijay Ramani, Peter Robson, Narayanan Nair Srihari, Jacob Tanguay, Mahesh K. B. Parmar, Matthew R. Sydes. Formal analysis: Christopher D. Brawley, Adrian Cook, Matthew R. Sydes. Funding acquisition: Chris C. Parker, Noel W. Clarke, David P. Dearnaley, Malcolm D. Mason, Mahesh K. B. Parmar, Matthew R. Sydes. Investigation: Chris C. Parker, Nicholas D. James, Christopher D. Brawley, Noel W. Clarke, Adnan Ali, Claire L. Amos, Gerhardt Attard, Adrian Cook, William Cross, David P. Dearnaley, Hassan Douis, Duncan C. Gilbert, Clare Gilson, Silke Gillessen, Alex Hoyle, Rob J. Jones, Ruth E. Langley, Zafar I. Malik, Malcolm D. Mason, David Matheson, Robin Millman, Mary Rauchenberger, Hannah Rush, J Martin Russell, Hannah Sweeney, Amit Bahl, Alison Birtle, Lisa Capaldi, Omar Din, Daniel Ford, Joanna Gale, Ann Henry, Peter Hoskin, Mohammed Kagzi, Anna Lydon, Joe M. O’Sullivan, Sangeeta A. Paisey, Omi Parikh, Delia Pudney, Vijay Ramani, Peter Robson, Narayanan Nair Srihari, Jacob Tanguay, Mahesh K. B. Parmar, Matthew R. Sydes. Methodology: Chris C. Parker, Nicholas D. James, Noel W. Clarke, Malcolm D. Mason, Mahesh K. B. Parmar, Matthew R. Sydes. Project administration: Chris C. Parker, Nicholas D. James, Christopher D. Brawley, Noel W. Clarke, Adnan Ali, Claire L. Amos, Clare Gilson, Alex Hoyle, Hannah Rush, Hannah Sweeney, Mahesh K. B. Parmar, Matthew R. Sydes. Software: Mary Rauchenberger. Supervision: Chris C. Parker, Nicholas D. James, Noel W. Clarke, Claire L. Amos, Gerhardt Attard, Simon Chowdhury, Adrian Cook, William Cross, David P. Dearnaley, Duncan C. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 18 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease Gilbert, Silke Gillessen, Rob J. Jones, Ruth E. Langley, Zafar I. Malik, David Matheson, Robin Millman, J Martin Russell, Mahesh K. B. Parmar, Matthew R. Sydes. Validation: Christopher D. Brawley, Noel W. Clarke, Adnan Ali, Adrian Cook, Alex Hoyle, Matthew R. Sydes. Visualization: Chris C. Parker, Christopher D. Brawley, Adrian Cook, Mahesh K. B. Parmar, Matthew R. Sydes. Writing – original draft: Chris C. Parker, Nicholas D. James, Christopher D. Brawley, Noel W. Clarke, Adrian Cook, Matthew R. Sydes. Writing – review & editing: Chris C. Parker, Nicholas D. James, Christopher D. Brawley, Noel W. Clarke, Adnan Ali, Claire L. Amos, Gerhardt Attard, Simon Chowdhury, Adrian Cook, William Cross, David P. Dearnaley, Duncan C. Gilbert, Clare Gilson, Silke Gillessen, Alex Hoyle, Ruth E. Langley, Zafar I. Malik, Malcolm D. Mason, David Matheson, Robin Millman, Mary Rauchenberger, Hannah Rush, J Martin Russell, Hannah Sweeney, Amit Bahl, Alison Birtle, Lisa Capaldi, Omar Din, Daniel Ford, Joanna Gale, Ann Henry, Peter Hoskin, Mohammed Kagzi, Anna Lydon, Joe M. O’Sullivan, Sangeeta A. Paisey, Omi Parikh, Delia Pudney, Vijay Ramani, Peter Robson, Narayanan Nair Srihari, Jacob Tanguay, Mahesh K. B. Parmar, Matthew R. Sydes. References 1. Parker C, Castro E, Fizazi K, Heidenreich A, Ost P, Procopio G, et al. Prostate cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann Oncol. 2020; 31(9):1119–34. Epub 2020/07/01. https://doi.org/10.1016/j.annonc.2020.06.011 PMID: 32593798. 2. Parker CC, James ND, Brawley CD, Clarke NW, Hoyle AP, Ali A, et al. Radiotherapy to the primary tumour for newly diagnosed, metastatic prostate cancer (STAMPEDE): a randomised controlled phase 3 trial. Lancet. 2018; 392(10162):2353–66. Epub 2018/10/26. https://doi.org/10.1016/S0140-6736(18) 32486-3 PMID: 30355464; PubMed Central PMCID: PMC6269599. 3. Parker CC, Sydes MR, Mason MD, Clarke NW, Aebersold D, de Bono JS, et al. Prostate radiotherapy for men with metastatic disease: a new comparison in the STAMPEDE trial. Clin Oncol. 2013; 25 (5):318–20. Epub 2013/03/16. https://doi.org/10.1016/j.clon.2013.01.005 PMID: 23489869. 4. Cox JD, Stetz J, Pajak TF. Toxicity criteria of the Radiation Therapy Oncology Group (RTOG) and the European organization for research and treatment of cancer (EORTC). Int J Radiat Oncol Biol Phys. 1995; 31 (5):1341–6. https://doi.org/10.1016/0360-3016(95)00060-C PMID: 7713792 5. Sweeney CJ, Chen YH, Carducci M, Liu G, Jarrard DF, Eisenberger M, et al. Chemohormonal Therapy in Metastatic Hormone-Sensitive Prostate Cancer. N Engl J Med. 2015; 373 (8):737–46. https://doi.org/ 10.1056/NEJMoa1503747 PMID: 26244877. 6. Royston P, Parmar MKB. The use of restricted mean survival time to estimate the treatment effect in randomized clinical trials when the proportional hazards assumption is in doubt. Stat Med. 2011; 30:2409, Epub 2421–21. https://doi.org/10.1002/sim.4274 PMID: 21611958 7. Morris TP, Jarvis CI, Cragg W, Phillips PPJ, Choodari-Oskooei B, Sydes MR. Proposals on Kaplan- Meier plots in medical research and a survey of stakeholder views: KMunicate. BMJ Open. 2019; 9(9): e030215. Epub 2019/10/03. https://doi.org/10.1136/bmjopen-2019-030215 PMID: 31575572; PubMed Central PMCID: PMC6773317. 8. Gray R. A Class of K-Sample Tests for Comparing the Cumulative Incidence of a Competing Risk. Ann Stat. 1988; 16 (3):1141–54. 9. Rush HL, Murphy L, Morgans AK, Clarke NW, Cook AD, Attard G, et al. Quality of life for men with pros- tate cancer contemporaneously randomly allocated to receive either docetaxel or abiraterone in the STAMPEDE trial. J Clin Oncol. 2021. https://doi.org/10.1200/JCO.21.00728 PMID: 34757812 10. Boeve LMS, Hulshof M, Vis AN, Zwinderman AH, Twisk JWR, Witjes WPJ, et al. Effect on Survival of Androgen Deprivation Therapy Alone Compared to Androgen Deprivation Therapy Combined with Con- current Radiation Therapy to the Prostate in Patients with Primary Bone Metastatic Prostate Cancer in a Prospective Randomised Clinical Trial: Data from the HORRAD Trial. Eur Urol. 2019; 75(3):410–8. Epub 2018/09/30. https://doi.org/10.1016/j.eururo.2018.09.008 PMID: 30266309. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 19 / 20 PLOS MEDICINERadiotherapy to the prostate for metastatic disease 11. Boeve L, Hulshof M, Verhagen P, Twisk JWR, Witjes WPJ, de Vries P, et al. Patient-reported Quality of Life in Patients with Primary Metastatic Prostate Cancer Treated with Androgen Deprivation Therapy with and Without Concurrent Radiation Therapy to the Prostate in a Prospective Randomised Clinical Trial; Data from the HORRAD Trial. Eur Urol. 2021; 79(2):188–97. Epub 20200922. https://doi.org/10. 1016/j.eururo.2020.08.023 PMID: 32978014. 12. Wilkins A, Mossop H, Syndikus I, Khoo V, Bloomfield D, Parker C, et al. Hypofractionated radiotherapy versus conventionally fractionated radiotherapy for patients with intermediate-risk localised prostate cancer: 2-year patient-reported outcomes of the randomised, non-inferiority, phase 3 CHHiP trial. Lan- cet Oncol. 2015; 16(16):1605–16. Epub 2015/11/03. https://doi.org/10.1016/S1470-2045(15)00280-6 PMID: 26522334; PubMed Central PMCID: PMC4664817. 13. Ali A, Hoyle A, Haran AM, Brawley CD, Cook A, Amos C, et al. Association of Bone Metastatic Burden With Survival Benefit From Prostate Radiotherapy in Patients With Newly Diagnosed Metastatic Pros- tate Cancer: A Secondary Analysis of a Randomized Clinical Trial. JAMA Oncol. 2021; 7(4):555–63. Epub 2021/02/19. https://doi.org/10.1001/jamaoncol.2020.7857 PMID: 33599706; PubMed Central PMCID: PMC7893550. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1003998 June 7, 2022 20 / 20 PLOS MEDICINE	Could not heal snippet
10.7554_elife.86784.pdf	Data availability Cryo- EM maps and atomic coordinates for the GC- C- Hsp90- Cdc37 complex have been deposited in the EMDB (EMD- 29523) and PDB (8FX4). Material availability: The plasmids used in this study are uploaded in (Supplementary file 1).	Data availability Cryo-EM maps and atomic coordinates for the GC-C-Hsp90-Cdc37 complex have been deposited in the EMDB (EMD-29523) and PDB (8FX4). Material availability: The plasmids used in this study are uploaded in (Supplementary file 1). The following datasets were generated:	RESEARCH ARTICLE Structural insight into guanylyl cyclase receptor hijacking of the kinase–Hsp90 regulatory mechanism Nathanael A Caveney1, Naotaka Tsutsumi1,2†, K Christopher Garcia1,2 1Departments of Molecular and Cellular Physiology, and Structural Biology, Stanford University School of Medicine, Stanford, United States; 2Howard Hughes Medical Institute, Stanford University School of Medicine, Stanford, United States Abstract Membrane receptor guanylyl cyclases play a role in many important facets of human physiology, from regulating blood pressure to intestinal fluid secretion. The structural mechanisms which influence these important physiological processes have yet to be explored. We present the 3.9 Å resolution cryo- EM structure of the human membrane receptor guanylyl cyclase GC- C in complex with Hsp90 and its co- chaperone Cdc37, providing insight into the mechanism of Cdc37 mediated binding of GC- C to the Hsp90 regulatory complex. As a membrane protein and non- kinase client of Hsp90–Cdc37, this work shows the remarkable plasticity of Cdc37 to interact with a broad array of clients with significant sequence variation. Furthermore, this work shows how membrane receptor guanylyl cyclases hijack the regulatory mechanisms used for active kinases to facilitate their regulation. Given the known druggability of Hsp90, these insights can guide the further development of membrane receptor guanylyl cyclase- targeted therapeutics and lead to new avenues to treat hypertension, inflammatory bowel disease, and other membrane receptor guanylyl cyclase- related conditions. eLife assessment In this important study, the human membrane receptor guanyl cyclase GC- C was expressed in hamster cells, co- purified in complex with endogenous HSP90 and CDC37 proteins, and the struc- ture of the complex was determined by cryo- EM. The study shows that the pseudo- kinase domain of GC- C associates with CDC37 and HSP90, similarly to how the bona fide protein kinases CDK4, CRAF and BRAF have been shown to interact. The methodology used is state of the art and the evidence presented is compelling. Introduction Cyclic guanosine monophosphate (cGMP) is an important second messenger for signaling in mamma- lian physiology, with roles in platelet aggregation, neurotransmission, sexual arousal, gut peristalsis, bone growth, intestinal fluid secretion, lipolysis, phototransduction, cardiac hypertrophy, oocyte maturation, and blood pressure regulation (Potter, 2011). Largely, cGMP is produced in response to the activation of guanylyl cyclases (GC), a class of receptors that contains both heteromeric soluble receptors (α1, α2, β1, and β2 in humans) and five homomeric membrane receptors (GC- A, GC- B, GC- C, GC- E, and GC- F in humans). Of note are the membrane receptor guanylyl cyclases (mGC) GC- A and GC- B, also known as natriuretic peptide receptors A and B (NPR- A and NPR- B), respectively, and GC- C, all of which have been a focus of therapeutic development. In the case of NPR- A and B, their role in regulating blood pressure in response to natriuretic peptide hormones (ANP, BNP, and *For correspondence: [email protected] (NAC); [email protected] (KCG) Present address: †Graduate School of Medicine, Dentistry and Pharmaceutical Sciences, Okayama University, Okayama, Japan Competing interest: The authors declare that no competing interests exist. Funding: See page 9 Sent for Review 25 February 2023 Preprint posted 16 March 2023 Reviewed preprint posted 02 May 2023 Reviewed preprint revised 12 July 2023 Version of Record published 03 August 2023 Reviewing Editor: Mohamed Trebak, University of Pittsburgh, United States Copyright Caveney et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 1 of 12 Research article CNP) has led to the exploration of agonists for use in the treatment of cardiac failure (Kobayashi et al., 2012). Meanwhile, GC- C is the target of clinically approved laxative agonists, linaclotide, and plecanatide (Miner, 2020; Yu and Rao, 2014), which increase intestinal fluid secretion. These membrane receptor GCs consist of an extracellular ligand binding domain (ECD), which acts as a conformational switch to drive intracellular rearrangements to activate the receptor (He et al., 2001) a transmembrane region (TM); a kinase homology domain or pseudokinase domain (PK); a dimerization domain; and a GC domain, which acts to produce cGMP. The PK domain is largely thought to be involved in scaffolding and physical transduction of the extracellular rearrangements to the GC domain, in some respects similar to the role of the PK domain in the Janus kinases of the cytokine signaling system (Glassman et al., 2022). In addition, the PK domains of mGCs are regulated through phosphorylation (Potter and Garbers, 1992; Potter and Hunter, 1998; Vaandrager et al., 1993) and via association with heat shock proteins (Hsp) (Kumar et al., 2001). While the role of the phosphorylation state on mGC activity has been explored in relative detail, how the heat shock protein 90 (Hsp90) is able to regulate mGC activity is largely unknown. It has been shown that GC- A activity can be regulated through the association of Hsp90 and the co- chap- erone Cdc37 (Kumar et al., 2001). The chaperone Cdc37 is known to assist in the Hsp90 regulation of around 60% of active kinases, both in soluble and membrane receptor form (Taipale et al., 2012). Given the sequence and structural similarities between the PK domains of mGCs and the active kinase domains which Hsp90–Cdc37 regulates, it is possible that mGCs have evolved to hijack the regulatory mechanisms that are more broadly deployed for active kinases. Here, we report the 3.9 Å resolution structure of the GC- C–Hsp90–Cdc37 regulatory complex. In this structure, the core dimer of Hsp90 forms its canonical closed conformation, while Cdc37 and the C- lobe of the GC- C PK domain asymmetrically decorate the complex. The client (GC- C) is unfolded into the channel formed at the interface between the Hsp90 dimers. To our knowledge, this is the first structure of a membrane protein client of Hsp90 and the first structure of a non- kinase client of the Hsp90–Cdc37 regulatory system. This work provides a pivotal understanding of the mechanism and structural basis of kinase fold recruitment to the Hsp90–Cdc37 regulatory complex. This increased understanding can guide the further development of mGC- targeted therapeutics and lead to new avenues to treat hypertension, inflammatory bowel disease (IBD), and other mGC- related conditions. In addition, the general insights into the recruitment of Hsp90–Cdc37 clients can guide the further development of Hsp90 targeting therapeutics in cancer treatment. Results Structure of the GC-C–Hsp90–Cdc37 regulatory complex Membrane receptor guanylyl cyclases have been largely recalcitrant to structural analysis by x- ray crystallography and electron microscopy, apart from various crystal structures of both liganded and unliganded ECDs (He et al., 2001; He et al., 2006; Ogawa et al., 2004; Ogawa et al., 2010; van den Akker et al., 2000). Given the relative disparity of our structural understanding, we sought to develop a stable construct to image and gain a crucial understanding of the regulatory and functional aspects of mGCs which occur intracellularly. By replacing the ligand- responsive ECD with a homod- imeric leucine zipper, we mimic the ligand- activated geometry of the ECD (He et al., 2001), while reducing complexity of the imaged complex and increasing stability (Figure 1A). This complex was recombinantly expressed in mammalian cells, purified with anti- FLAG affinity chromatography, and vitrified on grids for cryo- EM analysis. The purified sample had a substantial portion of imaged particles for which the native regulatory heat shock protein, Hsp90, and its co- chaperone, Cdc37, are bound. The Cricetulus griseus HSP90β and Cdc37 show remarkable sequence conservation in comparison to the human equivalents, at 99.7 and 94.2% identity, respectively. This native pulldown strategy contrasts with the structures of Hsp90– Cdc37 in complex with soluble kinases (García- Alonso et al., 2022; Oberoi et al., 2022; Verba et al., 2016), for which Hsp90 and Cdc37 had to be overexpressed to obtain complex suitable for imaging. Three- dimensional reconstruction of our GC- C–Hsp90–Cdc37 particles generated a 3.9 Å resolution map of the regulatory complex (Figure 1, Figure 1—figure supplements 1 and 2). A second, unsharpened map from subsequent heterogeneous refinement resolves additional density Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 2 of 12 Biochemistry and Chemical Biology Research article A ECD TM TM PK DD GC B GC-C Gn, Uro zipper zipper Cdc37 Cdc37 TMTM TM TM PK PK DD DD GC GC PK PK DD DD GC GC Hsp90 Hsp90 Hsp90 Hsp90 GTP cGMP signal C GC-C 90Å D NTD NTD Cdc37 coiled- coil MD Hsp90 Hsp90 Cdc37 coiled- coiled- coil coil MD CTD DDDD PK PK C-lobe C-lobe E NTD NTD Cdc37 coiled- coil MD MD GC-C PK C-lobe CTD 90° NTD MD MD MD GC-C PK PK C-lobe C-lobe MD 90° GC-C PK C-lobe CTD CTD CTD MD CTD Figure 1. Composition and cryo- EM structure of the GC- C–Hsp90–Cdc37 regulatory complex. (A) Cartoon representation of the components of guanylyl cyclase C (GC- C) signaling and Hsp90–Cdc37 regulation and the zippered and activated GC- C. GC- C is colored in red, guanylin/uroguanylin (Gn/Uro) in yellow, Hsp90 in blue and teal, and Cdc37 in purple. Extracellular domains (ECD), transmembrane domain (TM), pseudokinase domain (PK), dimerization domain (DD), and guanylyl cyclase domain (GC) are labeled. In the rightmost cartoon, the regions unobserved in the cryo- EM density are in a lighter shade with a dashed outline. (B) The refined and sharpened cryo- EM density map of GC- C–Hsp90–Cdc37, colored as in A, with a transparent overlay of an unsharpened map with additional DD density resolved. Cdc37 coil- coiled and middle domain (MD) are labeled. (C) Reference- free 2D averages for the GC- C–Hsp90–Cdc37 complex. (D) The refined and sharpened cryo- EM density map of GC- C– Hsp90–Cdc37, colored as in A and B, labeled with all domains as in A and B, with the addition of Hsp90 N- terminal domain (NTD), middle domain (MD), and C- terminal domain (CTD). (E) Ribbon representation of a model of GC- C– Hsp90–Cdc37 complex, colored and labeled as in A, B, and C. The online version of this article includes the following figure supplement(s) for figure 1: Figure 1 continued on next page Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 3 of 12 Biochemistry and Chemical Biology Research article Figure 1 continued Figure supplement 1. GC- C–Hsp90–Cdc37 complex cryo- EM data processing. Figure supplement 2. Representative density of GC- C–Hsp90–Cdc37. for the dimerization domain, extending outward from the PK domain (Figure 1B, Figure 1—figure supplement 1). The resultant GC- C–Hsp90–Cdc37 complex is a hetero- tetramer formed by one resolved monomer of the GC- C receptor bound to a dimer of Hsp90 and one Cdc37 co- chaperone (Figure 1D). As observed with most Hsp90–client structures, the bulk of the complex is composed of the C2 pseudo- symmetric, ATP bound, closed state Hsp90 dimer. Building on this dimeric core, the Cdc37 protrudes outward from one side with its characteristic long, coiled- coil, α-hairpin. On one face of the Hsp90 dimer core, Cdc37 interacts with the PK domain of GC- C, while an extended β-sheet wraps around to the other face, lying across and extending a β-sheet in the middle domain (MDHsp90) of one Hsp90 monomer. At the opposite face, the globular and α-helical Cdc37 middle domain (MDCdc37) is formed. The C- lobe of the GC- C PK domain packs against the N- terminal region of Cdc37 on one face of the dimeric Hsp90 core, with the N- lobe unfolding through the dimer core to interface with the MDCdc37 on the opposite face. N- terminal to the PK N- lobe is the TM region, the density for which was unob- served in our reconstructions. C- terminal to the PK C- lobe, we observe some poorly resolved density for the likely mobile dimerization domain in our unsharpened map. This would precede the GC domain, which is not observed in the density of our reconstructions (Figure 1B). Together, we can use our understanding of mGC topology and our reconstruction to orient the complex as it would sit on a membrane (Figure 1B), providing insight into how Hsp90 is able to access and regulate membrane protein clients. No density is observed for the second GC- C of the dimer, though it is sterically unlikely that an additional regulatory complex is forming on the second GC- C in a concurrent fashion, given the large size of the first Hsp90–Cdc37 and the requisite proximity of the second GC- C. In addi- tion, this disruption of the native state of GC- C, as observed in our structure, would likely leave GC domains out of each other’s proximity, precluding their catalytic activity while Hsp90 is bound. Cdc37 mediated GC-C recruitment and Hsp90 loading Despite the recognized plasticity of Cdc37 co- chaperone binding to approximately 60% of kinases (Taipale et al., 2012), the importance of the Hsp90–Cdc37 complex for pseudokinase domain- containing proteins in the human proteome is not well studied. Thus, the structural basis for how Cdc37 can recruit GC- C to the Hsp90 regulatory complex is of particular interest. In our structures, we see that Cdc37 is displacing the N- lobe of the pseudokinase domain of GC- C, binding to the C- lobe at the N–C interface, and guiding the unfolded N- lobe into the Hsp90 dimer (Figure 2). The Cdc37–GC- C interface is relatively modest in size, with a calculated mean surface area of 689 Å2 (as calculated by PISA Krissinel and Henrick, 2007). This interface is partly driven to form via charge complementarity, with positive contributions from a cluster of arginine residues on Cdc37 (R30, R32, R39) at the periphery of the interaction interface interacting with D609 and the polar residues Y580 and T586 (Figure 2B). Beyond this, the interface is likely largely driven via shape- complementarity, due to a minimal contribution from hydrogen bonding, salt- bridge formation, and aromatic packing contributions – in line with the ability of Cdc37 to chaperone such a diverse array of clients and client sequences. As the unfolded PK N- lobe extends away from Cdc37, it enters the channel formed at the interface between the dimer of Hsp90 (Figure 2C). Here, GC- C residues 528–544 (VKLDTMIFGVIEYCERG) lie across the upper region of the Hsp90 CTDs, which form the floor of the channel. These CTDs form the bulk of the interaction interface as the unfolded N- lobe passes through this channel, yet there are minor contributions from the loop regions of the β-sheet from the MDHsp90 which extend downward into this channel region. The unfolded region is relatively poorly resolved in the density, with some reconstructions from earlier refinement having no resolvable density in this channel region – indicative of the low stability and high mobility of the unfolded N- lobe as it passes through this region. Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 4 of 12 Biochemistry and Chemical Biology Research article A Hsp90 Hsp90 NTD Hsp90 NTD Cdc37 coiled- coil B Cdc37 R39 B GC-C PK C-lobe MD C MD T586T586 L581 T608 Y580 K585 D609D609 W31W31 R30 R30 R32R32 L28L28 I23I23 E18E18 N22N22 H20H20 K606 V604V604 K576K576 GC-C E Y C EE II R G544G544 GC-CGC-C Hsp90 Hsp90 CTD CTD C Hsp90 Hsp90 V528V528 D TT KK LL I M G V F Figure 2. Cdc37 mediated guanylyl cyclase C (GC- C) recruitment and heat shock protein 90 (Hsp90) loading interfaces. (A) Ribbon representation of a model of GC- C–Hsp90–Cdc37 complex. GC- C is colored in red, Hsp90 in blue and teal, and Cdc37 in purple. Pseudokinase (PK), coil- coiled, middle (MD), C- terminal (CTD), and N- terminal (NTD) domains are labeled. (B) The Cdc37–GC- C interface in ribbon representation, with interacting residues drawn in sticks, colored as in A. (C) The unfolded N- lobe of GC- C PK domain as it passes between the Hsp90 dimer, in ribbon representation, with interacting residues drawn in sticks, colored as in A and B. This region’s sequence is: VKLDTMIFGVIEYCERG. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Conservation of Cdc37 mediated heat shock protein 90 (Hsp90) regulation. Figure supplement 2. Regulatory mechanisms for membrane receptor guanylyl cyclase (mGC) activity. Conservation of Cdc37 mediated Hsp90 regulation The core structural principles of Cdc37 mediated client recruitment to Hsp90 appear to remain constant across its large range of client diversity. Across other clients–Hsp90–Cdc37 complexes with canonical soluble kinase clients (Cdk4, RAF1, B- raf) (García- Alonso et al., 2022; Oberoi et al., 2022; Verba et al., 2016), we see a conserved role for Cdc37 in client recruitment by associating with the C- lobe at the N-, C- lobe interface (Figure 2—figure supplement 1A, B). In these complexes, we see high levels of structural conservation for the Hsp90–Cdc37 (Cα RMSDs of 1.4–3.3 Å for Hsp90 and 1.5–2.5 Å for Cdc37), while the client is structurally most homogenous at the interface with Cdc37, though less structurally conserved overall (Cα RMSDs of 3.5–11.6 Å). Perhaps unsurprisingly, GC- C is one of the most divergent of these clients from a sequence perspective (Figure 2—figure supplement 1C), with sequence homology between the GC- C PK domain and the other client kinase domains ranging from 19 to 25% identity and 31 to 41% homology. This highlights the plasticity required of this system which can service such a vast array of clients across a broad range of sequence variations, yet more restricted fold architecture. Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 5 of 12 Biochemistry and Chemical Biology Research article Discussion The present cryo- EM structure of GC- C–Hsp90–Cdc37 resolves the loading of GC- C, via its PK domain and interaction with Cdc37, to the Hsp90 core dimer (Figures 1 and 2). This complex shows signifi- cant structural similarity to the mechanism that regulates soluble active kinases (García- Alonso et al., 2022; Oberoi et al., 2022; Verba et al., 2016) and presumably membrane receptor kinases in the human proteome. This structural and mechanistic conservation is largely driven by the co- chaperone Cdc37, which serves as the central binding platform for these clients by associating to the fold of the kinase (or pseudokinase in the case of mGC) domain, relatively independent of sequence identity. A model whereby recruitment is largely driven by both the fold complementarity and the specific stability properties of the kinase fold has been proposed previously (Taipale et al., 2012). In this model, instability of a fully folded kinase domain results in partial unfolding of the C- lobe, leading Cdc37 to bind the partially unfolded state. Given the lack of functional and sequence conservation for GC- C as a client of Cdc37, our data largely fits with this model for client recruitment. It is likely that the pseudokinase domains of mGC have largely evolved to facilitate regulatory mechanisms for these receptors, both via their phosphorylation and by hijacking the regulatory mechanisms used by active soluble and membrane receptor kinases. In the case of GC- A, previous work has shown that it associates with the Hsp90–Cdc37 complex to regulate GC activity (Kumar et al., 2001). The authors showed that adding geldanamycin, an Hsp90 inhibitor, reduces the overall cGMP output of cells in response to ANP stimulation while also reducing the association of the Hsp90 to GC- A. While this initially may seem counterintuitive, this data fits with a model of ligand- induced activity potentiating the instability of the PK domain, which then facilitates binding of the regulatory complex to ‘re- fold’ GC- A for further catalysis and cGMP produc- tion – in a core regulatory complex structurally similar to that which we observe for GC- C in this work (Figure 2—figure supplement 2). In the case of the Hsp90 inhibitor, this would release the Hsp90 and only allow full catalytic activity for the receptor until the receptor falls into the partially unfolded state, as the Hsp90 would no longer be able to re- engage at the C- lobe when inhibited (Figure 2—figure supplement 2). Interestingly there may be an additional layer of regulation involved, with crosstalk between the phosphorylation and Hsp90 regulatory mechanisms of mGC. The phosphatase PP5 is known to interact with the Hsp90–Cdc37 system and dephosphorylate Hsp90, Cdc37, and the system’s kinase clients (Oberoi et al., 2022). PP5 has been implicated in this role for mGC (Chinkers, 1994), though this interaction was unable to be detected by a pull- down in a second study (Kumar et al., 2001). In this way, mGC association with the Hsp90–Cdc37 complex could result in multiple fates and resultant activity profiles for the receptor. When the PK of an activated mGC falls into a destabilized state, this would result in the recruitment of the Hsp90–Cdc37. First, the regula- tory complex could refold the receptor to maintain the activity of the receptor (Figure 2—figure supplement 2i). In another scenario, the Hsp90–Cdc37 complex could additionally recruit PP5 to dephosphorylate the mGC (Figure 2—figure supplement 2ii). Particularly in the case of GC- A and GC- B, and to some extent GC- C (Potter and Garbers, 1992; Potter and Hunter, 1998; Vaandrager et al., 1993), this would impair the signaling activity of the mGC, though this could be rescued through the kinase re- association and phosphorylation. In a final scenario, the binding of the Hsp90–Cdc37 complex could result in the association of ubiquitin E3 ligases (Schopf et al., 2017; Figure 2—figure supplement 2iii), which would ubiquitinate the mGC client, leading to the removal of the receptor. The regulation of mGC is influenced by a network of factors working in harmony to ensure proper signaling and physiological response for these important receptors. The structure of the core regu- latory complex shown in this work is key to many facets of mGC regulation. We hope that the struc- tural basis for the Hsp90 regulatory platform for mGC will drive renewed investigation into these diverse mechanisms and lead to the therapeutic manipulation of these mechanisms to improve mGC targeting therapies. Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 6 of 12 Biochemistry and Chemical Biology Research article Key resources table Reagent type (species) or resource Methods Designation Source or reference Identifiers Additional information Cell line (Cricetulus griseus) Chinese hamster ovary kidney cells GIBCO ExpiCHO Recombinant DNA reagent pD649- GCN4- TM- GC- C_ICD (plasmid) This paper See: Methods - Cloning and protein expression Software, algorithm Data collection software SerialEM SerialEM Software, algorithm Data processing software Software, algorithm Data sharpening software Structura Biotechnology Inc. cryoSPARC Sanchez- Garcia et al., 2021 DeepEMhancer Software, algorithm Initial modeling software Jumper et al., 2021 AlphaFold Software, algorithm Graphics software Pettersen et al., 2021 UCSF ChimeraX Software, algorithm Software, algorithm Modeling and refinement software Adams et al., 2010 Phenix Modeling and refinement software Emsley and Cowtan, 2004 Coot Software, algorithm Model validation software Chen et al., 2010 MolProbity Cloning and protein expression For cryo- EM studies, a construct containing an HA secretion signal (MKTIIALSYIFCLVFA), a FLAG peptide (DYKDDDD), linker and 3 C cleavage site (KGSLEVLFQGPG), GCN4 homodimeric zipper ( RMKQ LEDK VEEL LSKN YHLE NEVA RLKK LVGER), human GC- C regions corresponding to the small extracellular linker region, TM, and intracellular domains (residues 399–1,053), a second linker and 3 C cleavage site (AAALEVLFQGPGAA), a Protein C epitope tag (EDQVDPRLIDGK), and an 8 x His tag were cloned into a pD649 mammalian expression vector. This construct contains all domains of the native GC- C, with the exception of the ECD (Supplementary file 1). Protein was expressed using ExpiCHO Expression System Kit (Thermo Fisher). Briefly, ExpiCHO cells were maintained in ExpiCHO Expression Media at 37 °C with 5% CO2 and gentle agitation, and transiently transfected by the expression construct and cultured according to the manufacturer’s protocol. Cells were pelleted and stored at –80 °C. Protein purification Cells were resuspended in 20 mM HEPES- Na pH 8.0, 300 mM NaCl, 1 mM TCEP, protease inhibitor cocktail (Sigma), and benzonase (Sigma). Cells were lysed by Dounce homogenizer and cellular debris was pelleted by low- speed centrifugation at 500 × g. Membranes were collected by centrifugation at 46,000 × g and stored at –80 °C until use. Membranes were thawed and solubilized with the addition of 1% n- dodecyl β-D- maltoside (DDM) and 0.1% cholesteryl hemisuccinate (CHS) (10:1) (Anatrace). Debris and unsolubilized membranes were pelleted by centrifugation at 46,000 × g. The superna- tant was subsequently used in FLAG affinity chromatography. The supernatant was applied to M1 anti- FLAG resin. The resin was washed with 20 bed volumes of 20 mM HEPES- Na pH 8.0, 300 mM NaCl, 1 mM TCEP, 0.005% lauryl maltose neopentyl glycol (LMNG), 0.0005% CHS (10:1) (Anatrace), and 5 mM ATP. The protein complex was eluted with the addition of 200 μg/mL of FLAG peptide (DYKDDDD) (GenScript). Protein was subsequently concentrated to >2 mg/mL and used for cryo- EM imaging. Cryo-electron microscopy Aliquots of 3 μL of complex were applied to glow- discharged 300 mesh UltrAuFoil (1.2/1.3) grids. The grids were blotted for 3 s at 100% humidity with an offset of 3 and plunged frozen into liquid ethane using a Vitrobot Mark IV (Thermo Fisher). Grid screening and dataset collection occurred at Stanford cEMc on a 200 kV Glacios microscope (Thermo Fisher) equipped with a K3 camera (Gatan). Movies Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 7 of 12 Biochemistry and Chemical Biology Research article Table 1. Cryo- EM data collection, refinement, and validation statistics. GC- C–Hsp90–Cdc37 complex PDB 8FX4 EMD- 29523 GC- C–Hsp90–Cdc37 complex with DD density Data collection and processing Nominal magnification Acceleration voltage (kV) Electron exposure (e-/Å2) Defocus range (µm) Pixel size (Å) Symmetry imposed Final particle images 165,635 Map resolution FSC threshold Map resolution (Å) 3.9 Refinement Initial model used (PDB) 5FWK, 7ZR5, AlphaFold 45,000 200 58.8 0.8–2.0 0.9273 C1 0.143 48,283 6.3 Model resolution FSC threshold (Å) Model resolution (Å) Model Composition Non- hydrogen atoms Protein residues Ligands B- factors (Å2) Protein Ligand R.m.s. deviations Bond lengths (Å) Bond angles (°) Validation MolProbity score Clashscore Rotamer outliers (%) Ramachandran plot Favored (%) Allowed (%) Outliers (%) 0.5 4.2 13,478 1,654 2 119.49 102.85 0.004 0.914 2.14 13.88 0.67 92.0 7.6 0.4 Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 8 of 12 Biochemistry and Chemical Biology Research article were collected at a magnification corresponding to a 0.9273 Å per physical pixel. The dose was set to a total of 58.8 electrons per Å2. Automated data collection was carried out using SerialEM with a nominal defocus range set from –0.8 to –2.0 μM. Image processing All processing was performed in cryoSPARC (Punjani et al., 2017) unless otherwise noted (Figure 1— figure supplement 1). 8788 movies were motion- corrected using patch motion correction. The contrast transfer functions (CTFs) of the flattened micrographs were determined using patch CTF and an initial stack of particles was picked using Topaz picker (Bepler et al., 2019). Successive rounds of reference- free 2D classification were performed to generate a particle stack of 165,635 particles. These particles were then used in ab- initio reconstruction, followed by non- uniform refinement (Punjani et al., 2020) and finally local refinement with a loose mask around the entire complex. This resulted in a 3.9 Å reconstruction of the GC- C–Hsp90–Cdc37 complex which was sharpened with deepEMhancer (Sanchez- Garcia et al., 2021). These particles were also used in a 4- class heterogeneous refinement to pull out a volume containing some resolved density for the dimerization domain of GC- C. Model building and refinement The Cdk4–Hsp90β–Cdc37 (PDB 5FWK), PP5–B- Raf–Hsp90β–Cdc37 (PDB 7ZR5), and AlphaFold models for GC- C (Jumper et al., 2021; Mirdita et al., 2022) were docked into the map using UCSF Chimera X (Pettersen et al., 2021). A resultant hybrid model was then manually curated to contain the correct Cricetulus griseus sequences for Hsp90β–Cdc37 and run through Namdinator (Kidmose et al., 2019). This was followed by automated refinement using Phenix real space refine (Adams et al., 2010) and manual building in Coot (Emsley and Cowtan, 2004). The final model produced a favorable MolProbity score of 2.14 (Chen et al., 2010) with 0.4% Ramachandran outliers (Table 1). Model building and refinement software was installed and configured by SBGrid (Morin et al., 2013). Acknowledgements We thank Liz Montabana and Stanford cEMc for microscope access for data collection. We thank Paul LaPointe and Kevin Jude for their insightful discussion of the Hsp90 structure and regulatory mech- anisms. NAC is a CIHR postdoctoral fellow. KCG is an investigator with the Howard Hughes Medical Institute. KCG is supported by National Institutes of Health grant R01- AI51321, the Mathers Founda- tion, and the Ludwig Foundation. Additional information Funding Funder Canadian Institutes of Health Research National Institutes of Health Mathers Foundation Ludwig Foundation Grant reference number Author Postdoctoral Fellowship Nathanael A Caveney R01-AI51321 K Christopher Garcia K Christopher Garcia K Christopher Garcia The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Nathanael A Caveney, Conceptualization, Formal analysis, Investigation, Methodology, Writing - orig- inal draft, Writing – review and editing; Naotaka Tsutsumi, Formal analysis, Investigation, Writing – review and editing; K Christopher Garcia, Supervision, Funding acquisition, Project administration, Writing – review and editing Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 9 of 12 Biochemistry and Chemical Biology Research article Author ORCIDs Nathanael A Caveney Naotaka Tsutsumi K Christopher Garcia http://orcid.org/0000-0003-4828-3479 https://orcid.org/0000-0002-3617-7145 https://orcid.org/0000-0001-9273-0278 Peer review material Reviewer #1 (Public Review): https://doi.org/10.7554/eLife.86784.3.sa1 Reviewer #2 (Public Review): https://doi.org/10.7554/eLife.86784.3.sa2 Reviewer #3 (Public Review): https://doi.org/10.7554/eLife.86784.3.sa3 Author Response https://doi.org/10.7554/eLife.86784.3.sa4 Additional files Supplementary files • MDAR checklist • Supplementary file 1. Plasmids used in this study. Data availability Cryo- EM maps and atomic coordinates for the GC- C- Hsp90- Cdc37 complex have been deposited in the EMDB (EMD- 29523) and PDB (8FX4). Material availability: The plasmids used in this study are uploaded in (Supplementary file 1). The following datasets were generated: Author(s) Caveney NA, Garcia CK Year 2023 Caveney NA, Garcia KC 2023 Dataset title Dataset URL Database and Identifier Cryo- EM maps and atomic coordinates for the GC- C- Hsp90- Cdc37 complex have been deposited Cryo- EM maps and atomic coordinates for the GC- C- Hsp90- Cdc37 complex have been deposited https://www. ebi. ac. uk/ emdb/ EMD- 29523 EMDataResource, EMD- 29523 https://www. rcsb. org/ structure/ 8FX4 RCSB Protein Data Bank, 8FX4 References Adams PD, Afonine PV, Bunkóczi G, Chen VB, Davis IW, Echols N, Headd JJ, Hung LW, Kapral GJ, Grosse- Kunstleve RW, McCoy AJ, Moriarty NW, Oeffner R, Read RJ, Richardson DC, Richardson JS, Terwilliger TC, Zwart PH. 2010. PHENIX: A comprehensive python- based system for macromolecular structure solution. Acta Crystallographica. Section D, Biological Crystallography 66:213–221. DOI: https://doi.org/10. 1107/S0907444909052925, PMID: 20124702 Aiyer S, Zhang C, Baldwin PR, Lyumkis D. 2021. Evaluating local and directional resolution of cryo- EM density maps. Methods in Molecular Biology 2215:161–187. DOI: https://doi.org/10.1007/978-1-0716-0966-8_8, PMID: 33368004 Bepler T, Morin A, Rapp M, Brasch J, Shapiro L, Noble AJ, Berger B. 2019. Positive- unlabeled convolutional neural networks for particle picking in cryo- electron micrographs. Nature Methods 16:1153–1160. DOI: https:// doi.org/10.1038/s41592-019-0575-8, PMID: 31591578 Chen VB, Arendall WB, Headd JJ, Keedy DA, Immormino RM, Kapral GJ, Murray LW, Richardson JS, Richardson DC. 2010. MolProbity: all- atom structure validation for macromolecular crystallography. Acta Crystallographica. Section D, Biological Crystallography 66:12–21. DOI: https://doi.org/10.1107/ S0907444909042073, PMID: 20057044 Chinkers M. 1994. Targeting of a distinctive protein- serine phosphatase to the protein kinase- like domain of the atrial natriuretic peptide receptor. PNAS 91:11075–11079. DOI: https://doi.org/10.1073/pnas.91.23.11075, PMID: 7972012 Emsley P, Cowtan K. 2004. Coot: model- building tools for molecular graphics. Acta Crystallographica. Section D, Biological Crystallography 60:2126–2132. DOI: https://doi.org/10.1107/S0907444904019158, PMID: 15572765 García- Alonso S, Mesa P, de la Puente Ovejero L, Aizpurua G, Lechuga CG, Zarzuela E, Santiveri CM, Sanclemente M, Muñoz J, Musteanu M, Campos- Olivas R, Martínez- Torrecuadrada J, Barbacid M, Montoya G. 2022. Structure of the RAF1- HSP90- CDC37 complex reveals the basis of RAF1 regulation. Molecular Cell 82:3438–3452. DOI: https://doi.org/10.1016/j.molcel.2022.08.012, PMID: 36055235 Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 10 of 12 Biochemistry and Chemical Biology Research article Glassman CR, Tsutsumi N, Saxton RA, Lupardus PJ, Jude KM, Garcia KC. 2022. Structure of a Janus kinase cytokine receptor complex reveals the basis for dimeric activation. Science 376:163–169. DOI: https://doi.org/ 10.1126/science.abn8933, PMID: 35271300 He XI, Chow DC, Martick MM, Garcia KC. 2001. Allosteric activation of a spring- loaded natriuretic peptide receptor dimer by hormone. Science 293:1657–1662. DOI: https://doi.org/10.1126/science.1062246, PMID: 11533490 He X, Dukkipati A, Garcia KC. 2006. Structural determinants of natriuretic peptide receptor specificity and degeneracy. Journal of Molecular Biology 361:698–714. DOI: https://doi.org/10.1016/j.jmb.2006.06.060 Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, Tunyasuvunakool K, Bates R, Žídek A, Potapenko A, Bridgland A, Meyer C, Kohl SAA, Ballard AJ, Cowie A, Romera- Paredes B, Nikolov S, Jain R, Adler J, Back T, et al. 2021. Highly accurate protein structure prediction with AlphaFold. Nature 596:583–589. DOI: https://doi.org/10.1038/s41586-021-03819-2, PMID: 34265844 Kidmose RT, Juhl J, Nissen P, Boesen T, Karlsen JL, Pedersen BP. 2019. Namdinator - automatic molecular dynamics flexible fitting of structural models into cryo- EM and crystallography experimental maps. IUCrJ 6:526–531. DOI: https://doi.org/10.1107/S2052252519007619, PMID: 31316797 Kobayashi D, Yamaguchi N, Takahashi O, Deshpande GA, Fukui T. 2012. Human atrial natriuretic peptide treatment for acute heart failure: A systematic review of efficacy and mortality. The Canadian Journal of Cardiology 28:102–109. DOI: https://doi.org/10.1016/j.cjca.2011.04.011, PMID: 21908161 Krissinel E, Henrick K. 2007. Inference of macromolecular assemblies from crystalline state. Journal of Molecular Biology 372:774–797. DOI: https://doi.org/10.1016/j.jmb.2007.05.022, PMID: 17681537 Kumar R, Grammatikakis N, Chinkers M. 2001. Regulation of the atrial natriuretic peptide receptor by heat shock protein 90 complexes. The Journal of Biological Chemistry 276:11371–11375. DOI: https://doi.org/10.1074/ jbc.M010480200, PMID: 11152473 Miner PB. 2020. Plecanatide for the treatment of constipation- predominant irritable bowel syndrome. Expert Review of Gastroenterology & Hepatology 14:71–84. DOI: https://doi.org/10.1080/17474124.2020.1722101, PMID: 31985305 Mirdita M, Schütze K, Moriwaki Y, Heo L, Ovchinnikov S, Steinegger M. 2022. ColabFold: making protein folding accessible to all. Nature Methods 19:679–682. DOI: https://doi.org/10.1038/s41592-022-01488-1, PMID: 35637307 Morin A, Eisenbraun B, Key J, Sanschagrin PC, Timony MA, Ottaviano M, Sliz P. 2013. Collaboration gets the most out of software. eLife 2:e01456. DOI: https://doi.org/10.7554/eLife.01456, PMID: 24040512 Oberoi J, Guiu XA, Outwin EA, Schellenberger P, Roumeliotis TI, Choudhary JS, Pearl LH. 2022. HSP90- CDC37- PP5 forms a structural platform for kinase dephosphorylation. Nature Communications 13:7343. DOI: https:// doi.org/10.1038/s41467-022-35143-2, PMID: 36446791 Ogawa H, Qiu Y, Ogata CM, Misono KS. 2004. Crystal structure of hormone- bound atrial natriuretic peptide receptor extracellular domain: rotation mechanism for transmembrane signal transduction. The Journal of Biological Chemistry 279:28625–28631. DOI: https://doi.org/10.1074/jbc.M313222200, PMID: 15117952 Ogawa H, Qiu Y, Philo JS, Arakawa T, Ogata CM, Misono KS. 2010. Reversibly bound chloride in the atrial natriuretic peptide receptor hormone- binding domain: possible allosteric regulation and a conserved structural motif for the chloride- binding site. Protein Science 19:544–557. DOI: https://doi.org/10.1002/pro.332, PMID: 20066666 Pettersen EF, Goddard TD, Huang CC, Meng EC, Couch GS, Croll TI, Morris JH, Ferrin TE. 2021. UCSF Chimerax: structure visualization for researchers, educators, and developers. Protein Science 30:70–82. DOI: https://doi.org/10.1002/pro.3943, PMID: 32881101 Potter LR, Garbers DL. 1992. Dephosphorylation of the guanylyl cyclase- A receptor causes desensitization. The Journal of Biological Chemistry 267:14531–14534 PMID: 1353076. Potter LR, Hunter T. 1998. Identification and characterization of the major phosphorylation sites of the B- type natriuretic peptide receptor. The Journal of Biological Chemistry 273:15533–15539. DOI: https://doi.org/10. 1074/jbc.273.25.15533, PMID: 9624142 Potter LR. 2011. Guanylyl cyclase structure, function and regulation. Cellular Signalling 23:1921–1926. DOI: https://doi.org/10.1016/j.cellsig.2011.09.001, PMID: 21914472 Punjani A, Rubinstein JL, Fleet DJ, Brubaker MA. 2017. cryoSPARC: algorithms for rapid unsupervised cryo- EM structure determination. Nature Methods 14:290–296. DOI: https://doi.org/10.1038/nmeth.4169, PMID: 28165473 Punjani A, Zhang H, Fleet DJ. 2020. Non- uniform refinement: adaptive regularization improves single- particle cryo- EM reconstruction. Nature Methods 17:1214–1221. DOI: https://doi.org/10.1038/s41592-020-00990-8, PMID: 33257830 Sanchez- Garcia R, Gomez- Blanco J, Cuervo A, Carazo JM, Sorzano COS, Vargas J. 2021. DeepEMhancer: a deep learning solution for cryo- EM volume post- processing. Communications Biology 4:874. DOI: https://doi. org/10.1038/s42003-021-02399-1, PMID: 34267316 Schopf FH, Biebl MM, Buchner J. 2017. The Hsp90 chaperone machinery. Nature Reviews. Molecular Cell Biology 18:345–360. DOI: https://doi.org/10.1038/nrm.2017.20, PMID: 28429788 Taipale M, Krykbaeva I, Koeva M, Kayatekin C, Westover KD, Karras GI, Lindquist S. 2012. Quantitative analysis of Hsp90- client interactions reveals principles of substrate recognition. Cell 150:987–1001. DOI: https://doi. org/10.1016/j.cell.2012.06.047, PMID: 22939624 Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 11 of 12 Biochemistry and Chemical Biology Research article Vaandrager AB, van der Wiel E, de Jonge HR. 1993. Heat- stable enterotoxin activation of immunopurified guanylyl cyclase C. Modulation by adenine nucleotides. The Journal of Biological Chemistry 268:19598–19603 PMID: 8103520. van den Akker F, Zhang X, Miyagi M, Huo X, Misono KS, Yee VC. 2000. Structure of the dimerized hormone- binding domain of a guanylyl- cyclase- coupled receptor. Nature 406:101–104. DOI: https://doi.org/10.1038/ 35017602, PMID: 10894551 Verba KA, Wang RYR, Arakawa A, Liu Y, Shirouzu M, Yokoyama S, Agard DA. 2016. Atomic structure of Hsp90- Cdc37- Cdk4 reveals that Hsp90 traps and stabilizes an unfolded kinase. Science 352:1542–1547. DOI: https:// doi.org/10.1126/science.aaf5023, PMID: 27339980 Yu SWB, Rao SSC. 2014. Advances in the management of constipation- predominant irritable bowel syndrome: the role of linaclotide. Therapeutic Advances in Gastroenterology 7:193–205. DOI: https://doi.org/10.1177/ 1756283X14537882, PMID: 25177366 Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 12 of 12 Biochemistry and Chemical Biology	null
10.1038_s41467-022-32646-w.pdf	Data availability The data that support the ﬁndings in this study are available from the corresponding authors upon reasonable request. Source data are provided with this paper. Code availability All code to reproduce the main simulation results can be found on GitHub (https://github.com/babaf/motor-adaptation-local-vs-input.git).	Data availability The data that support the findings in this study are available from the corresponding authors upon reasonable request. Source data are provided with this paper. Code availability All code to reproduce the main simulation results can be found on GitHub ( https://github.com/babaf/motor-adaptation-local-vs-input.git ).	nature communications Article https://doi.org/10.1038/s41467-022-32646-w Small, correlated changes in synaptic connectivity may facilitate rapid motor learning Received: 28 October 2021 Accepted: 8 August 2022 Barbara Feulner Lee E. Miller4,5,6, Juan A. Gallego 1,7 & Claudia Clopath 1,7 1, Matthew G. Perich 2, Raeed H. Chowdhury 3, Check for updates ; , : ) ( 0 9 8 7 6 5 4 3 2 1 ; , : ) ( 0 9 8 7 6 5 4 3 2 1 Animals rapidly adapt their movements to external perturbations, a process paralleled by changes in neural activity in the motor cortex. Experimental studies suggest that these changes originate from altered inputs (Hinput) rather than from changes in local connectivity (Hlocal), as neural covariance is largely preserved during adaptation. Since measuring synaptic changes in vivo remains very challenging, we used a modular recurrent neural network to qualitatively test this interpretation. As expected, Hinput resulted in small activity changes and largely preserved covariance. Surprisingly given the presumed dependence of stable covariance on preserved circuit connectivity, Hlocal led to only slightly larger changes in activity and covariance, still within the range of experimental recordings. This similarity is due to Hlocal only requiring small, correlated connectivity changes for successful adaptation. Simulations of tasks that impose increasingly larger behavioural changes revealed a growing difference between Hinput and Hlocal, which could be exploited when designing future experiments. Animals, particularly primates, can perform a great variety of beha- viours, which they are able to adapt rapidly in the face of changing conditions. Since behavioural adaptation can happen even after a single failed attempt1, the neural populations driving this process must be able to adapt equally fast. How this occurs remains unexplained2. Rapid motor learning is typically studied using external perturbations such as a visuomotor rotation (VR), which rotates the coordinates of the visual feedback with respect to those of the movement. Both humans and monkeys can learn to compensate for the resulting error between actual and expected visual feedback in a few tens of trials3,4. This behavioural adaptation is accompanied by changes in the activity of neurons in primary motor cortex (M1)5, and the upstream dorsal premotor cortex (PMd)3. It is unclear whether these neural activity changes are mediated by synaptic weight changes within the motor cortices or are driven by altered inputs from even further upstream areas. When learning a skill over many days, behavioural improvements are paralleled by rewiring between M1 neurons6–9. This seems not to be the case for rapid learning: throughout VR adaptation, the statistical interactions across neural populations in both M1 and PMd remain largely preserved10. These preserved interactions rule out any large synaptic changes within the motor cortices, as they would cause these models to degrade11,12. Instead, rapid VR adaptation may be driven by the cerebellum13–16 and/or posterior parietal cortex17,18. A pioneering Brain Computer Interface (BCI) study cast further doubt that signiﬁcant synaptic changes occurring within M1 are necessary for rapid learning19,20. In that study, monkeys controlled a computer cursor linked by a “decoder” to the activity of recorded M1 1Department of Bioengineering, Imperial College London, London, UK. 2Département de neurosciences, Université de Montréal, Montréal, QC, Canada. 3Department of Bioengineering, University of Pittsburgh, Pittsburgh, PA, USA. 4Department of Neuroscience, Northwestern University, Evanston, IL, USA. 5Department of Biomedical Engineering, Northwestern University, Evanston, IL, USA. 6Department of Physical Medicine and Rehabilitation, Northwestern University, and Shirley Ryan Ability Lab, Chicago, IL, USA. 7These authors contributed equally: Juan A. Gallego, Claudia Clopath. e-mail: [email protected]; [email protected] Nature Communications \| (2022) 13:5163 1 Article https://doi.org/10.1038/s41467-022-32646-w neurons. After learning to use a decoder that used the natural “intrinsic” mapping of neural activity onto cursor movements, the monkeys were exposed to one of two types of perturbations. When faced with a new decoder that preserved the statistical interactions (i.e., neural covariance) across M1 neurons, the monkeys could master it within minutes. In stark contrast, if the new decoder required changes in the neural covariance (an “out of manifold” perturbation), they could not learn it within one session—in fact, it required a pro- gressive training procedure spanning just over nine days on average21. Recording large scale synaptic changes in vivo remains challen- ging and has not been achieved during rapid motor learning. Alter- natively, recurrent neural network (RNN) models offer an exciting yet unexplored opportunity to test the effect of synaptic changes (in the model) on simulated activity during motor learning. RNNs trained on motor, cognitive and BCI tasks exhibit many striking similarities with the activity of neural populations recorded in animal studies22–28, suggesting a fundamental similarity between the two. Previous work using RNNs to model the BCI experiment described above19 showed that network covariance can be highly preserved even when learning is happening through weight changes within the network29. Thus, con- trary to intuition, functionally relevant synaptic weight changes may not necessarily lead to measurable changes in statistical interactions across neurons30. As a consequence, synaptic changes within PMd and M1 during VR adaptation may be very hard to identify through the analysis of neural population recordings. Here, we used RNN models to test whether VR adaptation might be mediated by synaptic changes within PMd and M1, yet with largely preserved neural covariance within these areas. We addressed this question by asking how adaptation based on connection weight changes within PMd and M1 (Hlocal) alters network activity compared to the corresponding activity changes if VR adaptation is based on altered inputs from upstream areas (Hinput)10,13–18 (Fig. 1A). To validate our modelling results, we compared our simulations to experimental recordings from PMd and M1 populations during the same VR task10. Under Hlocal, the changes in covariance following VR adaptation only slightly exceeded those under Hinput and were comparable to experimental observations. Thus, when using neural population recordings alone, it may be more challenging to disentangle these two hypotheses than previously thought. Moreover, for both Hinput and Hlocal, the learned connectivity changes were small and highly coor- dinated, which made them surprisingly robust to noise. To identify additional differences between Hinput and Hlocal, we examined learning during paradigms requiring larger behavioural changes. Covariance changes were larger for these paradigms in both PMd and M1 under Hinput, but only in M1 under Hlocal, thus providing a possible way to distinguish between the two hypotheses in future experiments. Our ﬁndings have implications for the interpretation of neural activity changes observed during learning, and suggest that tasks eliciting larger behavioural changes may be necessary to elucidate how neural populations adapt their activity during rapid learning. Results To understand whether motor adaptation could be driven by synaptic changes within PMd and M1, we simulated a VR adaptation task using a modular RNN that modelled these two areas, and compared the resulting changes in network activity to those of neural population recordings from PMd and M1 during the same VR task10. We quantiﬁed neural activity changes both in the experimental data and in the model using two measures (Fig. 1B): (1) the relative change in trial-averaged single neuron activity, and (2) the change in neural covariance (“Methods”). Combined, they capture aspects of single neuron as well as population-wide activity changes during adaptation. Small but measurable changes in neural activity within PMd and M1 during VR adaptation Monkeys were trained to perform an instructed delay task, in which they reached to one of eight visual targets using a planar manip- ulandum to receive a reward (“Methods”). After performing a block of Fig. 1 \| Competing hypotheses to explain where learning happens during a visuomotor rotation task. A To study the processes mediating motor cortical activity changes during adaptation in a visuomotor rotation task, we analyze and model the activity of neural populations within dorsal premotor cortex (PMd) and primary motor cortex (M1). We compare two hypotheses: plasticity upstream of PMd/M1 (Hinput) and plasticity within PMd/M1 (Hlocal). B Measures to quantify the changes in neural activity following adaptation: (1) relative change in trial-averaged single neuron activity; (2) change in neural covariance. Both measures compare baseline trials to late adaptation trials, after monkeys had adapted to the task. Data from a representative session from Monkey C. Nature Communications \| (2022) 13:5163 2 Article https://doi.org/10.1038/s41467-022-32646-w Fig. 2 \| Small but measurable changes in neural activity within PMd and M1 during VR adaptation. A Hand trajectories during the ﬁrst 30 trials of the base- line, and the early adaptation epoch (ﬁrst 150 adaptation trials), and the last 30 trials of the late adaptation epoch (last 150 adaptation trials). Trajectories are colour-coded by target. Data from Monkey C. B Angular error of the hand trajec- tories for the example session in A has the typical time course of adaptation. C Change in trial-averaged activity following adaptation. Data pooled across all sessions from the two monkeys for PMd and M1 separately (green markers, 11 ses- sions, 5 Monkey C, 6 Monkey M). Control sessions during which no perturbation was applied are shown for comparison (black markers, 3 sessions, 1 Monkey C, 2 Monkey M). Shaded area and horizontal bars, data distribution with mean and extrema (n = 11 experimental sessions). D Change in covariance following adapta- tion. Same format as (C). The monkey image was created by Carolina Massumoto who gave permission to use it under CC-BY license. unperturbed reaches (200–243 trials, depending on the session), visual feedback about the position of the hand was rotated by 30°, either clockwise or counterclockwise, depending on the session. Monkeys adapted rapidly to these perturbations: the curved reaches observed immediately after the perturbation onset became straighter after tens of trials, with the hand trajectories in the second (late) half of the adaptation block becoming more similar to the baseline trajec- tories (Fig. 2A). The angular error quantifying the difference between initial reach direction and target location decreased during adaptation (Fig. 2B). This error curve followed a similar trend for clockwise and counterclockwise perturbations, allowing us to analyze the different perturbations together. Behavioural adaptation was accompanied by changes in neural activity within both PMd and M1 (Fig. 2C)10. These changes exceeded those during control sessions, where no perturbation was applied (Fig. 2C black; linear mixed model analysis: t = 4.4, P = 0.0017). The amount of change was greater within PMd than M1 (t = 8.9, P < 0.0001). We also found small but detectable changes in neural covariance during VR perturbation, suggesting that the statistical interactions among neurons change slightly during adaptation (Fig. 2D). Again, these changes exceeded those of the control sessions (Fig. 2D black; t = 2.6, P = 0.026). A modular recurrent neural network model to study VR adaptation To test whether experimentally observed changes in motor cortical activity could be driven by rapid synaptic plasticity9 within PMd and M1, we trained a modular RNN model23,27 to perform the centre-out reaching task that we studied experimentally. To mimic broadly the hierarchical architecture of the motor cortical pathways, input signals were sent to the PMd module which then projected to the M1 module to produce the ﬁnal output signal (Fig. 3A; “Methods”). After initial training on the task, the model produced correct reaching trajectories to each of the eight different targets (Fig. 3B and Supplementary Fig. 1). These RNN-controlled movements had the same dynamics as those of monkeys (Fig. 3C). Furthermore, Principal Component Analysis revealed that the population activity of the PMd and M1 network modules was similar to that of the corresponding recorded neural populations (Fig. 3D, E). We used Procrustes analysis31 to quantify this apparent similarity between model and experimental population activity (Supplementary Fig. 2). This analysis conﬁrmed that the modular RNN captured the area-speciﬁc features in the neural data accurately, as the PMd and M1 modules better explained neural data from the respective brain area compared to a cross-area control (Supplementary Fig. 2). Motor adaptation through altered inputs matches neural recordings After having veriﬁed that our modular RNN recapitulates the key aspects of PMd and M1 population activity during reaching, we simu- lated the VR adaptation experiment. The model was retrained to pro- duce trajectories rotated by 30°, replicating the perturbation monkeys had to counteract. Having full control of the location of learning- related changes, we ﬁrst constrained it to happen upstream of PMd (Hinput). As anticipated from previous modelling18 and experimental work10, changes in areas upstream of the motor cortices can lead to successful adaptation: the hand trajectories produced after learning were correctly rotated by 30° to counteract the perturbation (Fig. 4A). When examining the activity of each of the PMd and M1 modules, the relative change in network activity was similar in magnitude to the changes observed in the corresponding neural population recordings (Fig. 4B and Supplementary Fig. 3). PMd activity changed slightly more than M1 activity (Fig. 4B), indicating a relation between the two mod- ules that was also present in the experimental data (Fig. 2C). With respect to interactions between neurons, covariance within each module was strongly preserved (Fig. 4C), as was the case for the Nature Communications \| (2022) 13:5163 3 Article https://doi.org/10.1038/s41467-022-32646-w Fig. 3 \| A modular recurrent neural network model to study VR adaptation. A A modular RNN that models key motor cortical areas to study adaptation. B Simulated (top) and actual (bottom) hand trajectories during 30 reaches to each target taken from one session from Monkey C. C Example simulated and actual hand trajectories to one target. Note the similarity in kinematics between the model and the experimental data. D Simulated PMd population activity recapitulates key features of actual PMd population activity. Neural trajectories extend from 600 ms before the go cue (black dots) to 600 ms after the go cue (coloured dots); go cue is indicated with coloured crosses. Reaching targets are colour-coded as in B. E Same as (D) for M1. experimental data (Fig. 2D). VR adaptation through altered inputs to the motor cortices thus is very similar to the neural activity changes observed in vivo. Learning through plastic changes within PMd and M1 modules occurs despite preserved the covariance Our simulation results so far are consistent with experimental10,13–15 and modelling18 studies proposing that VR adaptation is mediated by regions upstream of the motor cortices. But can our model rule out the alternative that adaptation is instead implemented by recurrent con- nectivity changes within PMd and M1 (Hlocal)? To address this question, we implemented Hlocal by constrain- ing learning to happen only within PMd and M1, a process which also led to successful adaptation (Fig. 4D). Interestingly, the activity changes produced under Hlocal differed both from those of Hinput and the experimental data: there were larger changes in the M1 module than in the PMd module (Fig. 4E). However, learning based on recurrent weight changes within PMd and M1 did not lead to large changes in covariance, which was largely preserved (Fig. 4F), virtually as much as when no local plasticity was allowed (Hinput) (Fig. 4C). Thus, the intuition that preserved covariance should be interpreted as a sign of stable underlying connectivity may be misleading. Small but coordinated connectivity changes enable motor adaptation We wished to understand how the model can adapt to the VR pertur- bation by changing the recurrent connectivity within the PMd and Nature Communications \| (2022) 13:5163 4 Article https://doi.org/10.1038/s41467-022-32646-w Fig. 4 \| Activity changes following learning upstream (Hinput) and within (Hlocal) the motor cortices. A Hand trajectories after learning under Hinput (coloured tra- ces; baseline trajectories are shown in grey). B Changes in trial-averaged activity following adaptation under Hinput (green markers) for PMd and M1, and reference mean experimental values (black stars; same data as presented in Fig. 2). Shaded area and horizontal bars, data distribution with mean and extrema (n = 10 network initializations). C Change in covariance following adaptation under Hinput, and reference values for change in covariance following the initial de-novo training (dashed lines). Data presented as in B. D Hand trajectories after learning under Hlocal. E Change in trial-averaged activity following adaptation under Hlocal. F Change in covariance following adaptation under Hlocal. Data in D–F are pre- sented as in A–C.Source data. M1 modules without altering their covariance. Interestingly, the con- nectivity changes under both Hlocal and Hinput (Fig. 5A, B) were small relative to experimentally observed synaptic changes32: an average weight change of 1–2% was sufﬁcient regardless of whether they hap- pened upstream of (Fig. 5C and Supplementary Fig. 4) or within the motor cortical modules (Fig. 5E and Supplementary Fig. 5). These changes were smaller than those observed during initial training (4–31%), when the model learned to perform the reaching task from random connection weights (Supplementary Fig. 6). Thus, “functional connectivity” within the PMd and M1 modules, as measured here by their covariance, may be largely preserved after VR adaptation under Hlocal because network connection weights change very little (Supplementary Fig. 7). We next studied how such small changes in connection weights could nevertheless drive effective behavioural adaptation. Recent studies seeking to relate RNN activity and connectivity have high- lighted the importance of low-dimensional structures in connectivity, showing their explanatory power for understanding how tasks are solved33–36. Inspired by this work, we looked for low-dimensional structure in the connectivity changes emerging in the model during adaptation (“Methods”). Our analysis revealed that the connectivity change patterns of all plastic modules were low-dimensional, inde- pendent of where learning happened (Fig. 5B, D, F). We thus hypo- thesized that the small changes were effective because they were low- dimensional. To test this, we examined how random changes in the connection weights (noise), which are inherently high-dimensional, would affect the behaviour. Low-dimensional connectivity changes are highly robust to noise For both Hlocal and Hinput, the learned connectivity changes in the model were small and low-dimensional. When considering the biolo- gical plausibility of our model, this observation raises the question of how such small connectivity changes could compete with ongoing synaptic ﬂuctuation, which is a known challenge for actual brains37–40. To test the hypothesis that the low-dimensionality of the learned connectivity changes is what makes them highly effective, we tested how adding synaptic ﬂuctuations, which are inherently high-dimen- sional, would affect motor output. Simulating synaptic ﬂuctuations by applying random perturbations to the learned connectivity changes increased the dimensionality of the weight changes (Fig. 6B, G; “Methods”), but did not lead to any observable change in reaching kinematics (Fig. 6C) or network activity (Fig. 6D, E). This was the case even though the applied random perturbations in connectivity were ten times bigger in magnitude than the learned connectivity changes (Fig. 6F), completely masking them (Fig. 6A, B). Therefore, our model not only suggests that VR adaptation can be implemented based on coordinated synaptic weight changes within PMd and M1, but also that this type of learning would be highly effective due to its robustness to synaptic ﬂuctuation. Larger visuomotor rotations allow for a clearer distinction between Hinput and Hlocal Although neural activity changes during VR adaptation were better reproduced by a model in which learning happens upstream of the Nature Communications \| (2022) 13:5163 5 Article https://doi.org/10.1038/s41467-022-32646-w Fig. 5 \| Small but coordinated connectivity changes enable motor adaptation. A Example connection weights for the M1 module after initial training. Top: esti- mated dimensionality. B Example changes in M1 connection weights following VR adaptation under Hlocal. Same format as (A). C Change in connection weights fol- lowing adaptation under Hinput. Each bar summarizes results for either one module of the network or a set of cross-module connections; bars and error bars, mean and s.d. (n = 10 network initializations). D Estimated dimensionality of connection weight changes for each network module and cross-module connections; data presented as in (C). E Change in connection weights following adaptation under Hlocal. F Dimensionality of connection weight changes under Hlocal. Data in E, F are presented as in C, D, respectively. Source data. motor cortices (Hinput), activity changes following learning through weight changes within the motor cortices (Hlocal) were also in good agreement with the experimental data. To verify that the stable cov- ariance (Fig. 4C, F) is not a general feature of the model but reﬂects task-speciﬁc demands, we modelled tasks for which we would expect larger changes. We ﬁrst asked the network to learn larger VRs of 60° and 90° instead of the original 30° rotation (Fig. 7A). The model was able to compensate for these larger perturbations under both Hinput and Hlocal (Fig. 7B, E). As expected, larger perturbations led to changes in net- work activity and covariance that increased with rotation angle (Fig. 7B, C, F, G). For the 90° rotation, we found a clear difference between Hinput and Hlocal: Hinput produced larger activity changes in PMd compared to M1, opposite that under Hlocal (Fig. 7C, F). Larger rotation angles also increased the learning-related difference in covariance between Hinput and Hlocal. Under Hinput, the increase in covariance was similar for the PMd and M1 modules as the rotation increased (Fig. 7D). In contrast, under Hlocal, the M1 covariance chan- ged more with increasing rotation angle than did that of PMd (Fig. 7G). These model predictions could help differentiate between Hinput and In fact, preliminary M1 population Hlocal recordings obtained during larger VRs (45° and 60°) seemed to match the model predictions for the covariance change under Hinput (Fig. 7D stars), but not Hlocal (Fig. 7G stars). in future experiments. A visuomotor reassociation task can differentiate between Hlocal and learning through remapping of input signals Although larger visuomotor rotations help differentiate between upstream learning and learning within PMd and M1, we sought to identify a task that would lead to an even clearer distinction. To this Nature Communications \| (2022) 13:5163 6 Article https://doi.org/10.1038/s41467-022-32646-w Fig. 6 \| Low-dimensional connectivity changes are highly robust to noise. A Example changes in M1 connection weights following VR adaptation. Same data as in Fig. 5B. B Same connection weight changes as in A, but with random con- nectivity changes added. Note the dramatic increase in the dimensionality of the connection weight changes. C Root mean squared error between target and pro- duced hand trajectories following adaptation in models with and without random weight changes; bars and error bars, mean and s.d. (n = 10 network initializations), as in all the panels in this ﬁgure. Dashed line, error under VR perturbation without any learning. D Change in trial-averaged activity for PMd and M1 module, without (solid) and with (empty) random weight changes. E Change in covariance. Same format as (D). F Change in connection weights for each network module and cross- module connections in models with (green bars) and without (dashed lines) noise in connectivity. G Dimensionality of connection weight changes. Same format as (F). end, we implemented a reassociation task where the model had to learn a new, random mapping between cues and reaching directions (Fig. 8A; “Methods”). This task allowed us to test a very speciﬁc change in the input signal to the motor cortices that could implement adaptation20,41: instead of adjusting the connectivity in an upstream network (Hinput), which allows for highly unconstrained modulation of input signals, the target-related input signals were manually reordered to compensate for the reassociation of cue-reaching direction pairs (Fig. 8B). This “learning through input reassociation” resulted in large changes in network activity (Fig. 8C), comparable in magnitude to those under Hlocal (Fig. 8F). Nevertheless, it did not cause any change in covariance (Fig. 8D), which clearly distinguished it from Hlocal (Fig. 8G) and the standard Hinput (Supplementary Fig. 8). This was the case because, in contrast to the standard Hinput during VR adaptation, the input signals did not change per se, but were only reassigned to dif- ferent targets, thereby entirely preserving the network activity patterns. Discussion Rapid motor learning is associated with neural activity changes in the motor cortices. The origin of these changes remains elusive, due to the current challenge of measuring synaptic strength in vivo. Here, we have used modular RNNs to simulate the motor cortices and to explore whether learning to counteract a visuomotor rotation within tens of Nature Communications \| (2022) 13:5163 7 Article https://doi.org/10.1038/s41467-022-32646-w Fig. 7 \| Larger visuomotor rotations allow for a better distinction between Hinput and Hlocal. A To verify that the modelled perturbation does not always produce small activity changes, we tested adaptation to larger VR perturbations (60° and 90°). B Root mean squared error between target and produced hand trajectories without (grey) and with learning (black) under Hinput. Dashed line, error after initial training, with no perturbation applied; shaded area and horizontal bars, data distribution with mean and extrema (n = 10 network initializations). Same data presentation in all panels. C Change in trial-averaged activity for PMd and M1 under Hinput. A few experimental sessions from Monkey C with larger rotations are shown as comparison (stars). D Change in covariance following adaptation. Data are presented as in C. A few experimental sessions from Monkey C with larger rotations are again shown as comparison (stars). Note the similarity between PMd (light green) and M1 (dark green) across all rotation angles. E Error without (grey) and with learning (black) under Hlocal. Same format as (B). F Change in trial-averaged activity for PMd and M1 under Hlocal. G Change in covariance following adaptation. Data in E–G are presented as in B–D. minutes could be mediated by local synaptic changes (Hlocal). By comparing the modelled network activity changes under Hlocal to the modelled changes observed during learning upstream of the motor cortices (Hinput), we have shown how the two hypotheses could be distinguished based on neural population recordings during beha- viour. Critically, despite the intuition that learning through plastic changes should lead to detectable changes in neural interactions within and across PMd and M1 populations, both Hlocal and Hinput (Fig. 4) largely preserved the covariance within these two regions, closely matching experimental observations (Fig. 2). This likely hap- pened because adaptation under Hlocal was achieved through small, coordinated weight changes within the PMd and M1 network modules (Fig. 5). Finally, using our model, we propose tasks for which we anticipate a more dramatic difference between these contrasting hypotheses (Figs. 7, 8) which can potentially help to interpret experi- mental data in the future. Electrophysiological10,13–15 and modelling studies18, as well as psy- chophysical evidence1,42 suggest that VR adaptation is driven by areas upstream of the motor cortices. Neurophysiological evidence is largely based on the observation that the statistical interactions within PMd and M1 populations remain preserved throughout adaptation10. This conclusion is in good agreement with studies showing that learning to generate neural activity patterns that preserve the covariance structure only takes a few tens of minutes19. Our direct comparison between Hinput and Hlocal lends further support to this observation. However, it also paints the intriguing picture that small, globally organized changes in synaptic weights could enable rapid learning without changing the neural covariance, a result that was robust across model initializations (Fig. 4), parameter settings (Supplemen- tary Fig. 9 and Supplementary Fig. 10), architectural design choices (Supplementary Fig. 11) and learning algorithms (Supplementary Fig. 12). Even implementing the modular RNN as a spiking neural net- work, bringing it closer to biology, did not change this result (Sup- plementary Fig. 13). Our simulations thus robustly show that covariance stability is not as directly linked to stable local connectivity as previously thought, as changes in covariance were comparable between Hinput and Hlocal for a 30° VR perturbation (Fig. 4). Instead, the change in neural covariance seemed to be more related to the task itself, as it correlated with the size of the perturbation: the larger the initial error (e.g., caused by larger rotations), the larger the change in covariance (Fig. 7). However, the relation between initial error and change in covariance differed depending on where the learning hap- pened (Hinput or Hlocal). The main difference between the two learning hypotheses we have examined is where in the hierarchical RNN model the connectivity changes occur: within the motor cortices (Hlocal), or upstream of them Nature Communications \| (2022) 13:5163 8 Article https://doi.org/10.1038/s41467-022-32646-w Fig. 8 \| A visuomotor reassociation task can clearly differentiate between Hlocal and learning through reassociation of input signals. A We simulated a task in which the network had to learn new associations between target locations and reach directions (“Reassociation''). B Root mean squared error between target and produced hand trajectories without (grey) and with learning (black) through input reassociation. Dashed line indicates error during baseline trial. Shaded area and horizontal bars, data distribution with mean and extrema (n = 10 network initializations). Same data presentation in all panels. C Change in trial-averaged activity for PMd and M1 under input reassociation. D Change in covariance fol- lowing adaptation. Note that the covariance matrices do not change at all. E Root mean squared error between target and produced hand trajectories without (grey) and with learning (black) under Hlocal. F Change in trial-averaged activity for PMd and M1 under Hlocal. G Change in covariance following adaptation. Data in E–G are presented as in B–D. (Hinput). Although neural covariance was preserved similarly by Hlocal and Hinput, we found a key characteristic that distinguished the two. When local connectivity was allowed to be plastic, the largest activity changes happened within the M1 module, with only small changes in the PMd module (Fig. 4E). In contrast, when learning occurred upstream of the PMd and M1 modules, the activity changes were similar in PMd and M1 (Fig. 4B), even if some learning was also allowed within PMd and M1 (Supplementary Fig. 10 and Supplementary Fig. 14). The experimental data, with larger activity changes in PMd than M1, better matched the pattern produced by Hinput. This observation fur- ther supports the hypothesis that VR adaptation is mediated by plas- ticity upstream of the motor cortices. A more arbitrary visuomotor reassociation task allowed us to test an alternate way in which upstream learning could occur, with con- straints against input signals changing but simply being reassigned to different targets (Fig. 8). Comparing this learning to that mediated by local connectivity changes revealed a clear distinction: learning under Hlocal modiﬁed the covariance in both PMd and M1, whereas learning through input reassociation preserved it20. Thus, future experiments seeking to disentangle to which extent learning happens within the motor cortices and/or upstream could study this task. Studies of learning in RNNs have focused on how networks implement de-novo training23,24,27,36,43–49. However, our brain does not learn to perform any task from scratch; it has been “trained” over many generations throughout evolution50. Here we studied how neural net- works adapt a learned behaviour, as opposed to de-novo learning. Our work raises the intriguing possibility that rapid learning following a few tens of minutes of practice could be easily achieved through small but speciﬁc changes in circuit connectivity. Thus, initial training seems to provide a highly ﬂexible backbone to adapt behaviour as needed51. The fact that the connectivity changes during adaptation under both Hlocal and Hinput were small and low-dimensional (Figs. 5, 6) sug- gests that either one could mediate rapid learning. First, as every synaptic change is costly52, we would expect a constraint on the total amount of connectivity change in the brain. The VR task being solved with only minor weight changes reﬂects this; in fact, they could be well achieved through long-term potentiation or depression of existing synapses, as experiments have shown that synaptic strength can double within minutes32. Second, the low dimensionality of these weight changes is also important with respect to solving “credit assignment”, the problem of determining how each synapse should change in order to restore the desired behaviour53–56. Although it is still Nature Communications \| (2022) 13:5163 9 Article https://doi.org/10.1038/s41467-022-32646-w unclear how this is achieved in the brain, one possibility is that synaptic plasticity is guided by “teacher” signals57,58. Since neuromodulatory signals can regulate synaptic plasticity59,60, they seem ideal candidates to regulate biologically plausible learning41,61–63. The ﬁnding that the connectivity changes needed to adapt to the VR perturbation are “naturally” low-dimensional is promising, as it suggests that learning could be controlled through relatively few neuromodulatory signals. Such implementation would contrast dramatically with the daunting challenge of learning to regulate every single synapse independently. Lastly, the robustness against synaptic ﬂuctuations conveyed by the low-dimensional connectivity changes makes both Hlocal and Hinput attractive in terms of ensuring memory stability. Given the ﬂuctuating nature of brain connectivity37, it remains puzzling how animals remember anything38–40,64. That low-dimensional weight changes, much smaller than ongoing synaptic ﬂuctuations, can achieve suc- cessful behavioural adaptation provides a potential solution to this problem. Our model consistently underestimated the changes in trial- averaged activity observed during VR adaptation, despite closely matching the small covariance changes (Fig. 2,4). This is to be expec- ted, as the model only captures changes due to the motor adaptation process itself, whereas the actual neural activity contains signals rela- ted to other processes such as “impulsivity”65 or “engagement”66. In fact, the experimentally observed neural activity changes between the early and late trials of control reaching sessions with no perturbation were almost as large as the changes during adaptation in our model (Fig. 2C, black dots). How these changes that are not related to learning are combined with the learning-related changes studied here remains unclear. Our modelling predictions for the learning-related changes could help tackle this question in future studies. A second potential reason why our model consistently under- estimated the activity changes during adaptation could be the fact that we did not include visual or proprioceptive feedback signals in our modelling approach67–71. As those signals also change during adapta- tion, they might cause additional changes in trial-averaged neural activity, despite not being directly necessary to solve the motor adaptation task. This could explain why our model could solve the task with smaller changes in neural activity. On the other hand, feedback signals could also actively contribute to the adaptation process. From this view, we may presently overestimate the already small con- nectivity changes underlying VR adaptation (Fig. 5), as part of the learning process could have been instead driven by dynamic feedback signals. Thus, when taking feedback into account, rapid learning of a motor perturbation could potentially be realized with even smaller changes in underlying connectivity, or maybe even without any con- nectivity changes at all2. To this point, the concrete role of feedback for rapid motor learning remains unclear and it could be interesting to use our model to further investigate this question. Our simulations were not designed to study trial-by-trial learning: we were interested in the neural activity changes between the baseline and late adaptation phases when the subjects had largely learned to counteract the perturbation and reached stable behaviour (Fig. 2B). Given that motor adaptation seems to be mediated by two processes with different timescales72–74, our model mainly captures the slower of the two. The neural activity changes underlying the early phase adaptation may be driven by different processes10, which our model currently does not test. In conclusion, our comparison between the activity changes fol- lowing VR adaptation through plastic changes within or upstream of the motor cortices shows that local plasticity (Hlocal) leads to neural signatures that are unexpectedly similar to those of upstream learning (Hinput). Intriguingly, Hlocal not only largely preserved the covariance within PMd and M1 but also resulted in connectivity changes that seem biologically reasonable: they are small, make the network robust against synaptic ﬂuctuations, and can be controlled by relatively few Nature Communications \| (2022) 13:5163 teaching signals. Our simulations thus encourage caution when drawing conclusions from the analysis of neural population recordings during learning, and further suggest potential behavioural tasks that could make it easier to identify where learning is happening within the motor system. Methods Tasks We studied motor adaptation using a visuomotor rotation (VR) para- digm, previously described in Perich et al. 201810. Monkeys (macaca mulatta) performed an instructed delay centre-out-reaching task in which they had to reach to one of eight targets uniformly distributed along a circle. All targets were 2 cm squares. The delay period was variable and ranged between 500 and 1500 ms. For additional details on the task, see10. During the adaptation phase, visual feedback was rotated clockwise or counterclockwise by 30°, 45°, or 60°. All surgical and experimental procedures were approved by the Institutional Ani- mal Care and Use Committee (IACUC) of Northwestern University. Using our modular RNN model, we simulated both this task and a visuomotor reassociation task in which there was no consistent rota- tion of the visual feedback; instead, each target required reaching to a different direction, uniquely selected from the initial set of eight dif- ferent targets. Experimental recordings We analyzed eleven sessions from two monkeys (ﬁve for Monkey C, six for Monkey M) that were exposed to a clockwise or counterclockwise 30° rotation (data previously presented in10). In addition to these data, we also analyzed three control sessions (one for Monkey C, two for Monkey M) in which no perturbation was applied, as well as additional sessions with larger VR angles from Monkey C where only M1 data was collected (30°, nine sessions; 45°, two sessions; 60°, two ses- sions) (Fig. 7). The spiking activity of putative single neurons was binned into 10 ms bins and then smoothed using a Gaussian ﬁlter (s.d., 50 ms). Only successful trials, where monkeys received a reward at the end, were included in the analysis. We deﬁned the early and late adaptation epochs as the ﬁrst and last 150 trials of the perturbation phase, when the visuomotor rotation was applied, respectively. RNN model Architecture. The neural network contained three recurrent modules, each consisting of 400 neurons, which we refer to as upstream, PMd and M1, respectively (Fig. 3A). The PMd and the upstream modules received an identical three-dimensional input signal, with the ﬁrst two dimensions signalling the x and y target location of that trial, and the third dimension signalling go (1 until the go, and 0 from then on). The upstream module connects to the PMd module and the PMd module connects to the M1 module. The output is calculated as a linear readout of the M1 module activity. Recurrent, as well as feedforward connec- tions were all-to-all. The model dynamics are given by t + 1 = xUP xUP t + (cid:1) dt τ (cid:1)xUP t + WUP tanhðxUP t Þ + Win,UPst (cid:3) ð1Þ t + 1 = xPMd xPMd t + (cid:1) (cid:1)xPMd t dt τ + WPMd tanhðxPMd t Þ + WUP-PMd tanhðxUP t Þ + Win,PMdst (cid:3) t + 1 = xM1 xM1 t + (cid:1) dt τ (cid:1)xM1 t + WM1 tanhðxM1 t Þ + WPMd-M1 tanhðxPMd t (cid:3) Þ t = Wout tanhðxM1 xout t Þ + bout ð2Þ ð3Þ ð4Þ 10 Article https://doi.org/10.1038/s41467-022-32646-w where xUP describes the network activity in the upstream module, and xPMd and xM1 the network activity in the PMd and M1 module respec- tively. WUP, WPMd and WM1 deﬁne the recurrent connectivity matrix within the upstream module, the PMd module and the M1 module, respectively. WUP-PMd deﬁnes the connectivity matrix from the upstream module to the PMd module, and WPMd-M1 deﬁnes the con- nectivity matrix from the PMd module to the M1 module. The input connectivity matrices for the upstream and the PMd module are given by Win,UP and Win,PMd, respectively; st represents the three-dimensional input signal described above. The two-dimensional output xout is decoded from the M1 module activity via the output connectivity matrix Wout and the bias term bout. The time constant is τ = 0.05 s and the integration time step is dt = 0.01 s. Training. Each network was initially trained to produce planar reaching trajectories, mirroring the experimental hand trajectories. The training and testing data set were constructed by pooling the hand trajectories xtarget for successful trials during the baseline epochs from all experi- mental sessions, which resulted in 2238 trials of length 4 s (90%/10% randomly split into training/testing). The held out testing data was used to validate that the model had been trained successfully during the initial training period. Model simulations were implemented using PyTorch75 and training was performed using the Adam optimizer76 with a learning rate of 0.0001 (β1 = 0.9, β2 = 0.999). The initial training consisted of 500 training trials. The loss function was deﬁned as L = 1 BðT (cid:1) 50Þ2 B ∑ b T ∑ t= 50 ∑ d= x,y (cid:1) xout,d t,b (cid:1) xtarget,d t,b (cid:3) 2 + Eweights + Erates ð5Þ where the regularization term on the weights is given by (∣∣ . ∣∣ indicates L2 norm) Eweights = αð∣∣Win,UP∣∣ + ∣∣Win,PMd∣∣ + ∣∣Wout∣∣ + ∣∣WPMd∣∣ + ∣∣WM1∣∣ + ∣∣WPMd-M1∣∣ + ∣∣WUP∣∣ + ∣∣WUP-PMd∣∣Þ the regularization term on the rates is given by Erates = β 1 BTN B ∑ b T ∑ t (cid:4) (cid:1) N ∑ n (cid:1) tanh xPMd,n t,b (cid:3) (cid:3) 2 (cid:1) (cid:1) + tanh xM1,n t,b (cid:3) (cid:3) 2 (cid:1) (cid:1) + tanh xUP,n t,b ð6Þ (cid:5) (cid:3) (cid:3) 2 ð7Þ with batch size B = 80, time steps T = 400 and neurons N = 400. The regularization parameters were set to α = 0.001, β = 0.8. We clipped the gradient norm at 0.2 before we applied the optimization step. For the VR adaptation, we trained the initial network for another 100 trials with the target trajectory rotated 30° (or 60° or 90° for the case of the larger VRs). For the VR reassociation task we shufﬂed the stimulus s across reaching directions but kept the targets xtarget ﬁxed, as indicated in Fig. 8A (colours correspond to the given stimulus and sketched reaching trajectories correspond to the assigned target). The network had again 100 trials to adapt to this perturbation. Data analysis We quantiﬁed the changes in actual and simulated neural activity fol- lowing adaptation using two measures: changes in trial-averaged activity (or peristimulus time histogram, PSTH), and changes in cov- ariance. We calculated both metrics within a window that started 600 ms before the go signal and ended 600 ms after it. The change in activity was calculated by ∣PSTHLateadaptation (cid:1) PSTHBaseline∣ σBaseline ð8Þ model, 100 trials with similar go signal timing), PSTHLateadaptation is the trial-averaged activity in the late adaptation epoch (experimental data: last 150 trials of the adaptation epoch; simulation data: on a model trained to counteract the perturbation, 100 trials with similar go signal timing), and σBaseline is the neuron-speciﬁc standard deviation across time and targets during the baseline epoch. To summarize the change in trial averaged activity across all neurons, time points, and targets, we calculated their median; this provided one single value for each experimental session or simulation run. The change in covariance was calculated using the same trial-averaged data from the baseline and the late adaptation epoch. We calculated the covariance in each of these two epochs and then quantiﬁed the similarity by calculating the Pearson correlation coefﬁcient between the corresponding entries of the two matrices. The change in covariance is then deﬁned by 1 minus the correlation coefﬁcient. For the experimental sessions, we computed a lower bound for each measure using the control sessions in which monkeys were not exposed to a perturbation. To account for the fact that there could be activity changes unrelated to motor adaptation65,66, we compared the activity during 150 consecutive trials from the ﬁrst half of the control session with 150 consecutive trials from the second half of the control session. To compute the magnitude of the weight changes after networks learned to counteract the perturbation, we computed the average absolute weight change as dW = ∣WLateadaptation (cid:1) WBaseline∣ WBaseline ð9Þ . where ∣ ∣ indicates the element wise absolute value, WBaseline is deﬁned as the model parameter (either Win,PMd, Win,UP, WUP, WUP-PMd, WPMd, WPMd-M1, or WM1) after the initial training phase but before training on the VR perturbation, and WLateadaptation is deﬁned as the same model parameter after training on the VR perturbation. To obtain one sum- mary value for each simulation run, we calculated the median of all weight entries for a given parameter. To measure the dimensionality of weight change we calculated the singular values ki of WLateadaptation − WBaseline and deﬁned the dimensionality, using the participation ratio77: (cid:4) (cid:5) 2 ki ∑ i k2 i = ∑ i ð10Þ Statistics To statistically compare the change in activity found in the control sessions with the change found in the VR sessions, we performed a linear mixed model analysis using R (lmer package). The brain area (PMd or M1) and whether the experimental session included a per- turbation phase or not were included as ﬁxed effects, whereas monkey and session identity were included as random effects. A signiﬁcance threshold of P = 0.05 was used. Simulation of synaptic ﬂuctuation (Fig. 6) To simulate synaptic ﬂuctuation we added random values to the learned connectivity changes during adaptation. Those random values were drawn from a normal distribution with zero mean and s.d. ten times larger than the s.d. of the learned weight changes distribution. With that, we created synaptic noise which was completely unstruc- tured across connection sites. We did not add or delete any synapses in the model. where PSTHBaseline is the trial-averaged activity in the baseline epoch (experimental data: all baseline trials; simulated data: on a trained Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Nature Communications \| (2022) 13:5163 11 Article https://doi.org/10.1038/s41467-022-32646-w Data availability The data that support the ﬁndings in this study are available from the corresponding authors upon reasonable request. Source data are provided with this paper. Code availability All code to reproduce the main simulation results can be found on GitHub (https://github.com/babaf/motor-adaptation-local-vs-input.git). References 1. Thoroughman, K. A. & Shadmehr, R. Learning of action through adaptive combination of motor primitives. Nature 407, 742–747 (2000). Sohn, H., Meirhaeghe, N., Rajalingham, R. & Jazayeri, M. A network perspective on sensorimotor learning. Trends Neurosci. 44, 170–181 (2020). 2. 21. Oby, E. R. et al. New neural activity patterns emerge with long-term learning. Proc. Natl Acad. Sci. USA 116, 15210–15215 (2019). 22. Mante, V., Sussillo, D., Shenoy, K. V. & Newsome, W. T. Context- dependent computation by recurrent dynamics in prefrontal cor- tex. Nature 503, 78–84 (2013). 23. Sussillo, D., Churchland, M. M., Kaufman, M. T. & Shenoy, K. V. A neural network that ﬁnds a naturalistic solution for the production of muscle activity. Nature Neurosci. 18, 1025–1033 (2015). 24. Rajan, K., Harvey, C. D. & Tank, D. W. Recurrent network models of sequence generation and memory. Neuron 90, 128–142 (2016). 25. Song, H. F., Yang, G. R. & Wang, X.-J. Reward-based training of recurrent neural networks for cognitive and value-based tasks. Elife 6, e21492 (2017). 26. Wang, J., Narain, D., Hosseini, E. A. & Jazayeri, M. Flexible timing by temporal scaling of cortical responses. Nat. Neurosci. 21, 102–110 (2018). 3. Wise, S., Moody, S., Blomstrom, K. & Mitz, A. Changes in motor 27. Michaels, J. A., Schaffelhofer, S., Agudelo-Toro, A. & Scherberger, 4. 5. cortical activity during visuomotor adaptation. Exp. Brain Res. 121, 285–299 (1998). Krakauer, J. W., Pine, Z. M., Ghilardi, M.-F. & Ghez, C. Learning of visuomotor transformations for vectorial planning of reaching tra- jectories. J. Neurosci. 20, 8916–8924 (2000). Paz, R., Boraud, T., Natan, C., Bergman, H. & Vaadia, E. Preparatory activity in motor cortex reﬂects learning of local visuomotor skills. Nat. Neurosci. 6, 882–890 (2003). 6. Rioult-Pedotti, M.-S., Friedman, D., Hess, G. & Donoghue, J. P. 7. Strengthening of horizontal cortical connections following skill learning. Nat. Neurosci. 1, 230–234 (1998). Kleim, J. A. et al. Cortical synaptogenesis and motor map reorga- nization occur during late, but not early, phase of motor skill learning. J. Neurosci. 24, 628–633 (2004). 8. Xu, T. et al. Rapid formation and selective stabilization of synapses for enduring motor memories. Nature 462, 915–919 (2009). 9. Roth, R. H. et al. Cortical synaptic ampa receptor plasticity during motor learning. Neuron 105, 895–908 (2020). 10. Perich, M. G., Gallego, J. A. & Miller, L. E. A neural population mechanism for rapid learning. Neuron 100, 964–976 (2018). 11. Gerhard, F. et al. Successful reconstruction of a physiological cir- cuit with known connectivity from spiking activity alone. PLoS Comput. Biol. 9, e1003138 (2013). H. A goal-driven modular neural network predicts parietofrontal neural dynamics during grasping. Proc. Natl Acad. Sci. USA 117, 32124–32135 (2020). 28. Perich, M. G. et al. Inferring brain-wide interactions using data- constrained recurrent neural network models. Preprint at bioRxiv https://doi.org/10.1101/2020.12.18.423348 (2021). 29. Feulner, B. & Clopath, C. Neural manifold under plasticity in a goal driven learning behaviour. PLoS Comput. Biol. 17, e1008621 (2021). 30. Das, A. & Fiete, I. R. Systematic errors in connectivity inferred from activity in strongly recurrent networks. Nat. Neurosci. 23, 1286–1296 (2020). 31. Schönemann, P. H. A generalized solution of the orthogonal pro- crustes problem. Psychometrika 31, 1–10 (1966). 32. Froemke, R. C. & Dan, Y. Spike-timing-dependent synaptic mod- iﬁcation induced by natural spike trains. Nature 416, 433–438 (2002). 33. Aljadeff, J., Renfrew, D., Vegué, M. & Sharpee, T. O. Low- dimensional dynamics of structured random networks. Phys. Rev. E 93, 022302 (2016). 34. Mastrogiuseppe, F. & Ostojic, S. Linking connectivity, dynamics and computations in low-rank recurrent neural networks. Neuron 99, 609–623 (2018). 12. Rebesco, J. M., Stevenson, I. H., Koerding, K., Solla, S. A. & Miller, L. 35. Schuessler, F., Dubreuil, A., Mastrogiuseppe, F., Ostojic, S. & Barak, E. Rewiring neural interactions by micro-stimulation. Front. Syst. Neurosci. 4, 39 (2010). O. Dynamics of random recurrent networks with correlated low- rank structure. Phys. Rev. Res. 2, 013111 (2020). 13. Tseng, Y.-w, Diedrichsen, J., Krakauer, J. W., Shadmehr, R. & Bastian, A. J. Sensory prediction errors drive cerebellum-dependent adap- tation of reaching. J. Neurophysiol. 98, 54–62 (2007). 14. Rabe, K. et al. Adaptation to visuomotor rotation and force ﬁeld 36. Schuessler, F., Mastrogiuseppe, F., Dubreuil, A., Ostojic, S. & Barak, O. The interplay between randomness and structure during learn- ing in RNNs. Advances in neural information processing systems 33, 13352–13362 (2020). perturbation is correlated to different brain areas in patients with cerebellar degeneration. J. Neurophysiol. 101, 1961–1971 (2009). 15. Schlerf, J. E., Galea, J. M., Bastian, A. J. & Celnik, P. A. Dynamic 37. Calvin, W. H. & Stevens, C. F. Synaptic noise and other sources of randomness in motoneuron interspike intervals. J. Neurophysiol. 31, 574–587 (1968). modulation of cerebellar excitability for abrupt, but not gradual, visuomotor adaptation. J. Neurosci. 32, 11610–11617 (2012). 16. Tzvi, E., Koeth, F., Karabanov, A. N., Siebner, H. R. & Krämer, U. M. Cerebellar–premotor cortex interactions underlying visuomotor adaptation. NeuroImage 220, 117142 (2020). 17. Diedrichsen, J., Hashambhoy, Y., Rane, T. & Shadmehr, R. Neural correlates of reach errors. J. Neurosci. 25, 9919–9931 (2005). 18. Tanaka, H., Sejnowski, T. J. & Krakauer, J. W. Adaptation to visuo- motor rotation through interaction between posterior parietal and motor cortical areas. J. Neurophysiol. 102, 2921–2932 (2009). 19. Sadtler, P. T. et al. Neural constraints on learning. Nature 512, 423–426 (2014). 20. Golub, M. D. et al. Learning by neural reassociation. Nat. Neurosci. 21, 607–616 (2018). 38. Susman, L., Brenner, N. & Barak, O. Stable memory with unstable synapses. Nat. Commun. 10, 1–9 (2019). 39. Fauth, M. J. & van Rossum, M. C. Self-organized reactivation main- tains and reinforces memories despite synaptic turnover. Elife 8, e43717 (2019). 40. Rule, M. E. et al. Stable task information from an unstable neural population. Elife 9, e51121 (2020). 41. Legenstein, R., Chase, S. M., Schwartz, A. B. & Maass, W. A reward- modulated Hebbian learning rule can explain experimentally observed network reorganization in a brain control task. J. Neurosci. 30, 8400–8410 (2010). 42. Criscimagna-Hemminger, S. E., Bastian, A. J. & Shadmehr, R. Size of error affects cerebellar contributions to motor learning. J. Neuro- physiol. 103, 2275–2284 (2010). Nature Communications \| (2022) 13:5163 12 Article https://doi.org/10.1038/s41467-022-32646-w 43. Hennequin, G., Vogels, T. P. & Gerstner, W. Optimal control of transient dynamics in balanced networks supports generation of complex movements. Neuron 82, 1394–1406 (2014). 44. Stroud, J. P., Porter, M. A., Hennequin, G. & Vogels, T. P. Motor primitives in space and time via targeted gain modulation in cortical networks. Nat. Neurosci. 21, 1774–1783 (2018). 45. DePasquale, B., Cueva, C. J., Rajan, K., Escola, G. S. & Abbott, L. full- force: A target-based method for training recurrent networks. PLoS One 13, e0191527 (2018). 46. Kao, J. C. Considerations in using recurrent neural networks to probe neural dynamics. J. Neurophysiol. 122, 2504–2521 (2019). 47. Yang, G. R., Joglekar, M. R., Song, H. F., Newsome, W. T. & Wang, X.- J. Task representations in neural networks trained to perform many cognitive tasks. Nat. Neurosci. 22, 297–306 (2019). 48. Logiaco, L., Abbott, L. & Escola, S. Thalamic control of cortical dynamics in a model of ﬂexible motor sequencing. Cell Rep. 35, 109090 (2021). 49. Kao, T.-C., Sadabadi, M. S. & Hennequin, G. Optimal anticipatory control as a theory of motor preparation: A thalamo-cortical circuit model. Neuron 109, 1567–1581 (2021). 50. Zador, A. M. A critique of pure learning and what artiﬁcial neural networks can learn from animal brains. Nat. Commun. 10, 1–7 (2019). 51. Goudar, V., Peysakhovich, B., Freedman, D. J., Buffalo, E. A. & Wang, X.-J. Elucidating the neural mechanisms of learning-to-learn. Preprint at bioRxiv https://doi.org/10.1101/2021.09.02. 455707 (2021). 52. Li, H. L. & Van Rossum, M. C. Energy efﬁcient synaptic plasticity. Elife 9, e50804 (2020). 53. Whittington, J. C. & Bogacz, R. Theories of error back-propagation in the brain. Trends Cogn. Sci. 23, 235–250 (2019). 54. Lillicrap, T. P., Santoro, A., Marris, L., Akerman, C. J. & Hinton, G. Backpropagation and the brain. Nat. Rev. Neurosci. 21, 335–346 (2020). 55. Bellec, G. et al. A solution to the learning dilemma for recurrent networks of spiking neurons. Nat. Commun. 11, 1–15 (2020). 56. Payeur, A., Guerguiev, J., Zenke, F., Richards, B. A. & Naud, R. Burst- dependent synaptic plasticity can coordinate learning in hier- archical circuits. Nat. Neurosci. 24, 1–10 (2021). 57. Marr, D. & Thach, W. T. From the Retina to the Neocortex 11–50 (Springer, 1991). 58. Medina, J. F., Nores, W. L. & Mauk, M. D. Inhibition of climbing ﬁbres is a signal for the extinction of conditioned eyelid responses. Nature 416, 330–333 (2002). 59. Bailey, C. H., Giustetto, M., Zhu, H., Chen, M. & Kandel, E. R. A novel function for serotonin-mediated short-term facilitation in aplysia: Conversion of a transient, cell-wide homosynaptic Hebbian plasti- city into a persistent, protein synthesis-independent synapse-spe- ciﬁc enhancement. Proc. Natl Acad. Sci. USA 97, 11581–11586 (2000). 60. Nitsche, M. A. et al. Dopaminergic modulation of long-lasting direct current-induced cortical excitability changes in the human motor cortex. Eur. J. Neurosci. 23, 1651–1657 (2006). 61. Legenstein, R., Pecevski, D. & Maass, W. A learning theory for reward-modulated spike-timing-dependent plasticity with application to biofeedback. PLoS Comput. Biol. 4, e1000180 (2008). 62. Miconi, T. Biologically plausible learning in recurrent neural net- works reproduces neural dynamics observed during cognitive tasks. Elife 6, e20899 (2017). in a spiking neural network. Proc. Natl Acad. Sci. USA 118, e2111821118 (2021). 64. Kalle Kossio, Y. F., Goedeke, S., Klos, C. & Memmesheimer, R.-M. Drifting assemblies for persistent memory: Neuron transitions and unsupervised compensation. Proc. Natl Acad. Sci. USA 119, e2023832118 (2021). 65. Cowley, B. R. et al. Slow drift of neural activity as a signature of impulsivity in macaque visual and prefrontal cortex. Neuron 108, 551–567 (2020). 66. Hennig, J. A. et al. Learning is shaped by abrupt changes in neural engagement. Nat. Neurosci. 24, 727–736 (2021). 67. Omrani, M., Murnaghan, C. D., Pruszynski, J. A. & Scott, S. H. Dis- tributed task-speciﬁc processing of somatosensory feedback for voluntary motor control. Elife 5, e13141 (2016). 68. Stavisky, S. D., Kao, J. C., Ryu, S. I. & Shenoy, K. V. Motor cortical visuomotor feedback activity is initially isolated from downstream targets in output-null neural state space dimensions. Neuron 95, 195–208 (2017). 69. Kalidindi, H. T. et al. Rotational dynamics in motor cortex are con- sistent with a feedback controller. Elife 10, e67256 (2021). 70. Perich, M. G. et al. Motor cortical dynamics are shaped by multiple distinct subspaces during naturalistic behavior. Preprint at bioRxiv https://doi.org/10.1101/2020.07.30.228767 (2020). 71. Cross, K. P., Cook, D. J. and Scott, S. H. Convergence of proprio- ceptive and visual feedback on neurons in primary motor cortex. Preprint at bioRxiv https://doi.org/10.1101/2021.05.01. 442274 (2021). 72. Smith, M. A., Ghazizadeh, A. & Shadmehr, R. Interacting adaptive processes with different timescales underlie short-term motor learning. PLoS Biol. 4, e179 (2006). 73. Huberdeau, D. M., Krakauer, J. W. & Haith, A. M. Dual-process decomposition in human sensorimotor adaptation. Curr. Opin. Neurobiol. 33, 71–77 (2015). 74. Christou, A. I., Miall, R. C., McNab, F. & Galea, J. M. Individual dif- ferences in explicit and implicit visuomotor learning and working memory capacity. Sci. Rep. 6, 1–13 (2016). 75. Paszke, A. et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32, (2019). 76. Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. Preprint at arXiv https://doi.org/10.48550/arxiv.1412.6980 (2014). 77. Gao, P. et al. A theory of multineuronal dimensionality, dynamics, and measurement. Preprint at bioRxiv https://doi.org/10.1101/ 214262 (2017). Acknowledgements This work has been funded by BBSRC (BB/N013956/1 and BB/N019008/ 1), Wellcome Trust (200790/Z/16/Z), the Simons Foundation (564408) (all to C.C.), the EPSRC (EP/R035806/1 to C.C. and EP/T020970/1 to JAG), and the ERC (ERC-2020-STG-949660 to J.A.G.). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Author contributions Conceptualization: B.F., J.A.G., and C.C. Methodology: B.F., J.A.G., and C.C. Experimental data collection: M.G.P. Modelling: B.F. Analysis and interpretation of data: B.F., J.A.G., and C.C. Supervision: J.A.G. and C.C. Writing/review of the paper: all. 63. Liu, Y. H., Smith, S., Mihalas, S., Shea-Brown, E. & Sümbül, U. Cell- type-speciﬁc neuromodulation guides synaptic credit assignment Competing interests The authors declare no competing interests. Nature Communications \| (2022) 13:5163 13 Article https://doi.org/10.1038/s41467-022-32646-w Additional information Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-022-32646-w. Correspondence and requests for materials should be addressed to Juan A. Gallego or Claudia Clopath. Peer review information Nature Communications thanks Jonathan Michaels and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Reprints and permission information is available at http://www.nature.com/reprints Publisher’s note Springer Nature remains neutral with regard to jur- isdictional claims in published maps and institutional afﬁliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/. © The Author(s) 2022 Nature Communications \| (2022) 13:5163 14	null
10.1371_journal.pone.0219089.pdf	Data Availability Statement: All relevant data are within the paper and its Supporting Information files.	All relevant data are within the paper and its Supporting Information files.	RESEARCH ARTICLE Vibration of symmetrically layered angle-ply cylindrical shells filled with fluid Nurul Izyan Mat Daud1,2, K. K. ViswanathanID 1,3* 1 UTM Centre for Industrial and Applied Mathematics, Ibnu Sina Institute for Scientific & Industrial Research, Universiti Teknologi Malaysia, Johor Bahru, Johor, Malaysia, 2 Department of Mathematical Sciences, Faculty of Science, Universiti Teknologi Malaysia, Johor Bahru, Johor, Malaysia, 3 Department of Mathematics, Kuwait College of Science and Technology, Doha District, Safat, Kuwait * [email protected], [email protected] Abstract Vibrational behaviour of symmetric angle-ply layered circular cylindrical shell filled with qui- escent fluid is presented. The equations of motion of cylindrical shell in terms of stress and moment resultants are derived from the first order shear deformation theory. Irrotational of inviscid fluid are expressed as the wave equation. These two equations are coupled. Strain- displacement relations and stress-strain relations are adopted into the equations of motion to obtain the differential equations with displacements and rotational functions. A system of ordinary differential equation is obtained in one variable by assuming the functions in sepa- rable form. Spline of order three is applied to approximate the displacement and rotational functions, together with boundary conditions, to get a generalised eigenvalue problem. The eigenvalue problem is solved for eigen frequency parameter and associate eigenvectors of spline coefficients. The study of frequency parameters are analysed using the parameters the thickness ratio, length ratio, angle-ply, properties of material and number of layers under different boundary conditions. Introduction Composite materials are widely used in petroleum, chemical industry etc., due to the high stiff- ness and strength. Other characteristics of composite materials are resistance to corrosion and light in weight. The applications of these kinds of materials can be seen in industries like trans- portation, aircraft, construction, marine and consumer products. Hence, the investigation on the vibrational behaviour of the structure is carried out in order to determine its natural fre- quency. The important of finding frequency is to prevent the structure from resonance thus to improve the life-span of the structure. There are many literature studied the vibrational behav- iour of the shell structure especially in cylindrical shell itself by considering different theories and method of solution. Warburton and Higgs [1] used Rayleigh-Ritz method to solve a canti- lever cylindrical shell by considering Flugge’s shell theory. Using the same method, Song et al. [2] investigated the free vibration of symmetrically laminated composite cylindrical shell with arbitrary boundary conditions. a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Mat Daud NI, Viswanathan KK (2019) Vibration of symmetrically layered angle-ply cylindrical shells filled with fluid. PLoS ONE 14(7): e0219089. https://doi.org/10.1371/journal. pone.0219089 Editor: Fang-Bao Tian, University of New South Wales, AUSTRALIA Received: February 2, 2019 Accepted: June 14, 2019 Published: July 3, 2019 Copyright: © 2019 Mat Daud, Viswanathan. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the paper and its Supporting Information files. Funding: Funded by (K.K.Viswanathan) Grant number: PR17-16SM-05. Kuwait Foundation for the Advancement of Sciences. www.kfas.org/. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 1 / 18 Vibration of cylindrical shell with fluid Among the researchers those who used Finite Element Method (FEM) in their analysis were Sivadas and Ganesan [3], Lam and Wu [4]. The wave propagation method was applied by Li [5], Zhang [6] and Iqbal et al. [7]. Lopatin and Morozov [8] solved the problem of a cantile- ver composite cylindrical shell using Galerkin method to find the fundamental frequencies. In addition, Haar wavelet method was used by Xie et al. [9] to analyse the free vibration of cylin- drical shell based on Goldenveizer-Novozhilov shell theory. Meanwhile, Jin et al. [10] applied First Order Shear Deformation Theory (FSDT) to examine the vibration of functionally graded cylindrical shell. A spline strip method was adopted in studying the vibration of cross-ply lami- nated cylindrical shells [11]. Bickley-type spline method was applied to analyse the frequencies of free vibration of cylindrical shell [12–15]. Besides that, a spline-based differential quadrature method was used by Javidpoor et al. [16] and Ghasemi [17]. Ferreira et al. [18] used multi- quadric radial basis function method to determine the natural frequencies of doubly curved cross-ply composite shells. Moreover, the structure can also interact with fluid either the structure filled with, partially with, or submerged. The fluid can be flowing or non-flowing fluid. This type of investigation is known as fluid structure interaction. The characteristics of fluid itself affect the behaviour of the structure. Hence, the investigation has to consider both shell and fluid in order to get better results. Recently, the investigation on fluid structure interaction has get attention among the researchers. Finite Element Method (FEM) was applied in the problem free vibration of isotropic, verti- cal cylindrical shell partially and completely filled with stationary liquid [19]. The shell equa- tions were constructed using Sanders’ thin shell theory. Selmane and Lakis [20] applied FEM in solving the vibration of an anisotropic cylindrical shell submerged and subjected simulta- neously to an internal and external flow by considering Sanders’ thin shell theory. Other litera- tures applied FEM in their analysis are Kochupillai et al. [21], Kochupillai et al. [22], Toorani and Lakis [23], Toorani and Lakis [24]. Gunawan et al. [25] conducted a study on cylindrical shells filled with fluid based on elastic foundation by considering Sanders’ thin shell theory. Krishna and Ganesan [26] investi- gated the results of free vibration of cylindrical shells filled with fluid. The shell was governed by first order deformation theory. Lakis et al. [27] analysed the isotropic and anisotropic plates and shells with and without fluid for linear and nonlinear vibration with the shell equations were based on Sanders’ shell theory and dynamic pressure of fluid was derived from Bernoul- li’s equation. Galerkin method was used by Goncalves et al. [28] in solving the nonlinear dynamic behaviour of cylindrical shells filled with fluid with Donnell’s nonlinear shallow shell theory. A study on the vibration of vertical circular cylindrical shell partially filled by an incom- pressible, compressible, quiescent and inviscid fluid using Rayleigh-Ritz method was analysed by Amabili [29]. The shell was constrained by simply-supported boundary conditions. In addi- tion to that, Kwak et al. [30] studied the clamped-free cylindrical shell partially submerged in fluid using Sanders’ shell theory. Vibrational analysis of fluid filled double-walled carbon nanotubes using the wave propaga- tion method was carried out by Natsuki et al. [31] and the shell equations were based on sim- plified Flu¨gge shell theory. Iqbal et al. [32] applied Love’s thin shell theory to study the vibration of a Functionally Graded Material (FGM) circular cylindrical shell filled with fluid and the fluid was an incompressible non-viscous fluid. A study on cylindrical shells filled with fluid resting on elastic foundations was carried out by Shah et al. [33] using wave propagation method. The shell was constrained with simply supported at the both ends and the Love’s thin shell theory was applied to the problem. A nonlinear vibration of cantilevered circular PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 2 / 18 Vibration of cylindrical shell with fluid cylindrical containing quiescent fluid based on Flu¨gge’s shell theory and the fluid motion was modelled by linearized potential flow theory was investigated by Paak et al. [34]. A dynamic stiffness method was studied by Tran and Manh [35] to investigate the free vibration of cross-ply laminated composite circular cylindrical shells filled with fluid partially and also complete filling with fluid under clamped-free boundary conditions. Reissner-Mind- lin theory was used for the shell equations. The fluid considered as non-viscous and incom- pressible. In addition, spline method was applied to solve the free vibration of layered cylindrical shell filled with fluid using Love’s thin shell theory [36]. Nurul Izyan et al. [37] applied the spline method in their analysis to determine the frequencies of anti-symmetric angle-ply laminated composite cylindrical shell filled with fluid. The shells’ equations were for- mulated based on FSDT. According to a nonlocal theory, there are two types of models which are structural harden- ing and softening models. Li et al. [38] and Shen and Li [39] proved that both the hardening and softening models are correct and they are related to different types of surface effects, i.e. relaxation or tension of surface atomic lattices. Consequently, the two different models are caused by the long range attractive and repulsive interactions in pair potentials between atoms, respectively. Different approach study is done in order to study the fluid structure interaction. The aim of this study is to determine the vibrational behaviour of symmetric angle-ply laminated com- posite cylindrical shell. The quiescent fluid-filled shell is considered. The quiescent fluid means that the vibration of fluid is depends on the vibration of the shell. Once the shell vibrates, the fluid will vibrate accordingly. Bickley-type spline is used to approximate the dis- placements and rotational functions since it gives better accuracy and also uses of lower order approximation in solving boundary value problem [40]. The shells’ equations are based on FSDT. Three and five layered shell under clamped-clamped and simply-supported boundary conditions are studied. The material properties of Kevler-49 Epoxy (KGE) and AS4/3501-6 Graphite/Epoxy (AGE) are used. Parametric studies on shell geometry i.e, length and thickness of the shells, type of materials, ply orientations, number of layer of the materials, and boundary conditions on frequencies are studied. Formulation of the problem Equations of shell Fig 1 shows the geometry of the circular cylindrical shell with the shell coordinates defined as (x,θ,z) where x is along the axial direction, θ is in the circumferential and z is along the normal direction. The length of the shell is ‘, thickness is h and radius is r. The displacement compo- nents u,v,w are expressed under the first order shear deformation theory [41] as uðx; y; z; tÞ ¼ u0ðx; y; tÞ þ zcxðx; y; tÞ; vðx; y; z; tÞ ¼ v0ðx; y; tÞ þ zcyðx; y; tÞ; wðx; y; z; tÞ ¼ w0ðx; y; tÞ; ð1Þ where u,v,w are the displacements in x,θ,z directions. u0,v0,w0 are mid-plane displacements and ψx,ψθ are the rotational functions of the normal to the mid-plane with respect to the x- and θ- axes. Equations of motion for cylindrical shell are formulated using FSDT [42, 43]. The PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 3 / 18 Vibration of cylindrical shell with fluid Fig 1. A geometry and coordinate system of a circular cylindrical shell. https://doi.org/10.1371/journal.pone.0219089.g001 equations of cylindrical shell which included fluid are written as ¼ I1 @2u @t2 ; þ Qy ¼ I1 ; 1 r 1 r þ þ þ þ @Nx @x @Nxy @x @Qx @x @Mx @x @Mxy @x @Nyx 1 @y r @Ny 1 @y r @Qy 1 @y r @Mxy 1 @y r @My 1 @y r (cid:0) þ (cid:0) Qx ¼ I3 (cid:0) Qy ¼ I3 @2v @t2 @2w @t2 @2cx @t2 @2cy @t2 ; : Ny ¼ I1 (cid:0) p; ð2Þ Here, p is the fluid pressure. I1 and I3 are the normal and rotary inertia coefficients, given by [42,43]. R ðI1; I3Þ ¼ rðkÞð1;z2Þdz; where ρ(k) is the material density of the k-th layer of the shell. PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 ð3Þ 4 / 18 Stress and moment resultants are expressed as Vibration of cylindrical shell with fluid R z ðsx; sy; txy; txz; tyzÞdz; ðNx; Ny; Nxy; Qx; QyÞ ¼ ðMx; My; MxyÞ ¼ R z ðsx; sy; txyÞzdz: ð4Þ Using stress-strain relations and strain-displacement relations of the k-th layer by neglect- ing the transverse normal strain and stress, the stress and moment resultants are obtained as follows 0 B B B B B B B B B B B @ ¼ 1 C C C C C C C C C C C A 0 B B B B B B B B B B B @ Nx Ny Nxy Mx My Mxy A11 A12 A16 B11 A12 A22 A26 B12 A16 A26 A66 B16 B12 B22 B26 B16 B26 B66 B11 B12 B16 D11 D12 D16 B12 B22 B26 D12 D22 D26 B16 B26 B66 D16 D26 D66 1 0 C C C C C C C C C C C A B B B B B B B B B B B B B B B B B B B B @ @u0 @x @v0 1 @y r @u0 1 @y r @cx @x @cy 1 @y r @cx 1 @y r þ þ w r @v @x þ @cy @x 1 C C C C C C C C C C C C C C C C C C C C A ; and ! Qy Qx ¼ K A44 A45 A45 A55 0 ! cy þ B B @ cx þ (cid:0) @w @y 1 r @w @x v0 r 1 C C A: Here Aij, Bij and Dij are extensional rigidities, bending-stretching coupling rigidities and bending rigidities, respectively and defined as ð5Þ ð6Þ Aij ¼ XN(cid:0) 1 k¼1 QðkÞ ij ðzk (cid:0) zk(cid:0) 1Þ; Bij ¼ 1 2 XN(cid:0) 1 QðkÞ ij ðz2 k¼1 k (cid:0) z2 k(cid:0) 1Þ; Dij ¼ 1 3 and XN(cid:0) 1 QðkÞ ij ðz3 k¼1 k (cid:0) z3 k(cid:0) 1Þ; ði; j ¼ 1; 2; 6Þ; Aij ¼ K XN(cid:0) 1 k¼1 Qij ðkÞðzk (cid:0) zk(cid:0) 1Þ; ði; j ¼ 4; 5Þ: Here zk and zk−1 are boundaries of the k-th layer and K is the shear correction factor which is depends on lamination scheme [44, 45]. Since shell is considered to be a symmetric angle- ply, therefore, the coefficients A16,A26,A45,D16,D26 and Bij are identically zeroes [13]. The displacements u0, v0, w and shear rotations ψx,ψθ are assumed in the form of u0ðx; y; tÞ ¼ UðxÞcosnyeiot; v0ðx; y; tÞ ¼ VðxÞsinnyeiot; wðx; y; tÞ ¼ WðxÞcosnyeiot; cxðx; y; tÞ ¼ CxðxÞcosnyeiot; cyðx; y; tÞ ¼ CyðxÞsinnyeiot; ð7Þ 5 / 18 PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 Vibration of cylindrical shell with fluid in which ω,t and n are the angular frequency, time and circumferential node number respectively. Non-dimensional parameters used are as follows L ¼ X ¼ R ¼ H ¼ ‘ r x ‘ r ‘ h r l ¼ o‘ dk ¼ hk h ; a length parameter; ; a distance coordinate; ; a radius parameter; ; the thickness parameter; r ; a frequency parameter; ffiffiffiffiffiffiffi I1 A11 ; a relative layer thickness of k(cid:0) th layer: ð8Þ Substituting Eqs (5) and (6) into Eq (2), then, using Eqs (7) and (8), the differential equa- tions are obtained in terms of displacement and rotational functions and it can be written in the matrix form as follows 0 B B B B B B B B @ L11 L12 L13 L14 L15 L22 L22 L23 L24 L25 L31 L32 L33 L34 L35 L41 L42 L43 L44 L45 L51 L52 L53 L54 L55 1 0 C C C C C C C C A B B B B B B B @ 1 C C C C C C C A U V W CX Cy 1 C C C C C C C A ; 0 B B B B B B B @ 0 0 0 0 0 ¼ ð9Þ where Lij are differential operators given as follows L11 ¼ d2 dX2 (cid:0) S10 n2 R2 þ l2; L12 ¼ ðS2 þ S10Þ n R d dX L14 ¼ L15 ¼ 0; L21 ¼ (cid:0) ðS2 þ S10Þ n R d dX ; L22 ¼ S10 L23 ¼ (cid:0) ðS3 þ KS13Þ � ¼ L32; L24 ¼ 0; L25 ¼ KS13 � L33 ¼ KS14 (cid:0) S3 þ KS13 d2 dX2 n2 R2 � þ 1 þ rf I1 ; L13 ¼ S2 1 R � ; d dX n2 R2 S3 (cid:0) d2 dX2 1 R JnðRÞ nðRÞ J0 ; L31 ¼ (cid:0) S2 � l2; � 1 R2 þ l2; þ KS13 1 R d dX ; n R2 1 R2 n2 R2 1 R n2 R2 L34 ¼ KS14 d dX L44 ¼ S7 d2 dX2 (cid:0) ; L35 ¼ KS13 � n R ; L41 ¼ L42 ¼ 0; L43 ¼ (cid:0) KS14 � S12 þ KS14 þ l2p1; L45 ¼ ðS12 þ S8Þ ; d dX d n dX R ; L51 ¼ 0; L52 ¼ KS13 � ; L53 ¼ KS13 � L55 ¼ S12 d2 dX2 (cid:0) S9 þ KS13 þ l2p1; n R ; L54 ¼ (cid:0) ðS12 þ S8Þ n R d dX ; PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 6 / 18 with S2 ¼ A12 A11 ; S3 ¼ A22 A11 ; S7 ¼ D11 ‘2A11 ; S8 ¼ S12 ¼ D66 ‘2A11 ; S13 ¼ A44 A11 ; S14 ¼ A55 A11 ; S9 ¼ D12 ‘2A11 I3 ‘2I1 ; P1 ¼ : Vibration of cylindrical shell with fluid D22 ‘2A11 ; S10 ¼ A66 A11 ; Equation of fluid The fluid pressure, p is obtained as an explicit expression and coupled to the equations of the shell through radial displacement. The wave equation is obtained for irrotational flow of an inviscid fluid undergoing small oscillations in the cylindrical coordinates system (x, θ, r) [46] as @2p @r2 þ 1 r @p @r þ 1 r2 @2p @y2 þ @2p @x2 ¼ @2p c2@t2 ; where c is the speed sound of the fluid. The pressure is assumed in the separable form as pðx; y; r; tÞ ¼ cðxÞcosðnyÞJnðrÞeiot: ð10Þ ð11Þ Here Jn is the Bessel function of order n. The permeability condition on the fluid- shell interface ensures that the fluid remains in contact with the shell wall and written as (cid:0) 1 iorf @p @r � � � � ¼ r¼R � � � � @w @t ; r¼R ð12Þ where ρf is the density of the fluid. Eq (12) is solved by using the displacement component in the normal direction given in the Eq (7) together with the Eq (11), hence, the following relation is obtained cðxÞ ¼ o2rf nðrÞ J0 WðxÞ: Then, by applying Eq (13) into Eq (11), the fluid pressure, p is obtained as follows pðx; y; r; tÞ ¼ (cid:0) rf JnðrÞ nðrÞ J0 @2w @t2 : PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 ð13Þ ð14Þ 7 / 18 Vibration of cylindrical shell with fluid Solution procedure Spline collocation method The displacement functions U,V,W and rotational functions CX and Cθ are approximated using cubic splines as U �ðXÞ ¼ V �ðXÞ ¼ X2 i¼0 X2 aiXi þ ciXi þ XN(cid:0) 1 j¼0 XN(cid:0) 1 i¼0 j¼0 bjðX (cid:0) XjÞ3HðX (cid:0) XjÞ; djðX (cid:0) XjÞ3HðX (cid:0) XjÞ; W�ðXÞ ¼ X2 eiXi þ XN(cid:0) 1 i¼0 j¼0 fjðX (cid:0) XjÞ3HðX (cid:0) XjÞ: CX �ðXÞ ¼ X2 i¼0 giXi þ XN(cid:0) 1 j¼0 pjðX (cid:0) XjÞ3HðX (cid:0) XjÞ; Cy �ðXÞ ¼ X2 liXi þ XN(cid:0) 1 i¼0 j¼0 qjðX (cid:0) XjÞ3HðX (cid:0) XjÞ: ð15Þ Here, H(X−Xj) is the Heaviside step function and N is the number of intervals in the range of X2[0,1] is divided. These splines are lower order approximation and an effective method since it has fast convergence and high accuracy. The points of division X ¼ Xs ¼ s 0; 1; 2; . . .NÞ are chosen as the knots of the splines as well as the collocation points. Assuming that the differential equations given by Eq (9) are satisfied by these splines given in Eq (15) at the knots, a set of 5N+5 homogeneous equations into 5N+15 unknown spline coefficients ai,bj, ci,dj,ei,fj,gi,pj,li,qj (i = 0,1,2;j = 0,1,2,. . .,N−1) is obtained. ; ðs ¼ N Boundary conditions Two types of boundary conditions are used to analyse the problem i. Clamped-Clamped (C-C)(both the ends are clamped); U ¼ 0; V ¼ 0; W ¼ 0; CX ¼ 0; Cy ¼ 0 at X ¼ 0 and X ¼ 1: i. Simply supported–Simply supported (S–S)(both the ends are simply supported); V ¼ 0; W ¼ 0; Cy ¼ 0; Nx ¼ 0; Mx ¼ 0 at X ¼ 0 and X ¼ 1: The system will get ten more equations on spline coefficients by imposing any one of the boundary conditions. Combining these ten equations with the earlier 5N+5 homogeneous equations, one can get 5N+15 equations in the same number unknown coefficients. This can PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 8 / 18 Vibration of cylindrical shell with fluid be written as ½P�fqg ¼ l2½Q�fqg: ð16Þ Here, [P] and [Q] are the square matrices of order (5N+15)×(5N+15). {q} is the column matrix of the spline coefficients of order (5N+15)×1 and λ is the frequency parameter. Results and discussion Convergence and comparative studies The convergence study has been made in order to determine the number of iteration. This is done by fixing the circumferential node number, thickness ratio, length ratio, number of layers and orientation of the material under C-C and S-S boundary conditions. The number of knots, N is taken as 16 since the change in percentage for next following values of N is 0.29%. In the case of comparison studies, none of literature has been done on symmetric angle-ply shell with consideration of fluid. There are studies on symmetric angle-ply shells but limited to empty shells only [4, 13]. Hence, an investigation on the empty shell as well as the fluid-filled shells with respect to the length parameter is carried out as shown in Fig 2. The effects of the fluid on angular frequency ωm(m = 1,2,3) of three layered symmetric angle-ply shells using the materials Kevler-49 epoxy (KGE) and AS4/3501-6 Graphite/epoxy (AGE) are arranged in the order of KGE-AGE-KGE with the angle orientations at (30˚/0˚/30˚) are presented. The param- eters n = 2, H = 0.02 are fixed. The shell is clamped at both the ends. According to Fig 2, the Fig 2. Effect of length parameter on the angular frequency of three layered symmetric angle-ply shells under C-C boundary conditions. Layer materials: KGE-AGE-KGE. https://doi.org/10.1371/journal.pone.0219089.g002 PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 9 / 18 Vibration of cylindrical shell with fluid frequency of fluid-filled shell is lower than the frequency of the empty shell. This result is expected as the fluid in the shell gives an added mass to the shell. Analysis Three and five layered symmetric angle-ply circular cylindrical shells with fluid are analysed. The materials which are KGE and AGE are used [47]. The shear correction factor K = 5/6 is fixed throughout this analysis [44] and first three frequency parameters are studied for all the cases. Fig 3 shows the variation of angular frequencies ωm(m = 1,2,3) on length parameter is stud- ied for three layered shell with the combination of the materials KGE and AGE with C-C boundary conditions. The parameters n = 4 and H = 0.015 are fixed and the ply-angles are arranged as 30˚/0˚/30˚, 45˚/0˚/45˚ and 60˚/0˚/60˚ using the materials AGE and KGE and ordered as KGE-AGE-KGE. Since the frequency parameter λm is explicitly a function of length of the cylinder, hence when studying the influence of the length of the cylinder on its vibra- tional behaviour, the angular frequency ωm(m = 1,2,3) is considered instead of λm. The rela- tionship between the frequency and the length of the shell can be seen as in [48]. The natural frequency or equivalent rigidity increases with increasing the non-dimensional scale parame- ter [49]. From the analysis, it can be seen that the frequency decreases as the length parameter increases and the frequency is higher for higher modes. The frequency decreases fast at L rang- ing from 0.5 to 0.75. Later, it decreases slowly. It also can be seen that the frequency for m = 1 of Fig 3(A) is the lowest compared to Fig 3(B) and 3(C). As for m = 2,3, the values of the fre- quencies are the lowest for Fig 3(C), followed by Fig 3(B) and 3(A). Further, investigation on S-S boundary conditions is presented in Fig 4 by fixing the parameters as in Fig 3. The trend shown by the graph of Fig 4 is similar to Fig 3. The results show that the frequencies obtained by S-S boundary conditions are lower than C-C boundary conditions. Fig 3. Effect of length parameter on the angular frequency of three layered symmetric angle-ply shells under C-C boundary conditions. Layer materials: KGE-AGE-KGE. https://doi.org/10.1371/journal.pone.0219089.g003 PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 10 / 18 Vibration of cylindrical shell with fluid Fig 4. Effect of length parameter on the angular frequency of three layered symmetric angle-ply shells under S-S boundary conditions. Layer materials: KGE-AGE-KGE. https://doi.org/10.1371/journal.pone.0219089.g004 Fig 5 corresponds to the effect of length parameter on angular frequency ωm(m = 1,2,3) of five layered angle-ply arranged materials in the order KGE-AGE-KGE-AGE-KGE with ply- angles of 45˚/30˚/0˚/30˚/45˚, 30˚/45˚/0˚/45˚/30˚ and 60˚/30˚/0˚/30˚/60˚, respectively for C-C conditions. The values of n = 4 and H = 0.02 is fixed. The figure shows that the frequency decreases as the length parameter increases. The investigation is continued by applying S-S boundary conditions using the same parameters as shown in Fig 6. The trend of graph is Fig 5. Effect of the length parameter on the angular frequency of five layered symmetric angle-ply shells under C-C boundary conditions. Layer materials: KGE-AGE-KGE-AGE-KGE. https://doi.org/10.1371/journal.pone.0219089.g005 PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 11 / 18 Vibration of cylindrical shell with fluid Fig 6. Effect of the length parameter on the angular frequency of five layered symmetric angle-ply shells under S-S boundary conditions. Layer materials: KGE-AGE-KGE-AGE-KGE. https://doi.org/10.1371/journal.pone.0219089.g006 similar to the Fig 5. The values of ωm(m = 1,2,3) in Fig 6 are lower than the value of ωm(m = 1,2,3) in Fig 5. Fig 7 shows the effect of thickness parameter on frequencies λm(m = 1,2,3) by considering three layered shells arranged in KGE-AGE-KGE materials with n = 4 and L = 1.5 are fixed under C-C boundary conditions. The shell is arranged in the order of 30˚/0˚/30˚, 45˚/0˚/45˚ Fig 7. Effect of thickness parameter on the frequency for three layered symmetric angle-ply shells under C-C boundary conditions. Layer materials: KGE-AGE-KGE. https://doi.org/10.1371/journal.pone.0219089.g007 PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 12 / 18 Vibration of cylindrical shell with fluid Fig 8. Effect of thickness parameter on frequency for three layered symmetric angle-ply shells under S-S boundary conditions. Layer materials: KGE-AGE-KGE. https://doi.org/10.1371/journal.pone.0219089.g008 and 60˚/0˚/60˚. From the analysis, it can be seen clearly that the frequency rises as the thick- ness parameter rises. The frequencies are greater for larger modes. The values of λm(m = 1,2,3) in Fig 7(A) are the lowest compared to Fig 7(B) and 7(C). In addition, Fig 8 presents the varia- tion of frequencies λm(m = 1,2,3) under S-S boundary conditions. Results show that the values of λm(m = 1,2,3) in Fig 8 are lower than the values of λm(m = 1,2,3) in Fig 7. Fig 9 shows the variation of λm(m = 1,2,3) with respect to the thickness parameter for five layered angle-ply by fixing n = 4 and L = 1 for C-C boundary conditions. The materials are arranged as KGE-AGE-KGE-AGE-KGE with angle orientation at 45˚/30˚/0˚/30˚/45˚, 30˚/45˚/ Fig 9. Effect of thickness parameter on frequency for five layered symmetric angle-ply shells under C-C boundary conditions. Layer materials: KGE-AGE-KGE-AGE-KGE. https://doi.org/10.1371/journal.pone.0219089.g009 PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 13 / 18 Vibration of cylindrical shell with fluid Fig 10. Effect of thickness parameter on frequency for five layered symmetric angle-ply shells under S-S boundary conditions. Layer materials: KGE-AGE-KGE-AGE-KGE. https://doi.org/10.1371/journal.pone.0219089.g010 Fig 11. Effect of length parameter on frequency for three and five layered symmetric angle-ply shells under C-C boundary conditions. https://doi.org/10.1371/journal.pone.0219089.g011 PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 14 / 18 Vibration of cylindrical shell with fluid Fig 12. Effect of thickness parameter on frequency for three and five layered symmetric angle-ply shells under C-C boundary conditions. https://doi.org/10.1371/journal.pone.0219089.g012 0˚/45˚/30˚ and 60˚/30˚/0˚/30˚/60˚. Clearly, the frequency is higher for higher values of thick- ness parameter. Fig 9(A) has the lowest frequency compared to Fig 9(B) and 9(C). Applying the same parameters as in Fig 9, investigation on S-S boundary conditions is carried out as depicted in Fig 10. It can be seen that Fig 10(A) has the lowest frequency, followed by Fig 10 (B) and 10(C). It can be observed that the values of λm(m = 1,2,3) under S-S boundary condi- tions are lower compared to the values of λm(m = 1,2,3) under C-C boundary conditions. By fixing n = 3, the effects of length parameter for both three and five layered shells on fre- quencies under C-C boundary conditions are investigated as depicted in Fig 11. The materials are arranged as KGE-AGE-KGE with angle orientation at 45˚/0˚/45˚ as shown in Fig 11(A). Meanwhile, Fig 11(B) used KGE-AGE-KGE-AGE-KGE materials with angle orientation at 30˚/45˚/0˚/45/30˚. From Fig 11, it can be seen that the frequencies decrease as the length parameter increases. It decreases fast in the range 0.5 < ωm < 0.75. Also, the frequencies of five layered shell are higher than the frequencies of three layered shell. Next, the frequencies with respect to thickness parameter are analysed as shown in Fig 12. The frequencies of five layered shell in Fig 12(B) are higher than the frequencies of three lay- ered shell in Fig 12(A). Conclusion The vibrational behaviour of layered cylindrical shell with symmetric angle-ply is investigated using spline method. The shell contained a quiescent fluid. The equations of motion of the PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 15 / 18 Vibration of cylindrical shell with fluid shell are based on first order shear deformation theory. Investigations on empty and fluid- filled shells show that the frequency value reduces as the fluid term is included. This is due to the fluid in the shell that provides added mass to the shell. Results show that by increasing the length of the shell, the frequency decreases. In contrast, the frequency increases as the shell thickness increases. Meanwhile, frequency of C-C bound- ary conditions is higher than the frequency of S-S boundary conditions. It can be concluded that the geometric parameters, material properties, angle orientations, number of layers and boundary conditions significantly affects the frequency of the shell. Acknowledgments This work was funded partially by Kuwait Foundation for the Advancement of Sciences (KFAS) under Project code: PR17-16SM-05 and thanks to the President of KCST for his con- tinues support to complete this research work. Author Contributions Funding acquisition: K. K. Viswanathan. Methodology: Nurul Izyan Mat Daud, K. K. Viswanathan. Software: K. K. Viswanathan. Supervision: K. K. Viswanathan. Writing – original draft: Nurul Izyan Mat Daud. Writing – review & editing: K. K. Viswanathan. References 1. Warburton GB, Higgs J. Natural frequencies of thin cantilever cylindrical shells. J Sound Vib. 1970; 11 (3): 335–338. 2. Song X, Han Q, Zhai J. Vibration analyses of symmetrically laminated composite cylindrical shells with arbitrary boundaries conditions via Rayleigh-Ritz method. Compos Struct. 2015; 134: 820–830. 3. Sivadas KR, Ganesan N. Free vibration of circular cylindrical shells with axially varying thickness. J Sound Vib. 1991; 147(1): 73–85. 4. 5. 6. 7. 8. Lam KY, Wu Q. Free Vibration of symmetric angle-ply thick laminated composite cylindrical shells. Compos Part B. 2000; 31: 345–354. Li XB. Study on free vibration analysis of circular cylindrical shells using wave propagation. J Sound Vib. 2008; 311: 667–682. Zhang XM. Vibration analysis of cross-ply laminated composite cylindrical shells using the wave propa- gation approach. Appl Acous. 2001; 62 (11): 1221–1228. Iqbal Z, Naeem MN, Sultana N. Vibration characteristics of FGM circular cylindrical shells using wave propagation approach. Acta Mech. 2009; 208: 237–248. Lopatin AV, Morozov EV. Fundamental frequency of a cantilever composite cylindrical shell. Compos Struct. 2015; 119: 638–647. 9. Xie X, Jin G, Liu Z. Free vibration analysis of cylindrical shells using the Haar Wavelet method. Int J Mech Sci. 2013; 77: 47–56. 10. Jin G, Xie X, Liu Z. The Haar Wavelet method for free vibration analysis of functionally graded cylindrical shells based on the shear deformation theory. Compos Struct. 2014; 108: 435–448. 11. Mizusawa T, Kito H. Vibration of cross-ply laminated cylindrical panels by the Spline Strip method. Com- pos Struct. 1995; 57(2): 253–265. 12. Viswanathan KK, Aziz ZA, Amirah HZ, Javed S. Free vibration of symmetric angle-ply laminated circular cylindrical shells. IOP Conf Series; Earth and Environmental Science. 2014; 19: 1–7. PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 16 / 18 Vibration of cylindrical shell with fluid 13. Viswanathan KK, Javed S, Aziz ZA, Hossain I. Free vibration of symmetric angle-ply laminated cylindri- cal shells of variable thickness including shear deformation theory. Int J Phy Sci. 2011; 6(25): 6098– 6109. 14. Viswanathan KK, Javed S. Free vibration of anti-symmetric angle-ply cylindrical shell walls using first- order shear deformation theory. J Vib Control. 2016; 22(7): 1757–1768. 15. Grigorenko AY, Efimova TL, Sokolova LV. On the investigation of free vibrations of non-thin cylindrical shells of variable thickness by the Spline-Collocation method. J Math Sci. 2012; 181(4): 506–519. 16. Javidpoor S, Ali NA, Kabi A. Time integration of rectangular membrane free vibration using Spline- based Differential Quadrature. Journal of Appl Comp Mech. 2016; 2(2): 74–79. 17. Ghasemi M. High order approximations using Spline-Based Differential Quadrature method: Implemen- tation to the Multi-Dimensional PDEs. Appl Math Model. 2017; 46: 63–80. 18. 19. Ferreira AJM, Roque CMC, Jorge RMN. Natural frequencies of FSDT cross-ply composite shells by Multiquadrics. Compos Struct. 2007; 77(3): 296–305. Lakis AA, Paidoussis MP. Free vibration of cylindrical shells partially filled with liquid. J Sound Vib. 1971; 19(1): 1–15. 20. Selmane A, Lakis AA. Vibration analysis of anisotropic open cylindrical shells subjected to a flowing fluid. J Fluids Struct. 1997; 11: 111–134. 21. Kochupillai J, Ganesan N, Padmanabhan C. A semi-analytical coupled finite element formulation for shells conveying fluids. Compos Struct. 2002; 80: 271–286. 22. Kochupillai J, Ganesan N, Padmanabhan C. A semi-analytical coupled finite element formulation for composite shells conveying fluids. J Sound Vib. 2002; 258(2): 287–207. 23. 24. Toorani MH, Lakis AA. Dynamic analysis of anisotropic cylindrical shells containing flowing fluid. J Pres- sure Vessel Tech. 2001; 123: 454–460. Toorani MH, Lakis AA. Dynamics behavior of axisymmetric and beam-like anisotropic cylindrical shells conveying fluid. J Sound Vib. 2003; 259(2): 265–298. 25. Gunawan T, Mikami T, Kanie S, Sato M. Free vibrations of fluid-filled cylindrical shells on elastic founda- tions. Thin-Walled Struct. 2005; 43: 1746–1762. 26. Krishna BV, Ganesan N. Polynomial approach for calculating added mass for fluid-filled cylindrical shells. J Sound Vib. 2006; 29: 1221–1228. 27. Lakis AA, Toorani MH, Kerboua Y, Esmailzadeh M, Sabri F. Theory, analysis and design of fluid-shell structures. Tech Science Press. 2011; 6(3): 155–185. 28. Goncalves PB, Frederico MAS, Zenon JGNP. Transient stability of empty and fluid filled cylindrical shells. J Brazil Soc Mech Sci Engg. 2006; 28(3): 331–338. 29. Amabili M. Eigenvalue problems for vibrating structures coupled with quiescent fluids with free surface. J Sound Vib. 2000; 231(1): 79–97. 30. Kwak MK, Koo JR, Bae CH. Free vibration analysis of a hung clamped-free cylindrical shell partially submerged in fluid. J Fluids Struct. 2011; 27: 283–296. 31. Natsuki T, Ni QQ, Endo M. Vibrational analysis of fluid-filled carbon nanotubes using the wave propaga- tion approach. Appl Phys A. 2008; 90: 441–445. 32. Iqbal Z, Naeem MN, Sultana N, Shahid HA, Shah AG. Vibration characteristics of FGM circular cylindri- cal shells filled with fluid using wave propagation approach. Appl Math Mech. 2009; 30(11): 1393–1404. 33. Shah AG, Mahmood T, Naeem MN, Shahid HA. Vibration characteristics of fluid filled cylindrical shells based on elastic foundations. Acta Mech. 2011; 216: 17–28. 34. Paak M, Paidoussis MP, Misra AK. Nonlinear vibrations of cantilevered circular cylindrical shells in con- tact with quiescent fluid. J Fluids Struct. 2014; 49: 283–302. 35. Tran IT, Manh CN. Dynamic stiffness method for free vibration of composite cylindrical shells containing fluid. Appl Math Model. 2016; 40: 9286–9301. 36. Nurul Izyan MD, Viswanathan KK, Aziz ZA, Prabakar K. Free vibration of layered cylindrical shells filled with fluid. Appl Math Mech (Eng Ed). 2016; 37(6): 803–820. 37. Nurul Izyan MD, Aziz ZA, Viswanathan KK. Free vibration of anti- symmetric angle-ply layered circular cylindrical shells filled with quiescent fluid under first order shear deformation theory. Compos Struct. 2018; 193: 189–197. 38. Li C, Yao L, Chen W, Li S. Comments on nonlocal effects in nano-cantilever beams. Int J Engg Sci. 2015; 87: 47–57. 39. Shen JP, Li C. A semi-continuum-based bending analysis for extreme-thin micro/nano-beams and new proposal for nonlocal differential constitution. Compos Struct. 2017; 172: 210–220. PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 17 / 18 Vibration of cylindrical shell with fluid 40. Bickley WG. Piecewise cubic interpolation and two point boundary problems. Comp J. 1968; 11: 206– 208. 41. Hufenbach W, Holste C, Kroll L. Vibration and damping behaviour of multi-layered composite cylindrical shells. Compos Struct. 2002; 58: 165–174. 42. Reddy JN. Mechanics of laminated composite plates and shells; Theory and analysis. 2th ed. London: CRC Press; 2004. 43. Reddy JN. Theory and analysis of elastic plates and shells. 2th ed. London: CRC Press; 2007. 44. Pai PF, Schulz MJ. Shear correction factors and an energy-consistent beam theory. Inter J Solids Struct. 1999; 36: 1523–1540. 45. Whitney JM. Shear correction factors for orthotropic laminates under static load. J Appl Mech. 1973; 40: 302–304. 46. Zhang XM, Liu GR, Lam KY. Coupled vibration analysis of fluid filled cylindrical shells using the wave propagation approach. App Acous. 2001; 62: 229–243. 47. Bhimaraddi A. Large amplitude vibrations of imperfect antisymmetric angle-ply laminated plates. J Sound Vib. 1993; 162(3): 457–470. 48. Li C, Li S, Yao L, Zhu Z. Nonlocal theoretical approaches and atomistic simulations for longitudinal free vibration of nanorods/nanotubes and verification of different nonlocal models. Appl Math Modelling. 2015; 39(15): 4570–4585. 49. Xu XJ, Deng ZC, Zhang K, Xu W. Observations of the softening phenomena in the nonlocal cantilever beams. Compos Struct. 2016; 145: 43–57. PLOS ONE \| https://doi.org/10.1371/journal.pone.0219089 July 3, 2019 18 / 18	null
10.1242/dmm.041962	null	null	© 2020. Published by The Company of Biologists Ltd \| Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 RESEARCH ARTICLE Cell-specific and athero-protective roles for RIPK3 in a murine model of atherosclerosis Sarah Colijn1,2,, Vijay Muthukumar1,‡, Jun Xie1, Siqi Gao1,2 and Courtney T. Griffin1,2,§ ABSTRACT Receptor-interacting protein kinase 3 (RIPK3) was recently implicated in promoting atherosclerosis progression through a proposed role in macrophage necroptosis. However, RIPK3 has been connected to numerous other cellular pathways, which raises questions about its actual role in atherosclerosis. Furthermore, RIPK3 is expressed in a multitude of cell types, suggesting that it may be physiologically relevant to more than just macrophages in atherosclerosis. In this study, Ripk3 was deleted in macrophages, endothelial cells, vascular smooth muscle cells or globally on the Apoe−/− background using Cre-lox technology. To induce atherosclerosis progression, male and female mice were fed a Western diet for three months before tissue collection and analysis. Surprisingly, necroptosis markers were nearly undetectable in atherosclerotic aortas. Furthermore, en face lesion area was increased in macrophage- and endothelial-specific deletions of Ripk3 in the descending and abdominal regions of the aorta. Analysis of bone-marrow-derived macrophages and cultured endothelial cells revealed that Ripk3 deletion promotes expression of monocyte chemoattractant protein 1 (MCP-1) and E-selectin in these cell respectively. Western blot analysis showed upregulation of MCP-1 in aortas with Ripk3-deficient macrophages. Altogether, these data suggest that RIPK3 in macrophages and endothelial cells protects against atherosclerosis through a mechanism that involve necroptosis. This protection may be due to RIPK3-mediated suppression of pro- inflammatory MCP-1 expression in macrophages and E-selectin expression in endothelial cells. These findings suggest a novel and unexpected cell-type specific and athero-protective function for RIPK3. likely does not types, This article has an associated First Person interview with the first author of the paper. KEY WORDS: Necroptosis, Macrophages, Endothelial cells, MCP-1, Mouse 1Cardiovascular Biology Research Program, Oklahoma Medical Research Foundation, Oklahoma City, OK 73104, USA. 2Department of Cell Biology, University of Oklahoma Health Sciences Center, Oklahoma City, OK 73190, USA. Present address: Department of Cell Biology and Physiology, Washington University in St. Louis, School of Medicine, St. Louis, MO 63110, USA. ‡Present address: The Jackson Laboratory, Bar Harbor, ME 04609, USA. §Author for correspondence ([email protected]) C.T.G., 0000-0001-8100-3171 This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution and reproduction in any medium provided that the original work is properly attributed. Received 14 August 2019; Accepted 18 December 2019 INTRODUCTION Atherosclerosis is a highly complex pathology that is influenced by a multitude of environmental and genetic factors. The cellular mechanisms that have been linked to atherosclerosis development are vast in number and include metabolic imbalances, inflammation, fibrosis, mechanotransduction and cell death (Lusis, 2000; Tabas et al., 2015). Occasionally, these mechanisms are controversial when there are conflicting reports about the role of a particular gene in atherosclerosis. However, in order to safely translate new knowledge into a clinical setting, the benefits and disadvantages of targeting specific genes must be explored. Cell death is one of the leading causes of inflammation and necrotic core formation in atherosclerosis (Van Vré et al., 2012). Although apoptosis and secondary necrosis occur in a variety of cells in atherosclerotic plaques (Van Vré et al., 2012), recent reports propose that plaque macrophages also undergo necroptosis, a newly identified form of programmed cell death (Lin et al., 2013; Meng et al., 2015; Karunakaran et al., 2016; Leeper, 2016). Necroptosis relies on the essential executioner receptor-interacting protein kinase 3 (RIPK3) and phosphorylation of its downstream effector mixed lineage kinase domain-like protein (MLKL), which then acts to permeabilize the plasma membrane (Dhuriya and Sharma, 2018; Linkermann and Green, 2014). Similar to secondary necrosis, necroptosis culminates with a release of intracellular components that initiate an immune response. This type of programmed death is often associated with immune cells, and several studies show that wild-type macrophages are susceptible to necroptosis, whereas Ripk3−/− macrophages are protected (Karunakaran et al., 2016; Lin et al., 2013; Meng et al., 2015). As RIPK3 is necessary for necroptosis – and as necroptosis is considered to be inherently inflammatory – researchers have suggested that RIPK3 or MLKL should be targeted to decrease atherosclerosis severity in the clinical setting (Zhe-Wei et al., 2018; Coornaert et al., 2018). However, more recent work has revealed that RIPK3 has pleiotropic roles beyond necroptosis (Silke et al., 2015; Moriwaki and Chan, 2016; Vince and Silke, 2016; He and Wang, 2018; Weinlich et al., 2016). These new mechanisms include NF-κB- induced cytokine production and NLRP3 inflammasome-induced or caspase 8-induced IL-1β activation, which expand the pro- inflammatory capacity of RIPK3 activity beyond necroptosis. Surprisingly, RIPK3 has also been reported to promote aerobic metabolism through phosphorylation of several metabolic enzymes (Yang et al., 2018), thus it is also possible for RIPK3 to act in a non- inflammatory manner. Overall, these alternative functions for RIPK3 are often unacknowledged in disease studies. When Ripk3 is genetically deleted in a murine model of atherosclerosis, one report shows that atherosclerotic lesion area, necrotic area and macrophage infiltration are decreased (Lin et al., 2013). Another report shows that the necroptosis chemical inhibitor necrostatin-1 improves atherosclerosis severity (Karunakaran et al., 2016). However, these reports do not fully explore the pleiotropic 1 s m s i n a h c e M & s l e d o M e s a e s i D RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 roles of RIPK3, and instead propose that RIPK3 causes plaque macrophages to undergo inflammatory necroptosis. Moreover, as necrostatin-1 has many off-target effects and can inhibit apoptosis and necroptosis-independent inflammatory pathways (Vandenabeele et al., 2012), it is a non-ideal inhibitor for examining the specific effects of necroptosis. Furthermore, these studies do not address the fact that different cell types tend to use pro-inflammatory components very differently (Mussbacher et al., 2019), and thus RIPK3 could be playing alternative roles in each of the various cell types of the plaque. As RIPK3 is a widely expressed protein – as reported by the Human Protein Atlas (Uhlén et al., 2015) – there is potential for RIPK3 to have tissue- or cell-specific functions. To explore the cell- specific function of RIPK3 in the vasculature, and to confirm which cell types – if any – undergo necroptosis in atherosclerosis, we developed a conditional model of Ripk3 deletion that utilizes a Ripk3 locus integrated with loxP sites (Colijn et al., 2019). We conducted this study by using the Ripk3-floxed mice crossed with macrophage-, vascular smooth muscle cell- and endothelial cell- specific Cre recombinases on the Apoe−/− murine model of atherosclerosis. This conditional deletion of RIPK3 aids in understanding how cell-specific RIPK3 inhibition affects atherosclerosis and gives insight into the consequences of targeting components of the necroptosis pathway in a disease context. We now report that RIPK3 plays a biologically relevant role in atherosclerosis in macrophages and endothelial cells through an athero-protective – and likely non-necroptotic – mechanism. Our data indicate that RIPK3 plays an anti-inflammatory role in these cell types, possibly through the suppression of monocyte chemoattractant protein-1 (MCP-1; also known as CCL2) in macrophages and E-selectin (SELE) in endothelial cells. These results provide novel information about unexpected roles for RIPK3 in an inflammatory vascular disease, and raise questions about our previous understanding of the relationship between RIPK3, necroptosis, inflammation and atherosclerosis. RESULTS Ripk3 transcripts are present in atherosclerotic plaques at very low copy numbers To explore the role of RIPK3 in the various cell types of atherosclerosis, we first attempted to look at RIPK3 expression in the plaque regions. Unfortunately, as is fairly common for plaque immunostaining, all commercial antibodies that we used to detect RIPK3 showed widespread non-specific staining, which was confirmed by using Ripk3-knockout plaques as controls (data not shown). As an alternative to immunostaining, we used RNA in situ hybridization with RNAScope® technology to identify the expression pattern of Ripk3. After confirmation of the specificity of the Ripk3 probe (Fig. S1) and after identifying plaque areas with endothelial cells, macrophages and smooth muscle cells, we showed that Ripk3 transcripts were nearly undetectable in these regions (Fig. 1A-H). In fact, Ripk3 levels were even lower than Polr2a levels, which is a ubiquitously expressed positive control gene that is known to produce very low transcript copy numbers per cell (Fig. 1I-N) (Bingham et al., 2017). This very low copy number for Ripk3 transcripts prevented dual immunostaining to identify cell types with Ripk3 expression, as the additional steps would wash away the probe signal; however, we were able to immunostain sequential sections to confirm that all the pertinent cell types were present (Fig. 1A-D). Regardless, it is apparent that Ripk3 transcript copy number in our atherosclerotic plaques, and that immunostaining is not ideal for is very low in all cell types present detecting RIPK3 protein in plaque tissues. However, as previous studies have already linked RIPK3 and atherosclerosis (Lin et al., 2013; Meng et al., 2015), we suspected that this low Ripk3 transcript copy number did not accurately predict RIPK3 activity in the aorta. RIPK3 deletion in macrophages or endothelial cells exacerbates atherosclerosis Next, we wanted to determine the biological relevance of cell-specific RIPK3 in atherosclerosis, so we conditionally deleted Ripk3 in macrophages (Ripk3ΔMΦ-Cre), smooth muscle cells (Ripk3ΔSMC-Cre), endothelial cells (Ripk3ΔEC-Cre) and globally (Ripk3ΔGlobal) on the Apoe−/− background using Cre-lox technology and LysM (Lyz2)-Cre, SM22 (Tagln)-Cre, VE-Cadherin (Cdh5)-Cre and germline excision, respectively. For brevity and clarity, simplified nomenclature is used throughout the article in place of the complete genotypes. The full genotypes, associated nomenclature, and important notes are described in Table 1. These mice were then fed a Western diet for three months, at which point the tissues were collected for analysis. It should be noted as a caveat to our study that we did not obtain as many of the Ripk3ΔGlobal mice as we did the cell-specific knockouts owing to the difficulties associated with breeding and maintaining the Apoe−/− lineage. of by deletion As VE-Cadherin-Cre has been reported to delete in a portion of the adult bone marrow (Alva et al., 2006), and LysM-Cre is occasionally inefficient (Clausen et al., 1999), we investigated the efficiency bone-marrow-derived macrophages (BMDMs) and analyzing their RIPK3 expression levels (Fig. 2A-F). LysM-Cre showed robust deletion efficiency by transcript and protein analysis. As expected, VE-Cadherin-Cre showed a partial deletion of RIPK3 in BMDMs as well, meaning that we could not ignore a potential role for Ripk3-deficient macrophages in Ripk3ΔEC-Cre mice. isolating We also sought to confirm that the Ripk3-floxed allele – which results in removal of the first nine out of the ten exons of Ripk3 after Cre-mediated excision (Colijn et al., 2019) – does not produce a transcribed version of exon 10. This exon encodes a functional domain that is involved in several kinase-independent processes (Moriwaki et al., 2017), which, if still translated, could confound our results. However, transcript analysis using primers designed specifically for exon 10 show that there is no possibility for translation of a truncated form of RIPK3 (Fig. 2G). Furthermore, Ripk3ΔGlobal aortas show no detectable RIPK3 protein by immunoblotting (Fig. S2), altogether indicating that we were able to generate a true knockout allele. If are correct reports previous that RIPK3 increases atherosclerosis severity and that necroptotic macrophages are the main source of that increase in severity, then we would expect to see less atherosclerotic phenotypic severity in the Ripk3ΔGlobal, Ripk3ΔHet, and Ripk3ΔMΦ-Cre mice. En face Oil Red O staining revealed that Ripk3ΔMΦ-Cre and Ripk3ΔEC-Cre aortas showed an increase in percent atherosclerotic lesion area, whereas Ripk3ΔHet and Ripk3ΔGlobal aortas were no different from control, which conflicts with the previously published genetic study addressing the role of RIPK3 in atherosclerosis progression (Fig. 3A,B) (Lin et al., 2013). Instead, our data suggest that RIPK3 plays an athero- protective role in macrophages and endothelial cells. We also observed that the bulk of the increased plaque burden in Ripk3ΔMΦ-Cre and Ripk3ΔEC-Cre mice was detected in the descending and abdominal the aorta, which are typically athero-protected, as opposed to the athero-prone arch region (Fig. 3C-E). In addition, we saw no significant differences in lesion size, necrotic core size, macrophage area, smooth muscle cell regions of 2 s m s i n a h c e M & s l e d o M e s a e s i D RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 Fig. 1. Ripk3 transcripts are present in atherosclerotic plaques at very low copy numbers. (A-D) Sequential cross sections of control aortic roots were immunostained for the endothelial cell marker CD31 (green) and macrophage marker CD68 (red) (A,C) or for the smooth muscle cell marker aSMA (red) (B,D) and were co-stained to identify nuclei (Hoechst; blue). Two plaque regions that contain all three cell types are shown. (E-N) RNAScope® RNA in situ hybridization was performed on sequential sections of the two separate regions of control aortic roots shown in A-D. Ripk3 probe was used to identify Ripk3 transcripts ( pink) and co-stained with DAPI (blue) to identify nuclei (E,G). F and H show magnifications of the boxed regions (without DAPI) in E and G, respectively. Positive control probes Ubc ( pink; high transcript copy number) and Polr2a (red; low transcript copy number) were used to show the efficacy of the assay (I,L). J,K and M,N show magnifications of the boxed regions (without DAPI) in I and L, respectively. Positive signal is determined by the presence of ‘dots’ which are each meant to represent a single transcript. Images are representative of in situ hybridization experiments performed on sections from three separate control mice (n=3). Scale bars: 25 µm. area, collagen area or calcification in the athero-prone aortic roots within the arch region (Fig. 4A-R and Fig. S3). We also detected no increase in apoptosis in the aortic root, though apoptotic cells were infrequent enough that quantification could not be performed (data not shown). Interestingly, western blot analysis of the different aortic regions (arch, descending and abdominal; as demarcated in Fig. 3C) showed a significant increase in CD11b signal in the arch region of Ripk3ΔEC-Cre aortas (Fig. 4S,T). As CD11b is a in assists that monocyte/macrophage leukocyte trafficking (Schittenhelm et al., 2017), these data suggest that there is an increase in macrophage number in the arch receptor component region of Ripk3ΔEC-Cre aortas that is not reflected in the anti- CD68 immunostained aortic roots (Fig. 4R) or en face analyses (Fig. 3E). Triglyceride, cholesterol, glucose, and blood cell counts are unchanged between the genotypes To confirm that the changes in Ripk3ΔMΦ-Cre and Ripk3ΔEC-Cre mice were not due to a change in overall metabolism or blood content, we analyzed plasma triglyceride, cholesterol and glucose levels and found no difference between the genotypes (Fig. S4). We also found no difference in peripheral mononuclear cells and other blood cells (Fig. S5). Table 1. Genotypes, nomenclature, Cre-specificity and other important notes Genotype Nomenclature Cre specificity Notes Apoe−/−;Ripk3fl/fl Apoe−/−;Ripk3fl/Δ Apoe−/−;Ripk3Δ/Δ Apoe−/−;Ripk3fl/fl;LysM-Cre+ Apoe−/−;Ripk3fl/fl;SM22-Cre+ Apoe−/−;Ripk3fl/fl;VE-Cadherin-Cre+ Ripk3ΔHet and Ripk3ΔGlobal mice were generated by utilizing the germline excision phenomenon that occurs when Cre activity aberrantly turns on in the gonads of the breeders (Harno et al., 2013), thereby altering the Ripk3-floxed allele into a Ripk3-knockout allele. – – – Macrophages and neutrophils Vascular smooth muscle cells Endothelial cells and a subset of hematopoietic cells Cre-negative littermate controls Global heterozygous knockout of Ripk3 Global homozygous knockout of Ripk3 Clausen et al., 1999 Holtwick et al., 2002 Alva et al., 2006 Control Ripk3ΔHet Ripk3ΔGlobal Ripk3ΔMΦ-Cre Ripk3ΔSMC-Cre Ripk3ΔEC-Cre s m s i n a h c e M & s l e d o M e s a e s i D 3 RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 Fig. 2. Bone-marrow-derived macrophages exhibit significant reduction in RIPK3 levels in Ripk3ΔGlobal, Ripk3ΔMΦ-Cre and Ripk3ΔEC-Cre mice. (A,B) BMDMs were isolated, differentiated on chamber slides for 7 days, immunostained for macrophage markers CD68 (red; A) and CD11b ( pink; B), and co-stained for nuclei (Hoechst; blue). All cells from seven separate BMDM isolations were positive for these markers (n=7). (C) A phagocytosis assay was performed on BMDMs with fluorescent polystyrene beads. Phagocytic cells display TRITC+/Hoechst+ signal, and 85% of cells from two separate BMDM isolations were double positive (n=2). (D-F) Protein or RNA was collected from control, Ripk3ΔGlobal, Ripk3ΔMΦ-Cre and Ripk3ΔEC-Cre BMDMs. RNA was converted to cDNA and analyzed by qPCR for Ripk3 levels (D). Protein lysates were immunoblotted to identify CD11b, RIPK3 and β-actin (loading control) (E) and quantified (F). (G) RNA from control and Ripk3ΔGlobal BMDMs was converted to cDNA and analyzed by qPCR for Ripk3-Exon 10 levels. For panels D,F,G, each dot represents a BMDM isolation from an individual animal. Statistics for panels D and F were calculated using one-way ANOVA with Dunnett’s multiple comparisons test. Overall ANOVA P-values ( prior to the post hoc tests) are 0.0002 (D) and <0.0001 (F). Statistics for panel G were calculated using an unpaired t-test with Welch’s correction. P<0.05, P<0.01, P<0.001, **P<0.0001. Data are mean±s.d. Scale bars: 25 µm. The main role of RIPK3 in atherosclerosis is likely not to initiate necroptosis Owing to the unexpected exacerbation of atherosclerosis in Ripk3ΔMΦ-Cre and Ripk3ΔEC-Cre mice, we questioned whether necroptosis – which is considered to be inherently inflammatory and should increase atherosclerosis severity – was detectable in our atherosclerotic aortas. Because of widespread non-specific staining that occurred with the commercial antibodies that we used to immunostain phosphorylated MLKL ( p-MLKL), we alternatively used western blot techniques to detect this marker of necroptosis. We found that p-MLKL levels were so low as to be mostly undetectable by western blotting using two different anti-p-MLKL antibodies (Fig. 5A-C and Fig. S6). Thus, we suspect that the main role of RIPK3 in atherosclerosis is not to cause macrophage necroptosis, but rather to prevent macrophages and endothelial cells from becoming pro-inflammatory. for Furthermore, we found that RIPK3 protein levels in different aortic regions and genotypes were not significantly different from control levels by western blot (Fig. 5D,E). If macrophages, smooth muscle cells or endothelial cells predominantly expressed RIPK3 over the other cell types, then we would have expected to see a decrease in RIPK3 expression in the corresponding mutant genotype. As we did not, this suggests that the bulk of RIPK3 expression detected by western blotting in the aorta is not restricted to any one of these three cell types. Ripk3 deficiency impacts MCP-1 and IL-1β expression in BMDMs, and MCP-1 and E-selectin expression in cultured endothelial cells To try to understand how Ripk3 deletion in macrophages could lead to an increase in atherosclerosis severity, we isolated and cultured BMDMs from control and Ripk3ΔMΦ-Cre bone marrow and performed real-time quantitative PCR (qPCR) for a panel of genes that are related to necroptosis, inflammation, macrophage activation and endothelial cell activation (Fig. 6A and Fig. S7A). We found that Ccl2 (MCP-1) and Il1b (IL-1β) transcripts were significantly changed in Ripk3ΔMΦ-Cre BMDMs. Ccl2 was upregulated almost 12-fold in Ripk3ΔMΦ-Cre BMDMs over control BMDMs, and ELISA analysis of the medium confirmed the increase in MCP-1 and decrease in IL-1β (Fig. 6B,C). RIPK3 has already been implicated in IL-1β processing (Weinlich et al., 2016) and has been shown to promote Il1b transcription (Moriwaki et al., 2017), but this is the first report implicating RIPK3 in the suppression of Ccl2 expression. s m s i n a h c e M & s l e d o M e s a e s i D 4 RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 Fig. 3. Ripk3ΔMΦ-Cre and Ripk3ΔEC-Cre en face aortas exhibit increased lesion areas in the descending and abdominal aortic regions. (A) Representative images of control and Ripk3-deficient aortas from male and female mice that were fed the Western diet for 3 months. Aortas were then dissected, cut open en face and stained with ORO to visualize lesions. (B) Lesion area was quantified as the percent lesion area of the whole aortic area. (C) Aortic areas were analyzed separately and divided into the arch, descending and abdominal regions. The arch region extends from where the aorta enters the heart to the end of the curvature. The descending region extends to the renal arteries. The abdominal region extends to the iliac arteries. (D,E) Lesion area was quantified for the descending and abdominal regions (D) and arch region (E). For panels B,D, E, each dot represents an aorta from an individual animal. Statistics were calculated using one-way ANOVA and Dunnett’s multiple comparisons test for panels B and D. Overall ANOVA P-values ( prior to the post hoc tests) are 0.005 (B), 0.001 (D) and 0.33 (E). Data are mean±s.d. To try to understand how Ripk3 deletion in endothelial cells could lead to an increase in atherosclerosis severity, we utilized a cultured murine pancreatic endothelial cell line (MS1 cell line), which we chose for this study because it expresses high levels of endogenous RIPK3. By knocking down RIPK3 using siRNA oligos and looking at a similar panel of genes, we found that Ccl2 transcripts and MCP-1 protein levels were also elevated in endothelial cells, to the extent seen in BMDMs (Fig. 6D-F and although not Fig. S7B). Furthermore, Sele transcripts were (E-selectin) upregulated approximately fourfold in RIPK3-knockdown MS1 cells. This is relevant as E-selectin is a glycoprotein surface receptor on endothelial cells that assists in the recruitment of circulating leukocytes (McEver, 2015). Based on these data, we hypothesized that increased macrophage MCP-1 levels contributed to the more severe atherosclerosis we detected in Ripk3ΔMΦ-Cre mice, as macrophage-specific expression of MCP-1 is known to exacerbate atherosclerosis (Aiello et al., 1999). We also hypothesized that increased endothelial MCP-1 and E-selectin levels contributed to the elevated atherosclerosis severity in Ripk3ΔEC-Cre mice, as these molecules would increase the recruitment of leukocytes and therefore accelerate atherosclerosis. MCP-1 levels are elevated in Ripk3ΔMΦ-Cre abdominal aortas, and E-selectin levels trend higher in Ripk3ΔEC-Cre aortas To see whether the BMDM and MS1 cell functions of RIPK3 translated to our in vivo tissues, we analyzed MCP-1 and E-selectin levels in the arch, descending and abdominal regions of mouse aortas using western blotting. Ripk3ΔMΦ-Cre aortas exhibited an increase in MCP-1 in the abdominal region that was if variable, compared to controls (Fig. 7A,B). This robust, supports the hypothesis that loss of RIPK3 in macrophages increases macrophage MCP-1 levels in the abdominal region of Ripk3ΔMΦ-Cre aortas, which correlates with the increase in abdominal lesion area (Fig. 3D). In contrast, Ripk3ΔEC-Cre aortas did not show an increase in MCP-1 but showed an upward trend in E-selectin levels by western blot (Fig. 7C,D). We immunostained regions of the descending aorta to see whether the increase in E-selectin would be more apparent using another technique; however, we did not see an increase in E-selectin intensity (Fig. S8A,B), though this could be attributed to the fact that the E-selectin signal was already very high in control aortas and thus it was difficult to see a change in intensity. We also examined ICAM1 and VCAM1 levels because of their high expression in s m s i n a h c e M & s l e d o M e s a e s i D 5 RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 Fig. 4. Aortic root analyses reveal no differences in lesion area, necrotic area or macrophage area, though Ripk3ΔEC-Cre aortas exhibit higher CD11b signal. (A-R) After 3 months on a Western diet, hearts were dissected and aortic roots were sectioned and analyzed for lesion area, necrotic area and macrophage area. Aortic roots were stained with ORO and hematoxylin to visualize lesion area (black arrows) (A-E). Lesion area was reported as a fraction of the total aortic area (P). Aortic roots were stained with H&E to visualize necrotic areas (F-J). Necrotic areas were defined as acellular and are outlined in black. Necrotic area was reported as a percent of the lesion area (Q). Aortic roots were immunostained for CD68 (red) and nuclei (Hoechst; blue) to visualize macrophage-occupied area (K-O). Macrophage area was reported as the CD68+ percent of the lesion area (R). (S,T) Protein was collected from control (n=8), Ripk3ΔMΦ-Cre (n=5) and Ripk3ΔEC-Cre (n=4) aortas. Protein lysates were immunoblotted to identify CD11b and β-actin (loading control) (S) and quantified (T). For panels P,Q,R,T, each dot represents an individual animal. Statistics for panels P-R were calculated using one-way ANOVA. Overall ANOVA P-values ( prior to the post hoc tests) are 0.42 (P), 0.01 (Q) and 0.30 (R). Despite the significant overall ANOVA P-value for panel Q, Dunnett’s multiple comparisons test shows no significant difference from control aortic roots. Statistics for panel T were calculated using two-way ANOVA with Dunnett’s multiple comparisons test with an overall ANOVA P-value of 0.01. Data are mean±s.d. Scale bars: 200 µm. athero-prone regions and their connection to inflammation (Nakashima et al., 1998). However, neither gene was altered in Ripk3-deficient BMDMs or MS1 cells (Fig. S7); nor were ICAM1 in Ripk3ΔEC-Cre or and VCAM1 levels significantly different Ripk3ΔMΦ-Cre aortas (Fig. S8C-E). In addition, as RIPK1 is a homologous kinase upstream of RIPK3 and is often tied to RIPK3 function (Moriwaki and Chan, 2016), we immunoblotted for RIPK1 in the various regions of the aorta. When Ripk3 is deleted in a cell-specific manner, RIPK1 levels are not noticeably altered in the aorta lysate (Fig. S9A). s m s i n a h c e M & s l e d o M e s a e s i D 6 RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 Fig. 5. p-MLKL levels are nearly undetectable in advanced plaques. (A-C) After 3 months on a Western diet, protein was collected from control (n=7), Ripk3ΔMΦ-Cre (n=5), Ripk3ΔEC-Cre (n=3) and Ripk3ΔSMC-Cre (n=3) aortas. Protein lysates were immunoblotted to identify p-MLKL (Abcam; #ab196436), MLKL and β-actin (loading control) (A) and quantified (B,C). A faint positive p-MLKL signal can be seen in the representative blot in the control arch region; however, we could only detect p-MLKL in two out of the 18 aortas analyzed (B). Note that the same transfer membrane used for detecting CD11b (in Fig. 4S) was reprobed for p-MLKL and MLKL in A; the β-actin control blots are therefore the same in both figures. (D,E) Protein lysates were immunoblotted to identify RIPK3 and GAPDH (loading control) (D) and quantified (E). For panels B,C,E, each dot represents an individual animal. Statistics were calculated using two-way ANOVA. Overall ANOVA P-values are 0.76 (B), 0.07 (C) and 0.46 (E). Data are mean±s.d. However, aortas with heterozygous and global deletion of Ripk3 show an appreciable decrease in RIPK1 levels (Fig. S9B), indicating that RIPK1 protein levels are tied to RIPK3 levels. However, owing to the fact that RIPK1 plays roles in regulating necroptosis, apoptosis, autophagy, inflammation and cell survival pathways (Lin, 2014), it is difficult to speculate how the substantial decrease in RIPK1 affects atherosclerosis in Ripk3-deficient animals. Plasma IL-1β levels are increased in Ripk3ΔGlobal mice, whereas plasma MCP-1 levels are unaffected by Ripk3 deletion To determine whether the in vitro changes we detected in MCP-1 or IL-1β levels could be seen in the plasma of our mutant mice, we performed an ELISA assay on the plasma samples collected from the different genotypes. However, we detected no differences in MCP-1 levels (Fig. 8A). Although Ripk3ΔMΦ-Cre plasma did not show a decrease in IL-1β levels as would have been expected from the BMDM data, we did see a significant increase of IL-1β in the Ripk3ΔGlobal plasma (Fig. 8B). This suggests that, although RIPK3 has been reported to positively regulate IL-1β processing (Weinlich et al., 2016), it may also play a suppressive role in certain conditions or cell types. DISCUSSION Until now, RIPK3 has been associated primarily with necroptosis and – to a lesser extent – other pro-inflammatory pathways separable from necroptosis. In this study, which is the first to explore the cell- specific function of RIPK3 in the atherosclerotic aorta, we demonstrate that RIPK3 plays an unexpected athero-protective role in macrophages and endothelial cells. Here, we propose that the main role of RIPK3 in atherosclerosis is neither to promote IL-1β processing, which are the previously necroptosis nor described functions for RIPK3 (Weinlich et al., 2016). Instead, our data indicate that RIPK3 may play an anti-inflammatory role by suppressing MCP-1 expression in macrophages and E-selectin expression in endothelial cells (data summarized in Fig. S10). These results provide novel information about RIPK3 in an inflammatory vascular disease and raise questions about the efficacy and safety of targeting it in a clinical setting. We were surprised by the fact that our mice with global and heterozygous Ripk3 deletions did not resemble previously reported s m s i n a h c e M & s l e d o M e s a e s i D 7 RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 Fig. 6. Ripk3-deficiency impacts MCP-1 and IL-1β expression in BMDMs, and MCP-1 and E-selectin expression in cultured endothelial cells. (A-C) BMDMs were isolated from control (n=6) and Ripk3ΔMΦ-Cre (n=4) bone marrow, and RNA and medium were collected 7 days later. RNA was converted to cDNA and analyzed by qPCR (A). ELISA assays for MCP-1 (B) and IL-1β (C) were performed on the BMDM medium. Each dot (n value) represents a BMDM isolation from an individual animal. (D-F) MS1 endothelial cells were transfected with NS- and RIPK3-siRNA oligos for 24 h and were cultured in low-serum medium for another 24 h before RNA, medium and protein were collected (n=4 independent experiments). RNA was converted to cDNA and analyzed by qPCR (D). Protein lysates were immunoblotted for RIPK3 and the loading control, GAPDH (n=4 independent experiments) (E). An ELISA assay for MCP-1 was performed on the BMDM medium (n=3 independent experiments) (F). Statistics were calculated using unpaired t-tests, with Welch’s correction when necessary. P<0.05, **P<0.01. Data are mean±s.d. Ripk3 mutant mice that showed a dose dependent improvement in atherosclerotic lesion area (Lin et al., 2013). It is possible that this inconsistency may be due to different background strains in the mice that were analyzed. The Ripk3−/− mice used in the original study were derived from C57BL/6N embryonic stem cells (Newton et al., 2004), whereas our Ripk3fl/fl mice were derived from C57BL/6J embryonic stem cells (Colijn et al., 2019). Lin et al. reported that they performed their study on a C57BL/6J background; however, it is unclear for how many generations the original C57BL/6N strain was crossed onto the C57BL/6J strain. Although this distinction may appear to be minor, much work has been done to map the genetic and metabolic differences between the two backgrounds, and a difference of phenotype is entirely possible (Fontaine and Davis, 2016). Thus, in addition to revealing cell-specific roles for RIPK3, we may also have uncovered a background-specific regulation or function of RIPK3 that future studies should consider. Another potential cause of the discordant results from our genetic study and that of Lin et al. is the different environments in which the mice were housed and analyzed. The two studies were performed at different institutions and with different suppliers of the Western diet (Lin et al., 2013). This raises questions about the contribution of microbiota to RIPK3-dependent functions within the context of atherosclerosis. It has already been well established that the microbiome can influence atherosclerosis (Jie et al., 2017), thus microbiota that differentially affect inflammatory pathways could also differentially affect RIPK3 activation. As humans have such diverse and dissimilar microbiota, the effect of the microbiome on RIPK3 function should be explored before RIPK3 is considered as a therapeutic target. It is also possible that differences between the Ripk3−/− genetic model used by Lin et al. and our Ripk3fl/fl genetic model could contribute to the discrepancies between our studies. Specifically, whereas our Ripk3fl/fl mouse model results in removal of exons one through nine of Ripk3 upon Cre-mediated recombination, only the first three exons of Ripk3 are targeted for deletion in the Ripk3−/− mouse model (Colijn et al., 2019; Newton et al., 2004). Ripk3 has been reported to have two alternative splicing variants in addition to the full length form (Yang et al., 2005), and although part of the alternative spliced sequences include the first three exons and therefore may not be transcribed in the Ripk3−/− mouse, this was not addressed in the design of the Ripk3 knockout locus. Though unlikely, it is possible that a truncated or frame-shifted form of RIPK3 was present in the mice used by Lin et al. that could account for the difference between our results. Finally, it is worth noting that another study that looked at atherosclerosis in an Apoe−/−;Ripk3−/− mouse model did not completely replicate the study by Lin et al., although it did conclude that RIPK3-mediated necroptosis contributes to atherosclerosis pathology (Meng et al., 2015). The authors reported no difference in macrophage infiltration and did not observe a significant difference in aortic root lesion area in very advanced plaques, which is contrary to the study by Lin et al. Though Meng et al. did not bring attention to these differences, it is clear that there is some variation in RIPK3 atherosclerosis studies that should be considered. s m s i n a h c e M & s l e d o M e s a e s i D 8 RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 Fig. 7. MCP-1 levels are elevated in Ripk3ΔMΦ-Cre abdominal aortas, and E-selectin levels trend higher in Ripk3ΔEC-Cre aortas. (A-E) After 3 months on a Western diet, protein was collected from control (n=8), Ripk3ΔMΦ-Cre (n=5) and Ripk3ΔEC-Cre (n=4) aortas. Protein lysates were immunoblotted to identify MCP-1, E-selectin, GAPDH and β-actin (loading controls) (A,C) and quantified (B,D,E). The immunoblot in panel C shows non-contiguous lanes of a single gel. For panels B,D,E, each dot represents an individual animal. Statistics were calculated using two-way ANOVA, and Sidak’s multiple comparisons test in panel B. Overall ANOVA P-values ( prior to the post hoc tests) are 0.09 (B), 0.02 (D) and 0.99 (E). Data are mean±s.d. despite notable Surprisingly, atherosclerotic detecting phenotypes associated with Ripk3 deficiency, we were unable to detect much Ripk3 transcript expression in atherosclerotic plaques. Although the commercial RIPK3 antibodies we used were unsuitable for assessing patterns of RIPK3 expression in plaque tissues by immunostaining, we were able to determine by immunoblotting that neither macrophages, smooth muscle cells, nor endothelial cells are the main sources of RIPK3 expression in plaques, and that RIPK3 protein expression is not restricted to a particular region of the aorta. Nevertheless, the limitations in RIPK3 immunostaining prevented us from assessing Cre specificities and efficiencies directly in plaque tissues, which would have been ideal. As plaque tissue is difficult to immunostain owing to high levels of non-specific background signal, we propose that the best way to Fig. 8. Plasma IL-1β levels are increased in Ripk3ΔGlobal mice, whereas plasma MCP-1 levels are unaffected by Ripk3 deletion. (A,B) After mice were fed a Western diet for 3 months, plasma was collected and analyzed by ELISA for MCP-1 (A) and IL-1β (B). Each dot represents an individual animal. Statistics for both panels were calculated using Kruskal– Wallis tests, and Dunn’s multiple comparisons test for panel B. Overall ANOVA P-values ( prior to the post hoc tests) are 0.97 (A) and 0.01 (B). Data are mean±s.d. 9 s m s i n a h c e M & s l e d o M e s a e s i D RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 determine RIPK3 protein expression patterns accurately in the context of atherosclerosis would be to use a RIPK3-tagged reporter mouse, such as the RIPK3-GFP mouse developed by Moriwaki et al. (2017). A major finding from our study is that RIPK3 plays unexpected anti-inflammatory roles in atherosclerosis and contributes to the athero-protection of the descending and abdominal aortic regions. However, our study is not the first to indicate that RIPK3 activity can lead to the suppression of inflammation. One report demonstrates that promoting RIPK3 activation in cultured mouse fibroblasts – through TNF-treatment and caspase inhibition – results in a marked decrease in secretion of inflammatory cytokines (Kearney et al., 2015). Our findings that IL-1β is elevated in plasma from Ripk3ΔGlobal mice (Fig. 8B) and that MCP-1 is upregulated in the abdominal aortic region of Ripk3ΔMΦ-Cre mice (Fig. 7B) provide in vivo evidence of anti-inflammatory roles for RIPK3. In addition, the upward trends in E-selectin (Fig. 7D) and in CD11b (Fig. 4T) expression in Ripk3ΔEC-Cre aortas provide further indications that RIPK3 activity suppresses inflammation in a cell-type specific manner during atherosclerosis. Intriguingly, the marked elevation in pro-inflammatory IL-1β levels from Ripk3ΔGlobal plasma did not track with an increase in plaque severity. In fact, aortic plaques were less severe in our Ripk3ΔGlobal mice than in our Ripk3ΔEC-Cre and Ripk3ΔMΦ-Cre animals. Ultimately, we do not know the reason for the phenotypic differences between our mice with global and cell- specific Ripk3 deletion, but our findings indicate that RIPK3 may play opposing roles in different interacting cell types and that global deletion may negate the damaging effects of cell-specific RIPK3 deletion in the context of atherosclerosis. suppresses One important question that arises from our findings is whether RIPK3-mediated necroptosis inflammation during atherosclerosis progression, as has been suggested for the cultured mouse fibroblast study cited above (Kearney et al., 2015). It has been proposed that Mlkl deletion is a more definitive genetic approach than Ripk3 deletion for determining how necroptosis influences various pathologies, as RIPK3 has roles outside of necroptosis and MLKL is downstream of RIPK3 in the necroptosis pathway (Newton et al., 2016). As one published abstract hints that Mlkl-deficiency does not improve atherosclerosis (Rasheed et al., 2018), and as we were unable to detect p-MLKL in the atherosclerotic plaques we analyzed (Fig. 5A,B and Fig. S6), we propose that necroptosis contributes minimally to atherosclerosis progression and to the anti-inflammatory roles we have defined for RIPK3 in this study. Altogether, we believe this study challenges the field to reevaluate the prevalent idea that RIPK3 promotes pathologies exclusively through necroptotic or other pro-inflammatory mechanisms. Our data indicate that RIPK3 has cell-specific functions and that some of those may be beneficial and anti- inflammatory in the context of the vasculature. Exploring other protective roles for RIPK3 will help clarify the potential risks and benefits of targeting RIPK3 therapeutically in various diseases. MATERIALS AND METHODS Mice (Mus musculus) Apoe−/− mice on the C57BL/6J background (Piedrahita et al., 1992), SM22- Cre transgenic mice (Holtwick et al., 2002), LysM-Cre transgenic mice (Clausen et al., 1999), VE-Cadherin-Cre transgenic mice (Alva et al., 2006) and Ripk3-floxed (Ripk3fl/fl) mice on the C57BL/6J background (Colijn et al., 2019) have been described. Ripk3-knockout (Ripk3Δ/Δ) mice were generated by utilizing the germ-line excision phenomenon that occurs when Cre activity aberrantly turns on in the gonads of the breeders (Harno et al., 2013), thereby altering the Ripk3-floxed allele into a Ripk3-knockout allele. Mice were maintained and interbred on the C57BL/6J background at the Oklahoma Medical Research Foundation (OMRF) animal facility. Male and female breeders were used interchangeably for many generations to ensure all target genotypes were on a highly uniform background strain for the analyses. The OMRF Institutional Animal Care and Use Committee approved all animal use protocols. Genotyping Apoe−/−, SM22-Cre, LysM-Cre, VE-Cadherin-Cre and Ripk3-floxed mice were genotyped as previously described (Colijn et al., 2019; Holtwick et al., 2002; Goren et al., 2009; Mani et al., 2013; Wiley et al., 2015). Ripk3-knockout genotyping was performed by PCR using the forward primer 5′-CCATCCTCCCTTCATCAAAA-3′ and reverse primer 5′-GAATGCAAATGCAGGGTCTT-3′, which are located upstream of the 5′ LoxP site and downstream of the 3′ LoxP site, respectively. The PCR produced a 312-bp band to signify that Ripk3 excision had occurred. Atherosclerosis induction Apoe−/− mice were crossed with the Ripk3fl/fl line and either SM22-Cre, LysM-Cre or VE-Cadherin-Cre mice to generate the target genotypes: Apoe−/−;Ripk3fl/fl;SM22-Cre (n=18), Apoe−/−;Ripk3fl/fl;LysM-Cre (n=19), Apoe−/−;Ripk3fl/fl;VE-Cadherin-Cre (n=9), Apoe−/−;Ripk3Δ/Δ (n=9) and Apoe−/−;Ripk3fl/fl Cre-negative littermates as controls (n=36). Two-month-old male and female mice were fed a Western diet (21.2% fat, 0.2% cholesterol; TD.88137, Envigo) for three months to induce atherosclerosis progression. The mice were then euthanized. A cardiac puncture was performed to collect blood. Mice were then perfused with PBS followed by 4% paraformaldehyde (PFA). Aortas were cleaned of fat and removed along with the hearts. Bone marrow was collected from the femur and tibia as described (Amend et al., 2016). (n=14), Apoe−/−;Ripk3fl/Δ En face analysis The aorta was cut open longitudinally from the iliac arteries to the aortic root and pinned open on parafilm. After further fixation with 4% PFA, lesions were visualized with an Oil Red O stain (ORO, Newcomer Supply), and residual fat was removed. Individual high-resolution images were stitched together by hand to create the entire en face aorta images. Aortic images were further analyzed with NIS-Elements software for ORO staining, and data are reported as the percentage of the aortic surface covered by lesions. Aortic root analysis Hearts were fixed in 1% PFA overnight, placed in 20% sucrose, then incubated in a 1:1 mixture of 20% sucrose and Optimal Cutting Temperature compound (OCT; Sakura Finetek) overnight and frozen in OCT the next day. Whole aortic roots were sectioned at 10 µm with a Microm HM 505 E cryotome (Microm International) and adhered to Denville UltraClear Microscope Slides (Denville Scientific). Slides were frozen until further use. All analyses were performed using NIS-Elements software. Lesion area Aortic roots were stained with ORO to identify lesion area. Slides were first immersed in water, then 60% isopropanol for 30 s, ORO for 30 min, then 60% isopropanol for 30 s to rinse. Slides were co-stained with hematoxylin, mounted with 1,4-diazabicyclo[2.2.2]octane (DABCO) and coverslipped. Six aortic root sections per mouse were analyzed at 60 μm apart. The average lesion area from each mouse was normalized to the average total aortic area and reported as a fraction. Necrotic area Aortic roots were stained with hematoxylin and eosin (H&E) for necrotic core quantification. Three aortic root sections per mouse were analyzed at 120 μm apart. Necrotic cores were identified as acellular areas (negative for H&E). The average necrotic area was reported as the percent necrotic area of the total lesion area. Macrophage and smooth muscle cell area Macrophages and smooth muscle cells of the aortic root were visualized using rat anti-CD68 (1:250; #MCA1957; Bio-Rad) and Cy3-conjugated 10 s m s i n a h c e M & s l e d o M e s a e s i D RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 anti-aSMA (1:250; #C6198; Sigma-Aldrich) antibodies, respectively. Sections were blocked with 5% bovine serum albumin (BSA) for 2 h, incubated with primary antibody overnight, and incubated with secondary antibody Cy3 donkey anti-rat IgG (1:500; Jackson ImmunoResearch) for CD68 and Hoechst for nuclei (20 µg/ml) for 1 h. For macrophages, three aortic root sections per mouse were analyzed at 120 μm apart. For smooth muscle cells, one aortic root section per mouse was analyzed from the middle of the root. The average CD68+ area or aSMA+ area was reported as the percent area of the total lesion area. Collagen deposition Aortic roots were stained with the Picrosirius Red Stain Kit (Polysciences) for collagen area quantification. Sections were stained with Solution A for 2 min, Solution B for 1 h and Solution C for 2 min. Two aortic root sections per mouse were analyzed at 180 μm apart. Collagen was identified as the green, yellow and red colors that can be visualized under polarized light. The average collagen area was reported as the percent collagen area of the total lesion area. Calcification Aortic roots were stained with Alizarin Red S (Sigma-Aldrich) to detect calcification. Sections were stained with a 1% Alizarin Red S ( pH 4.2) solution for 5 min. Two aortic root sections were analyzed at 180 μm apart. Calcification was recorded as either present or absent. Plasma triglyceride, cholesterol and glucose analysis After blood was drawn by cardiac puncture, plasma was isolated using Microvette CB 300 LH tubes (Sarstedt). The concentrations of total cholesterol, LDL/VLDL cholesterol, HDL cholesterol and triglycerides were determined using enzymatic colorimetric assays (#E2HL-100 and #ETGA-200; BioAssay Systems) as per the manufacturer’s instructions. Glucose levels were detected using the Contour®Next EZ Blood Glucose Monitoring System with Contour®Next test strips. Complete blood count Blood was collected into Microvette 500 K3E tubes (Sarstedt). Complete blood counts were measured with a Hemavet 950 veterinary hematology analyzer (Drew Scientific). RNA in situ hybridization Hearts were fixed in 4% PFA for 24 h and passed through a sucrose gradient as per RNAScope® protocol, then embedded in OCT. After 10 μm aortic root sections were adhered to slides coated with Poly-L-Lysine (#P4707; Sigma-Aldrich), slides were then frozen until use. RNA in situ hybridization was performed using the RNAScope® Fluorescent Multiplex Reagent Kit (#320850; Advanced Cell Diagnostics). RNA retrieval was performed first by boiling the slides for 6 min in the Target Retrieval Buffer (#322000; Advanced Cell Diagnostics) and then 1 h with Protease III at 40°C. RNAScope® probe hybridization protocol was followed for the Ripk3 probe (#462541-C3; Advanced Cell Diagnostics). Positive control probes for Ubc and Polr2a (#320881; Advanced Cell Diagnostics) were used for each sample. After mounting with Prolong Gold (Thermo Fisher Scientific), images were obtained with a Nikon Eclipse Ti-E microscope. Immunofluorescence in tissue sections Hearts were fixed in 1% PFA overnight, placed in 20% sucrose, then incubated in a 1:1 mixture of 20% sucrose and OCT overnight and frozen in OCT the next day. Sections were cut at 10 µm and frozen until use. Sections were permeabilized in 0.1% Triton X-100 in PBS for 15 min then blocked in 5% BSA for 2 h at room temperature. Sections were incubated in primary antibody in 1% BSA overnight at 4°C, washed three times quickly in ice- cold 1% BSA, then incubated for 1 h at room temperature with secondary antibody in 1% BSA with Hoechst (20 µg/ml). Sections were quickly washed three times with cold 1% BSA and coverslipped with DABCO. Images were obtained using a Nikon Eclipse Ti-E microscope with NIS- Elements software. Primary antibodies used for immunofluorescence were: goat anti-PECAM-1 (1:100; #AF3628; R&D Systems), rat anti-CD68 (1:250; #MCA1957; Bio-Rad), Cy3-conjugated anti-aSMA (1:250; #C6198; Sigma-Aldrich) and chicken anti-E-selectin (1:100; #AF575; R&D Systems). Secondary antibodies used were: Cy3 donkey anti-rat IgG (#712-165-153), FITC donkey anti-goat IgG (#705-095-003) and Alexa 594 donkey anti-chicken IgY (#703-585-155). All secondary antibodies were acquired from Jackson ImmunoResearch and used at a 1:500 dilution. Isolation and differentiation of BMDMs BMDMs were isolated from tibias and femurs of Apoe−/−;Ripk3fl/fl, Apoe−/−; Ripk3fl/fl;LysM-Cre, Apoe−/−;Ripk3fl/fl;VE-Cadherin-Cre or Apoe−/−;Ripk3Δ/Δ mice and differentiated into macrophages using DMEM (#12-604F; Lonza) supplemented with 10% fetal bovine serum (FBS; HyClone), 1× antibiotic- antimycotic (#15240-062; Gibco), 2 mM L-glutamine (#95057-448; HiMedia), 1 mM sodium pyruvate (#25-000-Cl; Mediatech), 1× MEM non-essential amino acids (#11140-050; Gibco) and rhM-CSF (20 ng/ml; #78057; Stemcell Technologies) for 7 days. Differentiation was confirmed by a phagocytosis assay and by immunocytochemistry for the macrophage markers CD11b and CD68. For the phagocytosis assay, BMDMs were treated with FluoSpheres polystyrene microspheres (#F13082; Thermo Fisher Scientific) for 4 h, co-stained with Hoechst (10 µg/ml), washed, fixed and then visualized on a Nikon Eclipse Ti-E microscope. Phagocytic cells were identified as TRITC+ and quantified. Immunocytochemistry Bone marrow was plated on chamber slides (Lab-Tek) and differentiated into BMDMs as described above. On day 7, slides were fixed with 4% PFA for 10 min, permeabilized with 0.5% saponin for 10 min, blocked [1% BSA, 22.52 mg/ml glycine in PBS with 0.1% Tween 20 (PBST)] for 1 h and incubated in primary antibody diluted in 1% BSA in PBST overnight at 4°C. After washing with PBS, slides were incubated in secondary antibody with Hoechst (20 µg/ml) in 1% BSA in PBST for 1 h. After mounting with DABCO, images were obtained with a Nikon Eclipse Ti-E microscope. Primary antibodies used for immunocytochemistry were: rabbit anti-CD11b (1:250; #EPR1344; Abcam) and rat anti-CD68 (1:250; #MCA1957; BioRad). Secondary antibodies used were acquired from Jackson ImmunoResearch and used at a 1:500 dilution: Cy3 donkey anti-rat IgG (#712-165-153) and Cy5 donkey anti-rabbit IgG (#711-175-152). Cell culture and treatment The murine MS1 pancreas-derived endothelial cell line (#CRL-2279; ATCC) was maintained in DMEM supplemented with 5% FBS (HyClone) and 1× antibiotic-antimycotic (#15240-062; Gibco). MS-1 cells were tested quarterly for mycoplasma contamination by PCR. For RIPK3 knockdown, cells were treated with RNAiMAX transfection reagent (Thermo Fisher Scientific) and 50 nM of RIPK3 Silencer Select or non-targeting control siRNA oligonucleotides (#s80756 and #4390844, respectively; Ambion) in serum-free OptiMEM (Invitrogen). After 24 h, medium was replaced with low-serum medium. After another 24 h, cells and medium were harvested for subsequent transcript or protein analyses. On day 7 of BMDM differentiation, BMDMs and medium from either Apoe−/−;Ripk3fl/fl or Apoe−/−;Ripk3fl/fl;LysM-Cre mice were harvested for subsequent transcript or protein analyses. Western blot assays MS1 cells and BMDMs were lysed in RIPA buffer [50 mM Tris-HCl ( pH 7.4), 150 mM NaCl, 1% NP-40, 0.1% SDS, 1 mM EDTA, 0.5% sodium deoxycholate] with Protease Inhibitor Cocktail (#78430; Thermo Fisher Scientific) and Phosphatase Inhibitor Cocktail (#78420; Thermo Fisher Scientific). Aorta tissues were flash frozen and ground with a mortar and pestle then resuspended in RIPA buffer with the protease and phosphatase inhibitors. Protein concentration was determined using the Pierce BCA Protein Assay Kit (#23227, Thermo Fisher Scientific) and a NanoDrop 2000 from Thermo Fisher Scientific. Then, 5-10 μg protein was electrophoresed on a 12% SDS-PAGE gel, then transferred to a PVDF membrane, and then blocked in 5% nonfat dry milk in Tris-buffered saline with 0.1% Tween 20 (TBST) for 1 h. Membranes were cut into sections based on target protein molecular weights to optimize the number of proteins that could be detected. Primary antibodies (diluted in TBST for 11 s m s i n a h c e M & s l e d o M e s a e s i D RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 detection of phosphorylated proteins or 5% milk-TBST for all other proteins) were incubated at 4°C overnight with gentle agitation, and membranes were then washed three times (10 min each) in TBST. HRP- conjugated secondary antibodies (diluted in 5% milk-TBST) were applied at room temperature for 1 h with gentle agitation, and membranes were then washed four times (15 min each) in TBST. Secondary antibodies were detected using ECL Western Blotting Detection Reagent (#34076 or #34096; Thermo Fisher Scientific). Membranes were then stripped using Restore™ PLUS Western Blot Stripping Buffer (#46430; Thermo Fisher Scientific) and re-blotted with primary antibodies raised in a different residual HRP-signal. Densitometry analysis was species to prevent software. Primary antibodies used for performed using ImageJ anti-ICAM1 (1:5000; #AF796; R&D immunoblotting were: goat Systems), goat anti-VCAM1 (1:1000; #AF643; R&D Systems), rabbit anti-CD11b (1:3000; #EPR1344; Abcam), chicken anti-E-selectin (1:1000; rabbit anti-RIPK3 (1:4000; #NBP1-77299; #AF575; R&D Systems), NOVUS Biologicals), goat anti-MCP-1 (1:2000; #AF479; R&D Systems), rabbit anti-MLKL (1:1000; #orb32399; Biorbyt), rabbit anti- phosphorylated-MLKL (1:1000; #ab196436; Abcam), anti- phosphorylated-MLKL (1:1000; #37333; Cell Signaling Technology), rabbit anti-β-actin rabbit anti-RIPK1 (1:1000; #ab202985; Abcam), (1:10,000; #4970; Cell Signaling Technology) and rabbit-anti-GAPDH (1:20,000, #G9545; Sigma-Aldrich). rabbit qPCR assays To analyze transcript levels, total RNA isolated from MS1 endothelial cells or BMDMs was purified and treated with RNase-free DNaseI (Qiagen). cDNA was prepared using the iSCRIPT™ Reverse Transcriptase Kit (Bio- Rad), and qPCR was performed using 2X SYBR green qPCR master mix (Applied Biosystems) and the CFX96 Real-Time System (Bio-Rad) with gene-specific primers. Primers used for qPCR are listed and described in Table S1. The relative fold change in transcription was determined using the comparative CT method and three housekeeping genes (Gapdh, Rn18s and Actb) as internal controls. ELISA Maxisorb Immunolon 4HBX 96-well plates (Thermo Fisher Scientific) were used with mouse MCP-1 and mouse IL-1β ELISA kits (#432705 and respectively; BioLegend) according to the manufacturer’s #432605, instructions to detect MCP-1 or IL-1β levels in plasma samples or medium from cell culture experiments. Colorimetric values were detected using a FLUOstar Omega plate reader. Statistical analysis Data shown are mean±s.d. of n independent experiments or n animals (when necessary, n is reported in the corresponding figure legends). Statistical analyses were performed using GraphPad Prism 8. Data normality was determined by using the D’Agostino-Pearson omnibus test. Outliers were identified using the ROUT method (Q=1%). Statistical analyses, post-hoc tests and P-values are all described in corresponding figure legends. Significance was determined by a P-value of 0.05 or less. There were no statistically significant differences between male and female mice. Acknowledgements We thank current and past Griffin lab members for helpful discussions. Competing interests The authors declare no competing or financial interests. Author contributions Conceptualization: S.C., C.T.G.; Methodology: S.C., V.M., J.X., S.G., C.T.G.; Validation: S.C., S.G., C.T.G.; Formal analysis: S.C., S.G., C.T.G.; Investigation: S.C., V.M., J.X., S.G.; Resources: C.T.G.; Data curation: S.C., C.T.G.; Writing - original draft: S.C.; Writing - review & editing: S.C., C.T.G.; Visualization: S.C.; Supervision: C.T.G.; Project administration: C.T.G.; Funding acquisition: C.T.G. Funding This study was supported in part by National Institutes of Health grants HL134778 and HL144605 awarded to C.T.G., and GM114731 (to C.T.G., as part of a multi-investigator grant awarded to Rodger McEver, Oklahoma Medical Research Foundation), by a grant from the Oklahoma Center for the Advancement of Science and Technology (HR15-078) awarded to C.T.G., by grants from the American Heart Association awarded to C.T.G. (15GRNT25090015) and S.G. (19PRE34380708), and by grants from the Presbyterian Health Foundation awarded to C.T.G. Supplementary information Supplementary information available online at http://dmm.biologists.org/lookup/doi/10.1242/dmm.041962.supplemental References Aiello, R. J., Bourassa, P.-A. K., Lindsey, S., Weng, W., Natoli, E., Rollins, B. J. and Milos, P. M. (1999). Monocyte chemoattractant protein-1 accelerates atherosclerosis in apolipoprotein E-deficient mice. Arterioscler. Thromb. Vasc. Biol. 19, 1518-1525. doi:10.1161/01.ATV.19.6.1518 Alva, J. A., Zovein, A. C., Monvoisin, A., Murphy, T., Salazar, A., Harvey, N. L., Carmeliet, P. and Iruela-Arispe, M. L. (2006). VE-Cadherin-Cre-recombinase transgenic mouse: a tool for lineage analysis and gene deletion in endothelial cells. Dev. Dyn. 235, 759-767. doi:10.1002/dvdy.20643 Amend, S. R., Valkenburg, K. C. and Pienta, K. J. (2016). Murine hind limb long bone dissection and bone marrow isolation. J. Vis. Exp. 110, e53936. doi:10. 3791/53936 Bingham, V., Mcilreavey, L., Greene, C., O’doherty, E., Clarke, R., Craig, S., Salto-Tellez, M., Mcquaid, S., Lewis, C. and James, J. (2017). RNAscope in situ hybridization confirms mRNA integrity in formalin-fixed, paraffin-embedded cancer tissue samples. Oncotarget 8, 93392-93403. doi:10.18632/oncotarget. 21851 Clausen, B. E., Burkhardt, C., Reith, W., Renkawitz, R. and Fö rster, I. (1999). Conditional gene targeting in macrophages and granulocytes using LysMcre mice. Transgenic Res. 8, 265-277. doi:10.1023/A:1008942828960 Colijn, S., Gao, S., Ingram, K. G., Menendez, M., Muthukumar, V., Silasi-Mansat, R., Chmielewska, J. J., Hinsdale, M., Lupu, F. and Griffin, C. T. (2019). The NuRD chromatin-remodeling complex enzyme CHD4 prevents hypoxia-induced endothelial Ripk3 transcription and murine embryonic vascular rupture. Cell Death Differ 27, 618-631. doi:10.1038/s41418-019-0376-8 Coornaert, I., Hofmans, S., Devisscher, L., Augustyns, K., Van Der Veken, P., De Meyer, G. R. Y. and Martinet, W. (2018). Novel drug discovery strategies for atherosclerosis that target necrosis and necroptosis. Expert Opin Drug Discov. 13, 477-488. doi:10.1080/17460441.2018.1457644 Dhuriya, Y. K. and Sharma, D. (2018). Necroptosis: a regulated inflammatory mode of cell death. J. Neuroinflammation 15, 199. doi:10.1186/s12974-018-1235-0 Fontaine, D. A. and Davis, D. B. (2016). Attention to background strain is essential for metabolic research: C57BL/6 and the international knockout mouse consortium. Diabetes 65, 25-33. doi:10.2337/db15-0982 Goren, I., Allmann, N., Yogev, N., Schü rmann, C., Linke, A., Holdener, M., Waisman, A., Pfeilschifter, J. and Frank, S. (2009). A transgenic mouse model of inducible macrophage depletion: effects of diphtheria toxin-driven lysozyme M- specific cell lineage ablation on wound inflammatory, angiogenic, and contractive processes. Am. J. Pathol. 175, 132-147. doi:10.2353/ajpath.2009.081002 Harno, E., Cottrell, E. C. and White, A. (2013). Metabolic pitfalls of CNS Cre-based technology. Cell Metab. 18, 21-28. doi:10.1016/j.cmet.2013.05.019 He, S. and Wang, X. (2018). RIP kinases as modulators of inflammation and immunity. Nat. Immunol. 19, 912-922. doi:10.1038/s41590-018-0188-x Holtwick, R., Gotthardt, M., Skryabin, B., Steinmetz, M., Potthast, R., Zetsche, B., Hammer, R. E., Herz, J. and Kuhn, M. (2002). Smooth muscle-selective deletion of guanylyl cyclase-A prevents the acute but not chronic effects of ANP on blood pressure. Proc. Natl. Acad. Sci. USA 99, 7142-7147. doi:10.1073/pnas. 102650499 Jie, Z., Xia, H., Zhong, S.-L., Feng, Q., Li, S., Liang, S., Zhong, H., Liu, Z., Gao, Y., Zhao, H. et al. (2017). The gut microbiome in atherosclerotic cardiovascular disease. Nat. Commun. 8, 845. doi:10.1038/s41467-017-00900-1 Karunakaran, D., Geoffrion, M., Wei, L., Gan, W., Richards, L., Shangari, P., Dekemp, E. M., Beanlands, R. A., Perisic, L., Maegdefessel, L. et al. (2016). Targeting macrophage necroptosis for therapeutic and diagnostic interventions in atherosclerosis. Sci. Adv. 2, e1600224. doi:10.1126/sciadv.1600224 Kearney, C. J., Cullen, S. P., Tynan, G. A., Henry, C. M., Clancy, D., Lavelle, E. C. and Martin, S. J. (2015). Necroptosis suppresses inflammation via termination of TNF- or LPS-induced cytokine and chemokine production. Cell Death Differ. 22, 1313. doi:10.1038/cdd.2014.222 Leeper, N. J. (2016). The role of necroptosis in atherosclerotic disease. JACC Basic Transl. Sci. 1, 548-550. doi:10.1016/j.jacbts.2016.08.002 Lin, Y. (2014). RIP1-Mediated signaling pathways in cell survival and death control. In Necrotic Cell Death (ed. H.-M. Shen and P. Vandenabeele), pp. 23-43. New York, NY: Springer New York. Lin, J., Li, H., Yang, M., Ren, J., Huang, Z., Han, F., Huang, J., Ma, J., Zhang, D., Zhang, Z. et al. (2013). A role of RIP3-mediated macrophage necrosis in atherosclerosis development. Cell Rep. 3, 200-210. doi:10.1016/j.celrep.2012.12.012 Linkermann, A. and Green, D. R. (2014). Necroptosis. N. Engl. J. Med. 370, 455-465. doi:10.1056/NEJMra1310050 12 s m s i n a h c e M & s l e d o M e s a e s i D RESEARCH ARTICLE Disease Models & Mechanisms (2020) 13, dmm041962. doi:10.1242/dmm.041962 Lusis, A. J. (2000). Atherosclerosis. Nature 407, 233-241. doi:10.1038/35025203 Mani, S., Li, H., Untereiner, A., Wu, L., Yang, G., Austin, R. C., Dickhout, J. G., Lhoták, Š ., Meng, Q. H. and Wang, R. (2013). Decreased endogenous production of hydrogen sulfide accelerates atherosclerosis. Circulation 127, 2523-2534. doi:10.1161/CIRCULATIONAHA.113.002208 McEver, R. P. (2015). Selectins: initiators of leucocyte adhesion and signalling at the vascular wall. Cardiovasc. Res. 107, 331-339. doi:10.1093/cvr/cvv154 Meng, L., Jin, W. and Wang, X. (2015). RIP3-mediated necrotic cell death accelerates systematic inflammation and mortality. Proc. Natl. Acad. Sci. USA 112, 11007-11012. doi:10.1073/pnas.1514730112 Moriwaki, K. and Chan, F. K.-M. (2016). Necroptosis-independent signaling by the RIP kinases in inflammation. Cell. Mol. Life Sci. 73, 2325-2334. doi:10.1007/ s00018-016-2203-4 Moriwaki, K., Balaji, S., Bertin, J., Gough, P. J. and Chan, F. K.-M. (2017). Distinct role of RIPK3 in CD11c+ mononuclear phagocytes in kinase-independent cytokine-induced tissue repair. Cell Rep 18, 2441-2451. doi:10.1016/j.celrep. 2017.02.015 Mussbacher, M., Salzmann, M., Brostjan, C., Hoesel, B., Schoergenhofer, C., Datler, H., Hohensinner, P., Bası́lio, J., Petzelbauer, P., Assinger, A. et al. (2019). Cell type-specific roles of NF-κB linking inflammation and thrombosis. Front. Immunol. 10, 85. doi:10.3389/fimmu.2019.00085 Nakashima, Y., Raines, E. W., Plump, A. S., Breslow, J. L. and Ross, R. (1998). Upregulation of VCAM-1 and ICAM-1 at atherosclerosis-prone sites on the endothelium in the ApoE-deficient mouse. Arterioscler. Thromb. Vasc. Biol. 18, 842-851. doi:10.1161/01.ATV.18.5.842 Newton, K., Sun, X. and Dixit, V. M. (2004). Kinase RIP3 is dispensable for normal NF-κBs, signaling by the B-cell and T-cell receptors, tumor necrosis factor receptor 1, and toll-like receptors 2 and 4. Mol. Cell. Biol. 24, 1464-1469. doi:10. 1128/MCB.24.4.1464-1469.2004 Newton, K., Dugger, D. L., Maltzman, A., Greve, J. M., Hedehus, M., Martin- McNulty, B., Carano, R. A. D., Cao, T. C., van Bruggen, N., Bernstein, L. et al. (2016). RIPK3 deficiency or catalytically inactive RIPK1 provides greater benefit than MLKL deficiency in mouse models of inflammation and tissue injury. Cell Death Differ. 23, 1565-1576. doi:10.1038/cdd.2016.46 Piedrahita, J. A., Zhang, S. H., Hagaman, J. R., Oliver, P. M. and Maeda, N. (1992). Generation of mice carrying a mutant apolipoprotein E gene inactivated by gene targeting in embryonic stem cells. Proc. Natl. Acad. Sci. USA 89, 4471-4475. doi:10.1073/pnas.89.10.4471 circulating lipids but does not protect against atherosclerosis. Atheroscler. Suppl. 32, 100. doi:10.1016/j.atherosclerosissup.2018.04.306 Schittenhelm, L., Hilkens, C. M. and Morrison, V. L. (2017). β2 Integrins as regulators of dendritic cell, monocyte, and macrophage function. Front. Immunol. 8, 1866-1866. doi:10.3389/fimmu.2017.01866 Silke, J., Rickard, J. A. and Gerlic, M. (2015). The diverse role of RIP kinases in necroptosis and inflammation. Nat. Immunol. 16, 689. doi:10.1038/ni.3206 Tabas, I., Garcı́a-Cardeña, G. and Owens, G. K. (2015). Recent insights into the cellular biology of atherosclerosis. J. Cell Biol. 209, 13-22. doi:10.1083/jcb. 201412052 Uhlén, M., Fagerberg, L., Hallströ m, B. M., Lindskog, C., Oksvold, P., Mardinoglu, A., Sivertsson, Å., Kampf, C., Sjö stedt, E., Asplund, A. et al. (2015). Tissue-based map of the human proteome. Science 347, 1260419. doi:10.1126/science.1260419 Vandenabeele, P., Grootjans, S., Callewaert, N. and Takahashi, N. (2012). Necrostatin-1 blocks both RIPK1 and IDO: consequences for the study of cell death in experimental disease models. Cell Death Differ. 20, 185. doi:10.1038/ cdd.2012.151 Van Vré, E. A. V., Ait-Oufella, H., Tedgui, A. and Mallat, Z. (2012). Apoptotic cell death and efferocytosis in atherosclerosis. Arterioscler. Thromb. Vasc. Biol. 32, 887-893. doi:10.1161/ATVBAHA.111.224873 Vince, J. E. and Silke, J. (2016). The intersection of cell death and inflammasome activation. Cell. Mol. Life Sci. 73, 2349-2367. doi:10.1007/s00018-016-2205-2 Weinlich, R., Oberst, A., Beere, H. M. and Green, D. R. (2016). Necroptosis in development, inflammation and disease. Nat. Rev. Mol. Cell Biol. 18, 127. doi:10. 1038/nrm.2016.149 Wiley, M. M., Muthukumar, V., Griffin, T. M. and Griffin, C. T. (2015). SWI/SNF chromatin-remodeling enzymes Brahma-related gene 1 (BRG1) and Brahma (BRM) are dispensable in multiple models of postnatal angiogenesis but are required for vascular integrity in infant mice. J. Am. Heart Assoc. 4, e001972. doi:10.1161/JAHA.115.001972 Yang, Y., Hu, W., Feng, S., Ma, J. and Wu, M. (2005). RIP3 β and RIP3 γ, two novel splice variants of receptor-interacting protein 3 (RIP3), downregulate RIP3- induced apoptosis. Biochem. Biophys. Res. Commun. 332, 181-187. doi:10.1016/ j.bbrc.2005.04.114 Yang, Z., Wang, Y., Zhang, Y., He, X., Zhong, C.-Q., Ni, H., Chen, X., Liang, Y., Wu, J., Zhao, S. et al. (2018). RIP3 targets pyruvate dehydrogenase complex to increase aerobic respiration in TNF-induced necroptosis. Nat. Cell Biol. 20, 186-197. doi:10.1038/s41556-017-0022-y Rasheed, A., Geoffrion, M., Nguyen, M.-A., Lee, R. and Rayner, K. (2018). Whole body knockdown of mixed lineage kinase domain-like protein decreases Zhe-Wei, S., Li-Sha, G. and Yue-Chun, L. (2018). The role of necroptosis in cardiovascular disease. Front. Pharmacol. 9, 721. doi:10.3389/fphar.2018.00721 s m s i n a h c e M & s l e d o M e s a e s i D 13	null
10.1093_pnasnexus_pgad113.pdf	Data availability The processed experimental files for all single-cell data sets used in this study are available on Zenodo at https://doi.org/10.5281/ zenodo.7474099; Table 1 lists the Gene Expression Omnibus (GEO) accession numbers for each data set. The saved model weights for DELAY are available on Zenodo at https://doi.org/10. 5281/zenodo.7474115. All experimental logs from this study are available at https://tensorboard.dev/experiment/RBVBetLM RDiEvO7sBl452A. We have provided an open-source implementa- tion of DELAY in PyTorch with listed requirements and documen- tation at https://github.com/calebclayreagor/DE	Data availability The processed experimental files for all single-cell data sets used in this study are available on Zenodo at https://doi.org/10.5281/ zenodo.7474099 ; Table 1 lists the Gene Expression Omnibus (GEO) accession numbers for each data set. The saved model weights for DELAY are available on Zenodo at https://doi.org/10. 5281/zenodo.7474115 . All experimental logs from this study are available at https://tensorboard.dev/experiment/RBVBetLM RDiEvO7sBl452A . We have provided an open-source implementation of DELAY in PyTorch with listed requirements and documentation at https://github.com/calebclayreagor/DELAY .	PNAS Nexus, 2023, 2, 1–13 https://doi.org/10.1093/pnasnexus/pgad113 Advance access publication 30 March 2023 Research Report Depicting pseudotime-lagged causality across single-cell trajectories for accurate gene-regulatory inference Caleb C. Reagor a,b,c,, Nicolas Velez-Angel a and A. J. Hudspeth a aHoward Hughes Medical Institute and Laboratory of Sensory Neuroscience, The Rockefeller University, New York, NY 10065, USA bTri-Institutional PhD Program in Computational Biology and Medicine, New York, NY 10065, USA cPresent address: 1230 York Avenue, Campus Box 314, New York, NY 10065, USA To whom correspondence should be addressed: Email: [email protected] Edited By: Shibu Yooseph Abstract Identifying the causal interactions in gene-regulatory networks requires an accurate understanding of the time-lagged relationships between transcription factors and their target genes. Here we describe DELAY (short for Depicting Lagged Causality), a convolutional neural network for the inference of gene-regulatory relationships across pseudotime-ordered single-cell trajectories. We show that combining supervised deep learning with joint probability matrices of pseudotime-lagged trajectories allows the network to overcome important limitations of ordinary Granger causality-based methods, for example, the inability to infer cyclic relationships such as feedback loops. Our network outperforms several common methods for inferring gene regulation and, when given partial ground- truth labels, predicts novel regulatory networks from single-cell RNA sequencing (scRNA-seq) and single-cell ATAC sequencing (scATAC-seq) data sets. To validate this approach, we used DELAY to identify important genes and modules in the regulatory network of auditory hair cells, as well as likely DNA-binding partners for two hair cell cofactors (Hist1h1c and Ccnd1) and a novel binding sequence for the hair cell-specific transcription factor Fiz1. We provide an easy-to-use implementation of DELAY under an open- source license at https://github.com/calebclayreagor/DELAY. Keywords: gene-regulatory inference, gene-regulatory networks, deep learning, single-cell omics, auditory hair cells Significance Statement The sequencing of genes expressed in single cells provides detailed information about the developmental programs that define the identities and states of various cell types, but few computational methods can use the dynamic information encoded in these repre- sentations to identify causal mechanisms. By exploiting advances in machine learning, we develop a deep neural network that learns from temporal features of gene regulation to identify direct regulatory interactions between transcription factors and their target genes. Our method provides mechanistic insights into the development and function of cells by generating high-confidence predic- tions for interactions in complex gene-regulatory networks. Introduction Single-cell sequencing technologies can provide detailed data for the investigation of heterogeneous populations of cells collected at specific times—so-called snapshots—during cellular differenti- ation or dynamic responses to stimulation (1). However, owing to inherent delays in molecular processes such as transcription and translation, static measurements from individual cells cannot reveal the causal interactions governing cells’ dynamic re- sponses to developmental and environmental cues (2–4). Because population-level heterogeneity in tissues often reflects the asyn- chronous progression of single cells through time-dependent processes, observed patterns of gene expression can nonetheless indicate the stages of development to which individual cells be- long (5). Many algorithms exploit these cell-to-cell differences to infer dynamic trajectories and reconstruct cells’ approximate temporal progressions along inferred lineages in pseudotime (6, 7). Several methods for gene-regulatory inference rely on pseudo- time in Granger causality tests, which try to determine whether new time series can add predictive power to inferred models of gene regulation (8, 9). However, Granger causality-based methods can be error-prone when genes display nonlinear or cyclic interac- tions or when the sampling rate is uneven or too low (9–12). Because pseudotime these problems, Granger causality-based methods often underperform model-free approaches that exploit pure statistical dependencies in gene expression data (9, 13, 14). trajectories exhibit By contrast, deep learning-based methods make no assump- tions about the temporal relationships or connectivity between genes in complex regulatory networks; instead, these data-driven Competing Interests: The authors declare no competing interest. Received: September 23, 2022. Revised: March 21, 2023. Accepted: March 23, 2023 © The Author(s) 2023. Published by Oxford University Press on behalf of National Academy of Sciences. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited. 2 \| PNAS Nexus, 2023, Vol. 2, No. 4 approaches learn general features of regulatory interactions (15, 16). Here, we describe a deep learning-based method termed Depicting Lagged Causality (DELAY) that learns gene-regulatory interactions from discrete joint probability matrices of paired, pseudotime-lagged gene expression trajectories. Our data suggest that DELAY can address many shortcomings of current Granger causality-based methods and provide a useful, complementary approach to overcome common limitations in the inference of gene-regulatory networks from single-cell data. Results A convolutional neural network predicts pseudotime-lagged gene-regulatory relationships To predict gene-regulatory relationships from single-cell data, we developed a convolutional neural network based on Granger causality (17). The input to DELAY consisted of stacks of two-dimensional joint probability matrices for pairs of tran- scription factors, putative target genes, and highly correlated “neighbor” genes (18). We constructed the input matrices by aligning the gene expression trajectories of a transcription fac- tor A at several lagged positions in pseudotime relative to a target gene B to generate joint probability distributions from two-dimensional histograms of gene coexpression at each lag (Fig. 1A). The value of a given lag indicated the ordinal differ- ence between cells’ rank-ordered positions in pseudotime. Each input matrix consisted of the L2-normalized cell number counts from a two-dimensional gene coexpression histogram with 32 fixed-width bins in each dimension, spanning each gene’s minimum and maximum expression values. Although the marginal probability distributions for both A and B re- mained essentially unchanged at each lag except for cells lost at the leading and lagging edges of the shifted trajectories, realigning the gene expression in pseudotime altered key fea- tures of the resulting joint probability matrices. In other words, causally related genes share important pseudotime- lagged patterns of gene coexpression with nearby cells in single-cell trajectories. Using ground-truth labels from cell type–specific chromatin immunoprecipitation sequencing (ChIP-seq) data (14), we con- ducted supervised learning to train our neural network to pre- dict whether A directly regulates B. This procedure resembled a regression in a Granger causality test (17), in which values of a time series Y at timepoints yt are regressed against values from another time series X at timepoints xt, xt−1, xt−2, . . . , xt−T up to some maximum lag T to determine whether any time- lagged values of X add explanatory power to Y’s autoregres- sive model. DELAY likewise learned higher weights for gene coexpression matrices at specific pseudotime lags that indicated the true regulatory relationship between genes. After comparing several neural network architectures, we selected a six-layered convolutional network trained on pseudotime-aligned (T = 0) and five pseudotime-lagged (T ∈ {1, 2, 3, 4, 5}) gene coexpression matrices to predict direct gene-regulatory relationships (Fig. 1B). We also trained the network on lagged coexpression matrices of two highly corre- lated neighbor genes per gene, that is, the two transcription factors with the highest cross-correlation with A and B along the single-cell trajectory. Because the neighbor genes can pre- sent stronger alternatives to the primary hypothesis that A directly regulates B, including these matrices for A and B ver- sus their highly correlated transcription factors reduced false- positive predictions. DELAY outperforms several common methods of gene-regulatory inference We trained DELAY on gene expression data sets from human em- bryonic stem cells (hESCs) (19), mouse embryonic stem cells (mESCs) (20), and three lineages of mouse hematopoietic stem cells (mHSCs) (21). We generated separate training data sets for each of the hematopoietic lineages, and all data sets contained at least 400 cells per lineage (Table 1). Each trajectory was oriented according to known experimental timepoints or precursor cell types and lineages, and pseudotime values were inferred separ- ately with Slingshot (6) for each lineage (Fig. S1). We chose these data sets because the cell types and trajectories are well charac- terized and offer cell type–specific ChIP-seq data to generate ground-truth networks (14). Although the three hematopoietic data sets contained similar numbers of examples of true regula- tion and no interaction, both of the embryonic data sets were class-imbalanced and contained fewer examples of true regula- tion. To maximize the network’s generalizability, we trained DELAY simultaneously on all five gene expression data sets but validated it on each data set individually. We first generated ran- domly segregated 70–30% splits of all possible gene pair examples for each data set and then merged the 70% splits to create a combined training data set. In a separate analysis, we also cross- validated DELAY using inductive splits wherein specific transcrip- tion factors appeared in only the training or validation splits. After training DELAY on the combined 70% splits, the network outperformed eight of the most popular approaches for inferring gene-regulatory relationships, including six unsupervised meth- ods (8, 22–26 ) (Fig. 1C) and two deep convolutional neural net- works (18, 27) (Fig. 1D). We measured the performance of the six unsupervised methods across the combined training and valid- ation splits for each data set and then compared the results with the performance of the supervised methods on the held-out examples alone. With one exception, the deep learning-based methods outperformed all others according to the areas under both the precision–recall (PR) and receiver operating characteris- tic (ROC) curves (Fig. S2). Moreover, DELAY outperformed all the other methods according to both metrics, even though one of the deep learning-based methods, DeepDRIM, was trained on 5-fold as many neighbor gene matrices. Upon separate cross- validation, DELAY performed slightly worse on inductive splits but with a single exception outperformed all other methods on average (Fig. 1E). Across three of the training data sets, DELAY also outperformed a modified version of DeepDRIM that was trained on the same pseudotime-lagged input matrices as DELAY (Fig. S2). Together, these results suggest that DELAY out- performs other methods because it can better learn important gene-regulatory features from pseudotime-lagged gene coexpres- sion matrices. Transfer learning allows DELAY to predict novel gene-regulatory networks from new single-cell data sets To test whether DELAY generalizes to new data sets, we examined the human hepatocyte (hHep) gene-regulatory network using an additional data set with over 400 single cells and known ground- truth interactions from ChIP-seq data (14, 28). We inferred the network by the previous methods and found that tree- and mu- tual information-based methods performed slightly better than deep learning-based methods, which were not initially trained on the new data and consequently performed comparably to ran- dom predictors (Fig. 1F). To determine whether this lack of Reagor et al. \| 3 Fig. 1. Pseudotime-lagged causality allows accurate inference of gene-regulatory networks. A) Shifting the gene expression trajectory of transcription factor A forward in pseudotime with respect to that of target gene B generates a series of unique joint probability distributions. The resulting joint probability matrices contain gene coexpression signatures—in the upper-left, upper-right, and lower-right regions—that indicate A directly regulates B. B) The inverted architecture of DELAY is wide at the beginning of the network and progressively narrows to a one-dimensional vector followed by a linear classifier and sigmoid activation function to generate gene regulation probabilities. The network uses leaky ReLU activations and padded 3 × 3 convolutions throughout. C and D) DELAY outperforms eight of the most popular methods for inferring gene-regulatory interactions across several benchmark scRNA-seq data sets, including six unsupervised methods (C) and two supervised methods (D). E) All three supervised methods perform slightly worse upon cross-validation, but DELAY still outperforms all other methods on average, with the exception of PPCOR for a single metric. F and G) DELAY does not immediately generalize to a new testing data set (F) but outperforms all other methods if fine-tuned on a small percentage of the new data (G). Values for the area under the precision–recall curve (AUPRC) in C), D), F), and G) are normalized by the proportion of positive examples per data set (horizontal lines, E) and averaged across five model replicates for supervised methods. Boxes in E) show the first quartiles, means, and third quartiles for the per- transcription factor AUPRC values across the k = 5 validation folds. The statistical significance between DELAY and the next-best neural network was assessed using a one-sided Wilcoxon signed-rank test (P ≤ 0.05; P ≤ 0.01; *P ≤ 0.001). ABCEFGD4 \| PNAS Nexus, 2023, Vol. 2, No. 4 Table.1. Summary of the data sets used to train, test, and evaluate the neural network. Data sets Networks Cell type TC/S GEO Set hESC mESC mHSC-E mHSC-GM mHSC-L hHep mOHC (RNA) mOHC (ATAC) TC TC S S S TC TC S GSE75748 GSE98664 GSE81682 GSE81682 GSE81682 GSE81252 GSE137299 GSE157398 Train/Val Train/Val Train/Val Train/Val Train/Val Test Eval Eval Cells 758 421 1,071 889 847 425 1,563 391 TFs Targets Examples Edges Density 37 (4) 125 (17) 33 23 16 33 (3) 56 48 541 642 533 523 516 536 556 382 20,017 80,250 17,589 12,029 8,256 17,688 1,112 764* 3,166 20,580 9,592 6,587 4,026 6,171 166* 157* 15.82% 25.64% 54.53% 54.76% 48.76% 34.89% 14.93%* 20.55%* Columns cell type to cells describe the single-cell data sets used to train, test, and evaluate DELAY, and columns TFs to density describe the corresponding gene-regulatory networks. For three of the networks, differentially expressed transcription factors without target genes (in parentheses) were included only as target genes. Network density is equal to the number of edges from the ground-truth ChIP-seq data divided by the total number of examples. E, erythroid lineage; GM, granulocyte–monocyte lineage; L, lymphoid lineage; mOHC, mouse outer hair cell; S, snapshot; TC, time course; TF, transcription factor. Partial ground-truth network. generalizability arises from batch effects in the single-cell data, we used transfer learning to fine-tune the three deep learning- based methods on a small number of randomly segregated exam- ples from the new data set while validating on the rest of the examples (Fig. 1G). After training each network on only 1% of the new examples and validating it on the remaining 99%, the net- works’ performance matched or exceeded that of the unsuper- vised methods; training on up to 20% of the new examples yielded further performance increases. DELAY again outper- formed every other deep learning-based method. During fine- tuning, we also observed a monotonic increase in the value of the mean validation metric across PR and ROC curves, suggesting that the networks can learn useful features of gene regulation after longer training periods. Moreover, these results demonstrate that DELAY can accurately predict gene-regulatory relationships from new data sets with partially known ground-truth labels. single cells (Fig. 2C) or additional gene-dropout noise (Fig. 2D), we also discovered that the network was more sensitive to down- sampling of single cells than to gene expression losses in low- expressing cells alone. These results indicate that DELAY relies more heavily on highly expressing cells and therefore assigns lar- ger input weights for the first lagged input because on average that input contains stronger features of gene activation than other lags (Fig. 2E and F). To investigate this hypothesis, we used aug- mented input matrices with masked regions to show that DELAY relies heavily on the combined upper-left, upper-right, and lower-right regions (AON ∪ BON) across all lagged matrices (Fig. 2G). By performing a post hoc analysis of the statistical dis- persion across correctly inferred gene pair examples, we addition- ally found that DELAY generally performs better on transcription factors with stable gene expression along single-cell trajectories (Fig. S3). DELAY performs well with new input configurations, augmented matrices, and modified data sets Using different numbers of pseudotime-lagged gene coexpression matrices or neighbor gene matrices, as well as examples from modified data sets with fewer cells or additional gene dropouts, we next examined the performance of DELAY across various input configurations. We employed the original data sets to train new models on gene coexpression matrices of up to 10 pseudotime lags (Fig. 2A) or up to 10 neighbor genes, as well as on gene coex- pression matrices with varying dimensions and resolutions (Fig. S3). Training DELAY on at least one pseudotime-lagged ma- trix or with at least one neighbor gene greatly increased the net- work’s performance across all data sets. Although training DELAY on up to 10 pseudotime-lagged matrices resulted in com- parable or slightly better performance across all data sets, train- ing the network on more than two neighbor gene matrices per gene decreased performance in some instances. Adding channel masks for specific lagged matrices suggested that DELAY relies on redundancies across all available lagged inputs (Fig. S3). We also characterized DELAY’s performance on pseudotime- shuffled trajectories and observed a sharp decrease in perform- ance after reordering nearby cells in each trajectory (Fig. 2B). This result suggests that the network relies heavily on the specific ordering of adjacent cells in each trajectory. Upon examining DELAY’s performance on modified data sets containing fewer internal DELAY recognizes causal relationships in hierarchical and cyclic gene-regulatory interactions To further investigate how the selection of neighbor genes can alter DELAY’s representations and subsequent gene-regulatory inferences, we modified gene pair examples in the original 30% validation splits to exclude all neighbor genes C with known interactions across several classes of three-gene mo- tifs. First, we used DELAY to infer gene regulation for potential hierarchical and cyclic interactions across validation set exam- ples (Fig. 3A), including mutual interactions (MIs), feedback loops (FBLs), and three classes of feedforward loops (FFLs). Unlike ordin- ary Granger causality, DELAY performed well across cyclic inter- actions including 2-cycles (MIs) and 3-cycles (FBLs), which are ubiquitous among gene-regulatory networks (29). We next ex- cluded from the same validation set examples all input matrices for any neighbor gene of either A or B known from the ChIP-seq data to be involved in potential FBLs or FFLs, replacing the inputs for all omitted neighbors C with the matrices of other highly cor- related neighbor genes. Because we saw a significant decrease in performance upon excluding these neighbors C in most cases of shared and sequential gene regulation, but not for downstream targets, DELAY apparently recognizes the differences between in- put matrices of causally related genes and those of merely corre- lated genes. Moreover, these results suggest that although our network uses principles of Granger causality, it can achieve true Reagor et al. \| 5 DELAY performs best on zinc-finger proteins, bHLH factors, and other chromatin remodelers To explore how DELAY performs on classes of transcription fac- tors containing different types of DNA-binding domains, we ana- lyzed the enrichment of gene ontology (GO) terms across correctly inferred transcription factors (Fig. 3B). We found that DELAY per- forms best on zinc-finger proteins, bHLH factors, and other chro- matin remodelers. Although DELAY performs well on C2H2-type zinc-finger proteins in terms of the total number of correct predic- tions, we found that it performs better on other chromatin remod- elers and plant homeodomain (PHD) zinc-finger proteins by the overall term enrichment (Table S1). Interestingly, these results in- dicate that training the network on cell type–specific ChIP-seq data allows the network to identify regulatory relationships in- volving some non sequence-specific transcription factors and co- factors that nevertheless associate with specific targets at preferred chromatin conformations. Predicting the gene-regulatory network of auditory hair cells through multi-omic transfer learning We next sought to generate predictions for a cell type with com- plex but incompletely characterized gene-regulatory dynamics. We devised a pipeline for multi-omic transfer learning to infer the gene-regulatory network for developing mouse outer hair cells, the mechanical amplifiers of the inner ear. We used both gene expression (30) and chromatin accessibility (31) data sets to fine-tune our network. By first calculating the cell-by-gene acces- sibility scores across annotated genes, we were able to generate lagged input matrices for the single-cell ATAC sequencing (scATAC-seq) data set (Fig. 4A). Because the underlying gene ex- pression and gene accessibility distributions are both zero-inflated (8), the resulting coaccessibility matrices are qualita- tively similar to gene coexpression matrices. To determine whether pseudotime-lagged gene coaccessibility matrices contain features that indicate direct gene-regulatory re- lationships, we fine-tuned DELAY on the scATAC-seq matrices us- ing ground-truth examples collected from two cell type–specific data sets of Sox2 and Atoh1 target genes (32, 33). Seventy percent of the randomly segregated ground-truth examples were used for fine-tuning, and the remaining 30% were held out for validation (Fig. 4B). DELAY performed slightly worse by the metric of area under the PR curve than by area under the ROC curve, indicating that the network is better at discriminating false positives than selecting true positives. Although DELAY did not outperform all other deep learning-based methods, the network’s performance was comparable with previous training and validation on small fractions of the hHep single-cell RNA sequencing (scRNA-seq) data set, which suggests that gene coaccessibility matrices are also useful for inferring direct gene-regulatory interactions. We separately fine-tuned the network on all available Sox2 and Atoh1 ground-truth examples from an additional gene expression data set for mouse outer hair cells (30). With previously deter- mined hyperparameters for scRNA-seq data, training DELAY on two graphics processing units (GPUs) required 230 ± 1 min (mean ± SD) per model. We then compared the target gene rank correlations between the inferred transcription factor-only gene-regulatory networks to determine whether networks in- ferred from scRNA-seq data and scATAC-seq data were similar (Fig. 4C). Although we observed stronger correlations between predictions from data set-specific models than between average predictions across data sets, the target gene rank correlations Fig. 2. The neural network relies on cell order in pseudotime and gene expression strength. A) Training new models of DELAY on increasing numbers of pseudotime-lagged matrices gives the largest performance increase when using up to a single lag. B) Reordering single cells in pseudotime sharply decreases performance, suggesting that DELAY relies on the specific ordering of adjacent cells in each trajectory. C and D) The network is sensitive to random down-sampling of cells across data sets (C) but relatively more robust to induced, additional gene dropouts in weakly expressing cells (D), suggesting that DELAY relies heavily on highly expressing cells. E and F) The network learns larger input weights for lagged matrices of the first pseudotime lag (E), which also contain more cells in the combined “ON” region (AON ∪ BON; dotted outline, F) on average across training data sets (dotted line, E). The combined “ON” region is comprised of the upper-left “ON–OFF” quadrant (AON ∩ BOFF), upper-right “ON–ON” quadrant (AON ∩ BON), and lower-right “OFF–ON” quadrant (AOFF ∩ BON). G) Masking different regions of the input matrices shows that the network relies heavily on the combined “ON” region. The lines and shaded regions in A–E) show the average and full range of values across five model replicates, and the markers in G) show the average values across model replicates. The statistical significance in G) was assessed with a Kruskal–Wallis test (*P ≤ 0.001). causal inference for genes involved in several classes of hierarch- ical and cyclic regulatory interactions by avoiding several limiting assumptions of ordinary Granger causality. ABCDEFG6 \| PNAS Nexus, 2023, Vol. 2, No. 4 Fig. 3. DELAY uncovers causal, cyclic, and context-specific gene-regulatory relationships. A) The network’s performance when inferring putative transcription factor-target gene pairs across two- and three-gene motifs in the validation set suggests that DELAY can distinguish between different types of hierarchical and cyclic gene regulation. Upon exclusion of one or more neighbor genes C from the input features of the validation examples, the network’s performance declines significantly if C is a shared or sequential regulator of A and B, but not if C is a shared target. B) GO-term enrichment indicates that DELAY performs best on zinc-finger proteins (PHD-type, GATA-type, C2H2-type, NHR-type, and C5HC2-type), bHLH factors (Myc-type), and other chromatin remodelers (SNF2-related, chromodomain, bromodomain, HDA domain, and helicase). Markers and error bars in A) show the average values and full range of performance across five model replicates. The statistical significance in A) was assessed with a one-sided Wilcoxon signed-rank test (P ≤ 0.05) and in B) with a Fisher’s exact test (Padjusted ≤ 0.05). between the two data sets were highly variable, and the predic- tions of some transcription factors agreed better than others. Reasoning that the inferences with better agreement across data sets constituted the best predictions and highest confidence interactions, we derived the consensus hair cell gene-regulatory network from the average predictions across both data sets. The resulting network consisted of 347 predicted transcription factor–target interactions with gene regulation probabilities >0.5. DELAY identifies important genes, interactions, and modules in the hair cell gene-regulatory network We used hierarchical clustering to group transcription factors and target genes in the transcription factor-only gene-regulatory network for hair cells by similarities in their predicted targets and regulators, respectively (Fig. 4D). This procedure revealed two distinct developmental modules corresponding to prosen- sory genes such as Sox2, Id2, Hes1, and Prox1 and hair cell- specific genes such as Atoh1, Pou4f3, Gfi1, Lhx3, and Barhl1. We sought to validate the predicted interactions by comparing the locations of known transcription factor-binding sites (34, 35) to the locations of open-chromatin peaks (31) within 50 kb of target genes’ transcription start sites. Twenty-two of 28 predicted in- teractions were confirmed by the accessibility of transcription factor-binding sites. Of the six remaining interactions, three were instances of predicted target-gene regulation by transcrip- tional cofactors lacking true DNA-binding domains. We add- itionally identified 13 reported interactions that were not detected by DELAY (31, 36–54). The most notable feature of the inferred gene-regulatory network for hair cells is that Sox2 and Id2—in addition to their proteins’ well-known roles in regulating target genes and main- taining a prosensory cell fate—are themselves the targets of a wide variety of transcription factors including Sox proteins, homeobox factors, zinc-finger proteins, and Notch effectors. Eleven of the 22 interactions confirmed by our binding-site acces- sibility analysis represented direct regulation of Sox2 and Id2, in- cluding mutually activating interactions between Sox2 and Sox9, Sox11, Prox1, and Hes1, and mutual inhibition with Hist1h1c. One other study has suggested that Hes1 directly regulates Sox2 (55). Another notable feature of the inferred hair cell network is that the LIM-homeobox transcription factor Lhx3 regulates its own ex- pression as well as that of two other LIM-only transcription factors (Lmo4 and Lmo1). This result implies a role for Lhx3 in maintain- ing the later expression of these important early Sox2 inhibitors (31, 47, 49). Other key features of the network include Sox11 regu- lation by Hes1, Nfkbia regulation by Gfi1, and regulation of the spli- cing factor gene Srsf2 by the products of three different prosensory genes. Srsf2 was recently predicted to play a role in the splicing of an important deafness gene in humans (56). Discriminative motif analysis of target-gene enhancer sequences enables de novo discovery of DNA-binding motifs DELAY permits a complementary approach typical gene-regulatory inference workflows such as SCENIC (57) that use cis-regulatory sequences to identify and discard false-positive interactions resulting from indirect gene regulation (11, 30, 31). to ABReagor et al. \| 7 Fig. 4. DELAY accurately predicts gene-regulatory interactions in the auditory hair cell network through multi-omic transfer learning. A) An example of an empirical joint probability matrix from scATAC-seq trajectories. B) After fine-tuning the network on lagged scATAC-seq input matrices, DELAY performs comparably with other neural networks and to previous testing on hHep scRNA-Sseq data. C) Training DELAY separately on scRNA-seq and scATAC-seq data sets of hair cell development reveals a stronger correlation between predictions from model replicates than between data sets across the transcription factor-only network. D) Average gene regulation probabilities P across the hair cell gene-regulatory network accurately predict interactions between transcription factors (rows) and targets (columns), when comparing known binding sites to open-chromatin peaks in the target genes’ enhancer sequences. Hierarchical clustering with WPGMA reveals distinct gene modules for prosensory genes (Sox2, Id2, Hes1, and Prox1) and hair cell genes (Atoh1, Pou4f3, Gfi1, Lhx3, and Barhl1). Up- or down-regulation was deduced from the correlation in the gene expression data. Markers and error bars in B) show the average values and full range of performance across five model replicates for each neural network. Markers in C) show the median target gene rank correlations across comparisons. The statistical significance in C) was assessed with a Kruskal–Wallis test (*P ≤ 0.001). Because these methods rely on databases of known DNA-binding motifs, they necessarily overlook predictions for cofactors and pu- tative transcription factors with unknown binding sequences. Instead of using cis-regulatory sequences to refine our initial pre- dictions, we introduced enhancer sequence information post hoc to perform discriminative motif analysis within the enhancers of predicted targets in the hair cell network (Figs. 5A and S4). We identified enriched motifs that closely resembled known DNA-binding motifs for nine of 11 transcription factors with at least one predicted target gene in the transcription factor-only hair cell network (Table S2). We additionally sought to predict the mechanisms by which several identified hair cell cofactors accomplish sequence-specific gene regulation in the absence of true DNA-binding domains. Specifically, we compared sequences enriched in the enhancers of cofactors’ predicted targets to several databases of known DNA-binding motifs to identify the cofactors’ DNA-binding part- ners (so-called guilt by association). Through this analysis, we were able to determine that the most significantly enriched se- quence in the enhancers of the histone Hist1h1c’s predicted target genes closely matched known binding motifs for several SoxB2/A transcription factors (Fig. 5B). In addition, a sequence enriched in the enhancers of cyclin Ccnd1’s targets accorded with known motifs for the Pbx family of homeobox transcription factors (Fig. S5). These cofactors might form complexes with the identi- fied transcription factors to regulate their target genes in hair cells. As a final demonstration of DELAY’s high-confidence predic- tions, we used the fine-tuned model for the scRNA-seq data to generate regulatory predictions for transcription factors ex- pressed in at least 20% of developing hair cells (Table S3). After considering these additional transcription factors, we identified the putative C2H2 zinc-finger protein Fiz1 as a likely regulator of Sox2 and Hes6. Discriminative motif analysis of these genes’ en- hancer sequences (Fig. 5C) uncovered a likely Fiz1 consensus DNA-binding sequence (5′-CGCTGC-3′) similar to that of other Sox2 regulators from the Zic family of C2H2 zinc-finger proteins (58). Discussion Building upon several deep learning-based methods (18, 27), we have demonstrated that combining fully supervised deep learning with joint probability matrices of pseudotime-lagged single-cell trajectories can overcome certain limitations of current Granger causality-based methods of gene-regulatory inference (9, 12), ADBC8 \| PNAS Nexus, 2023, Vol. 2, No. 4 Fig. 5. Accurate target-gene predictions enable de novo discovery of DNA-binding motifs. A) Three examples of motifs enriched in the enhancer sequences of predicted target genes (bottom row) closely resemble known DNA-binding motifs (top row) for transcription factors with at least one predicted target gene in the transcription factor-only hair cell gene-regulatory network. B and C) Sequences for putative transcription factors or cofactors may represent novel DNA-binding motifs or indicate sequence-specific interactions through multi-protein complexes. Motifs for two hair cell genes closely match known DNA-binding sequences, suggesting that the linker histone Hist1h1c (B, top row) forms complexes with SoxB2/A factors (B, bottom row) and that the putative C2H2 zinc-finger protein Fiz1 (C, top row) recognizes motifs similar to Zic family C2H2 zinc-finger proteins (C, bottom row). The statistical significance of each motif alignment was estimated using the cumulative density function of all possible comparisons of known motifs across enriched sequences for A) or inferred motifs across all database motifs for B) and C) (P ≤ 0.01; *** P ≤ 0.001). such as their inability to infer cyclic regulatory motifs. Although our convolutional neural network DELAY remains sensitive to the ordering of single cells in pseudotime, the network can never- theless accurately infer direct gene-regulatory interactions from both time course and snapshot data sets, unlike many supervised methods that rely strongly on the number of available time course samples (15, 16). We suspect that DELAY is sensitive to the specific ordering of adjacent cells in trajectories because pseudotime in- ference methods such as Slingshot (6) infer lineages from min- imum spanning trees that directly depend on cell-to-cell similarities in gene expression values (7). We have presented a multi-omic paradigm for fine-tuning DELAY on both gene expression and chromatin accessibility data sets for the development of auditory hair cells. The network’s ABCaccurate predictions allow de novo inferences of known and unknown DNA-binding motifs, establishing DELAY as a comple- mentary approach to common methods of gene-regulatory infer- ence. Our method’s high-confidence predictions across the hair cell gene-regulatory network also support it as an attractive option for experimentalists with limited resources to predict true gene-regulatory relationships from complex, large-scale gene-regulatory networks while avoiding spurious, indirect inter- actions often introduced through unsupervised methods (14). We have additionally identified with high confidence interac- tions between cofactors and target genes in the regulatory net- work of hair cells that methods such as SCENIC (57) would have overlooked. Using “guilt by association,” we predict that the cyclin Ccnd1—a known transcriptional cofactor found at the enhancers of several Id genes during retinal development (59)—forms com- plexes with the Pbx family of homeobox transcription factors, of which at least one is a known target of Prox1 in the inner ear (60). Moreover, we predict that the retina-specific linker histone Hist1h1c (61) forms complexes with the SoxB2 family of transcrip- tion factors, which have known antagonistic effects on Sox2 ex- pression in hair cells (62). Finally, we predict that the putative hair cell-specific C2H2 zinc-finger protein Fiz1—which is also ex- pressed during retinal development (63)—is a regulator of Sox2 in hair cells and identified its likely DNA-binding sequence in the enhancers of predicted target genes. Believing that DELAY will be a valuable resource to the commu- nity, we have provided an easy-to-use and open-source imple- mentation of the algorithm as well as pre-trained model weights for subsequent fine-tuning on new single-cell data sets and partial ground-truth labels. Unlike other deep learning-based methods, our implementation of DELAY maximizes usability and prioritizes model flexibility. Although we chose to fine-tune DELAY on cell type–specific ChIP-seq data, future studies may choose to fine- tune DELAY on curated interactions from databases or from gain- or loss-of-function experiments, especially in the absence of ChIP-seq data (14). These additional interactions can also sup- plement smaller ground-truth data sets, such as those of Sox2 and Atoh1 target genes, to mitigate false-negative predictions. We be- lieve that the modest computational cost of training new models of DELAY will prove a worthwhile investment for future investiga- tions into complex gene-regulatory networks. Materials and methods Preparing single-cell RNA-seq data sets to train and test the convolutional neural network To train our convolutional neural network, we used scRNA-seq data sets from two time course studies of endodermal specifica- tion of hESCs (19) and mESCs (20) and an in vivo snapshot study of erythroid (E), granulocyte–monocyte (GM), and lymphoid (L) specification in mHSCs (21). We additionally employed a data set from a time course study of the differentiation of hHeps from induced pluripotent stem cells to test our network (28). For each of these studies, we collected the normalized gene expres- sion values, pseudotime values, and ground-truth networks from BEELINE’s supplementary data files (14). In BEELINE, Pratapa et al. collected the normalized gene expression values from the original studies or obtained the values themselves by log- transforming the transcripts per kilobase million. Moreover, those authors used Slingshot (6) to infer the pseudotime values separ- ately for each data set, orienting the inferred trajectories by known experimental timepoints or precursor cell types and Reagor et al. \| 9 lineages. Finally, Pratapa et al. selected genes with differential ex- pression across pseudotime by applying generalized additive models (GAMs) to compute gene variances and their associated P values. To train DELAY, we utilized their gene variances and cor- rected P values to select differentially expressed transcription fac- tors (Padjusted<0.01) and 500 additional genes with the highest variance for each data set. We also collected the ground-truth la- bels from BEELINE’s supplementary data files, which contained tables of transcription factor–target gene interactions curated from cell type–specific ChIP-seq experiments in the ENCODE (64), ChIP-Atlas (65), and ESCAPE (66) databases. Table 1 provides details for each data set including descriptions of the ground- truth networks. We also characterized our network on modified data sets with shuffled pseudotime, fewer cells, and additional gene-dropout noise. We shuffled pseudotime values across single-cell trajector- ies using np.random.normal in NumPy (v1.20.2) to select the indi- ces of single cells either leading or lagging successive cells in each trajectory at some length scale σ before swapping the cells’ posi- tions in pseudotime. We also used np.random.choice to select the indices of single cells to retain when down-sampling the num- ber of cells in a data set. Lastly, we induced additional gene- dropout noise in data sets by setting the gene expression values to 0 with a probability of p for the bottom p percent of cells ex- pressing each gene in a given data set. Generating examples of transcription factor– target gene pairs from single-cell data sets For each data set, we generated mini-batches of 512 transcription factor–target gene pair examples by first enumerating all possible combinations of differentially expressed transcription factors and potential target genes—including both transcription factors and highly varying genes. We then generated aligned and pseudotime- lagged gene coexpression matrices for up to five lags with the fol- lowing configurations as separate input channels: transcription factor–target, transcription factor–transcription factor, target– target, transcription factor–neighbor (for two neighbors), and tar- get–neighbor (for two neighbors). We concatenated the input channels for each example to form three-dimensional stacks of input matrices with dimensions 42 × 32 × 32 (channels × height × width). For each gene, we used np.correlate in NumPy to select the transcription factors with the highest absolute cross-correlation at a maximum offset of five pseudotime lags to use as neighbor genes. We trained five model replicates on unique, randomly seg- regated 70–30% splits of all possible gene pair examples generated with the random split function in PyTorch (v1.8.1). Each random split achieved a strict segregation of unique gene pair examples into either the training or validation sets, though gene pairs for the same transcription factor but different target genes appeared in both sets. For separate model cross-validation, we generated five inductive splits across each data set wherein specific tran- scription factors appeared in only the training (80%) or validation (20%) splits. For both analyses, training splits from individual data sets were subsequently merged to create combined training data sets. We later used the original randomly segregated validation splits to characterize our network on augmented coexpression matrices and matrices generated from modified single-cell data sets. To augment gene coexpression matrices, we masked with ze- ros either specific region across all input channels or entire input channels corresponding to specific pseudotime lags. 10 \| PNAS Nexus, 2023, Vol. 2, No. 4 Constructing a convolutional neural network to classify lagged gene coexpression matrices We designed a convolutional neural network based on an inverted VGGnet (67) that uses five convolutional layers to first expand the input to 1,024 channels and then successively halve the number of channels to 64 before classifying examples as either true regu- lation or no interaction with a fully connected linear layer. Each convolutional layer sums over the two-dimensional cross- correlations (⋆) of 3 × 3 kernels and input channels i to find the features for a given output channel j, as shown in Equation 1: output (Nk, Coutj ) = bias (Coutj ) + 􏽐Cin i=1 weights 3×3 (Cini , Coutj ) ⋆ input (Nk, Cini ), (1) in which the weights and bias are trainable parameters, N is the mini-batch size, C is the number of channels, Cout = Cin/2 for layers 2 through 5, and all convolutions are zero-padded at the edges of matrices. To preserve the gradient flow during training, we used leaky rectified linear units (ReLUs) with a negative slope of 0.2 (Equation 2) as our nonlinear activation function after each convolutional layer: Leaky ReLU (x) = max (0, x) + 0.2∗min (0, x). (2) As shown in Equation 3, we additionally used 2 × 2, unpadded max-pooling layers between convolutional layers to identify im- portant features and down-sample activation maps: 􏼓 􏼔 hin 2 , win 2 = stride 2×2 max m,n ∈ {0,1} 􏼕 input (Nk, Cinj , xm,n) , 􏼒 output Nk, Coutj , (3) in which x represents the 2 × 2 input windows and h, w are the height and width of the input channel j. After the final convolu- tional layer, we used global-average pooling (Equation 4) to reduce the remaining 64 feature maps j to a single, 64-dimensional vector x: output (Nk, Coutj ) = avg hin×win input (Nk, Cinj ). (4) We lastly used a fully connected linear layer (Equation 5) with a sigmoid activation function (Equation 6) to generate gene regula- tion probabilities: output (Nk) = bias FC + 􏽐64 i=1 weights FC (xi)∗input (Nk, xi), (5) Sigmoid (x) = 1 1 + exp(−x) , (6) where a probability of P > 0.5 indicates a true gene-regulatory interaction for the given gene pair example. Training and fine-tuning the network on pseudotime-lagged gene coexpression matrices We used PyTorch’s implementation of stochastic gradient descent (SGD) to optimize our network with respect to the class-weighted binary cross-entropy loss, summed across each mini-batch and scaled by the overall batch size 512, as shown in Equation 7: L(y, ˆy) = − wN 512 􏽘N n=1 yn∗log ˆyn + (1 − yn)∗log (1 − ˆyn), (7) in which y and ˆy are the target and predicted values (respectively), wN is the fraction of true examples in the mini-batch, and N is the size of the current mini-batch (which might be <512). Prior to optimization, we used He initialization (68) with uniform priors to set the weights for all convolutional and linear layers. We used a learning rate (LR) of 0.5 to train each model for up to 100 epochs, validating performance after each epoch and stopping training after 10 or more epochs without an improvement in the mean validation metric across PR and ROC curves. For fine-tuning, we trained each model for up to a maximum of 5,000 additional epochs due to the observed monotonic increase in the value of the mean validation metric. We occasionally reduced the LR to 0.25 or 0.1 if the training became unstable, and we tried several mini-batch sizes ≥24 for fine-tuning and separate model cross- validation. All training and testing was performed on two Nvidia RTX 8000 GPUs. Comparing DELAY to other top-performing gene-regulatory inference methods We compared the performance of our neural network to two other convolutional neural networks, as well as the top six best- performing methods identified in a previous benchmarking study. Using BEELINE, we inferred the gene-regulatory networks for all training and testing data sets with the tree-based methods GENIE3 (22) and GRNBoost2 (23), the mutual information-based method PIDC (24), and the partial correlation and regression- based methods PPCOR (25), SCODE (26), and SINCERITIES (8). We utilized the best parameter values identified in BEELINE for the partial correlation and regression-based methods. For the two deep learning-based methods [CNNC (27) and DeepDRIM (18)], we used the original studies to reconstruct the neural networks in PyTorch before training models on the same randomly segre- gated training examples as DELAY. In addition to training both neural networks on pseudotime-aligned gene coexpression matri- ces for primary transcription factor–target gene pairs, we also trained DeepDRIM on 10 neighbor gene matrices per gene as in the original study. Moreover, we separately trained DeepDRIM on five pseudotime-lagged matrices per gene pair with 10 neigh- bors across the same inductive splits as DELAY to compare the cross-validated performance between the networks. Analyzing enrichment of GO terms across correctly inferred gene pair examples We used Enrichr (69) (https://maayanlab.cloud/Enrichr/) to ana- lyze GO-term enrichment across correctly inferred transcription factors for terms related to InterPro DNA-binding domains (70). We used a Fisher’s exact test with Benjamini–Hochberg correction for multiple-hypothesis testing (Padjusted<0.05) to assess the statis- tical significance for each GO term. Preparing single-cell multi-omics data sets to infer the hair cell gene-regulatory network We used two single-cell data sets from developing mouse outer hair cells (30, 31) to predict the consensus hair cell gene- regulatory network. First, we collected the normalized gene expression values from the original scRNA-seq study (30) and then used reciprocal PCA to integrate single cells across four time- points of sensory epithelium development into a single data set in Seurat (71) (v3.1.4). Then, we used Slingshot (6) (v1.0.0) to infer pseudotime values across the outer hair cell trajectory and ap- plied LOESS regressions and GAMs to select the genes that were differentially expressed across both the sensory epithelium and pseudotime, respectively. We again selected all differentially ex- pressed transcription factors (P<0.01 after Bonferroni correction for multiple-hypothesis testing) and 500 additional genes with the highest variance and used the integrated gene expression as- say in Seurat to generate the corresponding gene coexpression matrices. In a separate analysis, we broadened our selection cri- teria to encompass all genes that were differentially expressed in at least 20% of the outer hair cells, regardless of their expression across the full sensory epithelium. For the scATAC-seq data set, we first collected the cell-by-peak chromatin accessibility data from the original study (31). Then, we used the function createGmatFromMat in SnapATAC (72) (v1.0.0) to calculate the cell-by-gene accessibility scores as the counts of each 5 kb bin per gene in the UCSC mouse genome mm10 (73) (TxDb.Mmusculus.UCSC.mm10.knownGene; v3.4.4) that contained at least one open-chromatin peak for a given cell. These values were then log-normalized. We again used Slingshot to infer pseudotime values across the outer hair cell trajectory and applied GAMs to the raw accessibility counts to select the dif- ferentially accessible genes across pseudotime. For the inferred scATAC-seq network, we selected all differentially accessible tran- scription factors and variable genes (Padjusted<0.01) that were also differentially expressed along the scRNA-seq trajectory. Because the scATAC-seq data set had fewer cells than the scRNA-seq data set, we used 24 fixed-width bins in each dimension to generate the corresponding gene coaccessibility matrices prior to fine-tuning. Discriminative motif analysis to discover de novo transcription factor-binding motifs We used the UCSC mouse genome mm10 and twoBitToFa to download the 100 kb enhancer sequences spanning 50 kb up- stream and downstream of all target genes’ transcription start sites in the transcription factor-only hair cell gene-regulatory net- work and then divided them into 100 bp fragments and sorted the resulting sequences into primary (predicted targets) and control (predicted no interaction) groups for each transcription factor with at least one target gene in the network. We then used STREME (74) (v5.4.1) to perform discriminative motif analysis with a P value threshold of 0.05 to identify enriched motifs in pre- dicted target genes’ enhancers, which we then compared with ei- ther the transcription factors’ known binding motifs from CIS-BP (34), or to all motifs from the JASPAR (75), UniPROBE (76) (mouse), and Jolma et al. (77) databases using TOMTOM (78) (v5.4.1) with an E value threshold of 10 for significant alignments. Acknowledgments The authors would like to acknowledge Junyue Cao, Viviana Risca, and Christina Leslie, as well as Adrian Jacobo, Agnik Dasgupta, Emily Atlas, and other members of the Laboratory of Sensory Neuroscience for helpful discussions and comments on the manuscript. Supplementary material Supplementary material is available at PNAS Nexus online. Funding Reagor et al. \| 11 Author contributions C.C.R. and N.V. conceived the study, and all authors designed the analysis. C.C.R. carried out all experiments and analysis and wrote the paper. A.J.H. supervised the project, and all the authors edited the paper. Data availability The processed experimental files for all single-cell data sets used in this study are available on Zenodo at https://doi.org/10.5281/ zenodo.7474099; Table 1 lists the Gene Expression Omnibus (GEO) accession numbers for each data set. The saved model weights for DELAY are available on Zenodo at https://doi.org/10. 5281/zenodo.7474115. All experimental logs from this study are available at https://tensorboard.dev/experiment/RBVBetLM RDiEvO7sBl452A. We have provided an open-source implementa- tion of DELAY in PyTorch with listed requirements and documen- tation at https://github.com/calebclayreagor/DELAY. References 1 Ocone A, Haghverdi L, Mueller NS, Theis FJ. 2015. Reconstructing gene regulatory dynamics from high-dimensional single-cell snapshot data. Bioinformatics 31:i89–i96. 2 Qian J, Dolled-Filhart M, Lin J, Yu H, Gerstein M. 2001. Beyond synexpression relationships: local clustering of time-shifted and inverted gene expression profiles identifies new, biologically relevant interactions. J Mol Biol. 314:1053–1066. Bar-Joseph Z, Gitter A, Simon I. 2012. Studying and modelling dy- namic biological processes using time-series gene expression data. Nat Rev Genet. 13:552–564. 3 4 Ding J, Sharon N, Bar-Joseph Z. 2022. Temporal modelling using single-cell transcriptomics. Nat Rev Genet. 23:355–368. https:// doi.org/10.1038/s41576-021-00444-7 5 Tritschler S, et al. 2019. Concepts and limitations for learning de- velopmental trajectories from single cell genomics. Development 146:dev170506. Street K, et al. 2018. Slingshot: cell lineage and pseudotime infer- ence for single-cell transcriptomics. BMC Genomics 19:477 . 6 7 Trapnell C, et al. 2014. The dynamics and regulators of cell fate decisions are revealed by pseudotemporal ordering of single cells. Nat Biotechnol. 32:381–386. Papili Gao N, Ud-Dean SMM, Gandrillon O, Gunawan R. 2018. SINCERITIES: inferring gene regulatory networks from time- stamped single cell transcriptional expression profiles. Bioinformatics 34:258–266. 8 9 Deshpande A, Chu L-F, Stewart R, Gitter A. 2022. Network infer- ence with Granger causality ensembles on single-cell transcrip- tomics. Cell Rep. 38:110333 . 10 Finkle JD, Wu JJ, Bagheri N. 2018. Windowed granger causal infer- ence strategy improves discovery of gene regulatory networks. Proc Natl Acad Sci. 115:2252–2257. 11 Michailidis G, d’Alché-Buc F. 2013. Autoregressive models for gene regulatory network inference: sparsity, stability and causal- ity issues. Math Biosci. 246:326–334. 12 Zhang Y, Chang X, Liu, X. 2021. Inference of gene regulatory net- works using pseudo-time series data. Bioinformatics. 37:btab099. https://doi.org/10.1093/bioinformatics/btab099. C.C.R. is supported by National Science Foundation Graduate Research Fellowship Grant No. 1946429. A.J.H. is an Investigator of Howard Hughes Medical Institute. 13 Qiu X, et al. 2020. Inferring causal gene regulatory networks from coupled single-cell expression dynamics using scribe. Cell Syst. 10:265–274.e11. 12 \| PNAS Nexus, 2023, Vol. 2, No. 4 14 Pratapa A, Jalihal AP, Law JN, Bharadwaj A, Murali TM. 2020. 35 Grant CE, Bailey TL, Noble WS. 2011. FIMO: scanning for occur- Benchmarking algorithms for gene regulatory network inference from single-cell transcriptomic data. Nat Methods 17:147–154. 15 Yuan Y, Bar-Joseph Z. 2021. Deep learning of gene relationships from single cell time-course expression data. Brief Bioinform. 22: rences of a given motif. Bioinformatics 27:1017–1018. 36 Zheng JL, Shou J, Guillemot F, Kageyama R, Gao WQ. 2000. Hes1 is a negative regulator of inner ear hair cell differentiation. Development 127:4551–4560. bbab142. 37 Zine A, et al. 2001. Hes1 and Hes5 activities are required for the 16 Xu Y, Chen J, Lyu A, Cheung WK, Zhang, L. 2022. dynDeepDRIM: a normal development of the hair cells in the mammalian inner dynamic deep learning model to infer direct regulatory interac- ear. J Neurosci. 21:4712–4720. tions using time-course single-cell gene expression data. Brief 38 Ikeda R, Pak K, Chavez E, Ryan AF. 2015. Transcription factors Bioinform. bbac424. 17 Granger CWJ. 1969. Investigating causal relations by economet- ric models and cross-spectral methods. Econometrica 37:424–438. 18 Chen J, et al. 2021. DeepDRIM: a deep neural network to recon- struct cell-type-specific gene regulatory network using single- cell RNA-seq data. Brief Bioinform. 22:bbab325. 19 Chu L-F, et al. 2016. Single-cell RNA-seq reveals novel regulators of human embryonic stem cell differentiation to definitive endo- derm. Genome Biol. 17:173. 20 Hayashi T, et al. 2018. Single-cell full-length total RNA sequen- cing uncovers dynamics of recursive splicing and enhancer RNAs. Nat Commun. 9:619 . 21 Nestorowa S, et al. 2016. A single-cell resolution map of mouse hematopoietic stem and progenitor cell differentiation. Blood 128:e20–e31. 22 Huynh-Thu VA, Irrthum A, Wehenkel L, Geurts P. 2010. Inferring regulatory networks from expression data using tree-based methods. PLoS One 5:e12776. 23 Moerman T, et al. 2019. GRNBoost2 and Arboreto: efficient and scalable inference of gene regulatory networks. Bioinformatics 35:2159–2161. 24 Chan TE, Stumpf MPH, Babtie AC. 2017. Gene regulatory network inference from single-cell data using multivariate information measures. Cell Syst. 5:251–267.e3. 25 Kim S. 2015. Ppcor: an R package for a fast calculation to semi- partial correlation coefficients. Commun Stat Appl Methods 22: 665–674. 26 Matsumoto H, et al. 2017. SCODE: an efficient regulatory network inference algorithm from single-cell RNA-seq during differenti- ation. Bioinformatics 33:2314–2321. 27 Yuan Y, Bar-Joseph Z. 2019. Deep learning for inferring gene rela- tionships from single-cell expression data. Proc Natl Acad Sci. 116: 27151–27158. 28 Camp JG, et al. 2017. Multilineage communication regulates hu- man liver bud development from pluripotency. Nature 546: 533–538. with conserved binding sites near ATOH1 on the POU4F3 gene enhance the induction of cochlear hair cells. Mol Neurobiol. 51: 672–684. 39 Du X, et al. 2013. Regeneration of mammalian cochlear and ves- tibular hair cells through Hes1/Hes5 modulation with siRNA. Hear Res. 304:91–110. 40 Kirjavainen A, et al. 2008. Prox1 interacts with Atoh1 and Gfi1, and regulates cellular differentiation in the inner ear sensory epithelia. Dev Biol. 322:33–45. 41 Yu KS. et al. Development of the mouse and human cochlea at single cell resolution. 739680 Preprint at https://www.biorxiv. org/content/10.1101/739680v2 (2019). 42 Benito-Gonzalez A, Doetzlhofer A. 2014. Hey1 and Hey2 control the spatial and temporal pattern of mammalian auditory hair cell differentiation downstream of hedgehog signaling. J Neurosci. 34:12865–12876. 43 Doetzlhofer A, et al. 2009. Hey2 regulation by FGF provides a Notch-independent mechanism for maintaining pillar cell fate in the organ of Corti. Dev Cell. 16:58–69. 44 Kamaid A, Neves J, Giraldez F. 2010. Id gene regulation and func- tion in the prosensory domains of the chicken inner ear: a link be- tween Bmp signaling and Atoh1. J Neurosci. 30:11426–11434. 45 Yu R, Wang P, Chen X-W. 2020. The role of gfi1.2 in the develop- ment of zebrafish inner ear. Hear Res. 396:108055 . 46 Jones JM. 2006. Inhibitors of differentiation and DNA binding (Ids) regulate Math1 and hair cell formation during the development of the organ of Corti. J Neurosci. 26:550–558. 47 Hertzano R, et al. 2007. Lhx3, a LIM domain transcription factor, is regulated by Pou4f3 in the auditory but not in the vestibular sys- tem. Eur J Neurosci. 25:999–1005. 48 Hertzano R, et al. 2004. Transcription profiling of inner ears from Pou4f3 ddl/ddl identifies Gfi1 as a target of the Pou4f3 deafness gene. Hum Mol Genet. 13:2143–2153. 49 Deng M, et al. 2014. LMO4 Functions as a negative regulator of sensory organ formation in the mammalian cochlea. J Neurosci. 29 Alon U. 2007. Network motifs: theory and experimental ap- 34:10072–10077. proaches. Nat Rev Genet. 8:450–461. 30 Kolla L, et al. 2020. Characterization of the development of the 50 Bae S, Bessho Y, Hojo M, Kageyama R. 2000. The bHLH gene Hes6, an inhibitor of Hes1, promotes neuronal differentiation. mouse cochlear epithelium at the single cell level. Nat Commun. Development 127:2933–2943. 11:2389 . 51 Fior R, Henrique D. 2005. A novel hes5/hes6 circuitry of negative 31 Wang S, Lee MP, Jones S, Liu J, Waldhaus J. 2021. Mapping the regulation controls Notch activity during neurogenesis. Dev Biol. regulatory landscape of auditory hair cells from single-cell 281:318–333. multi-omics data. Genome Res. 31:1885–1899 https://doi.org/10. 52 Matern MS, et al. 2020. GFI1 Functions to repress neuronal gene 1101/gr.271080.120 expression in the developing inner ear hair cells. Development 32 Kwan KY, Shen J, Corey DP. 2015. C-MYC transcriptionally amp- 147:dev186015. lifies SOX2 target genes to regulate self-renewal in multipotent 53 Hou K, et al. 2019. A critical E-box in Barhl1 3′ enhancer is essen- otic progenitor cells. Stem Cell Rep. 4:47–60. tial for auditory hair cell differentiation. Cells 8:458. 33 Cai T, et al. 2015. Characterization of the transcriptome of nas- 54 Chonko KT, et al. 2013. Atoh1 directs hair cell differentiation and cent hair cells and identification of direct targets of the Atoh1 survival in the late embryonic mouse inner ear. Dev Biol. 381: transcription factor. J Neurosci. 35:5870–5883. 401–410. 34 Weirauch MT, et al. 2014. Determination and inference of eukary- 55 Waldhaus J, et al. 2012. Stemness of the organ of Corti relates to otic transcription factor sequence specificity. Cell 158:1431–1443. the epigenetic status of Sox2 enhancers. PLoS One 7:e36066. Reagor et al. \| 13 56 Booth KT, et al. 2020. Novel loss-of-function mutations in COCH cause autosomal recessive nonsyndromic hearing loss. Hum Genet. 139:1565–1574. 57 Aibar S, et al. 2017. SCENIC: single-cell regulatory network infer- ence and clustering. Nat Methods 14:1083–1086. 58 Marchal L, Luxardi G, Thomé V, Kodjabachian L. 2009. BMP inhib- ition initiates neural induction via FGF signaling and Zic genes. Proc. Natl. Acad. Sci. 106:17437–17442. 59 Bienvenu F, et al. 2010. Transcriptional role of cyclin D1 in devel- opment revealed by a genetic–proteomic screen. Nature 463: 374–378. 60 Luo Z, Zhang J, Qiao L, Lu F, Liu Z. 2021. Mapping genome-wide binding sites of Prox1 in mouse cochlea using the CUT&RUN ap- proach. Neurosci Bull. 37:1703–1707. 61 Popova EY, et al. 2013. Developmentally regulated linker histone H1c promotes heterochromatin condensation and mediates structural integrity of rod photoreceptors in mouse retina. J Biol Chem. 288:17895–17907. 62 Freeman SD, Daudet N. 2012. Artificial induction of Sox21 regu- lates sensory cell formation in the embryonic chicken inner ear. PLoS One 7:e46387. 63 Mali RS, et al. 2008. FIZ1 Is part of the regulatory protein complex on active photoreceptor-specific gene promoters in vivo. BMC Mol Biol. 9:87. 64 The ENCODE Project Consortium. 2012. An integrated encyclope- dia of DNA elements in the human genome. Nature 489:57–74. 65 Oki S, et al. 2018. ChIP-Atlas: a data-mining suite powered by full integration of public ChIP-seq data. EMBO Rep. 19:e46255. 66 Xu H, Ang Y-S, Sevilla A, Lemischka IR, Ma’ayan A. 2014. Construction and validation of a regulatory network for pluripo- tency and self-renewal of mouse embryonic stem cells. PLoS Comput Biol. 10:e1003777. 67 Simonyan K, Zisserman, A. 2015. Very deep convolutional net- works for large-scale image recognition. in. doi:arXiv:1409. 1556v6 68 He K, Zhang X, Ren S, Sun J. 2015. Delving deep into rectifiers: surpassing human-level performance on ImageNet classifica- tion. Preprint at http://arxiv.org/abs/1502.01852. 69 Kuleshov MV, et al. 2016. Enrichr: a comprehensive gene set en- richment analysis web server 2016 update. Nucleic Acids Res. 44: W90–W97. 70 Blum M, et al. 2021. The InterPro protein families and domains database: 20 years on. Nucleic Acids Res. 49:D344–D354. 71 Stuart T, et al. 2019. Comprehensive integration of single-cell data. Cell 177:1888–1902.e21. 72 Fang R, et al. 2021. Comprehensive analysis of single cell ATAC-seq data with SnapATAC. Nat Commun. 12:1337. 73 Navarro Gonzalez J, et al. 2021. The UCSC Genome Browser data- base: 2021 update. Nucleic Acids Res. 49:D1046–D1057. 74 Bailey TL. 2021. STREME: accurate and versatile sequence motif discovery. Bioinformatics 37:2834–2840. 75 Khan A, et al. 2018. JASPAR 2018: update of the open-access data- base of transcription factor binding profiles and its web frame- work. Nucleic Acids Res. 46:D260–D266. 76 Hume MA, Barrera LA, Gisselbrecht SS, Bulyk ML. 2015. UniPROBE, update 2015: new tools and content for the online database of protein-binding microarray data on protein–DNA in- teractions. Nucleic Acids Res. 43:D117–D122. 77 Jolma A, et al. 2013. DNA-binding specificities of human tran- scription factors. Cell 152:327–339. 78 Gupta S, Stamatoyannopoulos JA, Bailey TL, Noble W. 2007. Quantifying similarity between motifs. Genome Biol. 8:R24.	null
10.1371_journal.pone.0294847.pdf	Data Availability Statement: All relevant data are within the paper and its Supporting information files.	Data Availability Statement: All relevant data are within the paper and its Supporting information files. Funding: The ATL1102 in DMD clinical trial was funded in its entirety by the sponsor Antisense Therapeutics Ltd. Authors Ms Desem and Dr Tachas are employees of the sponsor and so received payment for services from the sponsor as employees. Ms. Desem and Dr Tachas hold an equity interest in the sponsor. Dr Tachas and Ms Desem along with other sponsor employees and sub-contracted specialists were involved in the study design and data analysis. Authors Dr Woodcock and Dr Ryan were at the time of the trial employees of the Royal Children's Hospital and Murdoch Children's Research Institute and are not affiliated with the sponsor in any way and have not received any direct personal payment or honoraria from the sponsors, nor do they or their family members hold a financial interest or stock in the sponsor company. Dr Woodcock is still an employee of the above institutions, but Dr Ryan has since left the employment to take up public office. Dr Woodcock and Dr Ryan were involved in the trial design as unpaid consultants. As this was a clinical trial, publication was always planned from trial inception. No employees of the sponsor were involved in the data collection, although Ms Desem did liaise closely with the MCRI/RCH site staff and Clinical Trial Organisation throughout the trial. Author Dr Button was paid for services as the study statistician. None of the other authors received any payment from the sponsor to conduct this study. All other authors had input into writing or revising this manuscript. ATL1102 is a second-generation immunomodulatory 2'MOE gapmer antisense oligonucleotide which specifically targets human CD49d RNA. After binding to the RNA of CD49d, intracellular RNase H attaches resulting in downregulation of CD49d RNA. ATL1102 has previously been trialled to treat Relapsing Remitting Multiple Sclerosis (RRMS), noting that CD49d high expressing T cells are the effector and central memory T cells in RRMS. In this phase 2 RRMS clinical trial, ATL1102 dosed 200mg three times weekly in the first week and twice weekly to 8 weeks substantially reduced inflammatory brain lesions by 88.5% and circulating lymphocytes and T lymphocytes by 25% [10] . Reported here, collaborators from the same institution ran two separate but complementary studies. The initial trial, a phase 2 clinical trial examining for the first time the safety and efficacy of a low dose ATL1102 treatment for 24 weeks in non-ambulant patients with DMD on concomitant corticosteroid treatment. The second, a pre-clinical study in the mdx mouse model for DMD, conducted using a mouse specific second generation CD49d ASO (ISIS 348574) to show that monotherapy treatment can reduce CD49d mRNA expression in muscle and decrease contraction induced muscle damage.	RESEARCH ARTICLE A phase 2 open-label study of the safety and efficacy of weekly dosing of ATL1102 in patients with non-ambulatory Duchenne muscular dystrophy and pharmacology in mdx mice 1,2,3, George Tachas4, Nuket Desem4, Peter J. Houweling2,3, Ian R. WoodcockID Michael Kean5, Jaiman Emmanuel5, Rachel KennedyID 1,2,6, Kate Carroll1,2, Katy de Valle1,2,6, Justine Adams2, Shireen R. Lamande´ 2,3, Chantal Coles2, Chrystal Tiong2, Matthew Burton2, Daniella Villano1, Peter Button7, Jean-Yves Hogrel8, Sarah Catling- Seyffer1,2, Monique M. Ryan1,2,3, Martin B. Delatycki9,10, Eppie M. Yiu1,2,3 1 Department of Neurology, The Royal Children’s Hospital, Melbourne, Australia, 2 The Murdoch Children’s Research Institute, Melbourne, Australia, 3 Department of Paediatrics, University of Melbourne, Melbourne, Australia, 4 Antisense Therapeutics Ltd, Melbourne, Australia, 5 Department of Medical Imaging, The Royal Children’s Hospital, Melbourne, Australia, 6 Department of Physiotherapy, University of Melbourne, Melbourne, Australia, 7 McCloud Consulting Group, Sydney, Australia, 8 Institut de Myologie, GH Pitie´ - Salpêtrière, Paris, France, 9 Victorian Clinical Genetics Service, Melbourne, Australia, 10 Murdoch Children’s Research Institute, Bruce Lefroy Centre for Genetic Health Research, Melbourne, Australia a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS [email protected] Citation: Woodcock IR, Tachas G, Desem N, Houweling PJ, Kean M, Emmanuel J, et al. (2024) A phase 2 open-label study of the safety and efficacy of weekly dosing of ATL1102 in patients with non-ambulatory Duchenne muscular dystrophy and pharmacology in mdx mice. PLoS ONE 19(1): e0294847. https://doi.org/10.1371/ journal.pone.0294847 Editor: Julie Dumonceaux, UCL: University College London, UNITED KINGDOM Received: June 29, 2023 Accepted: October 19, 2023 Published: January 25, 2024 Peer Review History: PLOS recognizes the benefits of transparency in the peer review process; therefore, we enable the publication of all of the content of peer review and author responses alongside final, published articles. The editorial history of this article is available here: https://doi.org/10.1371/journal.pone.0294847 Copyright: © 2024 Woodcock et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Abstract Background ATL1102 is a 2’MOE gapmer antisense oligonucleotide to the CD49d alpha subunit of VLA- 4, inhibiting expression of CD49d on lymphocytes, reducing survival, activation and migra- tion to sites of inflammation. Children with DMD have dystrophin deficient muscles suscepti- ble to contraction induced injury, which triggers the immune system, exacerbating muscle damage. CD49d is a biomarker of disease severity in DMD, with increased numbers of high CD49d expressing T cells correlating with more severe and progressive weakess, despite corticosteroid treatment. Methods This Phase 2 open label study assessed the safety, efficacy and pharmacokinetic profile of ATL1102 administered as 25 mg weekly by subcutaneous injection for 24 weeks in 9 non- ambulatory boys with DMD aged 10–18 years. The main objective was to assess safety and tolerability of ATL1102. Secondary objectives included the effect of ATL1102 on lymphocyte numbers in the blood, functional changes in upper limb function as assessed by Perfor- mance of Upper Limb test (PUL 2.0) and upper limb strength using MyoGrip and MyoPinch compared to baseline. PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 1 / 16 PLOS ONEData Availability Statement: All relevant data are within the paper and its Supporting information files. Funding: The ATL1102 in DMD clinical trial was funded in its entirety by the sponsor Antisense Therapeutics Ltd. Authors Ms Desem and Dr Tachas are employees of the sponsor and so received payment for services from the sponsor as employees. Ms. Desem and Dr Tachas hold an equity interest in the sponsor. Dr Tachas and Ms Desem along with other sponsor employees and sub-contracted specialists were involved in the study design and data analysis. Authors Dr Woodcock and Dr Ryan were at the time of the trial employees of the Royal Children’s Hospital and Murdoch Children’s Research Institute and are not affiliated with the sponsor in any way and have not received any direct personal payment or honoraria from the sponsors, nor do they or their family members hold a financial interest or stock in the sponsor company. Dr Woodcock is still an employee of the above institutions, but Dr Ryan has since left the employment to take up public office. Dr Woodcock and Dr Ryan were involved in the trial design as unpaid consultants. As this was a clinical trial, publication was always planned from trial inception. No employees of the sponsor were involved in the data collection, although Ms Desem did liaise closely with the MCRI/RCH site staff and Clinical Trial Organisation throughout the trial. Author Dr Button was paid for services as the study statistician. None of the other authors received any payment from the sponsor to conduct this study. All other authors had input into writing or revising this manuscript. Competing interests: The ATL1102 in DMD clinical trial was funded in its entirety by the commercial sponsor Antisense Therapeutics Ltd. Antisense Therapeutics Ltd is a publicly traded company, listed on the Australian ASX. At the time of the trial, authors Ms Desem and Dr Tachas were employees of the sponsor and so received payment for services from the sponsor as employees. Ms. Desem and Dr Tachas hold an equity interest in the sponsor. Dr Tachas and Ms Desem along with other sponsor employees and sub-contracted specialists were involved in the study design and data analysis. Ms Desem has since left the company and no longer is employed by the sponsor. Authors Dr Woodcock and Dr Ryan were at the time of the trial employees of the Royal Children’s Hospital and Murdoch Children’s Research Institute and are not affiliated with the sponsor in any way and have not received any direct personal payment or honoraria from the sponsors, nor do they or their family members Phase 2 clinical trial of ATL1102 in non-ambulatory DMD Results Eight out of nine participants were on a stable dose of corticosteroids. ATL1102 was gener- ally safe and well tolerated. No serious adverse events were reported. There were no partici- pant withdrawals from the study. The most commonly reported adverse events were injection site erythema and skin discoloration. There was no statistically significant change in lymphocyte count from baseline to week 8, 12 or 24 of dosing however, the CD3+CD49d+ T lymphocytes were statistically significantly higher at week 28 compared to week 24, four weeks past the last dose (mean change 0.40x109/L 95%CI 0.05, 0.74; p = 0.030). Func- tional muscle strength, as measured by the PUL2.0, EK2 and Myoset grip and pinch mea- sures, and MRI fat fraction of the forearm muscles were stable throughout the trial period. Conclusion ATL1102, a novel antisense drug being developed for the treatment of inflammation that exacerbates muscle fibre damage in DMD, appears to be safe and well tolerated in non- ambulant boys with DMD. The apparent stabilisation observed on multiple muscle disease progression parameters assessed over the study duration support the continued develop- ment of ATL1102 for the treatment of DMD. Trial registration Clinical Trial Registration. Australian New Zealand Clinical Trials Registry Number: ACTRN12618000970246. Introduction Duchenne muscular dystrophy (DMD), a severe, progressive, X-linked genetic muscle disease is the most common muscle disorder in boys, affecting 1 in 5000 live male births worldwide [1]. Boys with DMD have onset of progressive muscle weakness in the first decade of life, with death due to cardiorespiratory failure expected in the late third or early fourth decades [2]. Currently the only disease modifying medical treatment is corticosteroid therapy, which delays loss of ambulation by a median 3 years, to 13 years of age [3–5] but carries a significant treat- ment burden of adverse effects [5]. DMD is associated with absence of dystrophin from muscle. This causes increased suscepti- bility to contraction-induced muscle damage, with activation of the innate immune macro- phages in turn activating the adaptive immune system T lymphocytes, leading to upregulation of pro-inflammatory cytokines, including the extracellular structural protein osteopontin, resulting in chronic inflammation, fibrosis and reduced muscle strength [6]. CD49d, the alpha chain subunit of integrin very late antigen 4 (VLA-4) is expressed widely on immune cells in this cascade and can bind osteopontin [7].In patients with DMD, the number of CD49d high expressing T lymphocytes is inversely proportional to ambulation speed, with highest concen- tration seen in non-ambulant patients [8]. Patients with higher concentrations have more severe weakness and are more likely to lose ambulation before 10 yrs of age despite corticoste- roid use, suggesting CD49d may be a biomarker of disease severity or activity [9]. In ex vivo studies a monoclonal antibody to VLA-4 prevented patient T-cell binding to muscle cells and transendothelial migration, highlighting a potential therapeutic avenue [9]. PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 2 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD hold a financial interest or stock in the sponsor company. Dr Woodcock is still an employee of the above institutions, but Dr Ryan has since left the employment to take up public office as a Member of the Australian Parliament. Dr Woodcock and Dr Ryan were involved in the trial design as unpaid consultants. Dr Woodcock has received honoraria for work performed including educational activities and attendance at advisory board meetings from pharmaceutical companies Biogen, Novartis, Roche and Avidity and an educational travel bursary to attend an international conference in 2016 from Biogen. Dr Woodcock has received grants for research work from FSHD Global Research Foundation, FSHD Society and Fulcrum Therapeutics. Dr Woodcock has been principal investigator on a number of industry-sponsored clinical trials. None of these disclosures affected the work Dr Woodcock performed on this clinical trial. Dr Ryan has received honoraria for work performed including educational activities and attendance at advisory board meetings from pharmaceutical companies Biogen, Novartis, Roche. Dr Ryan has been principal investigator on a number of industry-sponsored clinical trials. None of these disclosures affected the work Dr Ryan performed on this clinical trial. Dr Yiu has received advisory board honoraria from Biogen and Roche, and has received research support from Biogen, Roche, Pfizer and PTC therapeutics unrelated to the content of this manuscript. Dr Yiu has been principal investigator on a number of industry-sponsored clinical trials. None of these disclosures affected the work Dr Yiu performed on this clinical trial. Prof. Delatycki has received grant awards from NHMRC and is principal investigator in industry sponsored clinical trials including trials sponsored by Rearta and PTC. As this was a clinical trial, publication was always planned from trial inception. No employees of the sponsor were involved in the data collection, although Ms Desem did liaise closely with the MCRI/RCH site staff and Clinical Trial Organisation throughout the trial. As the study statistician, author Dr Button was paid a consultancy fee for his services from the trial sponsor commercial company Antisense Therapeutics Ltd. Authors Dr Houweling, Dr Coles and Dr Tiong were recipients of a grant to perform the MDX studies. This grant was paid by the sponsor Antisense Therapeutics Ltd. None of the other authors received any payment from the sponsor to conduct this study. All other authors had input into writing or revising this manuscript. The authors confirm that the involvement of employees of the sponsor Antisense Therapeutics Ltd in the trial design, data analysis and decision to ATL1102 is a second-generation immunomodulatory 2’MOE gapmer antisense oligonucle- otide which specifically targets human CD49d RNA. After binding to the RNA of CD49d, intracellular RNase H attaches resulting in downregulation of CD49d RNA. ATL1102 has pre- viously been trialled to treat Relapsing Remitting Multiple Sclerosis (RRMS), noting that CD49d high expressing T cells are the effector and central memory T cells in RRMS. In this phase 2 RRMS clinical trial, ATL1102 dosed 200mg three times weekly in the first week and twice weekly to 8 weeks substantially reduced inflammatory brain lesions by 88.5% and circu- lating lymphocytes and T lymphocytes by 25% [10]. Reported here, collaborators from the same institution ran two separate but complementary studies. The initial trial, a phase 2 clinical trial examining for the first time the safety and effi- cacy of a low dose ATL1102 treatment for 24 weeks in non-ambulant patients with DMD on concomitant corticosteroid treatment. The second, a pre-clinical study in the mdx mouse model for DMD, conducted using a mouse specific second generation CD49d ASO (ISIS 348574) to show that monotherapy treatment can reduce CD49d mRNA expression in muscle and decrease contraction induced muscle damage. Methods Ethics statement The clinical trial received approval from the Royal Children’s Hospital Human Research Ethics Committee with assigned number HREC/17/RCHM/121. The trial was subsequently regis- tered at the Australian New Zealand Clinical Trials Registry (ACTRN12618000970246). An independent Data Safety Monitoring Board (DSMB) was established to provide safety over- sight for the trial. Participant consent to participate in the trial was sort from parents. Pre-clin- ical mdx mouse analyses were approved by the Murdoch Children’s Research Institute (MCRI) animal care and ethics committee (ACEC; approval number A899). Pre-clinical studies in mice The mdx mouse model is commonly used to study DMD. We tested efficacy of the second gen- eration 2’MOE gapmer mouse specific CD49d ASO ISIS 348574 as ATL1102 is specific to human CD49d RNA and not homologous to mouse. Mdx mice do not have circulating lym- phocytes with high CD49d but have high CD49d expressing lymphocytes in the lymph nodes at 9 weeks [11]. Symptomatic 9 week old mdx mice were treated for 6 weeks to determine the effect of ISIS 348574 on CD49d mRNA expression and muscle function measures. Male mdx and age matched C57Bl10/J wild-type controls were purchased from the Jackson laboratories at 5 weeks of age and acclimatised to the MCRI facility for a total of 4 weeks. Ani- mals were housed in a specific-pathogen-free environment at a constant ambient temperature of 22˚C and 50% humidity on a 12 h light-dark cycle, with ad libitum access to food and water. The mdx mice (n = 12 /group) were randomly assigned to 4 treatment groups (saline, low (5mg/kg) and high (20mg/kg) dose ISIS 348574, and a control gapmer oligonucleotide with the same 20 nucleotides scrambled such that it is not complementary to CD49d or other RNA (20mg/kg)). Mdx Mice received weekly subcutaneous injections of either saline, ISIS348574 and control oligonucleotide for a total of 6 weeks. Wild-type controls (n = 12) received saline only. The high dose (20mg/kg/week) equates to 1.6mg/kg/w dose in human equivalent body surface area (BSA) and the low dose (5mg/kg/week) equates to 0.4mg/kg/week dose on BSA. After treatment mice were anaesthetised using inhaled isoflurane (0.6 ml per min) and muscle function was examined using the Arora Scientific 1300A whole mouse test system and 701C stimulator as previously published [12]. Mice were then euthanised by cervical dislocation and the spleen and skeletal muscle (quadriceps) were collected for further analysis. PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 3 / 16 PLOS ONEpublish this data does not alter our adherence to PLOS ONE policies on sharing data and materials. Phase 2 clinical trial of ATL1102 in non-ambulatory DMD Flow cytometry The mdx spleen was perforated to isolate the splenic sub-cellular content and incubated in red blood cell lysis buffer (Thermofisher) for ten minutes at 4˚C. Splenocytes were centrifuged 1500 g for five minutes at 4˚C. Cell pellets were washed in wash buffer (PBS:1% BSA (bovine albumin serum)). CD4+ and CD8+ T cell populations were identified using anti-CD4-V450 (Biolegend, San Diego, CA, USA) and anti-CD8a-allophycocyanin–cyanine 7 (anti-CD8a- APC-Cy7) (Biolegend, San Diego, CA, USA). Stained cells were analysed using BD LSRFor- tessa™ X-20 Cell Analyzer to identify populations of CD4+ and CD8+ T cells. Clinical trial design This was a phase 2 open-label clinical trial assessing the safety of ATL1102 in non-ambulant boys with DMD concomitantly with their usual corticosteroid therapy, in all but one partici- pant who discontinued corticosteroids years prior to the study. CONSORT flow diagram of study design is shown in Fig 1. Eligibility criteria for the clinical trial are included as S1 Fig. Participants were recruited from a single site in Melbourne, Australia with the study running from August 2018 until January 2020. After providing written informed consent, participants received weekly subcutaneous injec- tions of 25mg of ATL1102 for twenty-four weeks. Injections were administered into subcuta- neous fat of the abdomen by a registered nurse or trained parent. Injection sites were rotated in quadrants around the umbilicus. Parents monitored injection sites for cutaneous reactions and participant discomfort for fourty-eight hours post-administration. Adverse events were recorded in a diary which was returned to the study coordinators at each fortnightly visit. Participants underwent fortnightly venepuncture for exploratory and safety blood tests. This included monitoring of haematology, biochemistry and inflammatory markers and at point of care urinalysis dipstick to monitor kidney function. Participants were seen monthly by the study team for physical examination and respiratory function assessments. Participant safety was monitored routinely and at regularly scheduled meetings by the inde- pendent DSMB. Fig 1. CONSORT flow diagram of study design. https://doi.org/10.1371/journal.pone.0294847.g001 PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 4 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD Outcome measurements The primary endpoint of the trial was safety of ATL1102 as assessed by the frequency and intensity of adverse events, including injection site reactions and any laboratory value derangement. Secondary outcome measures included both laboratory and functional efficacy endpoints. Laboratory efficacy outcome measures included lymphocyte-modulation activity determined by cell surface flow cytometry measuring variation in the number and percentage of total lym- phocytes as well as those CD4 and CD8 T lymphocytes expressing high levels of CD49d (CD49dhi) to week twenty-four of treatment and to week twenty-eight, four weeks past the last treatment. Changes in upper limb muscle strength and function were measured at baseline and again at weeks five, eight, twelve and twenty-four. Muscle function was assessed by a questionnaire- based outcome measure of disease burden (Egen Klassifikation Scale version 2—EK2) and per- formance-based measures (the Performance of the Upper Limb scale version 2 (PUL 2.0), and the Moviplate 30 second finger tapping score of the MyoSet tool). The Myoset tool also mea- sured distal upper limb strength as determined by the MyoPinch (key pinch strength) and MyoGrip (hand grip strength) scores [13–15]. Other outcome measures assessed were respira- tory function (forced vital capacity (FVC) and forced expiratory volume in one second (FEV1)), and quality of life assessed using the neuromuscular module of the Pediatric Quality of Life Instrument (PedsQL NMD™). Muscle Magnetic Resonance Imaging (MRI) Participants underwent MRI of the dominant forearm at baseline, week twelve and week twenty-four. Unilateral upper-limb MRI was performed at 1.5T (Siemens Aera; Siemens, Erlangen, Germany) using a flexible surface matrix coil (4-Channel Flex Coil) wrapped around the forearm. Participants lay in the scanner in the head-first supine position, with the arm to be imaged lying in a comfortable position on the scanner bed alongside the torso. Two point- Dixon images were acquired (3D gradient-echo TE1/TE2/TR = 2.39/4.44/6.99ms, flip angle 10˚, nine 6mm axial slices, slice gap 0mm, FOV 18x18cm, matrix 320×320, pixel size 0.56×0.56mm, NEX = 4). Fat fraction maps were obtained using on scanner tools. Change over time in muscle composition (atrophy, oedema and fatty infiltration) was mea- sured on the muscles of the central, proximal, and distal forearm using Short Tau Inversion Recovery (STIR) and 3-point Dixon sequences on MRI. Changes in muscle composition were scored using the semi-quantitative visual scoring Mercuri method and by quantitative fat frac- tion analysis [16–18]. Due to fatty infiltration, identification of individual muscles was chal- lenging, such that a compartment composite score of volar, dorsal and ECRLB Br (extensor carpi radialis longus/brevis and brachioradialis) compartments was used as per a previous published study [19]. The lean muscle mass was calculated using previously published meth- ods: Cross-sectional muscle compartment area x ((100 –total muscle compartment fat percent) / 100) [19]. Statistical analysis Based upon data from a previous study of ATL1102 in RRMS patients analyzing blood 3 days after the last dose in week 8, the laboratory efficacy end point of lymphocyte modulation potential was established as a reduction in total lymphocyte count of 0.47x109/L (25% reduc- tion) [10]. For the sample size calculation, the level of significance was set to 0.05 with a 2-sided paired t-test, mean difference of 0.47 (x109/L) from baseline to end of treatment, and standard deviation of 0.428 (x109/L). Using nQuery (Version 8.5.2.0, Table MOT1-1 Paired t- PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 5 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD test for differences in means, Statistical Solutions Ltd.), a sample size of 9 participants was cal- culated as required to achieve a power of 80%. Nine participants were considered sufficient to investigate the safety, tolerability and PK and PD profile of ATL1102 in this rare target patient population. Data were analysed using SAS1 Version 9.4. primary and secondary efficacy mea- sures were analysed using the paired t-test and the non-parametric Wilcoxon sign-rank test. The study was not powered to see a change on the secondary efficacy endpoints from baseline to end of treatment. The study protocol and statistical analysis plan are available as supplemen- tary data in “Protocol” and eSAP1 and eSAP2 respectively. Repeated measures analysis of lym- phocytes and T lymphocyte and NK lymphocyte subsets in a post hoc analysis was conducted comparing baseline to a linear combination of values measured three days post-dose at weeks 8, 12 and 24. Associations between variables was tested using a Pearson correlation test. Preclinical mdx mouse analyses were performed in Graphpad Prism (V9, Graphpad Soft- ware Inc.). For in situ muscle function, one-way ANOVA with Tukey correction for multiple testing was performed (n = 9 animals / treatment). Unpaired T-Tests were used for CD4 and CD8 T-cell analyses (n = 5–6 samples / treatment). All data shown as mean with 95% confi- dence intervals, unless otherwise stated in the figure legends. Results Proof of concept preclinical study using mdx mice Ex vivo analyses of monocytes collected from mdx mice (n = 3), showed that the ASO to mouse CD49d, ISIS 348574 can reduce the expression of CD49d mRNA (Fig 2A). We then examined the in vivo response to ISIS 348574 in mdx mice treated for 6 weeks which showed that CD49d mRNA expression was reduced in skeletal muscle by approximately 40% when treated with either a low (5mg/kg/week) or high (20mg/kg/week) dose of ISIS 348574, com- pared to saline controls (Fig 2B, One-way ANOVA, summary p<0.01, with Tukey correction displayed on the graphs, p = * <0.05, ** <0.01). This study also found that mdx mice treated with the 20mg/kg/week dose of ISIS 348574 showed a reduction in the percentage of splenic CD4+ (30%, p<0.05) and CD8+ (21%, p = 0.058) T lymphocytes, compared to saline treated mice (Fig 2C and 2D). Furthermore in situ muscle funcation analyses show that the high dose (20mg/kg/week) treated mdx mice were protected from the effects of eccentric muscle damage, producing 72% of the original muscle force (P<0.01). This was in contrast to mdx mice treated with either saline, scrambled control or low dose (5mg/kg) ISIS 348574, which generated approximately 50% of the original force following eccentric muscle contractions. (Fig 2E and 2F, One-way ANOVA summary p = <0.001, with Tukey correction displayed on the graphs, p = * <0.05, <0.01, * <0.001). Clinical trial results Eleven adolescent males with DMD were screened for participation. All had been non-ambu- lant for at least six months prior to screening. Two screened participants were excluded (one participant had started cardioprotective medication within three months of the initial visit and the other participant exceeded the pre-determined weight limit for inclusion). Nine participants were enrolled into the open-label study. All had a confirmed pathogenic variant in DMD, with a clinical phenotype consistent with DMD as assessed by their treating/ referring clinician and the study investigator. Participant demographics are summarized in Table 1. Safety. There were no serious adverse events (SAEs) or suspected or unexpected serious adverse reactions (SUSARs). A total of 136 adverse events were recorded (Table 2), with all PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 6 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD Fig 2. Pre-clinical data using the mdx mouse model of DMD to test the effects of ISIS 348574 (mouse specific Cd49d oligonucleotide to ATL 1102) in vivo. A) Monocytes were isolated from the spleens of mdx mice (n = 3) and exposed to a single dose of ISIS 348574 for 48hrs in vitro to show that we could achieve a reduction in CD49d mRNA using ISIS 348574 to mouse CD49d RNA. B) Following 6 weeks of treatment CD49d mRNA expression was reduced in mice treated with either the low (5mg/kg/week) or high (20mg/kg/week) dose of ISIS 348574, compared to saline controls (One-way ANOVA with Tukey correction, p = * <0.05, ** <0.01). C and D) Proportion of CD4+ and CD8+ T cells from the spleens of mdx mice with and without ISIS 348574 drug treatment. Cells are expressed as a proportion of total live cells isolated from the spleen. One way ANOVA with Fishers LSD test, * p < 0.05. E and F) In situ muscle physiology analyses shows that mdx mice treated with either saline (red, ~45% force recovery), scrambled (grey, ~45% force recovery) or low dose (orange, ~ 50% force recovery) ISIS 348574 were susceptible to eccentric muscle contraction damage compared to wild-type (black) controls, whereas the mice treated with a high dose of ISIS 348574 were resistant to the effects of eccentric muscle damage and produced 72% of the original force following the eccentric PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 7 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD muscle damaging protocol. This was still significantly less than the 95% force recovery seen in WT mice, however this improvement in force following a muscle damage protocol suggests that the use of a 20mg/kg/week dose of ISIS 348574 was able to protect the muscles of mdx mice. One-way ANOVA with Fishers LSD test, p = * <0.05, <0.01, * <0.001), ***<0.0001). https://doi.org/10.1371/journal.pone.0294847.g002 participants reporting at least one adverse event. Sixty-three percent of the reported adverse events were injection site related, with all but one participant experiencing transient erythema within twenty-four hours of the administration of ATL1102. Six (67%) participants had mild post-inflammatory hyperpigmentation of the skin of their abdomen which was persistent; four had resolved and two were ongoing but improving at the post completion follow-up study visit. The hyperpigmentation was noticed in the first participant after receiving eleven weekly doses of ATL1102. The DSMB was made aware and the participant informed consent form updated. In the five subsequent participants who had a similar reaction it was seen after four to eleven doses. The hyperpigmentation was not regarded as a clinical safety concern by the DSMB. Pain, discomfort or atrophy of the subcutaneous tissues were not reported, and there were no signs of systemic involvement. No participants withdrew from the study. There were no other significant adverse events felt to be related to ATL1102 or its administration. Efficacy. Lymphocyte Count: There was no statistically significant decrease in lymphocyte count from baseline to week eight, week twelve or week twenty-four of dosing (Table 3). There was no statistically significant decrease in CD49d+CD3+CD8+ or CD49d+CD3+CD4+ T lym- phocytes seen between baseline, weeks 8, 12 or 24 (Table 3). This 9 participant trial did not achieve the pre-specified laboratory activity outcome measure of a significant -0.47x109/L (25% reduction) in total lymphocyte count. There was, though, a consistent trend toward declines in the mean number of lymphocytes, and CD49d+ T lymphocytes measured 3 days post-dose at week 8, 12 and 24. The mean num- ber of CD3+CD49d+ T lymphocytes (i.e. CD3+CD4+ and CD3+CD8+ expressing CD49d) measured at week 28 was statistically significantly higher compared to end of dosing at week 24 (mean change 0.40x109/L 95%CI 0.05, 0.74; paired T-Test, p = 0.030) (Table 3). Repeated measures analysis of CD3-CD49d+ NK lymphocytes in a post hoc analysis comparing baseline to a linear combination of values measured three days post-dose at weeks 8, 12 and 24 was sig- nificantly lower compared to baseline (p = 0.018), with comparable NK lymphocyte numbers at week 28 (Fig 3). Functional outcome measures. There were no statistically significant changes in any upper limb functional outcome measures at week 24 compared to baseline (Table 4). The PUL2.0 score remained stable with no significant change between baseline and week twenty- Table 1. Summary of participant demographics. Characteristic Sex Age (years) Weight (kg) Height (cm) BMI Time since non-ambulant (years) Corticosteroid Medication https://doi.org/10.1371/journal.pone.0294847.t001 Category Male Yes Prednisolone Deflazacort Statistic n (%) Mean (SD) Median (range) Mean (SD) Mean (SD) Mean (SD) Median (range) n (%) ATL1102 N = 9 9 (100) 14.9 (2.1) 14.0 (12–18) 52.7 (9.8) 141.1 (10.0) 27.1 (7.4) 2.2 (0.6–9.2) 8 (88.9) 3 (33.3) 5 (55.6) PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 8 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD Table 2. Treatment emergent adverse events reported in at least two participants. SYSTEM ORGAN CLASS Preferred Term Participants reporting any AEs All N = 9 Participants (%) [No. of Events] 9 (100.0%) [136] GENERAL DISORDERS AND ADMINISTRATION SITE CONDITIONS Injection site erythema Injection site pain Injection site swelling Injection site bruising Pyrexia SKIN AND SUBCUTANEOUS TISSUE DISORDERS Skin discolouration GASTROINTESTINAL DISORDERS Vomiting Constipation RESPIRATORY, THORACIC AND MEDIASTINAL DISORDERS Cough Nasal congestion Oropharyngeal pain INFECTIONS AND INFESTATIONS Lower respiratory tract infection Nasopharyngitis NERVOUS SYSTEM DISORDERS Migraine https://doi.org/10.1371/journal.pone.0294847.t002 8 (88.9%) [59] 5 (55.6%) [7] 3 (33.3%) [6] 4 (44.4%) [4] 2 (22.2%) [4] 6 (66.7%) [7] 2 (22.2%) [4] 2 (22.2%) [2] 2 (22.2%) [2] 2 (22.2%) [2] 2 (22.2%) [2] 2 (22.2%) [2] 2 (22.2%) [2] 2 (22.2%) [2] four with a mean increase in PUL2.0 score of 0.9 (95%CI -1.33, 3.11, where a higher score indi- cates better function) EK2 scores were stable throughout the trial period. ATL1102 treatment showed no significant effect on lung function throughout the 24 week trial period (Table 4). The components of the Myoset all reflected stable grip and pinch strength over the course of the trial. Table 3. Summary of lymphocytes mean change from baseline to weeks 24 and 28. White blood cell type (X109 cells per litre) Mean Baseline count (x109 cells per litre) Mean Change from baseline Median percentage change from baseline (%) Paired T-Test (p value) of mean change Lymphocytes CD3+ T cells CD3+ CD49d+ T cells CD4+ T cells CD4+ CD49d+ T cells CD8+ T cells CD8+ CD49d+T cells Week 8 Week 12 Week 24 Week 28 Week 8 Week 12 Week 24 Week 28 between week 28 and 24 -0.56 -0.53 -0.50 -0.30 -0.28 -0.20 -0.22 -0.53 -0.33 -0.39 -0.20 -0.22 -0.09 -0.12 -0.28 -0.18 -0.28 -0.15 -0.19 -0.02 -0.05 +0.19 +0.25 +0.11 +0.11 +0.01 +0.14 +0.11 -4.63 -10.9 -12.3 -12.3 -12.5 -9.35 -10.9 -7.14 -5.46 -10.0 -5.23 -7.09 -5.21 -7.32 -4.22 +11.81 0.86 +17.11 -9.78 -1.12 -16.7 -2.62 -5.79 +9.93 +16.50 +1.73 +17.99 +13.37 0.051 0.056 0.03 0.063 0.073 0.068 0.064 3.68 2.93 2.44 1.57 1.20 1.22 1.17 The Lymphocyte mean number of cells at week 24 (at the end of dosing) is trending significantly lower vs week 28 (p = 0.051 paired T test). The mean number of CD3+CD49d+T lymphocytes (CD4+CD49d+ and CD8+CD49d+ T lymphocytes) at week 24 is statistically significantly lower vs week 28 (p = 0.030 paired T test). https://doi.org/10.1371/journal.pone.0294847.t003 PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 9 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD Fig 3. Showing lymphocyte baseline, week 8, 12 and 24 week data 3 days post each dose and the week 28 data, four week past end of dosing data is shown as a bee swarm plot with mean expression values of lymphocytes. https://doi.org/10.1371/journal.pone.0294847.g003 MRI. In this trial, there was some variation in the quality of proximal and distal slices due to variable positioning of the participant’s forearm in consecutive scans. The quality of the cen- tral slices through the muscle body in the forearms muscles was not compromised and so this measurement was chosen for detailed comparison between baseline and week 24 scans for each participant. No significant change was apparent between baseline and week twenty-four for the mean Mercuri semi-quantitative score of fatty infiltration (0.1 point change), atrophy (0 point change) and muscle oedema (0.3 point change) measuring the central slice around the elbow (Table 5). There was a trend towards minor improvement in the percentage fat fraction in all muscle groups measuring the central slice, although statistical significance was not achieved. There was no significant pattern of change in cross-sectional muscle area of any muscle group (Table 6). Table 4. Change in functional outcome measures from baseline to week 24. Change from Baseline to Week 24 MyoGrip (dom) (% Pred) MyoPinch (dom) (Kg) MyoPinch (dom) (% Pred) MoviPlate Score (dom) % Predicted FVC % Predicted PEF EK2 Patient No. PUL 2.0 01–001 01–002 01–003 01–004 01–006 01–008 01–009 01–010 01–011 +2 +2 0 +2 -3 +7 0 0 -2 MyoGrip (dom) (Kg) -0.63 0.22 0.68 1.09 -0.27 1.00 -0.33 0.05 0.11 -4.49 0.49 1.02 1.01 -0.60 1.11 -3.75 0.11 -1.31 0.03 -0.02 -0.40 0.37 0.07 0.30 -0.22 0.06 -0.18 -0.62 -0.29 -6.59 2.99 0.94 2.77 -4.97 0.72 -3.63 Mean Change (95% CI): +0.9 (-1.33, 3.11) +0.2 (-0.25, 0.67) -0.7 (-2.33, 0.90) 0.0 (-0.18, 0.19) -1.0 (-3.56, 1.63) *Higher score = greater disability. #Reduction in Fat Fraction (%) = improvement. https://doi.org/10.1371/journal.pone.0294847.t004 -14.0 13.0 -3.0 7.0 8.0 3.0 7.0 -15.0 11.0 1.9 (-6.08, 9.85) -3.20 -14.8 -9.10 0.80 -6.50 -7.70 -9.10 -0.40 -1.10 6.30 -17.3 8.70 7.20 6.90 -18.2 -4.30 9.20 2.00 +1 +1 +2 +2 -6 -1 +2 -1 +2 -5.68 (-9.60, -1.76) 0.06 (-8.33, 8.44) 0.2 (-1.80, 2.25) PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 10 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD Table 5. Mercuri visual semi-quantative score in whole forearm compartment from baseline to week 24. Fatty Infiltration Atrophy Oedema Whole Forearm Baseline 5.2 1.6 2.6 Wk 24 Change 5.3 1.6 2.9 0.1 0.0 0.3 https://doi.org/10.1371/journal.pone.0294847.t005 Table 6. Change in the MRI Central reading fat fraction, cross sectional area and lean muscle mass from baseline to week 24. Mean Change (95% CI) from Screening/Baseline to Week 24 MRI Parameter Fat Fraction (%) Volar Muscle Dorsal Muscles ECRLB-Br Average Fat Fraction Cross Sectional Muscle Area (mm2) Volar Muscle Dorsal Muscles ECRLB-Br Total Area Lean Muscle Mass (mm2) N MRI Central Reading Mean (95%CI) 9 9 9 9 9 9 9 9 9 -0.57 (-7.81, 6.68) -0.88 (-3.41, 1.65) -0.12 (-6.42, 6.17) -0.52 (-5.62, 4.58) 22.78 (-31.2,76.73) 0.89 (-18.9,20.65) -1.33 (-8.94, 6.28) 22.33 (-36.8,81.42) 13.9 (72.6, 100.4) ECRLB-Br = extensor carpi radialis longus/brevis and brachioradialis. Volar Muscles; flexor digitorum profundus and flexor pollicis longus (FDP), flexor digitorum superficialis and palmaris longus (FDS), flexor carpi ulnaris (FCU), flexor carpi radialis (FCR). Dorsal Muscles: Extensor carpi ulnaris (ECU), extensor digiti minimi (EDM), extensor digitorum (ED), extensor pollicis longus (EPL), abductor pollicis longus (APL), extensor carpi radialis longus/brevis and brachioradialis (ECRLB-Br), but the ECRL-BR are not included in the Dorsal muscle measurement in the Central Reading. Change in the MRI Proximal reading average Fat Fraction (%) from baseline to week 24 was -2.14 [95%CI -7.60; 3.3] for the 9 patients. https://doi.org/10.1371/journal.pone.0294847.t006 Correlation of parameters assessed in the Phase 2 study. Correlation analyses were per- formed across assessment measures PUL2.0, Myoset and MRI. Positive correlations were observed in the Phase 2 study between the different measures of muscle function of Moviplate scores and the PUL 2.0 scores of the distal domain (r = 0.664) which support the consistency of the observed changes across the measures assessed in the study over the 24 week ATL1102 treatment period (S2 Fig). Positive correlations were also observed in the Phase 2 study between the MRI results of the lean muscle area (non-fat) and MyoGrip results (r = 0.604), suggesting a consistency of results across the different parameters of muscle structure and muscle strength (S3 Fig). Discussion Safety This open-label phase 2 clinical trial met its primary safety end point. All but one participant experienced post-injection site erythema, swelling or discomfort suggesting that the investi- gation product is a mild irritant, as has been observed with other MOE antisense drugs, and PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 11 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD is commonly reported with subcutaneiously injected drugs. Future clinical trials of ATL1102 could consider including using ice as a pre-injection site treatment to minimise these reac- tions. Six participants experienced an unexpected post-inflammatory skin hyperpigmenta- tion which resolved or faded at the completion of the study. Interestingly this reaction has not been previously reported in other clinical trials of ATL1102 and was not viewed as a safety concern [10]. The dose chosen for this 24 week trial (25mg/week) was considered as a presumed safe dose in this patient population. In a previous phase 2 trial in individuals with RRMS a loading dose of 200mg every other day for one week was administered then 400mg/week (twice weekly 200mg) for seven weeks [10]. This DMD trial was the first clinical trial to investigate the safety of ATL1102 over a six-month period. Lymphocyte modulation Given the previously observed action of ATL1102 of reducing lymphocytes in the RRMS study, the trial sample size was calculated to see a 25% reduction in total lymphocyte count when the drug is at equilibrium from week 8. This activity endpoint was not met in this trial. There was however a consistent trend toward declines in the mean number of lymphocytes at week 8, 12 and 24 each measured 3 days past dosing, and statistically significant reductions in the mean number of CD49d+ NK lymphocytes at week 8, 12 and 24 weeks of treatment, using repeated measures analysis. The mean number of CD49d+ T lymphocytes (i.e CD3+CD4 + and CD3+CD8+ that are CD49d+) was statistically significantly higher at week 28 compared to week 24, indicating a rebound elevation of CD49d+ T lymphocytes four weeks post the last treatment dose. These results collectively suggest that ATL1102 suppresses CD49d expressing lymphocytes at a dose of 25mg per week. It is anticipated that higher doses will increase the level of lymphocyte reduction whilst maintaining a favourable safety profile in part due to sparing of the majority of T lymphocytes and NK lymphocytes. Future studies will look at dose escalation as supported by this study, and modelling with ATL1102. PUL2.0 and EK2 upper limb function PUL2.0 measures shoulder, elbow, and wrist finger dimensions of disease burden and is a reli- able measure of disease severity and progression in DMD where a lower score indicates loss of function. The mean increase from baseline to week twenty-four in the PUL2.0 was 0.9 (95% CI -1.33 to 3.11). Although the Minimal Clinical Important Difference (MCID) for the PUL2.0 has not been established, an external historical cohort with same inclusion criteria as in the ATL1102 phase 2 trial, showed a decrease in PUL2.0 score of 2.0 (standard deviation 3.02) from baseline over a six month period [20]. In the ATL1102 phase 2 study four of the nine patients achieved an increase in their PUL2.0 score of +2, and another three patients were sta- bilized in the PUL2.0 score (Table 4). This is an encouraging trend that warrants further inves- tigation. A previously published data from a historical cohort also reported that over a twelve month period a mean decrease in PUL2.0 score of 2.17 can occur, albeit in a cohort not directly comparable to the participants in the phase 2 study due to older age and larger propor- tion not on corticosteroids [21]. There was no change in the EK2 over the course of the trial period. This composite outcome measure encompasses multiple aspects of disease burden and as such is a useful clinical moni- toring tool (higher score equals greater disease burden) but is not likely to be as responsive as the PUL2.0 measure to small changes in upper limb function. As such, a stabilisation over the six month trial periods is encouraging and needs to be confirmed in future studies. PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 12 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD Myoset tests: MoviPlate, MyoGrip, and MyoPinch The MyoSet functional outcome measures consist of the MoviPlate muscle function assess- ment of repetitive flexion extension of the wrist and fingers, and MyoGrip and MyoPinch assessment of muscle grip and pinch strength [13,14]. These measures have been validated for use in clinical trials of DMD; MyoGrip and MyoPinch in particular have been shown to be sensitive to change in non-ambulant boys and to correlate well with lean muscle mass on MRI [15]. There was no change in any of these measures over the trial period. Previous natural his- tory studies have shown significant deterioration over six months (Grip -0.5kg [95%CI -1.01; 0.002] and Pinch -0.38kg [95%CI -0.53; -0.22]) [15,18]. Matching the MyoGrip and MyoPinch protocol with that of a previously published natural history cohort allowed for comparison of change in grip and pinch strength over a six month period, yielding a statistically significant improvement on grip (p = 0.03) and pinch (p = 0.003) strength [19]. The lack of decline in these measures with ATL1102 during the trial period is once again encouraging and warrants further investigation. MRI of upper limb Magnetic Resonance Imaging of muscle is increasingly used as a biomarker for disease stage and progression. The most widely used scoring method is the Mercuri Score, which requires a skilful investigator to visually score the chosen muscles based upon a standardised set of crite- ria encompassing degree of atrophy, oedematous changes and fatty infiltration of the muscle, to create an aggregate score. The more recent development of automated fat fraction analysis reduces the inter-user variability and provides a more quantitative measure of assessment. Matching the MRI protocol with that of a previously published natural history cohort allowed for direct comparison of change in fat fraction over a six month period [19]. From this pub- lished natural history data, disease is expected to progress with a mean increase of central fore- arm muscle fat fraction percentage of 3.9% (95%CI 1.9,5.7) over six months. The apparent trend towards a decrease in mean forearm muscle fat fraction of 0.52% (95% CI -5.62, 4.58; Median 1.4%) seen after six months of treatment with ATL1102 may suggest that ATL1102 could be modifying the rate of fatty infiltration into these muscles. These changes were repli- cated in the proximal and distal muscle groups. For future MRI studies it would be important to set a clear protocol with imaging tags placed over surface landmarks to ensure uniformity of subsequent scans. Conclusion The proof of concept pre-clinical data supports a potential protective effect of an antisense oli- gonucleotide to CD49d RNA in the mdx mouse model of DMD. This phase 2 open-label clini- cal trial has shown that ATL1102 has a good safety profile and is well tolerated with minor injection site reactions the only treatment-related adverse events reported. The positive obser- vations in functional efficacy outcomes suggesting stabilization, and results compared with historical natural history data, particularly the PUL2.0, MyoGrip, MyoPinch and MRI fat frac- tion analysis, justifies the ongoing drug development program of ATL1102 in non-ambulant boys with DMD and provides a rationale to proceed with larger placebo-controlled studies of this novel therapeutic agent. Supporting information S1 Checklist. TREND statement checklist. (PDF) PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 13 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD S1 Fig. Inclusion and Exclusion criteria for the clinical trial. (DOCX) S2 Fig. Scatter Plot showing the association between the Moviplate scores and the PUL 2.0 distal dimension scores over the 24 Week ATL1102 treatment period with a linear regres- sion line plotted. (TIF) S3 Fig. Scatter plot showing the association of the Grip Strength scores and the MRI data of the forearm lean muscle area (non-fat) over the 24 week ATL1102 treatment period: The MRI is the determined lean muscle mass compartment across the mid (central) domi- nant forearm. The linear regression line is also plotted. (TIF) S4 Fig. Gating strategy for identification of murine CD4+ and CD8+ T cell populations isolated from spleen of mdx mice treated with ISIS 348574. A) Singlets (FSC-H vs FSC-A), B) Live cells (Propidium Iodide vs FSC-A) and C) Lymphocytes (SSC-A vs FSC-A) were gated to remove doublets, dead cells, debris and large/granular cells. D) Anti-CD4-V450 and anti- CD8a APC-Cy7 to were used to gate populations of CD4+ and CD8+ T cells. (TIF) S5 Fig. Correlation of the Moviplate scores and the PUL 2.0 distal dimension scores over the 24 Week ATL1102 treatment period. (TIF) S6 Fig. Correlation of the Grip Strength scores and the MRI data of the lean muscle area (non-fat) across the mid (central) dominant forearm over the 24-week ATL1102 treatment period. (TIF) S7 Fig. Table of Participant specific genetic variant within Dystrophin Gene. (DOCX) S1 File. (PDF) S2 File. (PDF) S3 File. (PDF) S4 File. (PDF) Acknowledgments The authors wish to acknowledge the contribution of Isabelle Ledoux and Simone Birnbaum of the Insitute of Myology, Paris, France who provided support with the quality control and analysis of the MyoSet functional outcome measures; Valeria Ricotti of the Dubowitz Neuro- muscular Centre, London UK, who provided support on the comparative analysis of the MRI observations; and Annabell Leske and Vicky Beal and the team at Avance Clinical, Adelaide, Australia. PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 14 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD Author Contributions Conceptualization: George Tachas, Peter J. Houweling, Monique M. Ryan. Data curation: Ian R. Woodcock, Peter J. Houweling, Michael Kean, Jaiman Emmanuel. Formal analysis: Ian R. Woodcock, Nuket Desem, Jaiman Emmanuel, Chantal Coles, Chrystal Tiong, Peter Button, Jean-Yves Hogrel. Funding acquisition: George Tachas. Methodology: Ian R. Woodcock, George Tachas, Peter J. Houweling, Chrystal Tiong, Moni- que M. Ryan. Project administration: Ian R. Woodcock, Nuket Desem, Peter J. Houweling, Rachel Ken- nedy, Kate Carroll, Katy de Valle, Sarah Catling-Seyffer. Resources: George Tachas, Shireen R. Lamande´. Software: Michael Kean, Peter Button. Supervision: Peter J. Houweling, Kate Carroll, Shireen R. Lamande´, Daniella Villano, Moni- que M. Ryan, Martin B. Delatycki, Eppie M. Yiu. Writing – original draft: Ian R. Woodcock, George Tachas, Peter J. Houweling, Eppie M. Yiu. Writing – review & editing: Ian R. Woodcock, George Tachas, Nuket Desem, Peter J. Hou- weling, Michael Kean, Jaiman Emmanuel, Rachel Kennedy, Kate Carroll, Katy de Valle, Jus- tine Adams, Shireen R. Lamande´, Chantal Coles, Chrystal Tiong, Matthew Burton, Daniella Villano, Peter Button, Jean-Yves Hogrel, Sarah Catling-Seyffer, Monique M. Ryan, Martin B. Delatycki, Eppie M. Yiu. References 1. Crisafulli S, Sultana J, Fontana A, Salvo F, Messina S, Trifiro G. Global epidemiology of Duchenne mus- cular dystrophy: an updated systematic review and meta-analysis. Orphanet J Rare Dis. 2020; 15 (1):141. https://doi.org/10.1186/s13023-020-01430-8 PMID: 32503598 2. Landfeldt E, Thompson R, Sejersen T, McMillan HJ, Kirschner J, Lochmuller H. Life expectancy at birth in Duchenne muscular dystrophy: a systematic review and meta-analysis. Eur J Epidemiol. 2020; 35 (7):643–53. https://doi.org/10.1007/s10654-020-00613-8 PMID: 32107739 3. Bello L, Gordish-Dressman H, Morgenroth LP, Henricson EK, Duong T, Hoffman EP, et al. Prednisone/ prednisolone and deflazacort regimens in the CINRG Duchenne Natural History Study. Neurology. 2015; 85(12):1048–55. https://doi.org/10.1212/WNL.0000000000001950 PMID: 26311750 4. Bello L, Kesari A, Gordish-Dressman H, Cnaan A, Morgenroth LP, Punetha J, et al. Genetic modifiers of ambulation in the Cooperative International Neuromuscular Research Group Duchenne Natural History Study. Ann Neurol. 2015; 77(4):684–96. https://doi.org/10.1002/ana.24370 PMID: 25641372 5. Birnkrant DJ, Bushby K, Bann CM, Apkon SD, Blackwell A, Brumbaugh D, et al. Diagnosis and manage- ment of Duchenne muscular dystrophy, part 1: diagnosis, and neuromuscular, rehabilitation, endocrine, and gastrointestinal and nutritional management. Lancet Neurol. 2018; 17(3):251–67. https://doi.org/ 10.1016/S1474-4422(18)30024-3 PMID: 29395989 6. Rosenberg AS, Puig M, Nagaraju K, Hoffman EP, Villalta SA, Rao VA, et al. Immune-mediated pathol- ogy in Duchenne muscular dystrophy. Sci Transl Med. 2015; 7(299):299rv4. https://doi.org/10.1126/ scitranslmed.aaa7322 PMID: 26246170 7. Zanotti S, Gibertini S, Di Blasi C, Cappelletti C, Bernasconi P, Mantegazza R, et al. Osteopontin is highly expressed in severely dystrophic muscle and seems to play a role in muscle regeneration and fibrosis. Histopathology. 2011; 59(6):1215–28. https://doi.org/10.1111/j.1365-2559.2011.04051.x PMID: 22175901 8. Pinto-Mariz F, Carvalho LR, de Mello W, Araujo Ade Q, Ribeiro MG, Cunha Mdo C, et al. Differential integrin expression by T lymphocytes: potential role in DMD muscle damage. J Neuroimmunol. 2010; 223(1–2):128–30. https://doi.org/10.1016/j.jneuroim.2010.03.006 PMID: 20382434 PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 15 / 16 PLOS ONEPhase 2 clinical trial of ATL1102 in non-ambulatory DMD 9. Pinto-Mariz F, Rodrigues Carvalho L, Prufer De Queiroz Campos Araujo A, De Mello W, Goncalves Ribeiro M, Cunha Mdo C, et al. CD49d is a disease progression biomarker and a potential target for immunotherapy in Duchenne muscular dystrophy. Skelet Muscle. 2015; 5:45. https://doi.org/10.1186/ s13395-015-0066-2 PMID: 26664665 10. 11. Limmroth V, Barkhof F, Desem N, Diamond MP, Tachas G, Group ATLS. CD49d antisense drug ATL1102 reduces disease activity in patients with relapsing-remitting MS. Neurology. 2014; 83 (20):1780–8. https://doi.org/10.1212/WNL.0000000000000926 PMID: 25239835 Lagrota-Candido J, Canella I, Savino W, Quirico-Santos T. Expression of extracellular matrix ligands and receptors in the muscular tissue and draining lymph nodes of mdx dystrophic mice. Clin Immunol. 1999; 93(2):143–51. https://doi.org/10.1006/clim.1999.4749 PMID: 10527690 12. Hogarth MW, Houweling PJ, Thomas KC, Gordish-Dressman H, Bello L, Cooperative International Neuromuscular Research G, et al. Evidence for ACTN3 as a genetic modifier of Duchenne muscular dystrophy. Nat Commun. 2017; 8:14143. https://doi.org/10.1038/ncomms14143 PMID: 28139640 13. Seferian AM, Moraux A, Annoussamy M, Canal A, Decostre V, Diebate O, et al. Upper limb strength and function changes during a one-year follow-up in non-ambulant patients with Duchenne Muscular Dystrophy: an observational multicenter trial. Plos One. 2015; 10(2):e0113999. https://doi.org/10.1371/ journal.pone.0113999 PMID: 25643053 14. Servais L, Deconinck N, Moraux A, Benali M, Canal A, Van Parys F, et al. Innovative methods to assess upper limb strength and function in non-ambulant Duchenne patients. Neuromuscul Disord. 2013; 23 (2):139–48. https://doi.org/10.1016/j.nmd.2012.10.022 PMID: 23219352 15. Hogrel JY, Wary C, Moraux A, Azzabou N, Decostre V, Ollivier G, et al. Longitudinal functional and NMR assessment of upper limbs in Duchenne muscular dystrophy. Neurology. 2016; 86(11):1022–30. https://doi.org/10.1212/WNL.0000000000002464 PMID: 26888987 16. Mercuri E, Pichiecchio A, Counsell S, Allsop J, Cini C, Jungbluth H, et al. A short protocol for muscle MRI in children with muscular dystrophies. Eur J Paediatr Neurol. 2002; 6(6):305–7. https://doi.org/10. 1016/s1090-3798(02)90617-3 PMID: 12401454 17. Mercuri E, Pichiecchio A, Allsop J, Messina S, Pane M, Muntoni F. Muscle MRI in inherited neuromus- cular disorders: past, present, and future. J Magn Reson Imaging. 2007; 25(2):433–40. https://doi.org/ 10.1002/jmri.20804 PMID: 17260395 18. Fischer D, Kley RA, Strach K, Meyer C, Sommer T, Eger K, et al. Distinct muscle imaging patterns in myofibrillar myopathies. Neurology. 2008; 71(10):758–65. https://doi.org/10.1212/01.wnl.0000324927. 28817.9b PMID: 18765652 19. Ricotti V, Evans MR, Sinclair CD, Butler JW, Ridout DA, Hogrel JY, et al. Upper Limb Evaluation in Duchenne Muscular Dystrophy: Fat-Water Quantification by MRI, Muscle Force and Function Define Endpoints for Clinical Trials. Plos One. 2016; 11(9):e0162542. https://doi.org/10.1371/journal.pone. 0162542 PMID: 27649492 20. Tachas G, Desem N, Button P, Pane E, and Mercuri E, World Muscle Society 2020, P.284 ATL1102 treatment improves PUL2.0 in non-ambulant boys with Duchenne muscular dystrophy compared to a natural history control. Neuromuscular Disorders Volume 30, Supplement 1, S129–S130 October 1, 2020. 21. Pane M, Coratti G, Brogna C, Mazzone ES, Mayhew A, Fanelli L, et al. Upper limb function in Duchenne muscular dystrophy: 24 month longitudinal data. Plos One. 2018; 13(6):e0199223. https://doi.org/10. 1371/journal.pone.0199223 PMID: 29924848 PLOS ONE \| https://doi.org/10.1371/journal.pone.0294847 January 25, 2024 16 / 16 PLOS ONE	null
10.1107_s2052252522010612.pdf	null	null	research papers IUCrJ ISSN 2052-2525 CRYOjEM Received 17 June 2022 Accepted 3 November 2022 Edited by L. A. Passmore, MRC Laboratory of Molecular Biology, United Kingdom Keywords: cryoEM; automated cryoEM data collection; computer vision; microscope automation software; machine learning; deep learning; automation; single-particle cryoEM. Supporting information: this article has supporting information at www.iucrj.org Published under a CC BY 4.0 licence Learning to automate cryo-electron microscopy data collection with Ptolemy Paul T. Kim,a Alex J. Noble,a Anchi Chengb and Tristan Beplera* aSimons Machine Learning Center, Simons Electron Microscopy Center, New York Structural Biology Center, New York, NY USA, and bSimons Electron Microscopy Center, New York Structural Biology Center, New York, NY USA. *Correspondence e-mail: [email protected] Over the past decade, cryo-electron microscopy (cryoEM) has emerged as an important method for determining near-native, near-atomic resolution 3D structures of biological macromolecules. To meet the increasing demand for cryoEM, automated methods that improve throughput and efﬁciency of microscope operation are needed. Currently, the targeting algorithms provided by most data-collection software require time-consuming manual tuning of parameters for each grid, and, in some cases, operators must select targets completely manually. However, the development of fully automated targeting algorithms is non-trivial, because images often have low signal-to-noise ratios and optimal targeting strategies depend on a range of experimental parameters and macromolecule behaviors that vary between projects and collection sessions. To address this, Ptolemy provides a pipeline to automate low- and medium-magniﬁcation targeting using a suite of purpose-built computer vision and machine-learning algorithms, including mixture models, convolutional neural networks and U-Nets. Learned models in this pipeline are trained on a large set of images from real-world cryoEM data-collection sessions, labeled with locations selected by human operators. These models accurately detect and classify regions of interest in low- and medium-magniﬁcation images, and generalize to unseen sessions, as well as to images collected on different microscopes at another facility. This open-source, modular pipeline can be integrated with existing microscope control software to enable automation of cryoEM data collection and can serve as a foundation for future cryoEM automation software. 1. Introduction Cryo-electron microscopy (cryoEM) is a rapidly growing method for determining the structure of proteins in near- native conformations at high resolution (Bai et al., 2015). CryoEM structure determination typically starts with the application of a solution containing puriﬁed protein to an EM grid, a holey substrate supported by a thin metal mesh (Cheng et al., 2015). The sample droplet is then reduced to a thin liquid ﬁlm and the grid is plunged into a cryogen, converting the thin ﬁlm to a layer of vitriﬁed ice (Egelman, 2016). The grid is then transferred to a transmission electron microscope (TEM) to collect high-magniﬁcation (high-mag) micrographs of the particles suspended in vitreous ice within the holes. Vitreous ice containing particles is found in windows in the grid termed ‘squares’ [Fig. 1(a)]. Within these squares are circular ‘holes’ [Fig. 1(b)] and particle images are obtained by taking high- resolution micrographs of the ice within these holes [Fig. 1(c)]. Each micrograph will typically provide numerous individual 2D projections of the protein particles, and these images can be processed to produce a three-dimensional map of the protein of interest (Wu & Lander, 2020). Solving a protein structure to high resolution usually requires tens to many 90 https://doi.org/10.1107/S2052252522010612 IUCrJ (2023). 10, 90–102 research papers Figure 1 Example square, hole and highest-magniﬁcation exposure. Windows in the cryoEM grid containing ice are known as ‘squares’. Within squares are circular ‘holes’. Highest-magniﬁcation exposures are taken within holes. Ideal highest-magniﬁcation exposures contain particles (as shown) and have thin ice. thousands of hundreds of individual randomly oriented particle projection images which often requires collecting many thousands of high-quality high-resolution micrographs. Because EM grid preparation is not a well controlled process, the locations where highest-magniﬁcation data are to be collected must be identiﬁed from a series of successively increasing magniﬁcation images (Chua et al., 2022). The process of collecting high-magniﬁcation data begins by taking low-magniﬁcation images of the grid [Fig. 2(a)], typi- cally acquired at a pixel size of (cid:2)200–500 nm pixel(cid:3)1. Squares are selected from these images, and medium-magniﬁcation images [Fig. 2(b)] with a pixel size of (cid:2)10–100 nm pixel(cid:3)1 are taken within these squares. Holes and subsequent high- Figure 2 Example low- and medium-magniﬁcation images: carbon grid (left) and gold grid (right). First, low-magniﬁcation images of the grid, where squares are visible, are acquired at a pixel size of (cid:2)200–500 nm pixel(cid:3)1. Subsequently, medium-magniﬁcation images are acquired at a pixel size of (cid:2)10–100 nm pixel(cid:3)1 by imaging the regions inside the squares. magniﬁcation collection locations are identiﬁed from the medium-magniﬁcation images. Not all squares or holes will be suitable for collection; the goal is to identify squares and holes in the grid with vitreous ice of suitable quality, ice that is the right thickness (typically slightly thicker than the largest diameter of the particle) and that contains a reasonable number of particles, ideally oriented in a range of angles (Noble et al., 2018). Ultimately, the success of a data collection is determined by the quality of the resulting 3D reconstruc- tion, which is a function of the number of particles found, the range of orientation angles present among the 2D projections, and the maximum resolution and signal-to-noise ratio (SNR) of the micrographs. Automated data-collection software such as Leginon (Carragher et al., 2000; Suloway et al., 2005), SerialEM (Mastronarde, 2005) or EPU (made by ThermoFisher Scien- tiﬁc) provide several built-in tools for identifying potentially promising squares and holes. These tools include template-, correlation- or feature clustering-based image analysis algo- rithms and automated selection capabilities. However, none of these tools generalize out-of-the-box across the wide variety of grids that are encountered in practice. They can struggle to both detect holes and squares under contaminated or low- SNR conditions and to reliably prioritize good collection locations across different macromolecule specimens. This means that microscope operators must often manually identify squares in low-magniﬁcation images and tune parameters used for automated targeting of holes in medium-magniﬁcation images (https://em-learning.com). Additionally, existing hole targeting algorithms fail on a non-trivial percentage of cases, especially on noisy, contaminated or carbon grids with minimal contrast variation between the holes and substrate. Operators are then required to manually target holes in medium-magniﬁcation images which is a labor-intensive task. The human operator time required for targeting limits the efﬁciency of collection on expensive and over-subscribed cryo- transmission electron microscopes (cryo-TEMs). Additionally, with increasing detector speeds, automated targeting would IUCrJ (2023). 10, 90–102 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy 91 research papers allow for better utilization of microscope time that is other- wise wasted waiting for human input. Thus, fully automated targeting methods are needed to not only reduce the burden on human operators, but also to increase access and throughput for the entire scientiﬁc research community. However, human-free automation of cryoEM data collec- tion is challenging. Many types of EM grids exist, each with holes and squares of different shapes, sizes and spacings. The grids themselves are made from different materials (e.g. carbon or gold), which causes the resulting low- and medium- magniﬁcation images to have very different properties (Fig. S1 of the supporting information). Carbon grids, for example, have signiﬁcantly less contrast between collection regions of interest (ROIs) in holes and background substrate at medium magniﬁcation compared with gold grids (Fig. S1). This is further complicated by variable sample preparation conditions leading to variable ice thickness and empty regions of the grid, along with deformations, contamination and lesions, all of which introduce visual artifacts (Fig. S1). In addition, cryoEM images have low SNR, especially at high magniﬁcation, and microscope parameters such as electron beam dose (often between 40 and 80 e(cid:3) A˚ (cid:3)2) can signiﬁcantly alter image properties (Cheng et al., 2015). Furthermore, images at each magniﬁcation level may contain many collection ROIs or none at all (Lyumkis, 2019). Finally, even if good candidate collec- tion ROIs are found, it is challenging to ﬁnd ice containing many particles with enough diversity of projection orienta- tions to produce high-quality 3D reconstructions. In recent years, machine-learning techniques have trans- formed single-particle cryoEM data analysis. Tools such as MicAssess (Li et al., 2020) and MicrographCleaner (Sanchez- Garcia et al., 2020) allow for efﬁcient post-processing of high- resolution micrographs collected, whereas others such as Topaz (Bepler et al., 2019), CASSPER (George et al., 2021), Warp (Tegunov & Cramer, 2019) and crYOLO (Wagner et al., 2019) use deep learning to automate particle picking and image denoising. Machine-learning-based 3D reconstruction algorithms have also emerged, including cryoSPARC (Punjani et al., 2017) and CryoDRGN (Zhong et al., 2021). These methods have signiﬁcantly improved our ability to analyze high-resolution cryoEM data quickly and thoroughly after collection. However, comparatively little attention has been given to accelerating or automating data collection itself. Yokoyama et al. (2020) recently introduced a machine- learning method for detection and classiﬁcation of ROIs in medium-magniﬁcation images, but it requires retraining a model with an annotated medium-magniﬁcation image dataset for each data-collection session. To address these challenges in cryoEM data collection, we present Ptolemy, a pipeline that uses computer vision algo- rithms and pre-trained convolutional neural networks (CNNs) to navigate cryoEM grids at low and medium magniﬁcation and determine high-quality targeting locations without human input. We train the Ptolemy models on large datasets of low- and medium-magniﬁcation images with corresponding collection locations selected by operators from 55 different data-collection sessions. These sessions include carbon and Figure 3 Pipeline overview. High-magniﬁcation images are taken from holes in medium-magniﬁcation images, which come from squares in low- magniﬁcation images. (a) Ptolemy detects, crops and then classiﬁes squares in low-magniﬁcation images, which have a pixel size of approximately 200–500 nm pixel(cid:3)1. (b) Next, Ptolemy detects, crops and then classiﬁes holes in subsequent medium-magniﬁcation images, which have a pixel size of approximately 10–100 nm pixel(cid:3)1. gold holey grids and feature a variety of proteins, grid conditions, magniﬁcations and electron beam dosages. Rather than attempting to learn separate models for different grid types or for different particles, we develop a single uniﬁed pipeline to localize and classify ROIs to approximate user selection locations in low- and medium-magniﬁcation cryoEM images (Fig. 3). We demonstrate that Ptolemy can effectively detect and classify squares and holes in low- and medium-magniﬁcation images. We evaluate these predictions with comparison against operator-selected locations, while noting that opera- tors target incompletely. We validate the models by holding- out entire data-collection sessions to conﬁrm that the models generalize well to unseen sessions. Additionally, we compare our medium-magniﬁcation localization algorithm to an existing method (Yokoyama et al., 2020) that performs this is medium-magniﬁcation localization (in that study, referred to as low-magniﬁcation localization) and show that our method yields superior generalization performance. Finally, a separate companion paper has been published: ‘Fully automated multi-grid CryoEM screening using Smart Leginon’, where the utility of Ptolemy in real-world collection cases is demonstrated and analyzed (Cheng et al., 2022). The Ptolemy source code is freely available for academic use at (https://github.com/SMLC-NYSBC/ptolemy) under CC BY-NC 4.0 license. Ptolemy is designed to be modular and to integrate directly with existing microscope control software. More information on Ptolemy can be found in Appendix A. 2. Methods To automate microscope targeting for single-particle cryoEM data collection, we divide the problem into four sub-problems: (1) low-magniﬁcation square localization, (2) low-magniﬁca- 92 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy IUCrJ (2023). 10, 90–102 tion square classiﬁcation, (3) medium-magniﬁcation hole localization and (4) medium-magniﬁcation hole classiﬁcation. For low- and medium-magniﬁcation localization, the goal is to identify all possible collection ROIs: squares in low magniﬁcation and holes in medium magniﬁcation. The cropped ROIs are then fed into separate classiﬁcation models at each level that determine whether these ROIs should be collected. Low-magniﬁcation localization is solved by pixel- wise image segmentation using a mixture model, whereas medium-magniﬁcation localization is solved using a U-Net with a novel lattice-ﬁtting algorithm (Gupta & Sortrakul, 1998; Ronneberger et al., 2015). Classiﬁcation at low magniﬁcation is achieved using a feedforward CNN whereas medium-magni- ﬁcation classiﬁcation uses the U-Net localization probabilities because they outperformed a separate downstream classiﬁer in our experiments. Training and hyperparameter information for all trained models can be found in Appendix B. 2.1. Datasets and splits The data used to train and validate all models and algo- rithms come from 55 cryoEM data-collection sessions the Simons Electron Microscopy Center performed at (SEMC), a center within the New York Structural Biology Center (NYSBC), from 2018 to 2021. The sessions include gold and carbon grids, featuring regularly spaced and lacey holes, with tilted and untilted collection. Lacey-hole grids and tilted collection images were only used for low-magniﬁcation (square) localization and classiﬁcation. Positive labels repre- sent targeted selection locations used in these sessions. All labeled squares are manual operator selections on the low- magniﬁcation images. On the other hand, labeled holes on the medium-magniﬁcation images are generally automated selec- tions optimized by operators using template-correlation-based hole localization and ice-thickness-based hole classiﬁcation. A minority of labeled holes were also manually selected by operators. Data-collection sessions generally involve different samples, preparation methods and microscope settings (Weissenberger et al., 2021). This results in considerable variation between sessions in the appearance of collection locations especially at medium magniﬁcation, as well as in the characteristics that make for good collection locations (Fig. S1). Therefore, to ensure that our models can generalize to research papers unseen data-collection sessions with different experimental parameters, we primarily use session splits, where a set of sessions (termed ‘held-out sessions’) are withheld from the dataset used to train the models (training set). All perfor- mance metrics are reported from results on these held-out sessions. 2.2. Square localization The goal of square localization is to locate all squares (windows in the grid that may contain imageable ice) in low- magniﬁcation grid images. The input is a low-magniﬁcation image, and the output is a set of rectangular boxes tightly bounding the squares (Fig. 4). We ﬁnd these boxes using a mixture model-based image segmentation algorithm followed by a geometric algorithm for identifying the aligned minimum bounding rectangles surrounding each square. Pixels in the image are ﬁrst separated into two classes based on pixel intensity using a Poisson mixture model [Fig. 4(b)] (Forbes, 2018). Mixture model separation works because the distribu- tion of pixels in the image can be accurately decomposed into low-intensity pixels coming from the thick grid bars in the surrounding background and higher-intensity pixels coming from the much thinner squares (Fig. S2). This approach avoids the need for a user to set a speciﬁc intensity threshold for identifying squares, which changes from session to session. Next, we apply a ﬂood ﬁlling algorithm to identify discrete regions from the segmented square pixels and then ﬁnd a minimum bounding convex polygon to bound the pixels in each square [Fig. 4(c)]. Finally, we take advantage of the fact that the squares are axis-aligned to ﬁnd the angle (cid:2), for each low-magniﬁcation image where the minimum bounding rectangles aligned with (cid:2) bounding each minimum bounding convex polygon have the smallest total area. Formally, we seek i Ai; (cid:2) for N polygons, where Ai, (cid:2) is the area of the argmin(cid:2) minimum bounding rectangle around the ith polygon, aligned at angle (cid:2). We ﬁnd this angle (cid:2) using bounded optimization (Brent, 1973), and the resulting minimum bounding rectangles are used to obtain aligned crops of the squares in the low- magniﬁcation image [Fig. 4(d)]. This algorithm is applied to 1304 low-magniﬁcation images, resulting in 41 000 crops of squares. PN Figure 4 Example of the square localization procedure. (a) Original input image. (b) Mask recovered after segmenting pixels. (c) Finding convex polygons around the separate regions in the mask. (d) Aligned minimum bounding rectangles around the polygons, used for crops of the images. IUCrJ (2023). 10, 90–102 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy 93 research papers 2.3. Hole localization The goal of hole localization is to detect all hole locations in medium-magniﬁcation images. However, unlike in square localization, a mixture model-based segmentation approach intensities does not work, because the difference in pixel between the holes and the surrounding background is negli- gible, particularly for carbon grids. Some medium-magniﬁca- tion images even have ‘inverted’ holes where the pixel intensity within the hole is lower than the surrounding region (Fig. 5). Here, our choice of model is informed by the available data. Although we do not have a dataset of bounding boxes around holes, we do have a large dataset of 28k carbon and gold holey grid medium-magniﬁcation images with locations at or near the center of holes where operators collected high-magniﬁ- cation micrographs. Therefore, we seek to learn the hole centers in each pixel-normalized medium-magniﬁcation input image by training a U-Net model to output a map with the same dimensions as the input containing 1 at the locations where the operator collected and 0 everywhere else (Figs. 5 and 6). We choose a U-Net architecture, because the neurons in the bottleneck layer have large receptive ﬁelds, allowing them to capture needed context, while the output layers use the information propagated from the bottleneck, as well as high-resolution features, to ﬁnd the hole centers. The pixels in the medium-magniﬁcation image are normalized to control for variance in electron dose. Additionally, holes are known to lie on a regular square lattice, so we post-process the output of the U-Net to ﬁnd the best ﬁtting lattice. Given the lattice points in the image, we then crop around those points to extract hole images (Fig. 7). This helps to extend the predicted map from the U-Net to capture all holes in the image, not just the holes that the operators picked, while simultaneously cleaning erroneously detected regions. We ﬁnd the lattice from the U-Net output map by searching pairs of candidate anchor points and selecting the pair for which the lattice produced by these anchor points has the smallest pixelwise error against the output map. We ﬁnd argmina; b XN i (cid:3)1 oi (cid:3) li ð ð Þ 1 (cid:3) li ð Þ þ (cid:3)2 li (cid:3) oi Þ lið Þ; oi 2 O; li 2 La; b; where O is the output of the U-Net, N is the number of pixels in the image, La, b is the lattice generated by anchor points a and b, and (cid:3)1 and (cid:3)2 allow us to independently weight the cost for false positives and false negatives. Candidate anchor point pairs are found by taking centroids of high-probability regions in the U-Net output map, and for each centroid, pairing with Figure 5 Example medium-magniﬁcation images. Human operator-selected locations are marked in yellow. The input to the U-Net is the image without any marking for the selection locations, and the output is a map with a 1 at each pixel where a selection was made and 0 elsewhere. In rare cases (right), holes can be darker than the surrounding region. Figure 6 Example hole-center detection without lattice ﬁtting. For some cases, taking the centroids of the high-probability regions predicted by the U-Net is sufﬁcient. (a) Input image. (b) U-Net output. (c) Centroids from high-probability regions in U-Net output (red). 94 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy IUCrJ (2023). 10, 90–102 research papers Figure 7 Example hole-center detection on a difﬁcult image. On this low-contrast image, lattice ﬁtting extends to missed holes while removing erroneous detections. (a) Input image. (b) U-Net output. (c) Centroids from high-probability regions in U-Net output (red). Many locations outside holes are detected incorrectly, and one hole (orange) is missed. (d) Running the optimal-lattice-ﬁnding algorithm results in ﬁnding lattice anchor points (cyan). (e) The lattice generated by these anchor points (cyan) results in coverage of all holes and cleans the incorrect detections. the K closest to other centroids. K trades performance for run time. Here we use K = 6. Imposition of a regular square lattice over the U-Net output causes Ptolemy to be unable to handle tilted data collection because the tilt causes the hole centers to lie on a parallelo- gram-shaped (oblique) lattice rather than a regular square lattice. Support for tilted data collection will be added to a near-future update of Ptolemy. To improve training, we apply Gaussian blur to the model output before computing the loss (Shorten & Khoshgoftaar, 2019). This helps, because the exact location the operator selects in a hole is noisy: the selection location that is near the center of the hole but there is often deviation from the exact center pixel, and the direction and magnitude of displacement from the center varies between medium-magniﬁcation images. Therefore, this smoothing allows the model to learn the centers of these holes, rather than having to learn the displa- cement from the center for every hole image. We also perform gradient descent on the sigma parameter of the Gaussian blur simultaneously with training the U-Net weights to allow the model to learn the optimal level of smoothing over training time (Fig. S3). To improve generalization, we apply both random 90(cid:5) rotation augmentation to the images during training as well as random inversion of the normalized pixels. Inversion of pixels is helpful, because for some sessions, particularly with carbon grids, the pixels in the holes are darker than the background pixels. Although pixel inversion augmentation allows for better carbon grid hole targeting, it does not affect gold grid images which do not suffer from contrast inversions. 2.4. Square and hole classification models In square and hole classiﬁcation, we aim to obtain rankings of squares and holes in images to prioritize the ordering with which they are targeted. Although there are many possible parameters that may be important for determining whether a square or hole contains high-quality particles, experienced operators are able to consistently ﬁnd good locations, suggesting that at least some features of good target locations are identiﬁable in low- and medium-magniﬁcation images. Therefore, for each magniﬁcation we train a separate CNN to classify squares or holes as collected or not collected by operators. The input to our model is a cropped image of a square or hole, extracted using the square-localization method or hole-localization method above, and the output is a scalar probability. 2.4.1. Square classification. We train our square classiﬁca- tion model on a dataset containing 41k square crops, of which 11k were squares collected. Square images are normalized based on the intensity of all pixels within the bounding boxes for the squares in each low-magniﬁcation image to control for electron dose. We also include random forest (RF) and logistic regression (LR) models trained on summary statistics of the pixels extracted from the square image crops. This is because the operators typically use characteristics like the size/area and brightness of the squares to make their selections. Therefore, we include baselines which reﬂect this knowledge. The summary statistics used as features are mean intensity, maximum intensity, minimum intensity, variance in intensity, kurtosis, skew and crop area. IUCrJ (2023). 10, 90–102 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy 95 research papers 2.4.2. Hole classification. We compare the summed pixel- wise probabilities output by our localization U-net within each hole against two CNNs trained on a dataset containing 571k hole crops, of which 410k were targeted by operators. The dimensions of the holes and, therefore, the dimensions of the resulting crops vary widely between data-collection sessions. However, we do not want the model to use the size of the input image to decide if a hole is good or bad. We hypothesize that the location of image features within each hole (e.g. crystalline ice) is not important for classiﬁcation. Rather, the presence, absence or proportion of these features is the main concern. Therefore, we compare between a standard CNN model that pads all input images to the same dimension and one which averages over non-channel dimensions of the ﬁnal map before the fully connected layer, thereby treating the image as a bag of regions. Both CNNs normalize images based on pixels in the crop to control for difference in electron dose. Aside from the difference in padding versus average pooling, all other hyperparameters of the two models are identical. 3. Results and discussion Ptolemy can accurately locate and rank collection locations within low- and medium-magniﬁcation images, with each stage producing good performance metrics, and results that appear inspection. Furthermore, Ptolemy reasonable on visual generalizes well to new sessions without user intervention or retraining on a session-by-session basis. 3.1. Ptolemy localization of squares in low-magnification images The square localization algorithm successfully detects almost all operator-selected locations, as well as squares that were not collected, with few errors. Our algorithm successfully detects 98.8% of operator-selected locations (Table 1). An additional 30k unselected squares are detected, and visual inspection conﬁrms that these are real squares that were not selected by the operator (Fig. 8). Table 1 Statistics from running the square localization procedure on low- magniﬁcation images. No. of operator-selected locations No. of operator-selected locations detected Total no. of locations detected 10993 10857 41301 Table 2 Performance metrics of different ML models on the square classiﬁcation task measured on held-out sessions and held-out square images. We compare the performance of LR, an RF and a CNN based on two metrics. ROC AUC: area under receiver-operating characteristic curve. Session split Random split Model ROC AUC Average precision ROC AUC Average precision 0.539 LR RF–5 feature 0.603 0.608 CNN 0.258 0.344 0.331 0.499 0.867 0.733 0.259 0.734 0.489 3.2. Square classifier learns to rank squares effectively For square classiﬁcation, we explored three different models: an LR and RF on summary statistics (details in Appendix C) extracted from square images, and a CNN on the images themselves. Both the RF and the CNN perform simi- larly on this task. 3.2.1. Classifying squares for new sessions without prior knowledge. Square ranking without any information about the sample is a challenging task, but our models perform well and are signiﬁcantly better than random guessing (Table 2). The task of ranking squares given the session split is particularly difﬁcult, because the characteristics that make up good squares can vary from session to session. For optimum performance in this setting, a model would need to extract enough information from the square images to identify the optimal sample conditions and then predict the quality of the square accordingly. Although this is probably not possible for a classiﬁer such as ours, which makes predictions for each square independently, we hypothesize that there are, at least, characteristics of squares that are always bad and maybe squares that are always good that can still provide useful Figure 8 Example square localizations on low-magniﬁcation images. Ptolemy successfully segments bounding boxes around squares in the grid without human intervention. 96 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy IUCrJ (2023). 10, 90–102 research papers Figure 9 Example square classiﬁcations in low-magniﬁcation images. The model successfully prioritizes larger, brighter squares without cracks or contamination. Model-predicted probabilities for squares are in red. Colors from high to low score: dark blue, light blue, white, yellow, orange, red. one-to-one with an operator-collected hole, a false negative as an operator-collected hole that contains zero or more than one (2+) model selection locations, and a false positive as a model selection location that is not contained in any operator- collected holes, or is only contained in operator-collected holes that also contain other model selection locations. For more information on how we deﬁned ‘model selection loca- tions’ for each model in Table 3, see Appendix D. We ﬁnd our U-Net, without lattice ﬁtting, can learn operator-selected locations exceptionally well, and can iden- tify 98.4% of all operator-selected holes with 70.3% precision (Table 3, row 2). We compare this method with the Yolov5- based model trained by Yokoyama et al. (2020) on the same dataset and ﬁnd that the U-Net is superior in both recall and precision (Redmon & Farhadi, 2018). guidance. The difﬁculty of this task is evident in comparison with the performance of the models on random splits of the data, where without the difﬁculty of generalizing to unseen sessions, the RF and CNN models perform signiﬁcantly better (Table 2). Nonetheless, Ptolemy ranks squares signiﬁcantly better than random guessing even in the session split setting. On visual inspection of predictions on held-out sessions, we the CNN makes reasonable predictions, with ﬁnd that unbroken and larger squares prioritized over smaller, broken squares (Fig. 9). Additional example images with both model scores and user selection locations can be found in Fig. S4. 3.2.2. Simple extracted features exhibit good performance on square classification. The RF performing comparably to the CNN on an image classiﬁcation task is a surprising result for which there are several possible explanations. First, examining the feature importance of the extracted features for the RF models, using feature permutation, shows that area and mean pixel intensity are the most important features for predicting whether a square was selected (Fig. 10). This result aligns with our expectations, as operators usually use area and brightness of squares as primary criteria for selection. We hypothesize that the importance of area as a feature may partly explain the good performance of the RF relative to the CNN, as area may be a feature that is difﬁcult for a CNN to learn. Additionally, since our dataset is not exhaustively labeled, a more complex CNN model may be learning redundant, irrelevant features that do not generalize well from training set to test set. Nevertheless, the CNN does not require a burdensome computational cost (it runs very quickly on commodity CPU hardware), and likely has a higher potential performance if a larger and cleaner dataset is curated. We therefore use it as the default model for Ptolemy. 3.3. Ptolemy retrieves more holes with fewer false positives Next, we examine the performance of our methods for hole localization on medium-magniﬁcation images (Table 3). Since we do not have bounding box annotations for our dataset, we deﬁne a true positive as a model selection location that maps Figure 10 Feature importance for square classiﬁcation. Results obtained for feature importance of square image summary statistics to predict whether a square is selected using an RF model. Area and mean pixel intensity are the most important features. IUCrJ (2023). 10, 90–102 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy 97 research papers Table 3 Performance metrics of different methods on held-out sessions for hole localization from medium-magniﬁcation images. Reported metrics are aggregated by session and averaged. Deﬁnitions of true positive, false positive and false negative used for computing precision, recall, and F1 can be found in Section 3.3 and Appendix D. Model Precision Recall F1 Yolov5 (from Yokoyama et al., 2020) U-Net U-Net + Lattice Fitting U-Net + Lattice Fitting + Probability Threshold 0.395 0.703 0.549 0.802 0.669 0.984 0.993 0.891 0.459 0.815 0.702 0.837 3.3.1. Lattice fitting reduces false negative rate. With the addition of lattice ﬁtting (Table 3, row 3), we reduce the false negative rate by a factor of 2, from 1.6% to 0.7%. Although the precision is also reduced with lattice ﬁtting, we are primarily concerned with improving the false negative rate (1- recall) as our aim is to recover all the holes at this stage. In the manual inspection, we ﬁnd that the U-Net with lattice ﬁtting correctly locates holes across diverse images (Fig. 11). Because many holes are not selected by the operators, we expect that, for the goal of detecting all holes (selected and not selected), lattice ﬁtting is helpful. Additionally, only keeping lattice tiles based on a probability taken from the U-Net output (Table 3, row 4) signiﬁcantly improves precision at the cost of recall. 3.3.2. U-Net + Lattice Fitting generalizes hole localization to an external dataset. To further test generalization, we predict holes in the dataset by Yokoyama et al. (2020) using our U-Net + Lattice Fitting model. Yokoyama et al. (2020) reported 95–97% recall of their own Yolov5 based model on these data. We ﬁnd that the U-Net + Lattice Fitting model generalizes well, achieving 0.685 precision, 0.950 recall and 0.796 F1 score. The images in this dataset were collected from an external facility, using a different microscope (Ptolemy data were collected primarily on a TFS Krios and a Glacios; this dataset was collected on a JEM-Z300CF) and with different magniﬁcation compared with the images in our training dataset (Yokoyama et al., 2020), demonstrating that Ptolemy generalizes effectively. 3.4. Ptolemy successfully classifies holes in medium- magnification images after localization Next, we examine the performance of our hole classiﬁcation models (Table 4). Both the padded model and the average- pool model perform well on the hole classiﬁcation task. In example images, we see that the model can effectively sepa- rate good, unblemished holes from those with blemishes and artifacts on both gold and carbon grids (Fig. 12). 3.4.1. Average-pooling improves hole classification. The average-pool model slightly outperforms the padded model, which supports our hypothesis that the location in the image where features occur is not as important in determining hole quality, and that given the wide variance of hole sizes, a model that uses average pooling is preferable to padding all crops to the same size. Table 4 Performance of hole classiﬁcation CNNs on hold-out sessions. ROC AUC: area under receiver operating characteristic curve. Model Accuracy ROC AUC Average precision 0.748 CNN (padding) CNN (average pool) 0.758 U-Net + Probability Threshold 0.846 0.742 0.796 0.868 0.808 0.878 0.867 3.4.2. Classifying holes using the localization U-Net output. We compare the dedicated CNN classiﬁers against the sum of the hole localization U-Net probabilities within a crop to determine the probability of picking a hole (Table 4, row 3). Surprisingly, the U-Net score outperforms the dedi- cated CNN classiﬁers in accuracy and ROC AUC. This is probably because the U-Net uses the context around the hole to help predict where the hole was collected from. The U-Net is deep to allow large holes to ﬁt entirely within the receptive ﬁeld of the bottleneck-layer neurons. This means that, for Figure 11 Examples of hole localization using the U-Net + Lattice Fitting. Ptolemy successfully detects all holes across a wide range of hole sizes, brightness and contrast conditions, from easily visible gold grids (bottom left) to very low contrast carbon grids that are difﬁcult to see even for humans (top right). 98 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy IUCrJ (2023). 10, 90–102 research papers Figure 12 Example hole classiﬁcations in medium-magniﬁcation images using the average-pooling model. Ptolemy can successfully prioritize complete holes free of contaminants. Model-predicted probabilities for holes are in red. Colors from high to low score: dark blue, light blue, white, yellow, orange, red. grids with smaller holes, the U-Net will have information about the location of the hole on the grid and the character- istics of nearby holes that our classiﬁers, which use only hole crops, lack. 3.5. Training and evaluating human operator selections One of the major challenges in developing Ptolemy is the lack of fully annotated data for training and assessment of model performance. We rely on training data composed of incomplete expert operator selections. These selections only represent an expert guess at the best collection locations and our models are trained to recapitulate these operator selection decisions as a surrogate for selecting high-quality data. Ideally, we would train and evaluate our models based on the true end goal of cryoEM data collection, particle quantity and quality, as determined by the resulting 3D structures, but these data are currently unavailable. Furthermore, the operators do not exhaustively select all possible viable collection locations. Since our model evalua- tion considers locations that the operator did not select as ground-truth negatives, the reported precision values are likely to underestimate the true precision of Ptolemy models. This lack of exhaustive selection also creates biases in the training data, especially on the hole and square classiﬁcation tasks, which may be learned by our models. For example, due to the microscope setup at the SEMC/NYSBC, holes near the edge of medium-magniﬁcation images are often not collected even though they may be viable collection locations. 3.6. Future work A large fraction (the bulk majority at NYSBC) of cryoEM data-collection experiments involve gold or carbon holey grids generated via blotting and collected without tilting. Devel- opment, training and validation of Ptolemy were carried out with the goal of providing a system for automated collection on these grids. However, this leaves a long tail of less used but signiﬁcant grid types and preparation methods for which future work is required. These include lacy grids, grids generated via spray or spotting methods, and tilted collection. Furthermore, Ptolemy currently lacks the ability to update its predictions dynamically to adjust to a given collection session. Ptolemy is designed to allow for modular improvement over time, so future updates of Ptolemy can build on the current set of algorithms to cover these cases as well. Tilted images can be corrected with basic image processing, as the angle of the tilt will be known in advance, and then the un-tilted images can be processed as normal. Grids generated by spray or spotting methods may need new purpose-built algorithms, but they may also be effectively handled by an on-the-ﬂy learning algorithm. Lacy grids will likely require the most customiza- tion, including separate segmentation and prioritization algorithms at the medium-magniﬁcation level. 4. Conclusions Increasing throughput and reducing cost through automation is necessary to meet increasing demand for cryoEM. In this work, we present Ptolemy, an open-source, modular package IUCrJ (2023). 10, 90–102 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy 99 research papers using images for automatic targeting and classiﬁcation of cryoEM low- and purpose-designed medium-magniﬁcation computer vision and deep-learning algorithms. Ptolemy accurately localizes and ranks squares and holes in low- and medium-magniﬁcation images across a wide range of image and sample conditions. By training on large datasets of sparse microscope operator selection locations, the Ptolemy locali- zation algorithms generalize to diverse gold and carbon holey grids and rank potential collection locations effectively without session-speciﬁc parameters, as we have demonstrated on held-out collection sessions within the SEMC/NYSBC dataset and on an independent dataset from another facility. Additionally, Ptolemy has been integrated with the micro- scope control software Leginon to enable automated data collection in real-life use cases (Cheng et al., 2022). Ptolemy is similar to the recent works SmartScope (Bouv- ette et al., 2022) and CryoRL (Li et al., 2022), which both offer automated collection capabilities. SmartScope focuses on automated screening and providing a helpful user interface to view and control screening in real time, whereas CryoRL focuses on optimizing the cryoEM data collection as a path- planning problem and does not clearly address the square- and hole-detection problem. Both primarily repurpose existing deep-learning object detection and classiﬁcation algorithms. Ptolemy, on the other hand, uses novel algorithms that are purpose-built for detection and classiﬁcation of holes and squares in cryoEM images. Although Ptolemy identiﬁes and generally ranks ROIs remarkably well, it is unable to incorporate session-speciﬁc information to reprioritize targets on-the-ﬂy. In future work, we plan to use the current Ptolemy classiﬁcation models as prior models, and dynamically update these prior models during each collection session based on the quality of highest- magniﬁcation exposures that are collected from explored squares and holes in an active-learning framework. Ptolemy is a signiﬁcant advance in the automation of cryoEM data collection, allowing for fully unattended data collection, and increasing microscope and operator efﬁciency. To accelerate cryoEM collection for the whole community, Ptolemy is open source and freely available for academic use at https://github.com/SMLC-NYSBC/ptolemy. We anticipate that Ptolemy will become an integral part of the data-collec- tion pipeline and will serve as the basis for future work in cryoEM automation. The supporting information ﬁle pw5021sup1.xlsx contains the metadata for the various data-collection sessions used to train and test the models in Ptolemy. APPENDIX A Ptolemy implementation details and default models Ptolemy is an operating system and microscope control soft- ware agnostic and requires only Python (3+) and a few installed packages to run. Package dependencies can be found in the github repository at https://github.com/SMLC-NYSBC/ ptolemy. By default, Ptolemy is conﬁgured to use the mixture model for low-magniﬁcation square detection, the CNN for low-magniﬁcation square classiﬁcation, the U-Net with grid ﬁtting for medium-magniﬁcation hole detection and the average-pooled CNN for medium-magniﬁcation hole classiﬁ- cation. It runs fully on CPU and requires a maximum of only 2.7 GB RAM to detect and classify squares in a low-magniﬁ- cation image and 3.6 GB RAM to detect and classify holes in a medium-magniﬁcation image. After installation, running Ptolemy during data collection involves the following steps. (1) Microscope control software captures tiled low-magni- ﬁcation mrc(s) of the grid and saves the ﬁle. (2) Call python lowmag_cli.py (path to low-magniﬁ- cation mrc). This returns a JSON containing a list of diction- aries with the ﬁelds (i) vertices corresponding to vertex coordinates of squares, (ii) center corresponding to the centers of squares, (iii) area corresponding to areas of squares, (iv) brightness corresponding to mean brightness of pixels in squares, (v) score corresponding to CNN predicted score (between 0 and 1) of squares. (3) Use the output coordinates and scores to select the collection location(s) to collect medium-magniﬁcation images. (4) Microscope control software captures medium-magni- ﬁcation mrc(s) within squares and saves to ﬁle. (5) Call python medmag_cli.py (path to medium- magniﬁcation mrc). This returns a JSON containing a list of dictionaries with the ﬁelds (i) vertices corresponding to vertex coordinates of hole crops, (ii) center corresponding to the centers of holes, (iii) area corresponding to areas of hole crops, (iv) score corresponding to CNN predicted score (between 0 and 1) of holes. (6) Use the output coordinates and scores to select the the highest-magniﬁcation collection location(s) to collect exposures. (7) Return to step (3) to explore new medium-magniﬁcation collection locations from existing low-magniﬁcation images, or step (1) to select new low-magniﬁcation collection locations. For an example of integrating Ptolemy into cryoEM data- collection software in practice, please refer to the companion ‘Fully automated multi-grid CryoEM paper to Ptolemy, screening using Smart Leginon’ (Cheng et al., 2022). APPENDIX B Training and hyperparameters All deep-learning and machine-learning models were trained using default hyperparameters in PyTorch (Paszke et al., 2019) and scikit-learn (Pedregosa et al., 2011) except where stated otherwise. All deep-learning model parameters were ﬁtted with the default Adam, except for the sigma parameter of the Gaussian smoothing, which was initialized to e and trained with an Adam optimizer with a learning rate of 0.1 (Kingma & Ba, 2017). Binary-cross-entropy loss was used for all deep- learning models, and for the U-Net a positive weight of 100 was applied. The square classiﬁcation CNN was trained for two epochs and used two 5 (cid:6) 5 convolutional layers followed by three 3 (cid:6) 3 convolutional layers, with 64 channels per layer and with a 100 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy IUCrJ (2023). 10, 90–102 batch size of 128, whereas the hole classiﬁcation CNN was trained for ﬁve epochs and used one 5 (cid:6) 5 convolutional layer, followed by three 3 (cid:6) 3 convolutional layers with 128 channels per layer, with a batch size of 32. Both models used batch normalization, max-pooling and ReLU activations (LeCun et al., 2015; Ioffe & Szegedy, 2015). The U-Net for hole localization was trained for 6000 steps and used nine down-blocks and nine up-blocks, with each down-block and up-block using a 64 channel 3 (cid:6) 3 kernel convolutional layer. The model also used ReLU activations for down- and up-blocks, max-pooling for downsampling in down-blocks, and nearest interpolation for upsampling in up- blocks. Batch norm was not used. Also, the bias of the ﬁnal convolutional layer which produces the ﬁnal output of the model was initialized at (cid:3)10 to allow the model to initially predict all or mostly zeros in the output, since the target image contains zeros everywhere except for the few pixel locations where the operator made a selection. Random 90(cid:5) rotation augmentation is applied while training square and hole classiﬁcation CNNs. Random 90(cid:5) rotation augmentation is combined with random pixel inversion when training the U-Net for hole localization. APPENDIX C Random forest and logistic regression model details The RF and LR models were trained on the following features for the squares: mean pixel intensity, maximum pixel intensity, minimum pixel intensity, variance of pixel intensities, skew of pixel intensities, kurtosis of pixel intensities and area. Default hyperparameters from scikit-learn were used for both models. APPENDIX D Definition of ‘predicted collected regions’ for each model For U-Net + Lattice Fitting, crops are generated by creating squares centered at each lattice point with a side length equal to dl (cid:3) 60, where dl is the distance between lattice points. All crops are considered predicted collected regions. For U-Net alone, circles with a radius of 50 pixels around each centroid of high-probability regions in the U-Net output map are considered predicted collected regions. For U-Net + Lattice Fitting + Probability Threshold, we generated crops like in U-Net + Lattice Fitting, but then only kept the crops where the sum of pixel probabilities outputted by the U-Net within the crop was greater than 0.5. For the Yolov5 model in Yoneo-Locr, which outputs many bounding boxes at different conﬁdence levels, we had to decide how to set conﬁdence thresholds which determine the bounding boxes that are kept. We aimed to be generous to the model by picking the conﬁdence threshold that gave the maximum F1 score for each image independently and keeping all bounding boxes in that image that were predicted with conﬁdence greater than this threshold. research papers Acknowledgements We thank the SEMC Electron Microscopy Staff for their help in testing algorithms and models, in particular Mahira Aragon, Eugene Chua, Huihui Kuang, Kashyap Maruthi, Joshua Mendez, Anjelique Sawh and Hui Wei. Additionally, we thank Tohru Terada for providing us with hole localization data from Yokoyama et al. (2020). Finally, we thank members of the SEMC at NYSBC, in particular Bridget Carragher and Clint Potter, for useful discussions. Funding information This work was performed at the Simons Electron Microscopy Center located at the New York Structural Biology Center, supported by grants from the Simons Foundation (grant No. SF349247) and NIH National Institute of General Medical Sciences (grant No. GM103310). References Bai, X., McMullan, G. & Scheres, S. H. W. (2015). Trends Biochem. Sci. 40, 49–57. Bepler, T., Morin, A., Rapp, M., Brasch, J., Shapiro, L., Noble, A. J. & Berger, B. (2019). Nat. Methods, 16, 1153–1160. Bouvette, J., Huang, Q., Riccio, A. A., Copeland, W. C., Bartesaghi, A. & Borgnia, M. J. (2022). eLife, 11, e80047. Brent, R. P. (1973). Algorithms for Minimization Without Derivatives, 1st ed. Englewood Cliffs: Prentice-Hall. Carragher, B., Kisseberth, N., Kriegman, D., Milligan, R. A., Potter, C. S., Pulokas, J. & Reilein, A. (2000). J. Struct. Biol. 132, 33–45. Cheng, A., Kim, P. T., Kuang, H., Mendez, J. H., Chua, E. Y. D., Maruthi, K., Wei, H., Sawh, A., Aragon, M. F., Serbynovskyi, V., Neselu, K., Eng, E. T., Potter, C. P., Carragher, B., Bepler, T. & Noble, A. J. (2022). IUCrJ, 9, 77–89. Cheng, Y., Grigorieff, N., Penczek, P. A. & Walz, T. (2015). Cell, 161, 438–449. Chua, E. Y. D., Mendez, J. H., Rapp, M., Ilca, S. L., Tan, Y. Z., Maruthi, K., Kuang, H., Zimanyi, C. M., Cheng, A., Eng, E. T., Noble, A. J., Potter, C. S. & Carragher, B. (2022). Annu. Rev. Biochem. 91, 1–32. Egelman, E. H. (2016). Biophys. J. 110, 1008–1012. Forbes, F. (2018). Mixture Models for Image Analysis, edited by S. Fruhwirth-Schnatter, G. Celeux & C. P. Robert, pp. 397–418. New York: CRC press. George, B., Assaiya, A., Roy, R. J., Kembhavi, A., Chauhan, R., Paul, G., Kumar, J. & Philip, N. S. (2021). Commun. Biol. 4, 200. Gupta, L. & Sortrakul, T. (1998). Pattern Recognit. 31, 315–325. Ioffe, S. & Szegedy, C. (2015). arXiv:150203167. Kingma, D. P. & Ba, J. (2017). arXiv:14126980. LeCun, Y., Bengio, Y. & Hinton, G. (2015). Nature, 521, 436–444. Li, Y., Cash, J. N., Tesmer, J. J. G. & Cianfrocco, M. A. (2020). Structure, 28, 858–869.e3. Li, Y., Fan, Q., Cohn, J., Demers, V., Lee, J. Y., Yip, L., Cianfrocco, M. A. & Vos, S. M. (2022). bioRxiv, 2022.06.17.496614. Lyumkis, D. (2019). J. Biol. Chem. 294, 5181–5197. Mastronarde, D. N. (2005). J. Struct. Biol. 152, 36–51. Noble, A. J., Dandey, V. P., Wei, H., Brasch, J., Chase, J., Acharya, P., Tan, Y. Z., Zhang, Z., Kim, L. Y., Scapin, G., Rapp, M., Eng, E. T., Rice, W. J., Cheng, A., Negro, C. J., Shapiro, L., Kwong, P. D., Jeruzalmi, D., des Georges, A., Potter, C. S. & Carragher, B. (2018). eLife, 7, e34257. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Ko¨ pf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J. & Chintala, S. IUCrJ (2023). 10, 90–102 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy 101 research papers (2019). Advances in Neural Information Processing Systems 32, edited by H. Wallach, H. Larochelle, A. Beygelzimer, F. d’Alche´ - Buc, E. Fox & R. Garnett, pp. 8026–8037. Curran Associates Inc. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M. & Duchesnay, E´ . (2011). J. Mach. Learn. Res. 12, 2825–2830. Punjani, A., Rubinstein, J., Fleet, D. & Brubaker, M. A. (2017). Nat. Methods, 14, 290–296. Redmon, J. & Farhadi, A. (2018). arXiv:180402767. Ronneberger, O., Fischer, P. & Brox, T. (2015). arXiv:150504597. Sanchez-Garcia, R., Segura, J., Maluenda, D., Sorzano, C. O. S. & Carazo, J. M. (2020). J. Struct. Biol. 210, 107498. Shorten, C. & Khoshgoftaar, T. M. (2019). J. Big Data, 6, 60. Suloway, C., Pulokas, J., Fellmann, D., Cheng, A., Guerra, F., Quispe, J., Stagg, S., Potter, C. S. & Carragher, B. (2005). J. Struct. Biol. 151, 41–60. Tegunov, D. & Cramer, P. (2019). Nat. Methods, 16, 1146–1152. Wagner, T., Merino, F., Stabrin, M., Moriya, T., Antoni, C., Apelbaum, A., Hagel, P., Sitsel, O., Raisch, T., Prumbaum, D., Quentin, D., Roderer, D., Tacke, S., Siebolds, B., Schubert, E., Shaikh, T. R., Lill, P., Gatsogiannis, C. & Raunser, S. (2019). Commun. Biol. 2, 218. Weissenberger, G., Henderikx, R. J. M. & Peters, P. J. (2021). Nat. Methods, 18, 463–471. Wu, M. & Lander, G. C. (2020). Biophys. J. 119, 1281–1289. Yokoyama, Y., Terada, T., Shimizu, K., Nishikawa, K., Kozai, D., Shimada, A., Mizoguchi, A., Fujiyoshi, Y. & Tani, K. (2020). Biophys. Rev. 12, 349–354. Zhong, E. D., Bepler, T., Berger, B. & Davis, J. H. (2021). Nat. Methods, 18, 176–185. 102 Paul T. Kim et al. (cid:4) Learning to automate cryoEM data collection with Ptolemy IUCrJ (2023). 10, 90–102	null
10.1371_journal.pone.0243922.pdf	Data Availability Statement: All relevant data are within the paper and its Supporting Information files.	All relevant data are within the paper and its Supporting Information files.	RESEARCH ARTICLE The prevalence of esophageal cancer after caustic and pesticide ingestion: A nationwide cohort study Han-Wei MuID Dong-Zong HungID 1* 1,2, Chun-Hung Chen1,2, Kai-Wei Yang1,2, Chi-Syuan Pan1,2, Cheng-Li Lin3, 1 Division of Toxicology, China Medical University Hospital, Taichung, Taiwan, 2 Department of Emergency Medicine, China Medical University Hospital, Taichung, Taiwan, 3 Management Office of Health Data, China Medical University Hospital, Taichung, Taiwan a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 * [email protected] Abstract OPEN ACCESS Citation: Mu H-W, Chen C-H, Yang K-W, Pan C-S, Lin C-L, Hung D-Z (2020) The prevalence of esophageal cancer after caustic and pesticide ingestion: A nationwide cohort study. PLoS ONE 15(12): e0243922. https://doi.org/10.1371/journal. pone.0243922 Editor: Gianluigi Forloni, Istituto Di Ricerche Farmacologiche Mario Negri, ITALY Received: July 5, 2020 Accepted: November 30, 2020 Published: December 29, 2020 Peer Review History: PLOS recognizes the benefits of transparency in the peer review process; therefore, we enable the publication of all of the content of peer review and author responses alongside final, published articles. The editorial history of this article is available here: https://doi.org/10.1371/journal.pone.0243922 Copyright: © 2020 Mu et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the paper and its Supporting Information files. Habits such as smoking and alcohol drinking and existing esophageal malfunction are con- sidered the main risk factors for esophageal carcinogenesis. Caustic ingestion of acidic or alkaline agents or strong irritants can induce severe esophageal corrosive injury and increase esophageal cancer risk. We studied the relationship between esophageal carci- noma and acute detergent or pesticide poisoning by using nationwide health insurance data. Methodology/Principle findings: We compared a pesticide/detergent intoxication cohort (N = 21,840) and an age- and gender-matched control cohort (N = 21,840) identified from the National Health Insurance Research Database between 2000 and 2011. We used the multivariable Cox proportional model to determine esophageal carcinoma risk. The over- all incidence density of esophageal cancer was 1.66 per 10,000 person-years in the compar- ison cohort and 4.36 per 10,000 person-years in the pesticide/detergent intoxication cohort. The corresponding adjusted hazard ratio (HR) for esophageal cancer was 2.33 (95% confi- dence interval [CI] = 1.41–3.86) in the pesticide/detergent intoxication cohort compared with the control cohort. Patients with corrosive and detergent intoxication did not have a higher risk of esophageal cancer (adjusted HR = 0.98, 95% CI = 0.29–3.33) than those without pes- ticide/detergent intoxication. However, patients with pesticide intoxication had a significantly higher risk of esophageal cancer (adjusted HR = 2.52, 95% CI = 1.52–4.18) than those with- out pesticide/detergent intoxication. Conclusion: In the present study, after adjusting for conventional risk factors, we observed that pesticide intoxication could exert substantial effects through increased esophageal cancer risk. However, patients with detergent intoxi- cation may not have an increased risk of esophageal cancer. Introduction Essentially, self-ingestion of caustic agents, detergents, and pesticides is a serious public health problem in Taiwan. According to the Taiwan health statistics, 600 people ingested liquid tox- ins, including caustic agents and pesticides, for suicidal attempt in 1 year, and this is the third PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 1 / 11 PLOS ONEFunding: This work was supported by grants from the Ministry of Health and Welfare, Taiwan (MOHW108-TDU-B-212-133004), China Medical University Hospital (DMR-107-192, CMU107- ASIA-19), Academia Sinica Stroke Biosignature Project (BM10701010021), MOST Clinical Trial Consortium for Stroke (MOST 107-2321-B-039 -004-), Tseng-Lien Lin Foundation, Taichung, Taiwan, and Katsuzo and Kiyo Aoshima Memorial Funds, Japan. The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. No additional external funding was received for this study. Competing interests: The authors have declared that no competing interests exist. Esophageal cancer after pesticide ingestion common method for committing suicide. Caustic substance ingestion is most frequently encountered in children as a result of accidental swallowing or in adults as a result of self- harm. It often extensively injures the upper gastrointestinal tract and may lead to extensive necrosis, perforation, and death. Among the agents for pesticide poisoning in Taiwan, organo- phosphorus, herbicides, and other pesticides account for 45%, 23%, and 23%, respectively, according to the admission 2009 data from Taiwan National Health Insurance Database. Esophageal cancer accounts for >500,000 cancer deaths annually, and the incidence is rap- idly increasing worldwide [1]. In Taiwan, 2,630 new cases of esophageal cancer and 1,792 deaths caused by esophageal cancer occurred in 2013. The mean age at occurrence was 57 in men and 62 in women. Most esophageal cancer cases in Taiwan are of squamous cell carci- noma (93%), and the incidence is still increasing. The risk factors for esophageal cancer are smoking, alcohol consumption, dietary factors such as betel quid chewing and high tempera- ture beverage consumption, gastroesophageal reflux disease, and underlying esophageal dis- eases such as achalasia, and they are substantially different in various parts of the world [1–5]. Some studies have shown that caustic ingestion that induced severe esophageal corrosive injury might increase esophageal cancer risk [6–12]. Some studies with limited data even esti- mated a 1,000-fold higher risk [6]. However, these studies were based on a small number of case control studies; hence, the evidence is not strong. Pesticides protect plants from weeds, fungi, or insects. Pest control agents are usually applied through chemical dispersal in a hydrocarbon solvent-surfactant system to provide a homogeneous preparation. In addition to pesticides, these solvent-surfactants, such as the sur- factant of glyphosate, produce significant mucosal irritation effects. Some epidemiological studies have demonstrated high risks of certain cancers from exposure to some solvents [13]. Some pesticides are classified as carcinogenic or potentially carcinogenic to humans, such as captafol, diazinon, malathion, and glyphosate. Here, our study investigated the relationship between esophageal cancer and esophageal injuries after caustics ingestion and pesticide poisoning. Methods Data source This study used data from the National Health Insurance Research Database (NHIRD). The NHIRD was launched in Taiwan in 1995 and covers nearly 99% of the total population of Tai- wan with comprehensive healthcare benefits. For this study, we used the deidentified data of the residents to link two data files (subsets of the NHIRD), namely inpatient claims data and Registry of Beneficiaries. International Classification of Diseases-9-Clinical Modification (ICD-9-CM) codes were used to define diseases in the NHIRD. This study was approved by the Ethics Review Board of China Medical University (CMUH-104-REC2-115). Study population Patients with pesticide/detergent intoxication were identified from the NHIRD from January 1, 2000, to December 31, 2005, according to ICD-9-CM codes 983, 989.3–989.4, and 989.6. Patients diagnosed with cancer (ICD-9-CM codes 140–208) before pesticide/detergent intoxi- cation or those who lacked continuous health insurance coverage preceding cohort entry were excluded. Furthermore, all patients aged <20 years were excluded. Moreover, the comparison cohort of individuals without any history of pesticide/detergent intoxication was identified from the NHIRD. The comparison cohort also excluded those with cancer history, without health insurance before entering the study, or aged <20 years. In the final cohort, the pesti- cide/detergent intoxication cohort was matched to the comparison cohort at a 1:1 ratio by PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 2 / 11 PLOS ONEEsophageal cancer after pesticide ingestion gender, age, and the year of study entry. We designated 50 and 65 years as the age threshold. A consensus is lacking regarding the age at which an individual can be considered elderly, but the World Health Organization defines individuals >65 years as elderly in most developed countries. In less developed countries, for example in parts of Africa, >50 years old is consid- ered elderly. Thus, we classified participants into the age groups of <49, 50–64, and >65 to determine the difference between each group. The index date was defined as the date of first diagnosis of pesticide/detergent intoxication in the database. All participants were observed until they were diagnosed with esophageal can- cer (ICD-9-CM code 150), death, or the end of the study period (December 31, 2011). Outcome, comorbidity, and medication The primary clinical outcome was esophageal cancer (ICD-9-CM code 150). Furthermore, participants in the pesticide/detergent intoxication and control cohorts were compared for common comorbidities, including hypertension (ICD-9-CM codes 401–405), diabetes mellitus (ICD-9-CM code 250), chronic obstructive pulmonary disease (ICD-9-CM codes 491, 492, and 496), obesity (ICD-9-CM code 278), alcohol-related illness (ICD-9-CM codes 291, 303, 305, 571.0, 571.1, 571.2, 571.3, 790.3, A215, and V11.3), ischemic heart disease (ICD-9-CM codes 410–414), cerebrovascular disease (ICD-9-CM codes 430–438), and gastric disease (ICD-9-CM codes 530–534). Common comorbidities were identified according to the diagno- sis records in the inpatient file before the index date. Statistical analysis We used descriptive statistics to summarize the characteristics of the pesticide/detergent intox- ication cohort and matched comparison cohort. A continuous variable, such as age, was used in an independent t test to examine the mean ages between the two cohorts. Categorical vari- ables are presented as the number and percentage and included sex and common comorbidity assessed using the chi-square test. Univariable and multivariable Cox proportional hazard regression analyses were used to determine esophageal cancer risk, and the results are pre- sented as hazard ratios (HRs) with 95% confidence intervals (CIs). The differences in the cumulative incidence of esophageal cancer between the pesticide/detergent intoxication and control cohorts were estimated using the Kaplan–Meier method with the log-rank test. A two- tailed p value of <0.05 was considered statistically significant. We used SAS software (version 9.4 for windows; SAS Institute, Cary, NC, USA) for all statistical analyses and Kaplan–Meier survival curve plots. Results This study included 21,840 patients with pesticide/detergent intoxication and 21,840 control patients. The basic characteristics of the two cohorts are shown in Table 1. The mean ages of the pesticide/detergent intoxication cohort and comparison cohort were 52.1 ± 17.4 and 51.6 ± 17.6, respectively. No significant difference was noted in sex and age. The majority of pesticide/detergent intoxication patients were men (62.1%) and <49 years old (48.1%). In gen- eral, a high proportion of pesticide/detergent intoxication patients had hypertension, diabetes mellitus, gastric disease, ischemic heart disease, cerebrovascular disease, chronic obstructive pulmonary disease, alcohol-related illness, and obesity (all p < 0.001). The average follow-up duration was 5.25 ± 3.86 years for the pesticide/detergent cohort and 6.63 ± 3.29 years for the comparison cohort. The Kaplan–Meier curve showed that the cumulative incidence of esoph- ageal cancer was higher in the pesticide/detergent cohort than in the comparison cohort throughout the 12-year follow-up period (Fig 1). The cumulative incidence of esophagus PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 3 / 11 PLOS ONETable 1. Characteristics of patients with and without pesticide/detergent intoxication. Pesticide/Detergent intoxication Yes (N = 21840) No (N = 21840) Esophageal cancer after pesticide ingestion Age, year �49 50–64 � 65 Mean (SD) # Gender Female Male Comorbidity Hypertension Diabetes mellitus Gastric disease Ischemic heart disease Cerebrovascular disease Chronic obstructive pulmonary disease Alcohol-related illness Obesity Chi-square test. #t test. https://doi.org/10.1371/journal.pone.0243922.t001 n 10496 5386 5958 52.1 8269 13571 3454 2136 3027 1783 1875 1044 1071 16 % 48.1 24.7 27.3 17.4 37.9 62.1 15.8 9.78 13.9 8.16 8.59 4.78 4.90 0.07 n 10496 5386 5958 51.6 8269 13571 1870 957 1057 862 934 424 112 6 % p-value 48.1 24.7 27.3 17.6 37.9 62.1 8.56 4.38 4.84 3.95 3.82 1.94 0.51 0.03 0.99 0.004 0.99 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 cancer was significantly different between the pesticide/detergent and comparison cohorts (log-rank test; p < 0.001). The overall incidence densities of esophageal cancer were 1.66 and 4.36 per 10,000 person- years in the comparison and pesticide/detergent cohorts, respectively (Table 2). The corre- sponding adjusted HR for esophageal cancer was 2.33 (95% CI = 1.41–3.86) compared with controls after adjusting for age, sex, gastric disease, and alcohol-related illness. Compared with patients aged <49 years, those aged 50–64 and >65 years had 2.67-fold (95% CI = 1.54–4.64) and 3.18-fold (95% CI = 1.74–5.80) significantly higher risks of esophageal cancer, respectively. Compared with women, men had an adjusted HR of 19.8 (95% CI = 4.85–80.8) for esophagus cancer. Among various comorbidity types, significantly increased risk was observed in those with alcohol-related illness (adjusted HR = 7.14, 95% CI = 3.63–14.1). Table 3 presents the incidence and HR of esophageal cancer between patients with and without pesticide/detergent intoxication. Compared with patients without pesticide/detergent intoxication, men, patients aged <49 years, and those aged >65 years with pesticide/detergent intoxication had 2.22-fold (95% CI = 1.34–3.69), 2.84-fold (95% CI = 1.08–7.47), and 2.94-fold (95% CI = 1.19–7.26) increased esophagus cancer risks, respectively. For patients without comorbidity, those with pesticide/detergent intoxication had a significantly higher esophageal cancer risk than those without pesticide/detergent intoxication (adjusted HR = 2.32, 95% CI = 1.32–4.10). Among patients with non–alcohol-related illness, those with pesticide/deter- gent intoxication had a higher risk of esophageal cancer than controls (adjusted HR = 2.47, 95% CI = 1.46–4.16). Table 4 presents the incidence and adjusted HR of esophageal cancer between different groups of patients with pesticide/detergent intoxication. Patients with corrosive and detergent PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 4 / 11 PLOS ONEEsophageal cancer after pesticide ingestion Fig 1. Cummulative incidence comparison of esophagus cancer in patients with (dashed line) and without (solid line) pesticide/detergent intoxication. https://doi.org/10.1371/journal.pone.0243922.g001 intoxication (ICD-9-CM codes 983 and 989.6) did not have a higher risk of esophageal cancer (adjusted HR = 0.98, 95% CI = 0.29–3.33) than those without pesticide/detergent intoxication. Furthermore, patients with only pesticide intoxication (ICD-9-CM codes 989.3 and 989.4) had a significantly higher risk of esophageal cancer (adjusted HR = 2.52, 95% CI = 1.52–4.18) than those without pesticide/detergent intoxication. Discussion Several factors, including living habits and hobbies, contribute to esophageal cancer develop- ment. Esophageal cancer has two major subtypes, namely squamous cell carcinoma and ade- nocarcinoma, which have some same and different risk factors. Several genetic and epigenetic alterations are implicated in both the development and progression of esophageal cancer. Mucosal break, inflammation, and toxic injuries caused by excessive alcohol drinking and heavy smoking, two of the most important and common risk factors, are causes of esophageal carcinoma. Although the relationship between caustic ingestion and esophageal cancer and the mechanism of esophageal cancer development are unclear, lye-based cleaner burn has been found to complicate esophageal strictures and thus increase the risk of esophageal squa- mous cell carcinoma [6,7,11]. Fewer case series studies have shown that the esophageal cancer PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 5 / 11 PLOS ONEEsophageal cancer after pesticide ingestion Table 2. Incidence per 10,000 person-years of and risk factors for esophagus cancer. Variable Pesticide/Detergent intoxication Event PY Rate# Crude HR (95% CI) Adjusted HR& (95% CI) No Yes Age, year �49 50–64 � 65 Gender Female Male Comorbidity Hypertension No Yes Diabetes mellitus No Yes Gastric disease No Yes Ischemic heart disease No Yes Cerebrovascular disease No Yes Chronic obstructive pulmonary disease No Yes Alcohol-related illness No Yes Obesity No Yes 24 50 22 30 22 2 72 66 8 68 6 64 10 71 3 71 3 71 3 61 13 74 0 144761 114723 137357 66141 55986 100714 158769 238183 21300 247603 11880 242170 17314 248574 10910 248837 10646 254217 5266 254356 5128 259400 83 1.66 4.36 1.60 4.54 3.93 0.20 4.53 2.77 3.76 2.75 5.05 2.64 5.78 2.86 2.75 2.85 2.82 2.79 5.70 2.40 25.4 2.85 0.00 1.00 2.64(1.63, 4.30)�� 1.00 2.33(1.41, 3.86)�� 1.00 2.85(1.65, 4.95)�� 2.55(1.41, 4.61)�� 1.00 2.67(1.54, 4.64)�� 3.18(1.74, 5.80)�� 1.00 22.9(5.63, 93.5)�� 1.00 19.8(4.85, 80.8)�� 1.00 1.41(0.68, 2.95) 1.00 1.92(0.83, 4.42) 1.00 2.25(1.15, 4.38)� 1.00 0.99(0.31, 3.15) 1.00 1.02(0.32, 3.25) 1.00 2.13(0.67, 6.77) 1.00 1.00 1.00 0.83(0.40, 1.72) 1.00 1.00 1.00 1.00 10.9(5.96, 19.8)�� 1.00 7.14(3.63, 14.1)�� 1.00 - 1.00 Rate#: incidence rate per 10,000 person-years. Crude HR, relative hazard ratio. Adjusted HR&: Multivariable analysis including age, sex, gastric disease, and alcohol-related illness. �p < 0.05 ��p < 0.01 ��p < 0.001. https://doi.org/10.1371/journal.pone.0243922.t002 incidence caused by caustic ingestion is 1.4%–2.6% [6,7,14]. Although the incidence might be overestimated, most experts agree that corrosive injury might be a risk factor for esophageal carcinoma and have even alleged that the risk is 1,000 times that in the general population [15]. However, the results of this study are very different from those in the literature. This research is a nationwide, population-based cohort study designed to identify whether a PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 6 / 11 PLOS ONEEsophageal cancer after pesticide ingestion Table 3. Incidence and hazard ratio of esophageal cancer between patients with and without pesticide/detergent intoxication. Variables Event Gender Female Male Age, year �49 50–64 � 65 Comorbidity No Yes Alcohol-related illness No Yes 0 24 6 11 7 20 4 22 2 No PY 55515 89246 74309 36962 33489 130000 14760 144121 639 Pesticide/Detergent intoxication Rate# Event 0.00 2.69 0.81 2.98 2.09 1.54 2.71 1.53 31.3 2 48 16 19 15 30 20 39 11 Yes PY 45199 69523 63048 29178 22496 90783 23939 110234 4489 Rate# Crude HR (95% CI) Adjusted HR& (95% CI) 0.44 6.90 2.54 6.51 6.67 3.30 8.35 3.54 24.5 - 2.59(1.59, 4.23)�� - 2.22(1.34, 3.69)�� 3.16(1.24, 8.08)� 2.23(1.06, 4.68)� 3.19(1.30, 7.83)� 2.84(1.08, 7.47)� 1.63(0.74, 3.57) 2.94(1.19, 7.26)� 2.16(1.22, 3.80)�� 3.07(1.05, 8.98)� 2.32(1.32, 4.10)�� 2.77(0.92, 8.31) 2.33(1.38, 3.94)�� 2.47(1.46, 4.16)�� 0.78(0.17, 3.53) 1.26(0.27, 5.94) PY, person-years. Rate#: incidence rate per 10,000 person-years. Crude HR, relative hazard ratio. Adjusted HR†: Multivariable analysis including age, sex, gastric disease, and alcohol-related illness. �p < 0.05 ��p < 0.01 ��p < 0.001. https://doi.org/10.1371/journal.pone.0243922.t003 significant association exists between caustic ingestion and the risk of subsequent esophageal cancer. We defined conventional risk factors for esophageal cancer, such as age, sex, smoking, alcohol abuse, and gastric disease (such as achalasia and GERD), which were already well- established previously. In this 1-million-people cohort, 4,429 people were included in the detergent and corrosive intoxication group. The relative risk of esophageal cancer did not increase in patients with caustic agent and detergent poisoning compared with those without the poisoning after adjustment for these conventional risk factors. One of the reasons might be that our study included patients with exposure to detergents with less caustic characteristics. Detergents with acidic or alkaline characteristics are some of the most used toxic and corrosive Table 4. Incidence and adjusted hazard ratio of esophageal cancer between different entities of pesticide/detergent intoxication. Variable Without Pesticide/Detergent intoxication With Organophosphate/Carbamate + Pesticide (ICD-9-CM code 989.3, 989.4) With Detergent (ICD-9-CM code 983, 989.6) N 21840 17411 4429 No. of Events 24 47 3 Rate# 1.66 5.31 1.14 Adjusted HR† 1.00 2.52 0.98 95% CI (Reference) (1.52, 4.18) (0.29, 3.33) PY, person-years. Rate#, incidence rate per 10,000 person-years. Crude HR, relative hazard ratio. Adjusted HR†: Multivariable analysis including age, sex, gastric disease, and alcohol-related illness. �p < 0.05, ��p < 0.01, ��p < 0.001. https://doi.org/10.1371/journal.pone.0243922.t004 PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 7 / 11 PLOS ONEEsophageal cancer after pesticide ingestion chemicals at home. In general, detergents are classified into three categories according to their surfactant electrical charge: nonionic, anionic, and cationic. Nonionic and anionic detergents have low toxicity, although they may be mild to moderate irritants. Most serious toxins are cat- ionic detergents. Most of the detergents used at home are nonionic and anionic. Therefore, patients with ICD-9-CM codes 983.1 and 983.2 (acidic and alkali corrosive injury), 983.9 (caustic intoxication), or 989.6 (detergent intoxication) were identified, which expanded the dataset and weakened the results. The grade of esophagus corrosive injury of these cases is not available in the database. Thus, the true risk of esophageal cancer might be underestimated because, theoretically, esophageal cancer commonly occurs in patients with high-grade esoph- ageal corrosive injury. However, based on these data, the results still have considerable credibility. Another reason might be that the exposure interval after intoxication is shorter in this study than in previous studies (only 12 years with an average follow-up duration of only 5 years more). The results might be different if we increased the data and extended the study period. In previous studies, lye ingestion resulted in squamous cell carcinoma in the esophagus rather than adenocarcinoma [16]. Despite its uncertain etiology and pathogenesis, the mecha- nism of esophageal cancer after caustic agent and pesticide ingestion is probably similar to that of achalasia or esophageal diverticulum. The severe injury of esophagus after caustic ingestion causes lumen stricture or decreased esophageal motility. Subsequently, esophageal stasis occurs, which leads to local chronic inflammatory responses in the esophageal mucosa, which can lead to carcinogenesis. In cases of chronic irritation caused by foods and gastric fluid in achalasia, reflux esophagitis, or Barrett’s esophagus, the interval between disease diagnosis and esophageal carcinoma development was approximately 10–15 years [17]. However, the interval was considered to be shortened to 4 years for patients with aforementioned diseases who were exposed to airborne toxins that resulted from the terrorist attack of the World Trade Center [18]. Chemical hazard exposure can accelerate solid tumor development, such as esophageal carcinoma. In total, 287 chemicals or chemical groups with potential carcinogenic effects were identified in the field of the World Trade Center, including several organic solvents used in pesticide synthesis. In this study, the relative risk of esophageal cancer increased significantly by 2.52× in the pesticide group, and it was 2.47× even after excluding the comorbidity of alcohol-related ill- ness. Some pesticides are considered to become carcinogenic over a long time, including their main ingredients or organic solvents. However, such carcinogenicity was identified for most of them after chronic exposure in in vitro, in vivo, or epidemiological studies. No study has examined the relationship of acute large dose exposure with the occurrence of esophageal can- cer. However, some studies have reported that esophageal cancer is positively associated with intensive pesticide exposure. Jansson et al. found increased esophageal adenocarcinoma risk among people with high exposure to pesticides [19]. Meyer et al. showed that esophageal can- cer is correlated with pesticide exposure because of the high mortality caused by esophageal cancer in states in Brazil using a high proportion of pesticides [20]. Several pesticides have been identified as carcinogens, including their main ingredients or solvents. Animal studies have demonstrated strong genotoxicity for some pesticides, such as diazinon organophos- phates, malathion, and glyphosate herbicide, due to DNA and chromosomal damage. Further- more, numerous animal studies have shown strong cellular oxidative stress reactions for them. Glyphosate herbicide damages the retro-pharynx and esophagus more severely than other pes- ticides and causes a high rate of morbidity among patients because of its surfactant (poly- ethoxylated tallowamine) [21,22]. In this cohort study, a high proportion of patients in the pesticide/detergent intoxication cohort had hypertension, diabetes mellitus, gastric disease, ischemic heart disease, cerebrovascular disease, chronic obstructive pulmonary disease, PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 8 / 11 PLOS ONEEsophageal cancer after pesticide ingestion alcohol-relative illness, and obesity (all p < 0.001). Single severe direct esophageal mucosa damage as well as subsequent inflammation might be one of the causes of carcinogenicity in these patients with chronic systemic diseases and on long-term medication, with possible mal- function of the esophagus and stomach. However, in-depth animal experiments and studies are required to explore the possible mechanisms of the correlation. Our study has several limitations. The data were collected based on the ICD-9-CM codes in the database; therefore, some detailed information could not be obtained. First, the grade of esophageal corrosive injury after caustic ingestion is not provided in the database. This might underestimate the true risk of esophageal cancer because, theoretically, esophageal cancer occurs commonly in patients with high-severity esophageal corrosive injury. Second, although ICD-9-CM codes are used for acidic and alkali corrosive injury (983.1 and 983.2), most doc- tors in Taiwan refer such patients for the diagnosis of caustic intoxication (ICD-9-CM code 983.9) or detergent intoxication (ICD-9-CM code 989.6). It makes a huge difference in the case numbers between these diagnoses. Therefore, we cannot evaluate esophageal cancer risk in patients with acidic and alkaline caustic injury accurately. Third, although caustic, deter- gent, and pesticide intoxication in Taiwan are mostly through the oral route, using a diagnostic code to represent all oral-route intoxication could still slightly affect the results. Fourth, due to the limitation of the ICD-9-CM diagnostic codes, we could categorize the pesticides used for further detailed analysis. Furthermore, we were unable to extract the exact pathology reports from the database; thus, we could not further categorize the pathologies into premalignancy lesions, such as polyp or hyperplasia, or malignancies, such as adenocarcinoma or squamous cell carcinoma. Fifth, because a health insurance claims database was used, detailed informa- tion on certain general characteristics, such as obesity, body mass index, smoking, exercise, and dietary habits, was lacking. To compensate, we tried to use clinical examination-related morbidities to correct the individual examination index. Lastly, the present research involved only the Taiwanese general population, which includes 99.5% Han Chinese; thus, differences may be apparent in a stratified population. Conclusively, to determine the association between corrosive and detergent intoxication and esophageal cancer risk, the present study analyzed a population-based cohort from a nationwide claims database and adjusted for comorbidities to comprehensively assess corro- sive intoxication-related esophageal cancer risk. We observed that patients with preexisting corrosive poisoning did not exhibit a higher esophageal cancer risk than the general popula- tion. However, preexisting pesticide intoxication was associated with a 2.5-fold higher risk of esophageal cancer compared with the general population. Further investigations are required to delineate the association between esophageal carcinoma and esophageal corrosive injury or pesticide poisoning. Author Contributions Conceptualization: Han-Wei Mu, Chi-Syuan Pan. Data curation: Cheng-Li Lin. Formal analysis: Cheng-Li Lin. Methodology: Cheng-Li Lin. Supervision: Chun-Hung Chen, Dong-Zong Hung. Writing – original draft: Han-Wei Mu. Writing – review & editing: Kai-Wei Yang, Dong-Zong Hung. PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 9 / 11 PLOS ONEEsophageal cancer after pesticide ingestion References 1. Uhlenhopp DJ, Then EO, Sunkara T, Gaduputi V. Epidemiology of esophageal cancer: update in global trends, etiology and risk factors. Clin J Gastroenterol. 2020. Epub 2020/09/24. https://doi.org/10.1007/ s12328-020-01237-x PMID: 32965635. 2. Enzinger PC, Mayer RJ. Esophageal cancer. N Engl J Med. 2003; 349(23):2241–52. Epub 2003/12/06. https://doi.org/10.1056/NEJMra035010 PMID: 14657432. 3. Huang FL, Yu SJ. Esophageal cancer: Risk factors, genetic association, and treatment. Asian J Surg. 2018; 41(3):210–5. Epub 2016/12/18. https://doi.org/10.1016/j.asjsur.2016.10.005 PMID: 27986415. 4. Shaheen N, Ransohoff DF. Gastroesophageal reflux, barrett esophagus, and esophageal cancer: sci- entific review. JAMA. 2002; 287(15):1972–81. Epub 2002/04/19. https://doi.org/10.1001/jama.287.15. 1972 PMID: 11960540. 5. Arnold M, Colquhoun A, Cook MB, Ferlay J, Forman D, Soerjomataram I. Obesity and the Incidence of Upper Gastrointestinal Cancers: An Ecological Approach to Examine Differences across Age and Sex. Cancer Epidemiol Biomarkers Prev. 2016; 25(1):90–7. Epub 2015/10/24. https://doi.org/10.1158/1055- 9965.EPI-15-0753 PMID: 26494763; PubMed Central PMCID: PMC4713286. 6. Kiviranta UK. Corrosion carcinoma of the esophagus; 381 cases of corrosion and nine cases of corro- sion carcinoma. Acta Otolaryngol. 1952; 42(1–2):89–95. Epub 1952/01/01. https://doi.org/10.3109/ 00016485209120330 PMID: 14932967. 7. Appelqvist P, Salmo M. Lye corrosion carcinoma of the esophagus: a review of 63 cases. Cancer. 1980; 45(10):2655–8. Epub 1980/05/15. https://doi.org/10.1002/1097-0142(19800515)45:10<2655:: aid-cncr2820451028>3.0.co;2-p PMID: 7378999. 8. Hopkins RA, Postlethwait RW. Caustic burns and carcinoma of the esophagus. Ann Surg. 1981; 194 (2):146–8. Epub 1981/08/01. https://doi.org/10.1097/00000658-198108000-00005 PMID: 7259340; PubMed Central PMCID: PMC1345231. 9. Ti TK. Oesophageal carcinoma associated with corrosive injury—prevention and treatment by oeso- phageal resection. Br J Surg. 1983; 70(4):223–5. Epub 1983/04/01. https://doi.org/10.1002/bjs. 1800700413 PMID: 6831176. 10. Kim YT, Sung SW, Kim JH. Is it necessary to resect the diseased esophagus in performing reconstruc- tion for corrosive esophageal stricture? Eur J Cardiothorac Surg. 2001; 20(1):1–6. Epub 2001/06/26. https://doi.org/10.1016/s1010-7940(01)00747-3 PMID: 11423265. 11. Csikos M, Horvath O, Petri A, Petri I, Imre J. Late malignant transformation of chronic corrosive oeso- phageal strictures. Langenbecks Arch Chir. 1985; 365(4):231–8. Epub 1985/01/01. https://doi.org/10. 1007/BF01459611 PMID: 4087989. 12. Imre J, Kopp M. Arguments against long-term conservative treatment of oesophageal strictures due to corrosive burns. Thorax. 1972; 27(5):594–8. Epub 1972/09/01. https://doi.org/10.1136/thx.27.5.594 PMID: 5083728; PubMed Central PMCID: PMC470561. 13. Matic MG, Coric VM, Savic-Radojevic AR, Bulat PV, Pljesa-Ercegovac MS, Dragicevic DP, et al. Does occupational exposure to solvents and pesticides in association with glutathione S-transferase A1, M1, P1, and T1 polymorphisms increase the risk of bladder cancer? The Belgrade case-control study. PLoS One. 2014; 9(6):e99448. Epub 2014/06/11. https://doi.org/10.1371/journal.pone.0099448 PMID: 24914957; PubMed Central PMCID: PMC4051772. 14. Carver GM Jr., Sealy WC, Dillon ML Jr. Management of alkali burns of the esophagus. J Am Med Assoc. 1956; 160(17):1447–50. Epub 1956/04/28. https://doi.org/10.1001/jama.1956.02960520009003 PMID: 13306573. 15. Contini S, Scarpignato C. Caustic injury of the upper gastrointestinal tract: a comprehensive review. World J Gastroenterol. 2013; 19(25):3918–30. Epub 2013/07/11. https://doi.org/10.3748/wjg.v19.i25. 3918 PMID: 23840136; PubMed Central PMCID: PMC3703178. 16. Ruol A, Rampado S, Parenti A, Portale G, Giacomelli L, Battaglia G, et al. Caustic ingestion and oeso- phageal cancer: intra- and peri-tumoral fibrosis is associated with a better prognosis. Eur J Cardiothorac Surg. 2010; 38(6):659–64. Epub 2010/05/11. https://doi.org/10.1016/j.ejcts.2010.03.057 PMID: 20452236. 17. Meijssen MA, Tilanus HW, van Blankenstein M, Hop WC, Ong GL. Achalasia complicated by oesopha- geal squamous cell carcinoma: a prospective study in 195 patients. Gut. 1992; 33(2):155–8. Epub 1992/02/01. https://doi.org/10.1136/gut.33.2.155 PMID: 1541408; PubMed Central PMCID: PMC1373921. 18. Howard J. Minimum Latency & Types or Categories of Cancer: Centers for disease control and preven- tion; 2012 [updated Jan 6, 2015; cited 2012 Oct 17]. 5 p. Available from: https://www.cdc.gov/wtc/pdfs/ policies/WTCHP-Minimum-Cancer-Latency-PP-01062015-508.pdf. PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 10 / 11 PLOS ONEEsophageal cancer after pesticide ingestion 19. Jansson C, Plato N, Johansson AL, Nyren O, Lagergren J. Airborne occupational exposures and risk of oesophageal and cardia adenocarcinoma. Occup Environ Med. 2006; 63(2):107–12. Epub 2006/01/20. https://doi.org/10.1136/oem.2005.022467 PMID: 16421388; PubMed Central PMCID: PMC2078064. 20. Meyer A, Alexandre PC, Chrisman Jde R, Markowitz SB, Koifman RJ, Koifman S. Esophageal cancer among Brazilian agricultural workers: case-control study based on death certificates. Int J Hyg Environ Health. 2011; 214(2):151–5. Epub 2010/12/17. https://doi.org/10.1016/j.ijheh.2010.11.002 PMID: 21159552. 21. Chang CY, Peng YC, Hung DZ, Hu WH, Yang DY, Lin TJ. Clinical impact of upper gastrointestinal tract injuries in glyphosate-surfactant oral intoxication. Hum Exp Toxicol. 1999; 18(8):475–8. Epub 1999/08/ 26. https://doi.org/10.1191/096032799678847078 PMID: 10462358. 22. Hung DZ, Deng JF, Wu TC. Laryngeal survey in glyphosate intoxication: a pathophysiological investiga- tion. Hum Exp Toxicol. 1997; 16(10):596–9. Epub 1997/11/18. https://doi.org/10.1177/ 096032719701601007 PMID: 9363477. PLOS ONE \| https://doi.org/10.1371/journal.pone.0243922 December 29, 2020 11 / 11 PLOS ONE	null
10.1103_physrevresearch.5.013168.pdf	null	null	PHYSICAL REVIEW RESEARCH 5, 013168 (2023) Surrogate light curve models for kilonovae with comprehensive wind ejecta outﬂows and parameter estimation for AT2017gfo Atul Kedia ,1 Marko Ristic ,1 Richard O’Shaughnessy ,1 Anjali B. Yelikar ,1 Ryan T. Wollaeger ,2,3 Oleg Korobkin ,2,4 Eve A. Chase ,2,5 Christopher L. Fryer ,2,3,6,7,8 and Christopher J. Fontes 2,9 1Center for Computational Relativity and Gravitation, Rochester Institute of Technology, Rochester, New York 14623, USA 2Center for Theoretical Astrophysics, Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA 3Computer, Computational, and Statistical Sciences Division, Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA 4Theoretical Division, Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA 5Intelligence and Space Research Division, Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA 6University of Arizona, Tucson, Arizona 85721, USA 7Department of Physics and Astronomy, University of New Mexico, Albuquerque, New Mexico 87131, USA 8George Washington University, Washington, DC 20052, USA 9Computational Physics Division, Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA (Received 8 November 2022; accepted 20 January 2023; published 13 March 2023) The electromagnetic emission resulting from neutron star mergers have been shown to encode properties of the ejected material in their light curves. The ejecta properties inferred from the kilonova emission has been in tension with those calculated based on the gravitational wave signal and numerical relativity models. Motivated by this tension, we construct a broad set of surrogate light curve models derived for kilonova ejecta. The four-parameter family of two-dimensional anisotropic simulations and its associated surrogate explore different assumptions about the wind outﬂow morphology and outﬂow composition, keeping the dynamical ejecta component consistent. We present the capabilities of these surrogate models in interpolating kilonova light curves across various ejecta parameters and perform parameter estimation for AT2017gfo both without any assumptions on the outﬂow and under the assumption that the outﬂow must be representative of solar r-process abundance patterns. Our parameter estimation for AT2017gfo shows these surrogate models help alleviate the ejecta property discrepancy while also illustrating the impact of systematic modeling uncertainties on these properties, urging further investigation. DOI: 10.1103/PhysRevResearch.5.013168 I. INTRODUCTION Merging neutron stars have been demonstrated to emit both gravitational waves (GW) and a variety of accessible electromagnetic counterparts, as shown by the observation of GW170817 and the following transient AT2017gfo [1–7]. Also emitted during merger is the most unambiguous in- dication of matter in these systems: nuclear matter ejected due to the merger itself, which over time expands and heats through ongoing radioactive decay, producing a distinctive “kilonova” emission [8–11]. Particularly in conjunction with gravitational wave observations [1,2,12–17], kilonova discov- eries can provide insight into the physics and signiﬁcance of this radioactive ejecta. On the one hand, these counterparts probe uncertain nuclear physics [18–22]. On the other, these processes may be in part responsible for the production of r-process elements throughout the universe. [8,18,23]. Published by the American Physical Society under the terms of the Creative Commons Attribution 4.0 International license. Further distribution of this work must maintain attribution to the author(s) and the published article’s title, journal citation, and DOI. Kilonova observations in principle encode the amount and properties of the ejected material in multiwavelength light curves (LCs) and spectra [10,24,25]. Previous investigations have characterized this ejected material via two components, denoted as dynamical and disk wind ejecta, reﬂecting differ- ences in their formation and ejection mechanisms. Several studies of GW170817 attempted to infer the amount of mate- rial ejected [3–5,24–33]. However, the amount and properties of the ejected material from this event remain uncertain and in considerable tension with theoretical expectations for the amount of each type of ejecta [32,34–36], likely in part because of underestimated uncertainties in these theoretical ejecta estimates [37]. The tension between observation and expectations could in principle reﬂect modeling systematic errors. These obser- vations have historically been interpreted with semianalytic models [9,10,38] as they can be evaluated quickly and con- tinuously over the parameters, which characterize potential merger ejecta. However, it is well known that these semi- analytic models contain oversimpliﬁed physics of already simpliﬁed radiative transfer calculations [39–41] that ne- glect detailed full three-dimensional anisotropic radiative transfer, opacity, sophisticated nuclear reaction networks, and composition differences. That said, recent calculations using improved anisotropic radiative transfer and opacity 2643-1564/2023/5(1)/013168(15) 013168-1 Published by the American Physical Society ATUL KEDIA et al. PHYSICAL REVIEW RESEARCH 5, 013168 (2023) calculations still arrive at qualitatively similar conclusions for AT2017gfo [27,31–33,42,43]: a relatively high mass blue component outﬂowing along the poles, and a signiﬁcant mass in a red component outﬂowing preferentially towards the equator. Reference [32] also identiﬁes a larger velocity for the blue component, although the result reverses upon excluding bands that exhibit large uncertainties. However, recent nu- merical relativity simulations give surprisingly low dynamical outﬂow velocities when mass averaged, e.g., see Ref. [34] for example, and are not in as much disagreement with regards to the red component. (Some recent calculations with more complex outﬂow morphologies, however, recover wind speeds more consistent with prior expectations [26].) Essentially, the best-ﬁtting properties of the ejected material in our polar (blue) and equatorial (red) ejecta are in tension with theoreti- cal expectations for their physical origin. Motivated by this tension, in this paper we revisit our detailed anisotropic radiative transfer calculations and infer- ence [32,41], now allowing for a wide selection of outﬂow morphologies [36] to see impact on ejecta properties of AT2017gfo. The properties of the outﬂow that we vary in this paper are the outﬂow mass-density proﬁle for each com- ponent (referred to as the “morphology” for the rest of the paper) and the nuclear composition using the electron fraction Ye ≡ (np)/(np + nn) as a proxy, where np is the number of protons and nn is the number of neutrons in the ejecta. Effects of different dynamical ejecta morphologies have been studied in the past to show signiﬁcant light curve (LC) shifts [36], with a larger dependence on morphologies as compared to the anecdotal ejecta masses and velocities. Certain outﬂow models create a strong angular differential in the LC by suppressing the bluer bands along the equatorial plane (also known as lanthanide curtaining) [44–46]. In this paper we present a new set of surrogate models describing kilonova LCs resulting from a broad set of neu- tron star merger ejecta. This paper is organized as follows. In Sec. II we review our approach of adaptively generating kilonovae simulations, constructing the surrogate LC model using Gaussian process regression, and performing param- eter estimation (PE) for ejecta properties. We also discuss the speciﬁc alternate outﬂow morphologies and compositions employed in this paper. For context and to better highlight current tension with contemporary theoretical models, we also introduce estimates for the ejected material in each component applied to interpreting GW170817 given the well-constrained source distance and inclination. In Sec. III we discuss the new surrogate LC models and describe our PE results obtained using them. We further conclude this section with a discussion on the implications of this tension for joint multimessenger inference with contemporary ejecta models. We then conclude in Sec. IV. We provide all the simulation models, the surrogate models developed and employed in this paper, and sample codes to generate LCs at a public Zenodo repository [47]. The surrogates and sample codes are also available at a public GitHub repository [48]. initial simulations for emitted spectra using the radiative trans- fer code SuperNu. Then our active learning scheme iteratively chooses the “best” ejecta parameters to simulate until the simulation space is sufﬁciently large in size to produce a robust surrogate model. We then construct high-ﬁdelity mul- tidimensional surrogate LC models for each morphology and composition combination using Gaussian process regression (GP) as was performed in [32]. These surrogate models are now capable of making predictions of LCs along several model parameters. The surrogate models are then used to infer parameters for AT2017gfo and we compare these elec- tromagnetically inferred ejecta parameters with inferred ejecta parameters from the corresponding gravitational wave sig- nal GW170817 using a library of numerical relativity-based ejecta models. These steps are further discussed and detailed in the following subsections. A. Simulation methods and placement We perform simulations of kilonova LCs using the time- dependent radiative transfer code SuperNu [49,50], following the methodology presented in [32,36,39,41] and references therein. To determine nuclear heating rates, we use the detailed, time-dependent radioactive isotope composition re- sults from the nucleosynthesis simulations with the WinNet code [51,52] and include contributions of individual radiation species (such as α-, β-, γ -radiation, and ﬁssion products) for each isotope. The contributions are then weighted by thermalization efﬁciencies from [53] (a detailed description of the adopted nuclear heating can be seen in Ref. [39]). These heating rates and composition effects together along with the tabulated binned opacities resulting from the the Los Alamos suite of atomic physics codes [54,55], give the resul- tant kilonova LCs. The tabulated binned opacities used in this paper, however, are not calculated for all elements. Instead we produce opacities from representative proxy elements by com- bining pure-element opacities of nuclei with similar atomic properties [55]. The output we obtain from SuperNu are light spectra for 54 viewing angle bins uniformly distributed be- tween 0 and 180 deg in cosine space. For each viewing angle bin, the light rays are collected from across the face of the ejecta as seen from the direction of that viewing angle. To simplify comparison with previous paper, we adopted the same set of representative elements as in past work [32,41], however, a recent study [56] includes a more com- prehensive list of elements, including actinides. A comparison in Fig. 8 in Ref. [56] shows that when the actinide opacities are improved upon the luminosity varies by ∼15% at its maximum displacement. As will be discussed later in Sec. III, this displacement is currently signiﬁcantly smaller than our GP interpolation error making the changes with the new ele- mental approach less signiﬁcant for the purposes of this paper. However, any followup work should utilize the complete set of elemental opacities. II. METHOD The workﬂow pipeline is discussed in detail in this sec- tion and can be summarized as follows: We start by generating B. Initial conditions: Wind morphologies and compositions As initial conditions for the SuperNu code, we employ a two-component ejecta model, comprising an unbound dynam- ical ejecta from the merger, followed by the disk wind ejecta. 013168-2 SURROGATE LIGHT CURVE MODELS FOR KILONOVAE … PHYSICAL REVIEW RESEARCH 5, 013168 (2023) TABLE I. Ejecta morphologies and compositions studied in this paper. The composition of the dynamical component is ﬁxed at Ye = 0.04. In terms of this notation, the previous investigation studied a TPwind2 outﬂow [32]. Name TPwind1 TSwind1 TSwind2 Wind Morphology Peanut Spherical Spherical Ye 0.37 0.37 0.27 Dynamical Torus Torus Torus We set the outﬂow velocity based on prior understanding of expected ejecta [57,60,61]. The ejecta velocity proﬁle can alter the location of the photosphere for any given wavelength and thus affect the resultant LC. Speciﬁcally, the near-UV radiation peak time can be moved from an hour to a day postmerger by simply changing the velocity distribution. In this paper, however, we assume a homologous velocity proﬁle for the ejecta, which, i.e., it scales as r/t, where r is the distance from the center and t is the time postmerger. We do not include a third component that could attribute to the early blue peak. All ejecta components are assumed to be expanding homol- ogously with a prescribed velocity proﬁle, mass distribution versus angle (see Ref. [36] for their complete description), and prescribed uniform composition set by its initial electron fraction Ye. As discussed earlier, we use Ye as a proxy for ejecta composition in these simulations. We simulate a range of wind ejecta morphologies and compositions in order to study their inﬂuence on kilonova LCs, and build robust surrogate models that can be used for parameter inference of kilonovae observations. The outﬂow morphologies we model are selected from previously studied models [36,41] with a varying set of wind ejecta structure and electron fraction. Table I shows the morphology and composition combinations simulated. Here, for instance, the model labeled TPwind1 comprises a Torus shaped dynamical ejecta, and a Peanut shaped wind ejecta (S would indicate a Spherical wind ejecta) with their mathematical form given in Ref. [36]. The labels “wind1” and “wind2” denote the two choices of wind ejecta composition. wind1 corresponds to the less neutron rich conﬁguration with Ye = 0.37 whereas wind2 corresponds to a more neutron rich Ye = 0.27; wind2 with a Peanut morphology, i.e., TPwind2, was adopted in the previous paper [32] and hence a new surrogate model for the case is not generated in this study. As was in previous paper, we ﬁx the composition of the dynamical ejecta as Ye = 0.04. For each source morphology and composition, we place an initial coarse grid with ejecta masses (Me j/M(cid:4)) having val- ues 0.001, 0.003, 0.01, 0.03, 0.1 and ejecta velocities (ve j/c) having values 0.05, 0.15, 0.3 for each of the two compo- nents making our initial grid of (5 × 3)2 = 225 simulations (see Refs. [17,28,57–59] for a discussion on expected ejecta masses). The base grid [(5 × 3)2 = 225] for our simulations is the same as previously studied in Ref. [41] to which we add additional simulations doubling the grid space via the active learning technique discussed in Ref. [32]. The iterative process targets subsequent simulation parameters for investi- gation, based on the estimated uncertainty of a (simpliﬁed) Gaussian Process estimate for the emitted radiation from the existing simulation set. By the end of the iterative process we accumulate 450, 449, and 449 simulations for the three morphology and composition combinations, respectively. The TSwind1 and TSwind2 model families each have one fewer simulation in their training libraries due to isolated instances of single simulation processing error. We limit the iterative learning process to 450 simulations because in our experience from previous study (Ref. [32]) sufﬁciently accurate surrogate models can be obtained with a sample size this large. C. Motivation for alternate morphologies Inferences made about ejecta properties of the event AT2017gfo, derived from the kilonova LCs [32], are not con- sistent with contemporaneous forward-model predictions for ejecta masses from GW inferences of the masses deduced for GW170817. Our kilonova analysis suggests that much larger wind velocities (and masses) will be required, inconsis- tent with the modeling assumptions usually adopted for polar winds. Recently, Nedora et al. [34,62] have proposed an alter- native mechanism to generate disk outﬂows: a spiral-wave- driven wind, which generally distributes over a large solid angle. For the purposes of our phenomenological kilonova inference, which focuses on the nature of outﬂows rather than their origin, such an outﬂow will have a different mor- phology than the polar “peanut” wind and “torus” dynamical morphology assumed for outﬂows. Further, Breschi et al. [26] have performed inference using simpliﬁed LCs extracted from systematically explored one-, two- and three-component models with several isotropic and anisotropic morphologies. They ﬁnd a strong preference for anisotropic models with the inclusion of multiple components leading to largest evidence. However, their isotropic model (ISO-DV) makes even the dynamical components isotropic, whereas in this paper we choose an anisotropic dynamical ejecta shape (torus). Addi- tionally, Kawaguchi et al. [31] obtain a strong similarity in the LCs for AT2017gfo and for a two-component model with a postmerger wind ejecta that is spherically symmetric. The effect of morphologies of the ejecta was also studied extensively in the past [36], where morphologies were shown to have a greater impact on the LCs than the ejecta masses and velocities. The effect manifests heavily in the form of altering the peak luminosity and peak times. We therefore assemble an expanded archive of actively-learned kilonova models, with more options for the wind-driven outﬂow morphology, to as- sess whether we can better reconcile our kilonova inference with forward models for the ejecta. To maximize the impact of altering wind morphology in our study, we adopt spherical and peanut shaped wind model. Other detailed simulations and the studies above all naturally produce a morphology much more similar to our preferred (torus) morphology. D. Simulation interpolation Using the set of actively-learned simulations, we follow our previous approach [32] to interpolate the resulting AB magnitudes versus simulation parameters (md , vd , mw, vw ), time postmerger, viewing angle, and wavelength bands. 013168-3 ATUL KEDIA et al. PHYSICAL REVIEW RESEARCH 5, 013168 (2023) FIG. 2. Wavelength interpolation using our surrogate model for an off-sample wavelength bands CFHT (1451 nm, i.e., between the J- and H -bands) and F182M (1839 nm, between H - and K-bands), generated for the TPwind1 model with parameters (md /M(cid:4), vd /c, mw/M(cid:4), vw/c) = (0.014, 0.183, 0.085, 0.053) viewed along the sym- metry axis θ = 0. The dashed curves (red for CFHT and green for F182M) along with their respective bands indicate the predicted LC from our surrogate model and solid curves indicate the LC generated from the code SuperNu via the mechanism explained in Sec. II A. TPwind2 conﬁguration, as was performed in [32], whereas the bottom panel shows the same for TSwind2 model. Hence, apparent differences in the two predictions arise solely due to differences in the wind morphology. As noted earlier, our GP training scheme interpolates be- tween ﬁlter bands, over a nominal wavelength parameter. As a result, our GP approach can estimate LCs for ﬁlters both in- cluded and not included in our original training set. We verify the ﬁdelity of our wavelength interpolation LC prediction by comparison to SuperNu spectral output convolved with two new ﬁlters. Figure 2 illustrates our wavelength interpolation LC prediction in dashed curves compared to the SuperNu spectral output in solid curves convolved with two distinct real ﬁlters not included in our original training set. One of these ﬁlters is the the WIRCam 8105 broadband ﬁlter [63] from the Canada-France-Hawaii Telescope (CFHT), part of the Mauna Kea Observatory, with an effective wavelength of 1.45 µm. This ﬁlter has minimal overlap with the H-band ﬁlter, thus serving as an independent test of wavelength interpolation ability. We also show results for the JWST/NIRCam.F182M ﬁlter [64], which lies between the H- and K-bands with small overlap with the K band and has an effective wavelength of 1.84 µm [65]. Our GP models occasionally produce glitches as can be seen at t ∼ 10 days in Fig. 2. These glitches occur only for the wind1 composition models for both TS and TP mor- phologies (also visible in Fig. 5) and affect the quality of the predicted LC signiﬁcantly at times beyond 4 days. This error-prone training may be resolvable by developing a so- phisticated hyperparameter selection method that restricts the hyperparameters to the region of the preceding parameters hence avoiding large jumps. E. Ejecta parameter inference Following Ref. [32] we use conventional Bayesian tech- niques to reconstruct the ejecta parameters (md , vd , mw, vw ) and emission direction θ most consistent with GW170817, FIG. 1. Illustrating off-sample interpolations for two morpholo- gies. Top panel: The bottom panel from Fig. 4 of Ref. [32] showing interpolation of various LCs versus time for the TPwind2 morphol- ogy adopted in that paper. Different colors denote different ﬁlter bands, described in the legend. The dashed lines show full simulation output for each band. The colored points show our interpolated magnitude predictions at the evaluation times. The simulated param- eters and viewing angle for this conﬁguration are (md /M(cid:4), vd /c, mw/M(cid:4), vw/c) = (0.097, 0.198, 0.084, 0.298) at 0◦ [32]. Bottom panel: As above, but for the TSwind2 model morphology. The shaded region surrounding each solid curve shows our estimated GP ﬁtting uncertainty at each time. The differences between these two panels illustrate the impact of outﬂow morphology on our results. Interpolation over simulation parameters and for each band is performed by Gaussian process regression at ﬁducial ref- erence times and ﬁducial angles (0,30,45,60,75,90 degrees). A continuous LC over time and angle follows by stitching together these ﬁducial results with simple low-dimensional interpolation. Unless otherwise noted, we quantify the perfor- mance of our interpolation with the RMS difference between our prediction and the true value. Because of the substantial dynamic range of our many outputs, we interpolate in AB magnitudes using the LSST grizy and 2MASS JHK bands as our reference bands. Our raw LCs are calculated in absolute magnitudes, i.e., at a ﬁducial distance of 10 pc to the source. In Fig. 1 we show our surrogate model predicting LCs for off-sample outﬂow. The prediction is for outﬂow param- eters (md /M(cid:4), vd /c, mw/M(cid:4), vw/c) = (0.097, 0.198, 0.084, 0.298). These parameters are chosen such that they are neither on the primary training grid nor the actively chosen simulation parameters. Top panel shows the parameter LC predicted for 013168-4 SURROGATE LIGHT CURVE MODELS FOR KILONOVAE … PHYSICAL REVIEW RESEARCH 5, 013168 (2023) FIG. 3. Left panel: Illustration of systematics in modeling ejecta from binary mergers, compared to a ﬁducial inference of those ejecta parameters. Black curves are deduced from inferences on GW170817 mass estimates combined with different models for the ejecta; for example the black solid line is derived from Eqs. (1)–(3). To isolate the effect of ejecta modeling systematics, all predictions are shown under the optimistic assumption of a perfectly known nuclear equation of state, here APR4 [70], without adding additional uncertainties to account for estimated ﬁtting errors in the relations between ejecta and binary parameters. While the ejecta predictions for GW170817 are relatively consistent with one another, they are in tension with our interpretation of the GW170817-associated kilonova’s ejecta assuming a TPwind2 outﬂow form. Curves correspond to 90% credible regions. Right panel: Parameter inference for LC based on EM paper prior (uniform in velocities, log-uniform in masses) [32] (blue-solid) and r-process prior [69] (green-dashed), but allowing for relative systematic uncertainty exp(±0.3) in both ejecta masses. The inner and outer curves on the two-dimensional plots correspond to 68% and 95% credible regions. The ejecta estimate illustrated from GW includes no systematics, either from uncertainty in the EOS or from intrinsic uncertainty in the estimate itself; compare to Fig. 6. using the known source emission distance and taking into account prior information about the source orientation di- rection. As previously, we adopt uniform priors in velocity and log-uniform for masses over the limiting ranges. As a concrete example, the blue contours in Fig. 3 show previously-presented results from [32] for two-component ejecta properties assuming the TPwind2 morphology. While our default inference makes no additional assump- tions about the ejecta, we separately also require that the merger from AT2017gfo produces a full r-process element spectrum, consistent with the solar spectrum. Doing so con- strains the ratio of mw/md to be close to 13.90 (1.76) for the wind1 (wind2) model assuming FRDM2012 as the nuclear mass model [67] and ﬁssion model by Panov et al. (2010) [68] as described in [69]. The blue curves in the right panel of Fig. 3 show an analysis assuming this prior for the relative amounts of the two ejecta components. The inferred ejecta properties with these Torus/Peanut models are, however, somewhat surprising in the context of conventional theoretical expectations about the two compo- nents. Most notably, our analysis requires a substantial wind velocity (cid:6) 0.2c, consistent with previously-reported phe- nomenological estimates (see, e.g., Table 2 or Fig. 5 in [34]), but in contrast to most theoretical expectations [35]. This high wind velocity allows the peanut-morphology wind (which, in our calculations, explains the rapidly decaying blue compo- nent after its early peak) to have the short timescales required for the blue component, while simultaneously matching the target blue luminosity early on. By contrast, the “dynamical” red torus component can be consistent with GW170817’s associated emission with a wide range of velocities [26,35]. In earlier studies [36], it was noted that the dynamical ejecta mass was poorly constrained, due to the lanthanide curtain- ing effect [45]. Since lanthanide curtaining occurs in models with greater velocity difference in the dynamical and wind components, this makes viewing angle dependence (caused by the curtaining) more velocity dependent than mass dependent [36]. 013168-5 ATUL KEDIA et al. PHYSICAL REVIEW RESEARCH 5, 013168 (2023) TABLE II. Numerical relativity-based forward models employed in this work for GW170817 ejecta PE. The velocity ﬁts are common for all. Dynamical ejecta velocity ﬁt is given in Eq. (7) based on [28,59]. Wind velocity is set ﬁxed at vw = 0.08c [61]. Name CoDi19 DiCoSci20 KruFo20 Nedora21 md (cid:2) CoDi19 (cid:2) (cid:2) mw (cid:2) (cid:2) (cid:2) CoDi19 Reference [28] [17] [58] [60] F. Contemporary forward models for neutron star merger ejecta After using parameter inference to deduce the possible immediate outﬂow responsible for the kilonova associated with GW170817, we arrive at inferred ejecta masses and velocities for that outﬂow. For context and to help guide our investigations, we compare these inferred ejecta properties with the predictions provided by selected contemporary for- ward models, which attempt to estimate the ejecta properties from inferred binary properties [17,28,58,59,71]. We use four discrete models for this purpose, summarized in Table II, and present details about them here. We employ the following estimates for the dynamical and wind ejecta mass [28]. First, the disk mass is estimated as their Eq. (1), log10(mdisk[Mtot/Mthr]) (cid:2) (cid:2) = max −3, a 1 + b tanh (cid:4)(cid:5)(cid:5) (cid:3) c − Mtot/Mthr d (1) with a = −31.335, b = −0.9760, c = 1.0474, d = 0.05957. Mtot is the sum of neutron star masses, and Mthr is the thresh- old mass for the binary to undergo a prompt collapse to a BH after merger, which can be estimated by (e.g., Bauswein et al. 2013 [72]) Mth = 2.38 − 3.606 MTOV, (2) (cid:2) (cid:5) MTOV R1.6 where MTOV is the maximum gravitational mass of a stable nonrotating neutron star given the EOS and R1.6 is the radius of the neutron star for the EOS with a mass 1.6 M(cid:4). The estimated uncertainty on this relation is of roughly a factor of 2, or 0.02 M(cid:4). The wind ejecta mass mwind = ξ mdisk where ξ = 0.3, with an estimated uncertainty of O(1). The dynami- cal ejecta mass is estimated by their Eq. (2), (cid:2) (cid:4) (cid:3) log10 mdyn = a (1 − 2C1)m1 C1 + b m2 (cid:5) n m1 m2 + d 2 + [1 ↔ 2], (3) where in this expression m1 and m2 are the neutrons star masses, a = −0.0719, b = 0.2116, d = −2.42, n = −2.905, and Ci are the compactnesses of the two neutron stars. [1 ↔ 2] is a shorthand for the preceding terms in the expression with permuted subscripts. The estimated uncertainty is for this relation is 7 × 10−3M(cid:4) when calculated linearly or up to 36% when calculating for log10 mdyn. This model is labeled “CoDi19” in the relevant plots. For a more recent estimate, we also utilize the updated model from Ref. [17] that has been calibrated for a wider range of binary mass ratios. A new disk mass ﬁtting is calcu- lated in this study and is given in Eqs. (S4)– (S6) in Ref. [17] and is the same as Eq. (1) with updated constants. The ﬁtting parameters here are a = a0 + δa · ξ , b = b0 + δb · ξ , and 2 tanh (β( ˆq − ˆqtrans)), a0 = the resulting constants are ξ = 1 −1.5815, δa = −2.439, b0 = −0.538, δb = −0.406, c = 0.953, d = 0.0417, β = 3.910, ˆqtrans = 0.900. This model is labeled “DiCoSci20” in the relevant plots. As a complementary estimate to highlight their systematic error, we also provide postprocessing estimates provided with an alternative set of ﬁts from Kruger and Foucart [58]. In this approach, the disk mass is estimated as mdisk,KF = m1max[5 × 10−4, (aC1 + c)d ], where C1 is the compactness of the lighter of the two neutron stars, a = −8.1324, c = 1.4820, d = 1.7784 with an error of order 40%. The dynamical mass is estimated as (4) mdyn,KF 10−3M(cid:4) = (cid:2) a C1 (cid:2) (cid:5) n m2 m1 + b (cid:5) + c C1 m1 + [1 ↔ 2], (5) where a = −9.3335, b = 114.17, c = −337.56, and n = 1.5465, where negative values imply zero ejecta. This model is labeled “KruFo20” in the relevant plots. For broader context we also report estimates for the dynam- ical ejecta mass from GW170817 using the more inclusive and most recent ﬁtting method by Ref. [60]. Among the available choices of ﬁtting, we use the quadratic ﬁtting of the form log10(mdyn/M(cid:4)) = b0 + b1q + b2 ˜(cid:8) + b3q2 + b4q ˜(cid:8) + b5 ˜(cid:8)2, (6) where q is the mass ratio (m2/m1 mass), and ˜(cid:8) is the reduced tidal deformability. Among the new models introduced in that paper, this model provides the greatest ﬂexibility with both mass ratio and tidal deformability dependence, although we note the conspicuous absence of dependence on binary total mass. From this paper, we adopt the ﬁtting coefﬁcients obtained for the reference set + M0/M1 set (labelled as M0RefSet and M0/M1Set), which include the best available physics, speciﬁcally excluding analyses that omit neutrino re- absorption and have pertinent systematic differences with the reference calibration adopted here. The parameters thus used are (from Table 4 of Ref. [60]): b0 = −1.32, b1 = −3.82 × 10−1, b2 = −4.47 × 10−3, b3 = −3.39 × 10−1, b4 = 3.21 × 10−3, b5 = 4.31 × 10−7. We exclude a disk mass model from Ref. [60] due to large systematics and inconsistent deﬁni- tion between the original datasets. Instead we employ the ﬁt by [28]. This model is labeled “Nedora21” in the relevant plots. For ejecta velocities in each model we choose the results from Ref. [28,59]. Hence, for the dynamical ejecta we adopt the formula vd /c = a (1 + c C1) + a (1 + c C2) + b (7) (cid:2) (cid:5) m1 m2 (cid:2) (cid:5) m2 m1 013168-6 SURROGATE LIGHT CURVE MODELS FOR KILONOVAE … PHYSICAL REVIEW RESEARCH 5, 013168 (2023) FIG. 4. Angular dependence for the three surrogate models for each morphology and composition combination, i.e., TPwind1, TSwind1, and TSwind2. This ﬁgure compares the g-, y-, and K-band luminosity (top to bottom) at select times as a function of viewing angle. Different colors indicate different extraction times. Points show simulation AB magnitude results vs angle θ, while the solid curves and shaded regions show our prediction and its expected (statistical interpolation) uncertainty. All simulations and surrogate LCs use the parameters (md /M(cid:4), vd /c, mw/M(cid:4), vw/c) = (0.097, 0.198, 0.084, 0.298); because this conﬁguration has vw > vd (i.e., wind outside the dynamical ejecta), the wind ejecta’s emission and angular dependence could dominate trends versus angle. Left column: Angular dependence for TPwind1 light bands. This model shows the largest viewing angle dependence. Center column: Angular dependence for TSwind1. A change in angular dependence is highly apparent at late times particularly in the g band. Right column: Angular dependence for TSwind2. Differences from the ﬁrst two models are apparent. This conﬁguration shows the weakest angular dependence in its emission and the brightest emission. with a = −0.309, b = 0.657, and c = −1.879 with an esti- mated relative uncertainty of 20% [28]. For the velocity of wind ejecta, we adopt 0.08c as was adopted by [61], vw = 0.08c. (8) The solid-black contours in Fig. 3 show our estimated ejecta masses using these forward models for ejecta. As inputs to these expressions, we use the results of detailed parameter inference applied to GW170817, as described in Lange et al. (in preparation); see also [73]. Even allowing for reasonable uncertainties in these ejecta formulas, the tension with our inferred ejecta parameters is apparent: these models predict roughly 2× smaller ejecta masses for both components, much slower wind velocities, and an extremely narrow range of expected dynamical ejecta velocities. III. RESULTS A. Impact of morphology on interpolated light curves Using simpliﬁed and more realistic [26,74,75] [31,36,39,76,77] models for kilonovae and the associated radiative transfer, several previous studies have demonstrated that the angular distribution of the outﬂow can imprint its signature on the outgoing radiation. The ﬁrst and second columns of Fig. 4 illustrate the impact of outﬂow morphology on the interpolated LCs. The plots show g-, y-, and K-band luminosity as time proceeds along viewing angles for a sample model parameter (md /M(cid:4), vd /c, mw/M(cid:4), vw/c) = (0.097, 0.198, 0.084, 0.298). As anticipated from the aforementioned prior paper, we notice a strong impact from the choice of mor- phology. In particular, relative to our polar outﬂow illustrated in the ﬁrst column (TPwind1), the spherical wind outﬂow 013168-7 ATUL KEDIA et al. PHYSICAL REVIEW RESEARCH 5, 013168 (2023) in the second column (TSwind1) produces relatively less angular variability of the LC, particularly in the bluer bands. A noticeable feature of TPwind1 LCs is that their angular dependence ﬂips over in the late times, as can be seen by the apparent opposite curvature at t = 4 days and t = 8 days. This behavior is resulting from the dispersion of the dynamical torus component at late times. Compared to this result, the TS late time LCs are angle independent both for low and high electron fraction. The TPwind1 model is able to also exhibits a peak in the blue band in the early times (t = 0.5 days) as is noticeable in the top-left panel. A change in the morphology to TSwind1 changes this blue band LC to be monotonically diminishing, noticeable in the top-middle panel. This will again be seen in a Fig. 5 below where parameters ﬁtting AT2017gfo with TPwind1 and r process prior will produce the short-time blue kilonova peak. This can, however, be altered and other mor- phologies be made to produce the blue peak at desired times by ﬁne-tuning component parameters, thus highlighting the challenges in replicating this observation. The late time K band is rather similar for all morphologies and compositions, and most of the differences are present in the earlier times. This is because, at late times, the photo- sphere has moved inwards substantially so morphology effects are minimized and the temperature at the photosphere is near the K band. B. Impact of composition on interpolated light curves The second and third columns of Fig. 4 illuminate how ejecta composition (Ye) impacts emitted radiation. As ex- pected, we ﬁnd brighter LCs at early and late times for lower Ye (i.e., wind2 is brighter than wind1) in the g-, y-, and other lower wavelength bands, and somewhat in the K bands. As a result, the choice of wind composition (Ye) is partially degen- erate with wind ejecta mass mw, which is not surprising since products like Yemw enter naturally into the initial conditions. Further, their is almost no angular dependence in the TSwind2 and a small angular dependence in TSwind1 because the wind outﬂow surpasses the dynamical and dominates the kilonova producing a viewing angle-independent signal. At late times the wind ejecta has substantially outﬂown the dynamical ejecta due to having vd > vw and directs the emis- sion. The TSwind2 model, due to its lower Ye, has a higher wind-heating rates in comparison to the TSwind1 and results in a lower optical depth. The shallower optical depths of TSwind2 in comparison with TSwind1 causes its emission to be largely driven by the wind ejecta component, which is uniform along different viewing angles. On the other hand for TSwind1 the dynamical ejecta contribute relatively more to the emission and hence shows a stronger angular dependence. As was noticed in Sec. III A the early blue peak was miss- ing from the TSwind1 model. This peak is retained again here in the TSwind2 case. C. Estimated light curves Given the substantial ﬂexibility that we allow in our kilo- nova model families, each of which has ﬁve free parameters (two masses, two velocities and the viewing angle), unsurpris- ingly, our models can ﬁnd a good ﬁt for EM observations of TABLE III. Peak likelihood identiﬁed in parameter inference for GW170817. The mass ratios for the cases where the r-process prior was used during parameter inference were peaked at mw/md = 13.90 and 1.76 for the wind1 and wind2 models, respectively. The ﬁt likelihoods correspond to the LCs shown in Fig. 5. Model TPwind1 TPwind1 TPwind2 TPwind2 TSwind1 TSwind1 TSwind2 TSwind2 Prior ln Lmax ln L ln(L × p) Fit likelihood (MaP) uniform r process uniform r process uniform r process uniform r process 2.41 1.62 0.61 0.49 −9 −8.78 2.84 2.81 −0.15 −0.97 −0.62 −1.18 −11.22 −14.32 2.38 1.93 6.63 −2.20 6.77 4.23 −4.05 −11.26 9.09 6.77 GW170817’s kilonova LC (taken from [25], which compiles data from [3–7,23,78–82]). Figure 5 shows the inferred LCs deduced with each of our models. As previously [32] and unless otherwise noted henceforth, we adopt a strong angular prior based on afterglow observations [3,16,29,78,83–85]. In all cases, our unconstrained models ﬁt the observations reasonably well, albeit less well for the bluest bands. Con- sistent with prior paper [32], our posterior predictions for the bluest bands are consistently fainter and more rapidly decay- ing apart from the TPwind1 case, which ﬁts the ∼1 day peak of g band very well. Lacking large residuals, however, these ﬁgures provide relatively little insight into the overall good- ness of ﬁt. In Table III, we report the peak likelihood identiﬁed overall in each analysis (ln Lmax, a measure of the goodness of ﬁt of our unconstrained ﬁve-parameter kilonova model; and the likelihood associated with the maximum a posteriori (MaP) parameters [ln(L × p)], a goodness of ﬁt measure, which requires consistency with the prior (either uniform or r process). For each morphology and wind composition, the two estimates reported for ln Lmax derived from our unconstrained and r process constrained analyses are relatively consistent as expected: both measure quality of ﬁt of the unconstrained model. Keeping in mind differences in ln L of order unity correspond to odds ratio changes of order e1, this table shows that the TSwind1 models ﬁt the data very poorly, but that the remaining unconstrained models all ﬁt the data reasonably well. When the r-process prior is imposed, we disfavor the TPwind1 and TPwind2 models slightly in favor of TSwind2. D. Discussion of inferred parameters Despite similarity in inferred LCs, performing ejecta parameter inference with LC surrogate models from alter- nate morphologies and compositions unsurprisingly produces different inferences for ejecta parameters. In Fig. 6 we illus- trate parameter inference for AT2017gfo based on surrogate models for different ejecta models with and without includ- ing r-process priors. The inferred ejecta properties depend notably on the choice of the outﬂow’s morphology and com- position. For example, for our spherical models, the wind ejecta mass is much smaller than the dynamical ejecta mass, 013168-8 SURROGATE LIGHT CURVE MODELS FOR KILONOVAE … PHYSICAL REVIEW RESEARCH 5, 013168 (2023) FIG. 5. Posterior predicted LCs for AT2017gfo, compared with AT2017gfo observations. Points and error bars denote the observations, with each color denoting a different ﬁlter band. Following [32], the solid curves and shaded intervals show the median and 90% credible expected LC, deduced by ﬁtting our models to these observations. The left panel of ﬁgures denotes our unconstrained models; the right panels require each outﬂow to be consistent with Solar r-process abundance (i.e., we also adopt a prior on mw/md such that the ejected material; see Sec. II E). First row: TPwind1; second row: TPwind2 [32]; third row: TSwind1; fourth row: TSwind2. 013168-9 ATUL KEDIA et al. PHYSICAL REVIEW RESEARCH 5, 013168 (2023) FIG. 6. Morphology and composition dependent parameter inference. This ﬁgure shows parameter inference for ejecta properties (md /M(cid:4), vd /c, mw/M(cid:4), vw/c) derived from GW170817 and AT2017gfo. Colored solid and dotted curves show purely electromagnetic inferences derived using different families of two-component LC surrogate models and a prior on source inclination. Different solid colors indicate different choices of outﬂow and composition with a uniform logarithmic prior on ejecta masses, while colored dotted curves show inferences with the corresponding morphology/composition pairs while also requiring consistency with the Solar r-process abundance pattern. The black solid and dotted curves by contrast show three estimates inferred solely from GW measurements, combined with forward models for the outﬂow. The legend indicates the morphology pair (e.g., TP = Torus and Peanut), while the integer indicates the wind composition (1 refers to the wind1 sequence and so on). The inferred wind (and thus total) mass depends substantially on the assumed wind composition and morphology. The two-dimensional contours correspond to 90% credible intervals. and the (dominant) dynamical ejecta mass is strongly con- strained. However, despite broadening our model space, our ejecta inferences remain inconsistent with conventional prior ex- pectations about ejecta from binary mergers, suggesting that persistent modeling systematics remain. E. Discussion We ﬁnd that both the alternative morphologies and compo- sitions have a signiﬁcant impact on emission at t = 1 day and later in the g-, y-, and K-bands, and for emission in the equa- torial direction. Reassessing GW170817, we ﬁnd modestly different conclusions about ejecta parameters, conditioned on outﬂow morphology and composition. The alternate mor- phologies help alleviate the tension between observational and theoretical ejecta properties. However, the systematics remain insufﬁcient to completely reconcile our interpretation of the ejecta with our prior expectation about the material ejected from merger based on GW-only inference and theoretical ex- pectations for merger ejecta. After exploring a range of outﬂow morphologies, compo- sitions, ejecta masses, and ejecta velocities, we ﬁnd that the inferences developed from our two-component ejecta models continue to exhibit tension both with prior expectations of ejecta from the merger, using its GW-deduced parameters, and also with the observed LCs themselves missing a blue component. Our investigations here have omitted two directions, which deserve subsequent attention. First, we have adopted very sharp initial compositions Ye. Full numerical simulations have consistently suggested peaked Ye distributions with extended tails [86]. A small contribution of suitable material could help explain the bright extended blue component missing in some of our LCs. Second, and in a similar vein, we have restricted to a two-component model; an additional third small blue component enhancer, previously attributed to magnetically- driven winds could help improve our ﬁt as invoked in previous paper [75]. The recently observed kilonova and long gamma- ray burst, GRB211211A, has indicated the presence of an additional thermal component powered by either a GRB jet or a central magnetar [87–90]. Such a central power source has been proposed such as a long-lived magnetar, a black hole powered by accretion disk or a jet cocoon for previous observations as well [91–98]. Our current study excludes such a source to power the emission; however, based on the ﬁndings of those studies we anticipate that adding a central engine could help alleviate the early blue-component mismatch and should be investigated further in future work. Further, we notice that our Gaussian process modeling produces occasional glitches in the surrogate models. These glitches occur only for the wind1 composition models and affect the quality of the predicted LC largely at times beyond 013168-10 SURROGATE LIGHT CURVE MODELS FOR KILONOVAE … PHYSICAL REVIEW RESEARCH 5, 013168 (2023) 4 days. This error-prone training may get resolved by de- veloping a hyperparameter selection method, which carefully restricts the hyperparameters to preceding parameter neigh- borhood. Our inferences of ejecta properties are in modest tension with the expected ejecta properties, assuming GW-derived binary properties and NR-calibrated models for the outﬂows expected from different binary mergers. However, while we performed an analysis only of GW170817 where these NR- calibrated models are indeed most reliable, at larger mass ratios these models disagree more substantially [37]. These tensions could be challenging to resolve as a part of joint multimessenger GW-EM PE for generic neutron star mergers. IV. CONCLUSIONS In summary, we have developed families of surrogate mod- els for kilonova LCs resulting from binary neutron star merger events. As previously done, using our Gaussian process in- terpolation applied to an actively-extended simulation archive we develop surrogates for different ejecta conﬁgurations: three additional wind morphology and composition combi- nations in a two-component (dynamical and wind) ejecta approach. Our suite of four kilonova surrogate models now covers multiple outﬂow conﬁgurations comprising toroidal dynamical ejecta, and peanut and spherical shaped wind ejecta with two compositions. These broad set of surrogate models can be employed for parameter inference of ejecta compo- nents when kilonovae observations are made in the future and will assist in determining the physics of these events. These kilonovae simulations and their surrogates are available at a Zenodo repository [47] and the surrogates alone in a GitHub repository [48]. As expected, we ﬁnd that wind morphology and, to a lesser extent, composition (within our constrained set of composi- tions) have substantial impacts on the outgoing emission’s time, frequency, and angular dependence. To assess the impact of these differences, we calculate the likely ejecta properties associated with GW170817’s kilonova, conditioned on the assumption that the ejecta and radiation exactly reproduce one of our assumed ejecta models. We ﬁnd that the choice of model most strongly impacts the inferred ejecta masses. None of our inferences strongly constrains the ejected wind velocity. To better understand the context of these calculations, we compare and contrast them with the inputs and out- puts expected from GW170817’s merger. On the one hand, we compare our inferred ejecta masses with the expected ejecta masses deduced from GW170817’s GW-constrained masses, using several contemporary estimates for the ejecta. While the EM- and GW-deduced ejecta can be reconciled for some ejecta models, allowing for substantial ﬁt and EOS uncertainty, we generally see considerable tension in the in- ferred ejecta mass, with the EM-inferred masses generally larger than the GW-deduced ejecta masses when assuming a single plausible equation of state. Such tension between GW-deduced (and hence numerical-relativity informed) ejecta masses has been repeatedly highlighted in the literature [17,28,37,58,59,71] (a recent paper indicates the possibility of relieving this tension with the use of new heating rate ﬁtting formulas [99,100]). Our corroboration of this ﬁnding, even us- ing state-of-the-art ejecta calculations, suggests the resolution of this discrepancy must invoke other sources of systematic error, for example in the ejecta initial conditions. On the other hand, we repeat our ejecta mass inferences while requiring the ejected r-process mass abundances to be consistent with Solar r-process patterns. While this constraint sharply reduces the range of ejecta masses allowed, it does not signiﬁcantly change any of the aforementioned conclusions: the EM- deduced ejecta masses remain large, while EM-constrained ejecta velocities remain weakly constrained. Our r-process constraints on mw/md remain qualitatively consistent with the corresponding ratios expected from forward-modeling ejecta starting from GW inferences about binary masses. Our kilonova generation work has limitations in the fol- lowing ways, which should be addressed in future work. The two-ejecta component model is known to account for the lanthanide rich and lanthanide free outﬂow materials to a considerable degree. However, the LC being substantially different for differing morphologies implies that we would need a more exhaustive calculation of ejecta LCs with other morphologies or account for that with a third component. Hence, adding a third component or a continuous ejecta model could cause a shift in LCs. Similarly, our present and previ- ous work suggests that adjusting our assumptions about the composition, both in median and distribution, can also notably impact the LCs. We have also highly simpliﬁed the underly- ing nuclear physics, both in our assumptions about heating [18,19,53] and in using Ye as a proxy for the detailed isotopic composition. Finally, our results depend on our present un- derstanding of pertinent opacities, which, while dramatically improved over prior work, does not yet include the latest actinide opacities, let alone non-LTE physics at late times. For technical convenience, our discussion of the tension between our direct inferences about ejecta from EM and secondhand conclusions about ejecta derived from GW infer- ence of GW170817 adopted a ﬁxed nuclear equation of state (APR4) to predict ejecta properties from binary properties. By eliding uncertainty in the nuclear equation of state in our ﬁgure, we somewhat underestimate the systematic uncertainty in ejecta estimates derived from GW measurements, artiﬁ- cially strengthening the apparent tension between GW- and EM-deduced ejecta properties. ACKNOWLEDGMENTS The authors thank Erika M. Holmbeck and Matthew R. Mumpower for stimulating discussions throughout the devel- opment of this research. A.K. and M.R. acknowledge support from National Science Foundation (NSF) Grant No. AST- 1909534. R.O.S. acknowledges support from NSF Grant No. AST-1909534, NSF Grant No. PHY-2012057, and the Simons Foundation. A.B.Y. acknowledges support from NSF Grant No. PHY-2012057. R.T.W., O.K., E.A.C., C.L.F., and C.J.F. were supported by the U.S. Department of Energy through the Los Alamos National Laboratory (LANL). This research also used resources provided by LANL through the institu- tional computing program. Los Alamos National Laboratory is operated by Triad National Security, LLC, for the Na- tional Nuclear Security Administration of U.S. Department 013168-11 ATUL KEDIA et al. PHYSICAL REVIEW RESEARCH 5, 013168 (2023) of Energy (Contract No. 89233218CNA000001). Research presented in this paper was supported by the Laboratory Di- rected Research and Development program of Los Alamos National Laboratory under Project No. 20190021DR. [1] B. P. Abbott, R. Abbott, T. D. Abbott, F. Acernese, K. Ackley, C. Adams, T. Adams, P. Addesso, R. X. Adhikari, V. B. Adya et al., Multi-messenger observations of a binary neutron star merger, Astrophys. J. Lett. 848, L12 (2017). [2] B. P. Abbott, R. Abbott, T. D. Abbott, F. Acernese, K. Ackley, C. Adams, T. Adams, P. Addesso et al. (The LIGO Scientiﬁc Collaboration, the Virgo Collaboration), GW170817: Obser- vation of Gravitational Waves from a Binary Neutron Star Inspiral, Phys. Rev. Lett. 119, 161101 (2017). [3] N. R. Tanvir, A. J. Levan, C. González-Fernández, O. Korobkin, I. Mandel, S. Rosswog, J. Hjorth, P. D’Avanzo, A. S. Fruchter, C. L. Fryer et al., The emergence of a lanthanide-rich kilonova following the merger of two neutron stars, Astrophys. J. 848, L27 (2017). [4] S. J. Smartt, T. W. Chen, A. Jerkstrand, M. Coughlin, E. Kankare, S. A. Sim, M. Fraser, C. Inserra, K. Maguire, K. C. Chambers et al., A kilonova as the electromagnetic counter- part to a gravitational-wave source, Nature (London) 551, 75 (2017). [5] P. S. Cowperthwaite, E. Berger, V. A. Villar, B. D. Metzger, M. Nicholl, R. Chornock, P. K. Blanchard, W. Fong, R. Margutti, M. Soares-Santos et al., The electromagnetic counterpart of the binary neutron star merger LIGO/Virgo GW170817. II. UV, optical, and near-infrared light curves and comparison to kilonova models, Astrophys. J. Lett. 848, L17 (2017). [6] E. Pian, P. D’Avanzo, S. Benetti, M. Branchesi, E. Brocato, S. Campana, E. Cappellaro, S. Covino, V. D’Elia, J. P. U. Fynbo et al., Spectroscopic identiﬁcation of r-process nucleosynthe- sis in a double neutron-star merger, Nature (London) 551, 67 (2017). [7] I. Arcavi, G. Hosseinzadeh, D. A. Howell, C. McCully, D. Poznanski, D. Kasen, J. Barnes, M. Zaltzman, S. Vasylyev, D. Maoz, and S. Valenti, Optical emission from a kilonova following a gravitational-wave-detected neutron-star merger, Nature (London) 551, 64 (2017). [8] J. M. Lattimer and D. N. Schramm, Black-hole-neutron-star collisions, Astrophys. J. Lett. 192, L145 (1974). [9] L.-X. Li and B. Paczy´nski, Transient events from neutron star mergers, Astrophys. J. Lett. 507, L59 (1998). [10] B. D. Metzger, Kilonovae, Living Rev. Relativ. 23, 1 (2020). [11] S. Bernuzzi, Neutron star merger remnants, Gen. Relativ. Gravit. 52, 108 (2020). [12] B. P. Abbott, R. Abbott, T. D. Abbott, F. Acernese, K. Ackley, C. Adams, T. Adams, P. Addesso et al. (The LIGO Scientiﬁc Collaboration, the Virgo Collaboration), A standard siren mea- surement of the Hubble constant, Nature (London) 551, 425 (2017). [13] B. P. Abbott, R. Abbott, T. D. Abbott, F. Acernese, K. Ackley, C. Adams, T. Adams, P. Addesso et al. (The LIGO Scientiﬁc Collaboration, the Virgo Collaboration), GW170817: Mea- surements of Neutron Star Radii and Equation of State, Phys. Rev. Lett. 121, 161101 (2018). [14] B. P. Abbott, R. Abbott, T. D. Abbott, S. Abraham, F. Acernese, K. Ackley, C. Adams, V. B. Adya et al. (The LIGO Scientiﬁc Collaboration, the Virgo Collaboration), Model comparison from LIGO-Virgo data on GW170817’s binary components and consequences for the merger remnant, Class. Quantum Gravity 37, 045006 (2020). [15] A. Bauswein, S. Blacker, V. Vijayan, N. Stergioulas, K. Chatziioannou, J. A. Clark, N.-U. F. Bastian, D. B. Blaschke, M. Cierniak, and T. Fischer, Equation of State Constraints from the Threshold Binary Mass for Prompt Collapse of Neu- tron Star Mergers, Phys. Rev. Lett. 125, 141103 (2020). [16] K. Hotokezaka, E. Nakar, O. Gottlieb, S. Nissanke, K. Masuda, G. Hallinan, K. P. Mooley, and A. T. Deller, A Hub- ble constant measurement from superluminal motion of the jet in GW170817, Nat. Astron. 3, 940 (2019). [17] T. Dietrich, M. W. Coughlin, P. T. H. Pang, M. Bulla, J. Heinzel, L. Issa, I. Tews, and S. Antier, Multimessenger con- straints on the neutron-star equation of state and the Hubble constant, Science 370, 1450 (2020). [18] J. Barnes, Y. L. Zhu, K. A. Lund, T. M. Sprouse, N. Vassh, G. C. McLaughlin, M. R. Mumpower, and R. Surman, Kilo- novae across the nuclear physics landscape: The impact of nuclear physics uncertainties on r-process-powered emission, Astrophys. J. 918, 44 (2021). [19] Y. L. Zhu, K. A. Lund, J. Barnes, T. M. Sprouse, N. Vassh, G. C. McLaughlin, M. R. Mumpower, and R. Surman, Mod- eling kilonova light curves: Dependence on nuclear inputs, Astrophys. J. 906, 94 (2021). [20] N. Vassh, G. C. McLaughlin, M. R. Mumpower, and R. Surman, Markov chain Monte Carlo predictions of neutron- rich lanthanide properties as a probe of r-process dynamics, Astrophys. J. 907, 98 (2021). [21] C. J. Horowitz, Neutron rich matter in the laboratory and in the heavens after GW170817, Ann. Phys. 411, 167992 (2019). [22] K. Chatziioannou, Neutron-star tidal deformability and equation-of-state constraints, Gen. Relativ. Gravit. 52, 109 (2020). [23] M. R. Drout, A. L. Piro, B. J. Shappee, C. D. Kilpatrick, J. D. Simon, C. Contreras, D. A. Coulter, R. J. Foley, M. R. Siebert, N. Morrell et al., Light curves of the neutron star merger GW170817/SSS17a: Implications for r-process nucleosynthe- sis, Science 358, 1570 (2017). [24] M. W. Coughlin, T. Dietrich, Z. Doctor, D. Kasen, S. Coughlin, A. Jerkstrand, G. Leloudas, O. McBrien, B. D. Metzger, R. O’Shaughnessy, and S. J. Smartt, Constraints on the neutron star equation of state from AT2017gfo using radiative transfer simulations, Mon. Not. R. Astron. Soc. 480, 3871 (2018). [25] V. A. Villar, J. Guillochon, E. Berger, B. D. Metzger, P. S. Cowperthwaite, M. Nicholl, K. D. Alexander, P. K. Blanchard, R. Chornock, T. Eftekhari et al., The combined ultraviolet, op- tical, and near-infrared light curves of the kilonova associated with the binary neutron star merger GW170817: Uniﬁed data set, analytic models, and physical implications, Astrophys. J. Lett. 851, L21 (2017). 013168-12 SURROGATE LIGHT CURVE MODELS FOR KILONOVAE … PHYSICAL REVIEW RESEARCH 5, 013168 (2023) [26] M. Breschi, A. Perego, S. Bernuzzi, W. D. Pozzo, V. Nedora, D. Radice, and D. Vescovi, AT2017gfo: Bayesian inference and model selection of multicomponent kilonovae and con- straints on the neutron star equation of state, Mon. Not. R. Astron. Soc. 505, 1661 (2021). [27] J. Heinzel, M. W. Coughlin, T. Dietrich, M. Bulla, S. Antier, N. Christensen, D. A. Coulter, R. J. Foley, L. Issa, and N. Khetan, Comparing inclination-dependent analyses of kilonova tran- sients, Mon. Not. R. Astron. Soc. 502, 3057 (2021). [28] M. W. Coughlin, T. Dietrich, B. Margalit, and B. D. Metzger, Multimessenger Bayesian parameter inference of a binary neu- tron star merger, Mon. Not. R. Astron. Soc. 489, L91 (2019). [29] K. D. Alexander, E. Berger, W. Fong, P. K. G. Williams, C. Guidorzi, R. Margutti, B. D. Metzger, J. Annis, P. K. Blanchard, D. Brout et al., The electromagnetic counterpart of the binary neutron star merger LIGO/Virgo GW170817. VI. Radio constraints on a relativistic jet and predictions for late-time emission from the kilonova ejecta, Astrophys. J. Lett. 848, L21 (2017). [30] R. Chornock, E. Berger, D. Kasen, P. S. Cowperthwaite, M. Nicholl, V. A. Villar, K. D. Alexander, P. K. Blanchard, T. Eftekhari, W. Fong et al., The electromagnetic counterpart of the binary neutron star merger LIGO/Virgo GW170817. IV. Detection of near-infrared signatures of r-process nucleosyn- thesis with Gemini-South, Astrophys. J. Lett. 848, L19 (2017). [31] K. Kawaguchi, M. Shibata, and M. Tanaka, Diversity of kilo- nova light curves, Astrophys. J. 889, 171 (2020). [32] M. Ristic, E. Champion, R. O’Shaughnessy, R. Wollaeger, O. Korobkin, E. A. Chase, C. L. Fryer, A. L. Hungerford, and C. J. Fontes, Interpolating detailed simulations of kilonovae: Adaptive learning and parameter inference applications, Phys. Rev. Res. 4, 013046 (2022). [33] C. E. Collins, A. Bauswein, S. A. Sim, V. Vijayan, G. Martínez-Pinedo, O. Just, L. J. Shingles, and M. Kromer, 3D radiative transfer kilonova modelling for binary neutron star merger simulations, arXiv:2209.05246. [34] V. Nedora, S. Bernuzzi, D. Radice, B. Daszuta, A. Endrizzi, A. Perego, A. Prakash, M. Safarzadeh, F. Schianchi, and D. Logoteta, Numerical relativity simulations of the neutron star merger GW170817: Long-term remnant evolutions, winds, remnant disks, and nucleosynthesis, Astrophys. J. 906, 98 (2021). [35] D. Radice, S. Bernuzzi, and A. Perego, The dynamics of binary neutron star mergers and GW170817, Annu. Rev. Nucl. Part. Sci. 70, 95 (2020). [36] O. Korobkin, R. T. Wollaeger, C. L. Fryer, A. L. Hungerford, S. Rosswog, C. J. Fontes, M. R. Mumpower, E. A. Chase, W. P. Even, J. Miller, G. Wendell Misch, and J. Lippuner, Ax- isymmetric radiative transfer models of kilonovae, Astrophys. J. 910, 116 (2021). [37] A. Henkel, F. Foucart, G. Raaijmakers, and S. Nissanke, A study of the agreement between binary neutron star ejecta models derived from numerical relativity simulations, arXiv:2207.07658. [38] D. Grossman, O. Korobkin, S. Rosswog, and T. Piran, The long-term evolution of neutron star merger remnants - II. Ra- dioactively powered transients, Mon. Not. R. Astron. Soc. 439, 757 (2014). [39] R. T. Wollaeger, O. Korobkin, C. J. Fontes, S. K. Rosswog, W. P. Even, C. L. Fryer, J. Sollerman, A. L. Hungerford, D. R. van Rossum, and A. B. Wollaber, Impact of ejecta morphology and composition on the electromagnetic signatures of neutron star mergers, Mon. Not. R. Astron. Soc. 478, 3298 (2018). [40] W. Even, O. Korobkin, C. L. Fryer, C. J. Fontes, R. T. Wollaeger, A. Hungerford, J. Lippuner, J. Miller, M. R. Mumpower, and G. Wendell Misch, Composition effects on kilonova spectra and light curves. I, Astrophys. J. 899, 24 (2020). [41] R. T. Wollaeger, C. L. Fryer, E. A. Chase, C. J. Fontes, M. Ristic, A. L. Hungerford, O. Korobkin, R. O’Shaughnessy, and A. M. Herring, A broad grid of 2D kilonova emission models, Astrophys. J. 918, 10 (2021). [42] K. Kawaguchi, M. Shibata, and M. Tanaka, Radiative transfer simulation for the optical and near-infrared electromagnetic counterparts to GW170817, Astrophys. J. Lett. 865, L21 (2018). [43] M. Almualla, Y. Ning, M. Bulla, T. Dietrich, M. W. Coughlin, and N. Guessoum, Using neural networks to per- form rapid high-dimensional kilonova parameter inference, arXiv:2112.15470. [44] J. Barnes and D. Kasen, Effect of a high opacity on the light curves of radioactively powered transients from compact ob- ject mergers, Astrophys. J. 775, 18 (2013). [45] D. Kasen, R. Fernández, and B. D. Metzger, Kilonova light curves from the disc wind outﬂows of compact object mergers, Mon. Not. R. Astron. Soc. 450, 1777 (2015). [46] L. Nativi, M. Bulla, S. Rosswog, C. Lundman, G. Kowal, D. Gizzi, G. P. Lamb, and A. Perego, Can jets make the radioac- tively powered emission from neutron star mergers bluer? Mon. Not. R. Astron. Soc. 500, 1772 (2020). [47] https://zenodo.org/record/7335961. [48] https://github.com/markoris/surrogate_kne. [49] R. T. Wollaeger, D. R. van Rossum, C. Graziani, S. M. Couch, G. C. Jordan IV, D. Q. Lamb, and G. A. Moses, Radiation transport for explosive outﬂows: A multigroup hy- brid Monte Carlo method, Astrophys. J. Suppl. Ser. 209, 36 (2013). [50] R. T. Wollaeger and D. R. van Rossum, Radiation transport for explosive outﬂows: Opacity regrouping, Astrophys. J. Suppl. Ser. 214, 28 (2014). [51] C. Winteler, R. Käppeli, A. Perego, A. Arcones, N. Vasset, N. Nishimura, M. Liebendörfer, and F. K. Thielemann, Magne- torotationally driven supernovae as the origin of early galaxy r-process elements? Astrophys. J. Lett. 750, L22 (2012). [52] O. Korobkin, S. Rosswog, A. Arcones, and C. Winteler, On the astrophysical robustness of the neutron star merger r-process, Mon. Not. R. Astron. Soc. 426, 1940 (2012). [53] J. Barnes, D. Kasen, M.-R. Wu, and G. Martínez-Pinedo, Ra- dioactivity and thermalization in the ejecta of compact object mergers and their impact on kilonova light curves, Astrophys. J. 829, 110 (2016). [54] C. J. Fontes, H. L. Zhang, J. Abdallah, Jr., R. E. H. Clark, D. P. Kilcrease, J. Colgan, R. T. Cunningham, P. Hakel, N. H. Magee, and M. E. Sherrill, The Los Alamos suite of relativistic atomic physics codes, J. Phys. B: At. Mol. Opt. Phys. 48, 144014 (2015). [55] C. J. Fontes, C. L. Fryer, A. L. Hungerford, R. T. Wollaeger, and O. Korobkin, A line-binned treatment of opacities for the spectra and light curves from neutron star mergers, Mon. Not. R. Astron. Soc. 493, 4143 (2020). 013168-13 ATUL KEDIA et al. PHYSICAL REVIEW RESEARCH 5, 013168 (2023) [56] C. J. Fontes, C. L. Fryer, R. T. Wollaeger, M. R. Mumpower, and T. M. Sprouse, Actinide opacities for modeling the spectra and light curves of kilonovae, Mon. Not. R. Astron. Soc. 519, 2862 (2022). [57] D. Kasen, B. Metzger, J. Barnes, E. Quataert, and E. Ramirez- Ruiz, Origin of the heavy elements in binary neutron-star mergers from a gravitational-wave event, Nature (London) 551, 80 (2017). [58] C. J. Krüger and F. Foucart, Estimates for disk and ejecta masses produced in compact binary mergers, Phys. Rev. D 101, 103002 (2020). [59] T. Dietrich and M. Ujevic, Modeling dynamical ejecta from binary neutron star mergers and implications for electro- magnetic counterparts, Class. Quantum Grav. 34, 105014 (2017). [60] V. Nedora, F. Schianchi, S. Bernuzzi, D. Radice, B. Daszuta, A. Endrizzi, A. Perego, A. Prakash, and F. Zappa, Mapping dynamical ejecta and disk masses from numerical relativity simulations of neutron star mergers, Class. Quantum Grav. 39, 015008 (2022). [61] D. Radice, A. Perego, K. Hotokezaka, S. A. Fromm, S. Bernuzzi, and L. F. Roberts, Binary neutron star mergers: Mass ejection, electromagnetic counterparts, and nucleosynthesis, Astrophys. J. 869, 130 (2018). [62] V. Nedora, S. Bernuzzi, D. Radice, A. Perego, A. Endrizzi, and N. Ortiz, Spiral-wave wind for the blue kilonova, Astrophys. J. Lett. 886, L30 (2019). [63] P. Puget, E. Stadler, R. Doyon, P. Gigan, S. Thibault, G. Luppino, G. Barrick, T. Benedict, T. Forveille, W. Rambold, J. Thomas, T. Vermeulen, J. Ward, J.-L. Beuzit, P. Feautrier, Y. Magnard, G. Mella, O. Preis, P. Vallee, S.-y. Wang, C.-J. Lin et al., WIRCam: the infrared wide-ﬁeld camera for the Canada- France-Hawaii Telescope, Proc. SPIE 5492, 978 (2004). [64] JWST User Documentation (JDox), JWST User Documenta- tion Website, January (2016). [65] These ﬁlter response functions were taken from the SVO Filter Proﬁle Service [66]. [66] C. Rodrigo and E. Solano, The SVO ﬁlter proﬁle service, in XIV.0 Scientiﬁc Meeting (virtual) of the Spanish Astronomical Society (2020), p. 182. [67] P. Möller, A. J. Sierk, T. Ichikawa, and H. Sagawa, Nuclear ground-state masses and deformations: FRDM (2012), At. Data Nucl. Data Tables 109-110, 1 (2016). [68] I. V. Panov, I. Y. Korneev, T. Rauscher, G. Martínez-Pinedo, A. Keli´c-Heil, N. T. Zinner, and F. Thielemann, Neutron- induced astrophysical reaction rates for translead nuclei, Astron. Astrophys. 513, A61 (2010). [69] M. Ristic, E. M. Holmbeck, R. Wollaeger, O. Korobkin, E. Champion, R. O’Shaughnessy, C. L. Fryer, C. J. Fontes, M. R. Mumpower, and T. M. Sprouse, Constraining inputs to re- alistic kilonova simulations through comparison to observed r-process abundances, arXiv:2206.02273. [70] A. Akmal, V. R. Pandharipande, and D. G. Ravenhall, Equa- tion of state of nucleon matter and neutron star structure, Phys. Rev. C 58, 1804 (1998). [71] P. T. H. Pang, T. Dietrich, M. W. Coughlin, M. Bulla, I. Tews, M. Almualla, T. Barna, W. Kiendrebeogo, N. Kunert, G. Mansingh et al., NMMA: A nuclear-physics and multi- messenger astrophysics framework to analyze binary neutron star mergers, arXiv:2205.08513. [72] A. Bauswein, T. W. Baumgarte, and H.-T. Janka, Prompt Merger Collapse and the Maximum Mass of Neutron Stars, Phys. Rev. Lett. 111, 131101 (2013). [73] J. Wofford, A. B. Yelikar, H. Gallagher, E. Champion, D. Wysocki, V. Delfavero, J. Lange, C. Rose, V. Valsan, S. Morisaki, J. Read, C. Henshaw, and R. O’Shaughnessy, Improving performance for gravitational-wave parameter inference with an efﬁcient and highly-parallelized algorithm, Phys. Rev. D 107, 024040 (2023). [74] A. Perego, D. Radice, and S. Bernuzzi, AT 2017gfo: An anisotropic and three-component kilonova counterpart of GW170817, Astrophys. J. Lett. 850, L37 (2017). [75] M. Nicholl, B. Margalit, P. Schmidt, G. P. Smith, E. J. Ridley, and J. Nuttall, Tight multimessenger constraints on the neutron star equation of state from GW170817 and a forward model for kilonova light-curve synthesis, Mon. Not. R. Astron. Soc. 505, 3016 (2021). [76] M. Bulla, Possis: Predicting spectra, light curves, and po- larization for multidimensional models of supernovae and kilonovae, Mon. Not. R. Astron. Soc. 489, 5037 (2019). [77] K. Kawaguchi, S. Fujibayashi, M. Shibata, M. Tanaka, and S. Wanajo, A low-mass binary neutron star: Long-term ejecta evolution and kilonovae with weak blue emission, Astrophys. J. 913, 100 (2021). [78] E. Troja, L. Piro, H. van Eerten, R. T. Wollaeger, M. Im, O. D. Fox, N. R. Butler, S. B. Cenko, T. Sakamoto, C. L. Fryer et al., The x-ray counterpart to the gravitational-wave event GW170817, Nature (London) 551, 71 (2017). [79] S. Valenti, D. J. Sand, S. Yang, E. Cappellaro, L. Tartaglia, A. Corsi, S. W. Jha, D. E. Reichart, J. Haislip, and V. Kouprianov, The discovery of the electromagnetic counterpart of GW170817: Kilonova AT 2017gfo/DLT17ck, Astrophys. J. Lett. 848, L24 (2017). [80] M. M. Kasliwal, E. Nakar, L. P. Singer, D. L. Kaplan, D. O. Cook, A. Van Sistine, R. M. Lau, C. Fremling, O. Gottlieb, J. E. Jencson et al., Illuminating gravitational waves: A con- cordant picture of photons from a neutron star merger, Science 358, 1559 (2017). [81] B. J. Shappee, J. D. Simon, M. R. Drout, A. L. Piro, N. Morrell, J. L. Prieto, D. Kasen, T. W. S. Holoien, J. A. Kollmeier, D. D. Kelson et al., Early spectra of the gravita- tional wave source GW170817: Evolution of a neutron star merger, Science 358, 1574 (2017). [82] Y. Utsumi, M. Tanaka, N. Tominaga, M. Yoshida, S. Barway, T. Nagayama, T. Zenko, K. Aoki, T. Fujiyoshi, H. Furusawa et al., J-GEM observations of an electromagnetic counterpart to the neutron star merger GW170817, Publ. Astron. Soc. Jpn. 69, 101 (2017). [83] K. P. Mooley, A. T. Deller, O. Gottlieb, E. Nakar, G. Hallinan, S. Bourke, D. A. Frail, A. Horesh, A. Corsi, and K. Hotokezaka, Superluminal motion of a relativistic jet in the neutron-star merger GW170817, Nature (London) 561, 355 (2018). [84] P. A. Evans, S. B. Cenko, J. A. Kennea, S. W. K. Emery, N. P. M. Kuin, O. Korobkin, R. T. Wollaeger, C. L. Fryer, K. K. Madsen, F. A. Harrison et al., Swift and NuSTAR observations of GW170817: Detection of a blue kilonova, Science 358, 1565 (2017). [85] E. Troja, H. van Eerten, B. Zhang, G. Ryan, L. Piro, R. Ricci, B. O’Connor, M. H. Wieringa, S. B. Cenko, and T. Sakamoto, 013168-14 SURROGATE LIGHT CURVE MODELS FOR KILONOVAE … PHYSICAL REVIEW RESEARCH 5, 013168 (2023) A thousand days after the merger: Continued x-ray emis- sion from GW170817, Mon. Not. R. Astron. Soc. 498, 5643 (2020). [86] F. Foucart, M. D. Duez, R. Haas, L. E. Kidder, H. P. Pfeiffer, M. A. Scheel, and E. Spira-Savett, General relativistic simulations of collapsing binary neutron star mergers with Monte-Carlo neutrino transport, arXiv:2210.05670. [87] E. Troja, C. L. Fryer, B. O. Connor, G. Ryan, S. Dichiara, A. Kumar, N. Ito, R. Gupta, R. T. Wollaeger, J. P. Norris et al., A nearby long gamma-ray burst from a merger of compact objects, Nature (London) 612, 228 (2022). [88] J. Yang, S. Ai, B.-B. Zhang, B. Zhang, Z.-K. Liu, X. I. Wang, Y.-H. Yang, Y.-H. Yin, Y. Li, and H.-J. Lü, A long-duration gamma-ray burst with a peculiar origin, Nature (London) 612, 232 (2022). [89] J. C. Rastinejad, B. P. Gompertz, A. J. Levan, W.-fai Fong, M. Nicholl, G. P. Lamb, D. B. Malesani, A. E. Nugent, S. R. Oates, N. R. Tanvir et al., A kilonova following a long- duration gamma-ray burst at 350 mpc, Nature (London) 612, 223 (2022). [90] A. Mei, B. Banerjee, G. Oganesyan, O. S. Salaﬁa, S. Giarratana, M. Branchesi, P. D’Avanzo, S. Campana, G. Ghirlanda, S. Ronchini et al., Gigaelectronvolt emission from a compact binary merger, Nature (London) 612, 236 (2022). [91] Y.-W. Yu, B. Zhang, and H. Gao, Bright “merger-nova” from the remnant of a neutron star binary merger: A signature of a newly born, massive, millisecond magnetar, Astrophys. J. Lett. 776, L40 (2013). [92] B. D. Metzger and A. L. Piro, Optical and x-ray emission from stable millisecond magnetars formed from the merger of binary neutron stars, Mon. Not. R. Astron. Soc. 439, 3916 (2014). [93] I. Arcavi, The ﬁrst hours of the GW170817 kilonova and the importance of early optical and ultraviolet observations for constraining emission models, Astrophys. J. Lett. 855, L23 (2018). [94] T. Matsumoto, K. Ioka, S. Kisaka, and E. Nakar, Is the macronova in GW170817 powered by the central engine? Astrophys. J. 861, 55 (2018). [95] L. Piro, E. Troja, B. Zhang, G. Ryan, H. van Eerten, R. Ricci, M. H. Wieringa, A. Tiengo, N. R. Butler, S. B. Cenko et al., A long-lived neutron star merger remnant in GW170817: Constraints and clues from x-ray observations, Mon. Not. R. Astron. Soc. 483, 1912 (2019). [96] S.-Z. Li, L.-D. Liu, Y.-W. Yu, and B. Zhang, What powered the optical transient AT2017gfo associated with GW170817? Astrophys. J. Lett. 861, L12 (2018). [97] R. T. Wollaeger, C. L. Fryer, C. J. Fontes, J. Lippuner, W. T. Vestrand, M. R. Mumpower, O. Korobkin, A. L. Hungerford, and W. P. Even, Impact of pulsar and fallback sources on multifrequency kilonova models, Astrophys. J. 880, 22 (2019). [98] S. Ai and B. Zhang, On the binary neutron star post-merger magnetar origin of XRT 210423, Astrophys. J. Lett. 915, L11 (2021). [99] M. Bulla, The critical role of nuclear heating rates, thermal- ization efﬁciencies, and opacities for kilonova modelling and parameter inference, Mon. Not. R. Astron. Soc. 520, 2558 (2023). [100] S. Rosswog and O. Korobkin, Heavy elements and electro- magnetic transients from neutron star mergers, Ann. Phys., 2200306 (2022). 013168-15	null
10.1017_jfm.2023.90.pdf	Data availability statement. The codes used in this study are openly available in GitHub at https://github. com/Maplenormandy/qg-edgeofchaos.	Data availability statement. The codes used in this study are openly available in GitHub at https://github. com/Maplenormandy/qg-edgeofchaos .	J. Fluid Mech. (2023), vol. 958, A28, doi:10.1017/jfm.2023.90 Rossby waves past the breaking point in zonally dominated turbulence Norman M. Cao† Courant Institute of Mathematical Sciences, New York University, New York, NY 10012, USA (Received 10 August 2022; revised 21 January 2023; accepted 25 January 2023) The spontaneous emergence of structure is a ubiquitous process observed in ﬂuid and plasma turbulence. These structures typically manifest as ﬂows which remain coherent over a range of spatial and temporal scales, resisting statistically homogeneous description. This work conducts a computational and theoretical study of coherence in turbulent ﬂows in the stochastically forced barotropic β-plane quasi-geostrophic equations. These equations serve as a prototypical two-dimensional model for turbulent ﬂows in Jovian atmospheres, and can also be extended to study ﬂows in magnetically conﬁned fusion plasmas. First, analysis of direct numerical simulations demonstrates that a signiﬁcant fraction of the ﬂow energy is organized into coherent large-scale Rossby wave eigenmodes, comparable with the total energy in the zonal ﬂows. A characterization is given for Rossby wave eigenmodes as nearly integrable perturbations to zonal ﬂow Lagrangian trajectories, linking ﬁnite-dimensional deterministic Hamiltonian chaos in the plane to a laminar-to-turbulent ﬂow transition. Poincaré section analysis reveals that Lagrangian ﬂows induced by the zonal ﬂows plus large-scale waves exhibit localized chaotic regions bounded by invariant tori, manifesting as Rossby wave breaking in the absence of critical layers. It is argued that the surviving invariant tori organize the large-scale ﬂows into a single temporally and zonally varying laminar ﬂow, suggesting a form of self-organization and wave stability that can account for the resilience of the observed large-amplitude Rossby waves. Key words: geostrophic turbulence, jets, wave–turbulence interactions 1. Introduction In seeming contradiction to the conception of turbulence as a mixing phenomenon, turbulence often exhibits the spontaneous emergence of structured ﬂows. A prototypical † Email address for correspondence: [email protected] © The Author(s), 2023. Published by Cambridge University Press. This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/ licenses/by/4.0), which permits unrestricted re-use, distribution and reproduction, provided the original article is properly cited. 958 A28-1 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h N.M. Cao example is the formation of large-scale coherent vortices from small-scale stirring in two-dimensional Navier–Stokes turbulence, a process due to the inverse cascade posited by Kraichnan (1967). This structure can take many forms in other systems, but of particular interest to this work are the banded, zonally directed shear ﬂows known as zonal ﬂows, found ubiquitously in planetary atmospheres and magnetically conﬁned fusion plasmas (Diamond et al. 2005; Galperin & Read 2019). Prominent examples of zonal ﬂows occur in both natural and engineered systems, such as the majestic alternating bands of Jupiter (Vasavada & Showman 2005) and the sheared E × B ﬂows that form in high-conﬁnement-mode (H-mode) fusion plasmas (Wagner 2007). More than just an aesthetic quirk, zonal ﬂows are known to play a crucial in ﬂuid and plasma ﬂows which are rendered role in the regulation of transport quasi-two-dimensional by strong rotation or magnetization. Zonally directed ﬂows can accumulate signiﬁcant amounts of energy from the turbulent ﬂows, acting to regulate the turbulent mixing by limiting the extent of eddies and other structures transverse to the sheared zonal ﬂows; see for example the reviews in Galperin & Read (2019) and Diamond et al. (2005). Despite the importance of these ﬂows in climate and fusion modelling, many open questions remain about the dynamics of zonally dominated turbulence. Characterizing the dynamics of zonally dominated ﬂows requires an understanding of how the turbulence organizes over long time scales outside of statistical equilibrium. For example, Jupiter’s bands were observed to be nearly steady for centuries before suddenly losing one of the jets in the period 1939–1940 (Rogers 2009). Similarly, global gyrokinetic simulations and experimental evidence suggest that zonal ﬂows in tokamak plasmas also organize into long-lived discrete bands, which are only occasionally interrupted by mesoscale ‘avalanches’ (Dif-Pradalier et al. 2015, 2017). Recent work in Bouchet, Rolland & Simonnet (2019) and Simonnet, Rolland & Bouchet (2021) has suggested that zonal ﬂow states are in fact metastable, ﬁnding predictable noise-induced transition paths between different numbers of zonal bands in stochastically forced turbulent ﬂows. This work focuses on the organization of turbulence in the stochastically forced barotropic β-plane quasi-geostrophic model, referred to here as the QG model. This is a two-dimensional model for a single-layer rapidly rotating ﬂow in the presence of a background gradient of potential vorticity (PV) called the β effect (see Salmon 1998). The restoring force provided by the β effect supports the propagation of Rossby waves, which will prove to be crucial to the large-scale ﬂow organization. This model has long been used as a prototype for jet formation in the atmospheres of Jovian planets (Vasavada & Showman 2005). Furthermore, this model shares essential features with the Charney–Hasegawa–Mima equations, another prototypical model used to understand the dynamics of drift-Rossby wave turbulence relevant to various plasma and ﬂuid systems (Connaughton, Nazarenko & Quinn 2015). Zonal ﬂows are typically not driven directly by background gradients of density or pressure, instead requiring a net transfer of energy from non-zonal components of the ﬂow to sustain themselves against dissipation. This transfer process is frequently identiﬁed with the ‘inverse cascade of energy’ in two-dimensional turbulence introduced by Kraichnan (1967), and later Rhines (1975) in the context of atmospheric ﬂows. A great deal of work has gone into extending the ideas of the cascade to QG and related models; see for example the review in Connaughton et al. (2015) and references therein. However, two-dimensional cascades relevant to the QG model are generally found to be ‘non-local’, meaning that large-scale ﬂows can directly interact with small-scale ﬂuctuations without proceeding through intermediate scales in a self-similar cascade. Thus, this work considers 958 A28-2 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Rossby waves past the breaking point in turbulence the following question: once the large-scale ﬂows have formed, in lieu of a self-similar cascade picture, how are their dynamics and inﬂuence on smaller scales best described? Describing the dynamics of the large scales from a purely statistical perspective is challenging, as the coherence and memory of structured ﬂows impart non-trivial spatiotemporal correlations into the turbulence statistics. Finite-size effects also must be taken into account, which makes the application of scale separation arguments difﬁcult. As a concrete illustration of these challenges, consider the principle of spatially inhomogeneous mixing leading to ‘potential vorticity staircases’, used to understand the resilience of zonal jets to turbulence (Baldwin et al. 2007; Dritschel & McIntyre 2008). Spatial inhomogeneity implies that the two-point correlation function of the velocity will depend not only on the separation of the points, but also on the centre of mass of the points. Thus, even a complete description of the anisotropic Fourier energy spectrum E(kx, ky) of the ﬂows cannot describe this effect. This can be viewed as a form of spontaneous symmetry breaking, where the metastable banded zonal states lack the translational symmetry of the original system. The aim of this work is to study such departures from statistical homogeneity in a ﬁnite-size turbulent system primarily exhibiting large-scale wave behaviour atop a zonal ﬂow background. In the parameter regimes studied, it is shown that the strong zonal ﬂows act as a ‘waveguide’ supporting discrete Rossby wave eigenmodes, which have properties that distinguish them from Fourier modes. Surprisingly, even though the waves can contain energy comparable with the zonal ﬂows and induce perturbations to the ﬂow velocity comparable with their phase velocities, they retain a coherent linear character. To investigate this, the statistics of turbulence will be shown to correlate with deterministic properties of the Lagrangian ﬂow induced by the zonal ﬂows plus coherent waves. To study these wave-induced Lagrangian ﬂows, the discrete and coherent character of the waves will allow the usage of Poincaré section techniques. These techniques have been previously applied in a number of geophysical and plasma contexts (see e.g. Chirikov 1979; Behringer, Meyers & Swinney 1991; Pierrehumbert 1991; Del-Castillo-Negrete & Morrison 1993; Haynes, Poet & Shuckburgh 2007). Crucially, these techniques demonstrate that the wave-induced ﬂows can be understood as a single zonally and temporally varying laminar shear ﬂow that is organized by ‘invariant-torus-like Lagrangian coherent structures’ (LCSs) (Beron-Vera et al. 2010). It is also shown how this laminar ﬂow does not distort itself due to self-advection under the Poincaré map, which is interpreted as form of nonlinear wave stability that can account for the resilience of the observed large-amplitude coherent Rossby waves. The core arguments are organized in the paper as follows. First in § 2, the model equations are introduced and studied with direct numerical simulation (DNS). In a regime of weak forcing and damping relevant to turbulent ﬂows, it is shown that a signiﬁcant fraction of ﬂow energy, comparable with the zonal ﬂows, is organized into coherent large-scale Rossby wave eigenmodes. Next in § 3, a theorem is given for the Liouville integrability of Lagrangian ﬂows of time-periodic solutions to the QG equations, and a connection is given to laminar ﬂows. Flow conﬁgurations consisting of a superposition of Rossby wave eigenmodes atop zonal ﬂows are singled out as ‘optimally near-integrable’ perturbations to the integrable zonal ﬂow Lagrangian dynamics. Finally in § 4, the wave-induced Lagrangian ﬂows observed in the DNS are analysed using Poincaré sections. Lagrangian chaos induced by the wave ﬂows is shown to closely correlate to inhomogeneous mixing in the DNS, and the survival of invariant tori is linked to a certain nonlinear wave stability result. This leads to a proposal that the large-scale ﬂow dynamics in the observed regimes of Rossby wave turbulence self-organizes into ‘nearly-integrable Rossby waves past the breaking point in zonally dominated turbulence’. 958 A28-3 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 2. Identifying Rossby waves in simulations N.M. Cao 2.1. Model equations and simulation set-up This work focuses on the stochastically forced barotropic β-plane QG model (see Salmon 1998; Connaughton et al. 2015). The model equations describe the two-dimensional advection of a scalar q = q(x, y, t), the relative vorticity, by a non-divergent ﬂow which is in approximate geostrophic balance: ∂tq + {ψ, q + βy} = −αq − νh(−∇2)hq + (2.1) √ αη, (2.2) Here ψ is the streamfunction, β is a constant approximating the gradient of the Coriolis parameter and the Poisson bracket is deﬁned as follows: q = ∇2ψ. {f , g} := ∂xf ∂yg − ∂yf ∂xg. (2.3) In this model the PV q + βy, the sum of the relative plus planetary components of the vorticity, is a scalar material invariant, meaning its value is conserved along ﬂuid parcel trajectories in the absence of dissipation. Note that this system can be thought of as the Charney–Hasegawa–Mima equations in the limit of inﬁnite Rossby deformation radius or gyroradius. However, the inﬁnite gyroradius limit is not a physically relevant limit for magnetized plasmas, so the QG system is best considered in the context of atmospheric ﬂows. The boundary conditions are taken to be doubly-periodic in a square box D of side lengths L × L. In the absence of forcing and dissipation, these equations will conserve the quadratic invariants of energy E and enstrophy Q, given by (cid:2) E = − 1 2L2 Q = 1 2L2 (cid:4)ψq(cid:5) dx dy, D (cid:2) (cid:4)q2(cid:5) dx dy. D (2.4) (2.5) (2.6) The forcing is chosen to be white-in-time with a given correlation function: (cid:4)η(x, y, t)η(x (cid:6), y (cid:6), t (cid:6))(cid:5) = C(x − x (cid:6), y − y (cid:6))δ(t − t (cid:6)). Choosing mass units such that the ﬂuid density is 1, the average energy input rate per unit mass ε is given by (2.7) In the absence of (hyper)viscosity, the average kinetic energy of the ﬂow will be (cid:4)E(cid:5) = ε/α. ε = 1 2 αL2(∇−2C)(0). The DNS here focus on weakly damped and forced parameter regimes where inertial and Coriolis forces are expected to dominate, relevant for a Rossby wave turbulence regime. Parameters were chosen to be comparable with those used in previous studies of the Jovian atmosphere by Bouchet et al. (2019). To non-dimensionalize the equations, length units are chosen such that L = 2π and time units are chosen to ﬁx β = 8. In this case, the form of the equations as written does not change. The forcing was chosen to be uniform in an annular ring kf ∈ [14, 15], which is of small scale compared with the box size. The equations were solved pseudospectrally using the Dedalus framework (Burns et al. 2020) using the built-in second-order modiﬁed Crank–Nicolson Adams–Bashforth 958 A28-4 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Rossby waves past the breaking point in turbulence Name Resolution Box size Beta Friction Hyperviscosity Normalized forcing rate Symbol Nx × Ny L × L β α ν2 ε/α Value 2048 × 2048 2π × 2π 8 1.2 × 10−3 1.5 × 10−8 1 2π2 (case 2) 1 (case 1); Table 1. Parameters used in the DNS. (a) Case 1 3 2 1 0y –1 –2 –3 20 10 0 –10 –20 –30 –2 0 x 2 q + βy Case 2 3 2 1 0 –1 –2 –3 (b) 3 2 1 0 –1 –2 –3 –2 2 0 U(y) –2 0 x 3 2 1 0 –1 –2 –3 20 10 0 –10 –20 2 q + βy 0.5 0 U(y) Figure 1. Representative snapshots from the DNS of the PV q + βy and the zonally averaged zonal ﬂow proﬁle U( y) for (a) case 1 and (b) case 2. To emphasize the banded structure, the PV is displayed with a cyclic colour scheme that is shifted separately for each case to align with the observed zonal ﬂows. √ adaptive time-stepping scheme, which treats the linear terms implicitly and the nonlinear terms explicitly. The forcing was applied as a time-dependent function normalized by δt, where δt is the time step. De-aliasing was applied using the 2/3 rule to deal with 1/ the quadratic nonlinearity. Two cases were studied: case 1 is in a parameter regime studied by Bouchet et al. (2019) but with higher resolution and lower-order hyperviscosity, while case 2 is a similar case with greatly reduced forcing. These changes cleanly separate the large-scale discrete waves from the dissipation scales, ensuring that any observed effects of wave discreteness are not due to a premature truncation of the dynamic range of the turbulence. The exact values of parameters used are shown in table 1. After about three frictional relaxation times t = 2400 ≈ 3/α, the simulations reach a quasi-stationary state. Some representative snapshots from this state are shown in ﬁgure 1. This work primarily focuses on two sequences of 256 snapshots, one from each case taken after reaching this quasi-stationary state, spaced evenly apart with (cid:10)t = 0.25. The number and spacing of snapshots were chosen to resolve the linear oscillation frequency of the Rossby waves over several cycles. The length of the interval of observation [2400, 2463.75] is much shorter than a frictional relaxation time, and the zonal ﬂows did not evolve signiﬁcantly over this time interval. One-dimensional energy spectra E(k), energy ﬂuxes ΠE(k) and enstrophy ﬂuxes ΠQ(k) are shown in ﬁgure 2. Following Frisch (1995), these can be deﬁned in terms of the low-pass ﬁlter operator Pk which ﬁlters out components with wavenumber greater 958 A28-5 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h (a) Case 1 100 E(k) 10–4 10–8 ) 3 – 0 1 × ( ) k ( E Π 2 1 0 –1 –2 N.M. Cao (b) E(k) ) 4 – 0 1 × ( ) k ( E Π Case 2 k –5/3 k –4 100 10–3 10–6 10–9 1.0 0.5 0 –0.5 –1.0 k –5/3 k –4 4 2 0 –2 –4 ) 1 – 0 1 × ( ) k ( Q Π )2 2 – 0 1 × ( ) k ( Q Π 1 0 –1 –2 100 101 102 103 100 101 102 103 k k Figure 2. Top row: time-averaged one-dimensional energy spectra E(k) from (a) case 1 and (b) case 2. Bottom row: corresponding nonlinear energy ﬂux ΠE(k) (blue, solid) and enstrophy ﬂux ΠQ(k) (green, dashed). The forcing scale is marked with a red vertical line. than k: (cid:3) E(k) := d dk (cid:4) (cid:4)qPk[ψ](cid:5) dx dy , (cid:2) D − 1 2L2 (cid:2) ΠE(k) := − 1 L2 (cid:2) ΠQ(k) := 1 L2 (cid:4){ψ, q}Pk[ψ](cid:5) dx dy, D (cid:4){ψ, q}Pk[q](cid:5) dx dy. D (2.8) (2.9) (2.10) In practice, (cid:4)·(cid:5) is computed by time-averaging over all snapshots. For scales with k > kf , there is a ﬂux of enstrophy from the forcing scales to higher wavenumbers with relatively little ﬂux of energy, reminiscent of a direct cascade of enstrophy. The energy spectra appear to follow an approximately k−4 scaling, consistent with observations in for example Danilov & Gryanik (2004) and Scott & Dritschel (2012) of energy spectra in the strong zonal jet regime. Note that this k−4 is not necessarily indicative of a self-similar cascade, and has been proposed to be a result of spatially localized ‘sawtooth’ behaviour in the PV proﬁle (Danilov & Gurarie 2004). For scales with k < kf , there is a ﬂux of energy from the forcing scales to lower wavenumbers with relatively little transfer of enstrophy, reminiscent of an inverse cascade of energy. However, due to wavenumber discreteness, it is unclear if there is a single self-similar scaling law that describes the energy spectrum in this range. 2.2. Data-driven identiﬁcation of Rossby waves It has been observed that β-plane QG turbulence with an inﬁnite deformation radius typically exhibits a strong wave-like character above the Rhines scale k (cid:2) kRh := (β/2urms)1/2; see for example the studies and discussion in Rhines (1975), Shepherd (1987), Sukoriansky, Dikovskaya & Galperin (2008) and Suhas & Sukhatme (2015). The basic argument is that in the absence of any background ﬂows, the Rossby wave dispersion 958 A28-6 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Rossby waves past the breaking point in turbulence relation for Fourier modes at wavenumber k is given by ω0,k = − βkx k2 . (2.11) Meanwhile, the time scale for turbulent motions at a characteristic scale k should scale as 1/τturb,k ∼ kurms. The Rhines scale gives a rough estimate for what scales k linear effects should start to dominate over nonlinear effects, ω0,kτturb,k ∼ (k/kRh)−2. For the DNS snapshots, the Rhines scale can be computed directly using the time- and space-averaged root mean square velocities, resulting in kRh ≈ 1.46 for case 1 and kRh ≈ 6.5 for case 2. Noting the power-law decay of the energy spectrum, this suggests that a signiﬁcant fraction of turbulent ﬂow energy is concentrated into modes where the Rossby wave dispersion relation can play a signiﬁcant role. A natural question that arises is how to quantify how ‘good’ linear Rossby waves are at describing the observed ﬂows in the DNS. One way to do this in a data-driven manner is through proper orthogonal decomposition (POD) (Berkooz 1993), also known as the method of empirical orthogonal functions. Brieﬂy summarizing, given an observable ﬁeld w(x, t), POD ﬁnds a ranked set of r orthonormal spatial basis functions {w1(x), . . . wr(x)} and corresponding orthonormal temporal basis functions {a1(t), . . . , ar(t)}. The approximation ˆwn of w by the ﬁrst n ≤ r basis functions, ˆwn(x, t) := n(cid:5) k=1 σkak(t)wk(x), (2.12) will be ‘optimal’ in the sense that it minimizes the squared residual integrated over the spatial domain D and some ﬁxed time interval T: (cid:2) (cid:2) (cid:11)w − ˆwn(cid:11)2 := \|w − ˆwn\|2 d2x dt. (2.13) T D This minimization is taken over all possible sets of r orthonormal spatial and temporal basis functions. In practice, the observable is sampled over a regular spatial and temporal grid so the integrals are replaced by sums, and r is ﬁnite so the basis functions span only a subspace of the set of all possible functions. To apply POD to the DNS snapshots, the observable w = (−∇−2)1/2q is used so the (cid:6) standard inner product gives the energy norm for each snapshot E ∝ D w2 d2x. The resultant POD modes will be ranked in terms of their contribution to the time-integrated energy of the snapshots. These modes are ‘optimal’ in the sense that the projection of the snapshot data to the ﬁrst n modes will minimize the energy in the residual. The streamfunctions ψ = (−∇−2)−1/2w for a few selected POD modes are shown in ﬁgure 3, and the total energy captured by the ﬁrst n POD modes is shown in ﬁgure 4. Analysing the POD modes in case 1 ﬁrst, the zonal ﬂows are primarily captured by POD mode 0 and account for about 72 % of the total time-averaged energy. The next few POD modes form sine/cosine pairs in their temporal evolution and x (zonal) variation, showing a high degree of spatial and temporal coherence. The POD modes 1–4 form two pairs which account for about 24 % of the time-averaged energy, or about 86 % of the non-zonal energy. Higher POD modes are increasingly incoherent in both space and time, but make up only around 4 % of the time-averaged energy. Case 2 has a similar breakdown of energy, where zonal ﬂows account for 56 % of the time-averaged energy, and the next three pairs of POD modes account for 74 % of the non-zonal energy. Thus, despite the turbulent nature of the ﬂow, the vast majority of energy in both cases is organized into POD modes which exhibit regular structure in both space and time. 958 A28-7 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h x 0 Mode 0 2 Case 1 Mode 1 Mode 37 (a) –2 y 2 0 –2 N.M. Cao (b) 2 0 –2 Case 2 ψ(x, y) Mode 0 Mode 1 Mode 37 a(t) 0 20 40 t 60 0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60 Figure 3. A selection of POD modes from (a) case 1 and (b) case 2 displaying the streamfunction ψ(x, y) and time traces a(t) associated with each mode. Note that the POD modes and time traces are normalized, so the units are arbitrary. t o t E n E / 1.0 0.8 0.6 Case 1 Case 2 100 101 ntrunc Figure 4. Cumulative energy fraction En/Etot captured by the ﬁrst n POD modes for case 1 (blue, solid) and case 2 (orange, dashed). 102 2.3. Eigenmode character of Rossby waves The strong wave-like character of the spatiotemporal structures identiﬁed by POD suggests comparison with the Rossby wave dispersion relation frequencies ω0,k. However, the presence of nearly steady large-amplitude zonal ﬂows also suggests comparison of the POD modes with the eigenfrequencies and eigenmodes of the QG equations linearized around a zonal ﬂow background. For a reference zonal ﬂow state U( y), the equations of motion (2.1) and (2.2) in the absence of forcing and dissipation can be linearized and Fourier transformed in the symmetry direction x to derive an eigenvalue equation: (cid:6)(cid:6)( y))∇−2]˜q = iωeig ˜q. ikx[U( y) + (β − U (2.14) In the absence of background ﬂows U( y) = 0, the eigenfrequencies reduce to the dispersion relation frequencies ω0,k and the eigenmodes reduce to Fourier modes in both y and x. For the following analysis, the zonally averaged zonal ﬂow proﬁle time-averaged over the snapshots with no additional smoothing is used as a reference ﬂow proﬁle. The reference proﬁle satisﬁes the Rayleigh–Kuo criterion β − U(cid:6)(cid:6)( y) > 0 everywhere, ensuring that all eigenmodes will be neutrally stable (Kuo 1949). Figure 5 shows a comparison between selected pairs of POD modes with their corresponding dispersion relation frequencies, and corresponding Rossby wave eigenmodes and eigenfrequencies. The eigenmodes well capture the temporal frequency and spatial structure of the POD modes, which deviate from the dispersion relation. In particular, these eigenmodes have a strong spatial localization of ˜q( y) to regions of large 958 A28-8 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Rossby waves past the breaking point in turbulence (a) Eigenmode POD modes ψ(x, y) Case 1, modes 1 + 2 kx = 1, ky = 1 ω eig ª –4.33 POD ª –4.32 ω ω 0, k = –4.0 (b) Case 2, modes 1 + 2 kx = 1, ky = 3 ω eig ª –0.87 POD ª –0.86 ω 0, k = –0.8 ω (c) Case 2, modes 5 + 6 kx = 3, ky = 1 eig ª –2.53 ω POD ª –2.51 ω 0, k = –2.4 ω 2 0 y –2 2 0 y –2 2 0 y –2 0 5 0 2 0 –0.25 q˜ Figure 5. Streamfunction ψ(x, y) for eigenmodes and corresponding pairs of POD modes. The text on the left-hand side gives a comparison of the POD frequency with the eigenfrequency and dispersion relation frequency. The rightmost plots show a comparison of the eigenfunction envelope ˜q( y) for the eigenmode (orange, dashed) and the POD modes (blue, solid) at the average amplitudes observed in the DNS. PV gradient associated with the reference ﬂow ¯q(cid:6)( y) + β := −U(cid:6)(cid:6)( y) + β, giving them a somewhat ‘interfacial’ rather than ‘bulk’ character. This localization corresponds to a strong broadening in ky at ﬁxed kx, characteristic of ‘zonons’ identiﬁed in simulations of QG turbulence by Sukoriansky et al. (2008). The observation that the ﬁrst few POD modes correspond very closely to Rossby wave eigenmodes suggests that the QG dynamics linearized around the reference zonal ﬂow state plays a signiﬁcant role in organizing the large-scale coherent ﬂows. However, it is not immediately obvious from non-dimensional measures of wave amplitude why the linear character of the Rossby wave dynamics should persist in the DNS. For example, one estimate of the ratio of nonlinear to linear time scales for the waves is given by τnl/τlin ∼ uwave/uph. Here, uwave is the maximum zonally directed velocity induced by the wave and uph is the zonal component of the wave phase velocity. For the eigenmode with the largest amplitude, uwave/uph ≈ 0.24 in case 1 and uwave/uph ≈ 0.36 in case 2. The phase and amplitude of these modes remain coherent for much longer times than this simple estimate would suggest. This raises a key question of how ‘large amplitude’ can be reconciled with ‘linear coherence’ for the observed Rossby wave eigenmodes. 958 A28-9 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h N.M. Cao 2.4. Isolating linearly coherent Rossby waves While POD modes are orthogonal under the energy inner product, the Rossby wave eigenmodes are not. Thus, isolating the coherent Rossby wave ﬂow component requires expanding the snapshot data in the numerically computed eigenmode basis, then determining which of the eigenmodes has a strong linear character. Fourier transforming of the snapshot data in x and expanding each kx component in the eigenmode basis results in a complex eigenmode amplitude aj(tk) for each eigenmode j and snapshot time tk. To provide a quantitative cutoff between ‘linearly coherent’ and ‘not linearly coherent’ Rossby waves, the following coherence statistic is used: r2 min ( j) := min s∈S max β∈C ⎢ ⎢ ⎢ ⎢ ⎢ ⎣ ⎡ ⎤ N−s(cid:5) (cid:10) (cid:10) (cid:10)2 (cid:10)aj(tk+s) − βaj(tk) k=1 1 − (cid:10) (cid:10) (cid:10)2 (cid:10)aj(tk+s) N−s(cid:5) k=1 ⎥ ⎥ ⎥ ⎥ ⎥ ⎦ . (2.15) Here N is the number of snapshots and S = {1, 2, . . . , smax} is a set of time delays. The inner max of (2.15) is the r2 statistic of a least-squares ﬁt of the linear model ˆaj(tk+s) = βaj(tk), and can also be interpreted as an estimator for the squared lag-s autocorrelation coefﬁcient of a time-stationary signal. A signal with purely linear oscillatory dynamics described by aj(tk+s) = exp(−iωeig,js(cid:10)t)aj(tk) would have r2 = 1. The outer min of (2.15) selects the minimum r2 over the set of time delays. Choosing smax = 64, r2 min measures whether or not the signal remains linearly coherent over time windows of up to smax(cid:10)t = 16. Signals with r2 min close to 1 will only experience small ﬂuctuations in their amplitude and phase relative to a purely linear oscillatory signal, while signals with r2 min close to 0 experience relatively large ﬂuctuations. A plot of the r2 min statistic for all numerically computed eigenmodes with 1 ≤ kx ≤ 8 is > 0.4 is used to identify the coherent modes. These shown in ﬁgure 6. A cutoff of r2 coherent eigenmodes have zonally directed phase velocities which satisfy uph < U( y) everywhere, and hence correspond to non-singular eigenfunctions. Observing that uph ≈ −β/k2, modes with large negative phase velocities are seen to correspond to large-scale Rossby waves. These are referred to as ‘large-scale coherent modes’ in the following. min 3. Near-integrability of wave-induced Lagrangian ﬂows 3.1. Integrability and laminar ﬂow Following the text of Arnold (1989) on classical mechanics, a 2n-dimensional smooth continuous-time Hamiltonian system is said to be Liouville integrable if it has n independent, Poisson commuting ﬁrst integrals of motion, and furthermore the level sets of the integrals of motion are compact. The Liouville–Arnold theorem states that if a Hamiltonian system is Liouville integrable, then there exists a canonical transformation to action-angle coordinates in which the integrals of motion, including the Hamiltonian, depend only on the action coordinates. The action-angle coordinates are not necessarily global, and different regions of phase space may have different action-angle coordinates. One corollary of the existence of action-angle coordinates is that the level sets of the action variables (locally) foliate the phase space into tori which are invariant under the Hamiltonian ﬂow. In physical terms, a Hamiltonian system that is Liouville integrable has a phase space ﬂow which is ‘laminar’, with the ‘laminae’ consisting of the invariant tori. 958 A28-10 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h (a) n i 2m r (b) \| e v a w \| u p u s Rossby waves past the breaking point in turbulence Case 1 Case 2 –8 –6 –4 –2 0 2 1.00 0.75 0.50 0.25 0 1.00 0.75 0.50 0.25 0 1.00 0.75 0.50 0.25 0 0.3 0.2 0.1 0 –8 –6 –4 –2 0 2 –8 –6 –4 –2 0 2 –8 –6 –4 –2 0 2 uph Figure 6. (a) Eigenmode coherence r2 min versus zonally directed phase velocity uph for all eigenmodes with 1 ≤ kx ≤ 8. Green triangles denote the large-scale coherent modes used in the study. The red shaded region shows the range of the zonal ﬂows U( y). (b) Time-averaged eigenmode amplitude, measured by the maximum of the zonally directed velocity induced by the mode uwave, versus uph. Labelling is the same as in (a). uph In the absence of forcing and dissipation and with a constant Galilean shift c, the ideal QG PV advection equation can be written as ∂tq + {ψ + cy, q + βy} = 0. This can be thought of as a Hamiltonian evolution equation for q + βy, with time-dependent Hamiltonian ψ(x, y, t) + cy. Given a streamfunction ψ(x, y, t), (3.1) can be transformed into an autonomous Hamiltonian system by introducing a new momentum coordinate ξ conjugate to t and redeﬁning the Poisson bracket: (3.1) {f , g} := ∂xf ∂yg − ∂yf ∂xg + ∂tf ∂ξ g − ∂ξ f ∂tg. (3.2) With this new Poisson bracket, the advection equation then becomes a single Poisson commutation condition: {ψ(x, y, t) + cy − ξ, q(x, y, t) + βy} = { ˆψ(x, y, t, ξ ), ˆq(x, y, t)} = 0. (3.3) From (3.3), ˆψ and ˆq are seen to be independent integrals of motion for the dynamics generated by the Hamiltonian − ˆψ, which corresponds to ﬂow along Lagrangian ﬂow trajectories dx ds = {x, − ˆψ} = vx, dy ds = {t, − ˆψ} = 1. ⎫ ⎪⎪⎬ = {y, − ˆψ} = vy, ⎪⎪⎭ dt ds (3.4) In physical terms, the PV is a materially conserved scalar invariant, so the ﬂow will be conﬁned to contours of constant PV. This leads to the following theorem, originally reported in Brown & Samelson (1994), restated and proved here for clarity: 958 A28-11 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h N.M. Cao THEOREM 3.1 (Liouville integrability of time-periodic solutions). Given a smooth time-periodic solution q(x, y, t) = q(x, y, t + T) to the ideal QG equations (3.1) and (2.2), the Lagrangian ﬂow speciﬁed by (3.4) for z = (x, y, t, ξ ) ∈ T4 is integrable in the Liouville sense. Proof . We can identify ξ with ξ + Lξ such that the level sets of cy − ξ are compact. Then, since ψ and q are time-periodic, we can identify t with t + T so the 4-dimensional phase space (x, y, t, ξ ) can be identiﬁed with the 4-torus T4. Then, since ˆψ and ˆq are smooth, their level sets are compact, hence satisfying the requirements of the Liouville–Arnold (cid:3) theorem. In physical terms, the theorem implies that the Lagrangian ﬂow of smooth time-periodic solutions to the ideal QG equations is necessarily laminar due to the existence of a conserved scalar material invariant. This property follows from the principle of PV conservation, which can be seen as a general consequence of the particle relabelling symmetry of the Lagrangian formulation of the QG equations (Salmon 1988; Müller 1995). Such symmetries are also linked to Casimirs of functional Poisson brackets relevant to a variety of ﬂuid and plasma systems (Holm et al. 1985; Morrison 1998; Webb & Anco 2019). 3.2. Eigenmodes and near-integrability To establish the link between Rossby wave eigenmodes and near-integrability, note that time-independent zonal ﬂow states q = q0( y) satisfy the requirements of Theorem 3.1. The integrals of motion are ˆψ0 := ψ0 − ξ and ˆq0 := q0 + βy, which satisfy {ψ0( y) − ξ, q0( y) + βy} = 0. (3.5) Furthermore, if the Rayleigh–Kuo condition is satisﬁed everywhere, then the integrals of motion are independent over the entire domain, and hence the ﬂow is laminar globally. Now, consider periodic approximate solutions of the QG equations of the form Q = ˆq0( y) + (cid:16)q1(x, y, t), where (cid:16) is a formal perturbation parameter. These approximate solutions induce a Lagrangian ﬂow with Hamiltonian given by Ψ := ˆψ0( y, ξ ) + (cid:16)∇−2q1(x, y, t). (3.6) Since Q is no longer an exact solution to the equations of motion, the exact commutation condition (3.3) will no longer hold in general. Using the bilinearity of the Poisson bracket, Q can be shown to instead satisfy an approximate commutation condition: {Ψ, Q} = (cid:16)[{ψ0 − ξ, q1} + {∇−2q1, q0 + βy}] + (cid:16)2{∇−2q1, q1} =: (cid:16)R1 + (cid:16)2R2. To identify ﬂows Ψ which have the perturbed PV Q as an approximate integral of motion, the leading-order remainder term R1 is set to zero, which results in an equation that determines q1: (3.7) {ψ0 − ξ, q1} + {∇−2q1, q0 + βy} = [∂t + U∂x + (β − U (cid:6)(cid:6))∂x∇−2]q1 = 0. (3.8) Comparing (3.8) with (2.14) shows that the approximate integrability condition coincides with the linearized evolution equation describing Rossby wave eigenmodes. Thus, regular eigenfunctions are singled out as the perturbations to the zonal ﬂow dynamics which give both (i) approximate solutions to the equations of motion and 958 A28-12 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Rossby waves past the breaking point in turbulence (ii) nearly integrable perturbations to the zonal ﬂow Lagrangian dynamics. They are ‘optimal’ perturbations with respect to this formal perturbation expansion, since no O((cid:16)) terms remain in the approximate commutation condition (3.7). Note that the condition R1 = 0 can also be relaxed into a condition that R1 is small, corresponding to q1 which approximately satisfy the linear evolution equation. By appealing to Kolmogorov–Arnold–Moser (KAM) theory extended to non-twist systems (Delshams & Llave 2000), the near-integrability of eigenmodes suggests that the majority of invariant tori will survive ﬁnite-amplitude perturbation by the eigenmodes. In physical terms, the invariant tori in the unperturbed system correspond to contours of the PV ˆq0. These tori are material curves in the ﬂuid, meaning that ﬂuid parcels do not pass through them transversally. The KAM theory implies that under suitably small amplitude of perturbation, ‘many’ of these tori will experience only reversible deformations under the perturbed ﬂow. Tori with zonal velocities that are incommensurate with the wave phase velocities are the most robust, and can survive as transport barriers in the perturbed ﬂow. In many cases the surviving tori, known as ‘KAM tori’, are known to form sets of positive measure (ﬁnite area/volume). See Arnold (1989) for an introduction or de la Llave (2001) for a review of KAM theory. These transport barriers fall under the umbrella of LCSs, with most of the invariant tori corresponding to elliptic LCSs deﬁned in Haller (2015). The non-twist tori from the minima and maxima of the zonal ﬂow correspond to parabolic LCSs. These barriers are also called ‘invariant-torus-like Lagrangian coherent structures’, described in Beron-Vera et al. (2010). A key property of the elliptic LCSs is that rather than being isolated curves which separate regions of different ﬂow character, they instead form families that ﬁll space in regions of the same ﬂow topology, i.e. regions where a single continuous set of action-angle coordinates can be given. Note that the survival of separatrices to perturbation, typically corresponding to hyperbolic LCSs, can be studied through the lens of Melnikov theory (see e.g. Balasuriya, Jones & Sandstede 1998; Malhotra & Wiggins 1998). 4. Wave-induced Lagrangian ﬂows in turbulence 4.1. Construction of the Poincaré map Given a set of eigenmode amplitudes and relative phases (in the symmetry direction x), a wave-induced ﬂow ﬁeld can be constructed with streamfunction Ψ = ˆψ0 + ∇−2q1. Here ˆψ0( y, ξ ) = ψ0( y) − ξ is the reference zonal ﬂow streamfunction and q1(x, y, t) is the superposition of the Rossby wave eigenmodes. By taking the condition that R1 is small instead of zero, the quasi-periodic time variation of q1 can be approximated by a periodic one by replacing the eigenfrequencies with a rational frequency approximation Ω0ni + ckx,i ≈ ωeig,i. Here kx,i and ωeig,i are ﬁxed for each eigenmode i, while the other parameters Ω0 (the basic frequency), ni (an integer multiplier) and c (a Doppler shift) are ﬁtted to minimize an amplitude-weighted approximation error of the eigenfrequencies. Note that the earlier condition smax(cid:10)t ≈ 2π/Ω0 is a self-consistency requirement to ensure the eigenmodes remain linearly coherent over one basic period. In particular, eigenmodes with r2 min close to 1 will have amplitudes and relative phases which evolve very slowly compared to the the linear time scale 1/Ω0. For both cases 1 and 2, Ω0 ≈ 0.41, with the eigenfunctions typically completing between 1 and 10 oscillations within this period. Since this wave-induced ﬂow ﬁeld is time-periodic, Poincaré sections can be used to visualize and analyse the ﬂuid parcel trajectories. These sections are constructed by 958 A28-13 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h N.M. Cao initializing tracers at different spatial positions, evolving their positions under the ﬂow ﬁeld and then marking their position after each period T = 2π/Ω0 of the ﬂow ﬁeld. The map taking each tracer from its initial position to its position after one period is known as the Poincaré map. Poincaré section techniques have particular application to near-integrable Hamiltonian systems where they can indicate regions of phase space where invariant tori have survived perturbation. In the following, superpositions of the ‘large-scale coherent modes’ marked in ﬁgure 6 are used to construct periodic wave-induced ﬂow ﬁelds. For each individual DNS snapshot, eigenmode amplitudes and phases were extracted and used to construct a series of corresponding Poincaré sections. To identify the presence of chaotic orbits, an iterative Grassberger–Procaccia method is used to estimate the correlation dimension dC of the orbit of each tracer under the Poincaré map (see Rosenberg 2020). Orbits which trace out invariant tori and islands will have dC ≈ 1, while orbits which chaotically ﬁll space will have dC ≈ 2. 4.2. Localized PV mixing from wave-induced chaos Focusing ﬁrst on case 1, a representative Poincaré section is shown in ﬁgure 7(a). Visual inspection suggests that a large majority of invariant tori survive the perturbation by the waves, and that the chaotic orbits are mostly conﬁned to a particular region of space. To link the localization of chaos in the wave-induced passive tracer advection to spatially inhomogeneous mixing in the fully developed turbulence in the DNS, two different measures of mixing are considered here. The ﬁrst measure of mixing considered is the temporally and zonally averaged PV gradient ¯q(cid:6)( y) + β. Friction and viscosity tend to push ¯q(cid:6)( y) towards zero, pushing the the PV gradient towards the background value of β. Meanwhile, turbulent mixing tends to push the PV gradient towards zero. Figure 7(c) shows the PV gradient compared against the fraction of orbits which are chaotic at some y for cases 1 and 2. This fraction is computed by computing the average y of the particles, binning them into equally sized bins based on this average y, then computing the fraction of particles whose orbits have dC ≥ 1.5. In every region where there is a signiﬁcant fraction of chaotic orbits, the PV gradient is pushed below the background β, indicating a region of sustained turbulent mixing. Furthermore, in case 2 these chaotic regions and corresponding mixing regions are not in the centre of their respective broad westward ﬂow regions. The particular combination of Rossby waves produces an asymmetry in the chaotic regions that matches the asymmetry in the observed turbulent mixing regions. A second measure of mixing comes from the consideration that in two dimensions, enstrophy is typically understood to be anomalously dissipated by viscosity via the formation of small-scale structure (Boffetta & Ecke 2012). In real space, this small-scale structure can be formed by the chaotic advection of level set contours of the PV q + βy. The contours are stretched in directions with non-zero Lyapunov exponent, creating small-scale meanders and increasing the overall length of the contour. Eventually, the meanders reach a small enough scale that viscosity acts to reconnect and smooth the contours, limiting the growth of their length. To quantify this process in the DNS, the contours of q + βy which encircle the domain are identiﬁed, then their length is computed and time-averaged over all DNS snapshots (cid:19)q. For a pure zonal ﬂow ¯q( y) + βy, the undisturbed PV contours would be lines of constant y with the minimum possible length of (cid:19)q = L = 2π. 958 A28-14 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Rossby waves past the breaking point in turbulence (a) Case 1 dC 3 2 1 0y –1 –2 –3 3 2 1 0 –1 –2 –3 y –2 0 2 Case 2 –2 2 0 x (b) 3 2 1 0 –1 –2 –3 3 2 1 0 –1 –2 –3 2.0 1.8 1.6 1.4 1.2 1.0 2.0 1.8 1.6 1.4 1.2 1.0 –2 0 2 –2 2 0 x (cid:6) q 35 30 25 20 15 10 18 16 14 12 10 8 q¯(cid:2)(y) + β 20 40 (c) 3 0 2 1 0 –1 –2 –3 0 0 3 2 1 0 –1 –2 –3 0 0.5 1.0 10 20 30 1.0 0.5 fchaotic Figure 7. (a) Poincaré sections for the two cases, with tracers coloured by the computed correlation dimension dC of their orbit. Invariant tori bounding the major chaotic regions are highlighted with an orange dashed line. (b) Direct numerical simulation snapshots of the PV q + βy. The colour (cid:19)q is the time-averaged length of the level set contour of q + βy which encircles the domain. The invariant tori are again shown with an orange dashed line. (c) Comparison of the time- and zonally averaged PV gradient ¯q(cid:6)( y) + β (blue) with the fraction of orbits which are chaotic fchaotic (orange, ﬁlled). The dashed black line shows the background PV gradient β. Shown in ﬁgure 7(b) is the PV q + βy, coloured by (cid:19)q, for the DNS snapshot whose eigenmode amplitudes and phases were used to construct the Poincaré sections in ﬁgure 7(a). In both cases, all of the chaotic regions identiﬁed in the Poincaré sections correspond to a region of mixing identiﬁed in the DNS. These mixing regions only exist in the broad westward ﬂow regions, and are bounded on either side by regions consisting of shorter, less disturbed PV contours. The bulged shapes of the mixing regions also line up well with the invariant tori bounding the corresponding chaotic regions. Physically, this mixing process corresponds to the formation of elongated ﬁlaments of PV in chaotic ‘surf zones’. These chaotic regions occur in the absence of critical layers where uph = U( y), and are a result of Rossby waves which are standing in y. Observing the fate of these ﬁlaments in ﬁgure 1, some of these ﬁlamentary structures undergo reconnection to form isolated vortex patches. Cusps form in the PV contour at the point of reconnection. This is reminiscent of the vortex ﬁlamentation and ‘microbreaking’ process observed in Dritschel (1988) and Polvani & Plumb (1992), although here the breaking occurs in the ‘bulk’ rather than at sharp PV interfaces. Drawing a loose analogy with the overturning of stably stratiﬁed density layers by breaking gravity waves, this Rossby wave breaking process ultimately leads to an irreversible ‘overturning’ of a stably stratiﬁed PV gradient. Note that there also appears to be a breaking process occurring at the sharp PV 958 A28-15 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h N.M. Cao interfaces corresponding to the eastward jets as well, although this is not captured by the Poincaré section analysis. Together, these two measures of mixing provide strong evidence that localization of wave-induced chaos leads to a form of inhomogeneous PV mixing. This is broadly consistent with the results in for example Del-Castillo-Negrete & Morrison (1993) and Haynes et al. (2007), which show that the Lagrangian ﬂow induced by zonal ﬂows plus a few waves will generally create regions of mixing separated by invariant-torus-like transport barriers. Note from ﬁgure 7(c) that the PV gradient is actually close to or above the background value of β in many of the regions where the invariant tori survive, not just in the immediate vicinity of the transport barriers at the centre of the eastward jets. Thus, these tori can act as semi-permeable transport barriers over regions of ﬁnite area in the fully developed turbulence. 4.3. Self-organization into coherent Rossby waves The paradigm of inhomogeneous PV mixing leading to ‘potential vorticity staircases’ has been established as a way to understand how turbulence reinforces, rather than destroys, large-scale zonal structure (Baldwin et al. 2007; Dritschel & McIntyre 2008). Here, the parameter LRh/Lε = (U/β)1/2/(ε/β)1/5 identiﬁed in Scott & Dritschel (2012) is ≈ 6.5 for case 1 and ≈ 2.6 for case 2, suggesting the staircase regime is relevant. Chaos in the Poincaré sections was found to be localized near zonal ﬂow minima of broad jets with the strongest westward ﬂow, i.e. in the direction of propagation for the Rossby waves. Since the net y (meridional) PV ﬂux leads to a net zonal acceleration in the ideal QG model by the Taylor identity, this mixing would then reinforce the amplitude of the strongest westward jets. Note that this method of transfer of energy to the zonal ﬂows does not resemble either a local or a self-similar cascade. The PV perturbations at all scales, including the forcing scales, are exponentially stretched by the chaotic ﬂows induced by the box-scale waves. Again, this is broadly consistent with the picture in for example Haynes et al. (2007) linking the survival of invariant tori to jet formation in QG. Now moving beyond a zonally averaged sense, one key consequence of the survival of most invariant tori is that the majority of ﬂow energy can be understood as being organized into a temporally and zonally varying laminar ﬂow. Here, the remnants of the invariant-torus-like LCSs which ﬁll space in the DNS are the ‘laminae’ which organize the ﬂow. Appealing to the near-integrability established earlier, an argument can be constructed for why these laminar ﬂows predominantly organize into long-lived Rossby waves. This will show how the paradigm of inhomogeneous mixing can be extended to describe coherent ﬂow formation in non-zonal and temporally non-stationary components of the ﬂow. Since the perturbed PV ˆq0 + q1 for regular Rossby wave eigenmodes is an approximate integral of motion, the perturbed PV contours will nearly align with the surviving invariant tori. While these contours will not be mapped exactly back onto themselves by the Poincaré map, they will still be conﬁned by neighbouring invariant tori. This suggests that superpositions of Rossby waves with PV perturbations primarily localized to non-chaotic regions will enjoy a form of long-time nonlinear stability under the Poincaré map, as their PV perturbations are approximately mapped back onto themselves after any number of iterations of the map. This spatially localized stability is demonstrated in ﬁgure 8 for cases 1 and 2. In quiescent regions, contours initially aligned with the perturbed PV contours remain nearly unchanged after many iterations of the Poincaré map. The example contours shown in the ﬁgure for cases 1 and 2 have undergone 16 iterations of the Poincaré map. 958 A28-16 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Rossby waves past the breaking point in turbulence (a) Case 1 (cid:6) q (b) Case 2 (cid:6) q y 3 2 1 0 –1 –2 –3 –2 0 x 2 35 30 25 20 15 10 3 2 1 0 –1 –2 –3 –2 0 x 2 0 (cid:3)q˜ 2 10 (cid:4) coherent 18 16 14 12 10 8 (cid:4) 2 0 (cid:3)q˜ 2 coherent Figure 8. For (a) case 1 and (b) case 2, the left-hand part of each panel shows the image of a contour of the perturbed PV ˆq0 + q1 under several iterations of the Poincaré map. Light orange shows the initial contour and dark orange shows the ﬁnal contour. A different number of iterations are applied on each contour, as described (cid:5) carried by the coherent waves in the text. The right-hand part of each panel shows the ‘enstrophy’ (cid:4)˜q2 at each y. The area enclosed by the curve is coloured by the contour length (cid:19)q corresponding to the amount of mixing at each corresponding y. coherent Meanwhile in the chaotic regions, perturbed PV contours are rapidly distorted. Two iterations of the Poincaré map in case 1 and four iterations in case 2 are sufﬁcient to stretch the contours shown in the ﬁgure to approximately (cid:19)q. Lyapunov exponents in the chaotic regions can be estimated from the contour stretching, with λ ≈ 0.048 in case 1 and λ ≈ 0.018 in case 2, which are both a little slower than the Poincaré map frequency Ω0/(2π) ≈ 0.065. These nearly-integrable superpositions of Rossby waves can be thought of as generalizations of exact solutions to the QG equations consisting of pure plane waves in the absence of a background ﬂow. In the latter, contours of PV align exactly with invariant tori. Physically, ﬂuid parcels supporting the waves on invariant tori experience completely reversible perturbations from the large-scale waves. After one period of the Poincaré map, the material curves corresponding to invariant tori return to their original shapes. This allows the waves to survive for much longer times than suggested by the simple estimate τnl/τlin ∼ uwave/uph. In contrast, large-scale PV perturbations which induce non-wave-like ﬂows would produce signiﬁcant misalignment between the perturbed PV contours and the surviving tori. Examples are non-resonant beat modes between Rossby waves, or invariant hydrodynamic vortices. Under the action of the Poincaré map, either chaos from destroyed tori or frequency shear between neighbouring invariant tori would shear these perturbations to smaller scales. Since ψ ∼ k−2q, this damps the induced ﬂow perturbations, with laminar shear due to invariant tori resembling the Orr mechanism (Orr 1907). Thus, non-wave-like ﬂows are prevented from accumulating signiﬁcant amounts of energy, leaving only the coherent wave component of the ﬂow. Combining the factors of (i) inhomogeneous mixing leading to PV staircases, (ii) near-integrability of the Rossby waves and (iii) damping of non-wave-like perturbations together suggest that large-scale ﬂows will invariably organize into coherent Rossby waves on top of zonal ﬂows. Note that although the mechanism for transferring energy to the waves is not speciﬁed here, work in Bakas & Ioannou (2014) and Constantinou, Farrell & Ioannou (2016) has established that, using the turbulence closure provided by stochastic structural stability theory, non-zonal perturbations can extract energy from 958 A28-17 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h small-scale ﬂuctuations. This can be viewed as the small-scale stirring self-organizing into macroscopically coherent waves. N.M. Cao One piece of evidence supporting this picture of wave self-organization in the DNS is the observation that the most coherent waves tend to have PV perturbations localized to the regions of weakest turbulent mixing. As shown in ﬁgure 8, the zonally averaged squared magnitude of the PV perturbation (cid:4)˜q2 (cid:5) := 1 dx, where q1 is the superposition of the coherent Rossby wave eigenmodes, is localized to regions where the q contours are stretched the least. This mostly corresponds to the sharp eastward jets, although there also tend to be larger PV perturbations outside of the large-scale Rossby wave ‘surf zones’ than inside. Lx 0 q2 coherent (cid:6) Referring back to ﬁgure 5, this PV localization of coherent waves is also consistent with the observed ‘interfacial’ character of the most energetic eigenmodes. This suggests a possible ‘selection rule’ for determining which of the eigenmodes will maintain large amplitude and coherence in the fully developed turbulence. For case 1, the wavenumber (kx, ky) = (1, 1) of the most energetic eigenmode is the largest scale wave which can approximately align its maximum and minimum with the two sharp eastward jets. Similarly for case 2, the wavenumber (kx, ky) = (1, 3) of the most energetic eigenmode is the largest scale wave which can align its three maxima with the three sharp eastward jets. Meanwhile, waves with more of a ‘bulk’ character are unable to align their minima and maxima with the surviving invariant tori. Thus ‘bulk’ waves would experience stronger mixing, presumably reducing their amplitude and coherence. 5. Discussion and outlook 5.1. Open issues in the QG system Several questions remain regarding the turbulent mixing observed in the DNS. Referring back to ﬁgure 6, although a majority of the non-zonal ﬂow energy is captured in the large-scale Rossby waves, there are still a number of modes with signiﬁcant amplitude which were not included in the Poincaré map. These modes typically did not show a strong enough coherent character over several cycles to justify approximation by purely periodic ﬂows. While LCSs can be deﬁned for ﬁnite-time and aperiodic ﬂows (see e.g. Haller 2015), it is not clear if there is a generalization of the KAM theorem to such a situation. This issue also arises when attempting to generalize the arguments in § 4 to the ﬁnite deformation radius kD > 0 case, where the PV becomes ˆq = ∇2ψ − k2 ψ + βy. While the D manipulations in § 3 are also formally valid for the kD > 0 case, it is generally observed that as kD is increased, the ﬂows tend not to organize into large-scale coherent waves (Suhas & Sukhatme 2015). One possible avenue for future work is to better characterize the perturbative behaviour of elliptic LCSs in ﬁnite-time and aperiodic settings. This would clarify if the concept of near-integrability can be applied to identify ‘laminar’ character in ﬂows which are locally rather than globally organized. Additionally, the kinematics of transport driven by the wave-induced Lagrangian ﬂows are not well understood analytically. The zonal ﬂow is non-monotonic, so the Poincaré map does not satisfy the usual twist criterion (Meiss 1992). The chaos observed in the Poincaré sections also occurs in the complete absence of ﬁrst-order resonances, since uph < U( y) everywhere for the coherent waves. This distinguishes the picture of wave-driven transport in the QG system from the typical picture of wave-driven transport via overlapping wave–particle resonances in Vlasov turbulence (Chirikov 1979). It would be interesting to study how much of the observed route to chaos phenomenology in systems such as the Chirikov standard map or standard non-twist map (Del-Castillo-Negrete & Morrison 1993) 958 A28-18 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Rossby waves past the breaking point in turbulence carry over to this system. This would provide insight into how the chaotic regions form and change in response to changes in the wave amplitudes and phases. While this work does not develop a quantitative model for the long-time large-scale ﬂow evolution in the QG system, the relevance of coherent Rossby wave eigenmodes suggests some promising areas for development. The majority of the ﬂow energy can be described by ‘zonal ﬂows plus a few waves’, i.e. a mean zonal ﬂow proﬁle U( y), and a set of eigenmode amplitudes and reference phases {Ai, Θi}. Compared with the approach of discrete wave turbulence using Fourier modes (see e.g. L’vov & Nazarenko 2010), the role of low-order wave–wave resonances is de-emphasized. Wave–mean ﬂow interactions have an unmistakable inﬂuence on both the Rossby wave eigenfunctions and the small-scale turbulence statistics. The latter is best understood in a real-space Lagrangian picture. The clear impact of the wave-induced ﬂows on the turbulence statistics also suggests that purely ‘quasi-linear’ approaches, which discard all nonlinear interactions between non-zonal modes, do not account for all of the inhomogeneous mixing in the QG system. This effect has already been recognized in the development of more general ﬁltered quasi-linear approaches, such as in Constantinou et al. (2016) and Marston, Chini & Tobias (2016). However, the inﬂuence of small-scale ﬂuctuations on the zonal ﬂows and coherent waves appears to be small, changing U( y) and {Ai, Θi} slowly compared with the linear Rossby wave time scales. An intriguing possibility is if the time-scale separation arguments in Bouchet, Nardini & Tangarife (2013) and Woillez & Bouchet (2019) can be extended to allow the zonal ﬂows and coherent waves to play the role of ‘slow’ variables. This would represent a signiﬁcant simpliﬁcation compared to the general ﬁltered quasi-linear approach, as the dominant effect of the interaction between zonal modes and near-zonal modes is to organize the near-zonal modes into eigenmodes. 5.2. Integrability beyond QG Looking beyond the simple prototypical equations used in this work, waves with a strong linear character have been observed in turbulent settings in more complex and realistic ﬂows. One example is shallow water turbulence in the equatorial region, where the signature of linear waves has been seen both in simulations (Garﬁnkel et al. 2021; Schröttle et al. 2022) and in observational data (Wheeler & Kiladis 1999). Another example is the coexistence of internal gravity waves with turbulent balanced motions in the ocean, which has also been seen in idealized numerical simulations (Thomas & Yamada 2019; Thomas & Daniel 2020), global-scale ocean models (Richman et al. 2012; Torres et al. 2018) and in observational data (Rocha et al. 2016; Lien & Sanford 2019). Thus, it is interesting to ask if analogues of the results in §§ 3 and 4 hold in more complex, realistic ﬂows. The Rossby wave self-organization principle proposed in this work relied on three key ingredients: (i) the existence of enough materially conserved scalar invariants to guarantee Liouville integrability, i.e. a ‘laminar’ phase-space ﬂow; (ii) KAM theory, which characterized the persistence of the invariant tori ‘laminae’ under perturbations; and (iii) the PV inversion equation (2.2), which allowed the near-integrability conditions to be written into a form coinciding with the linearized evolution equations. Note that points (i) and (ii) are more ‘kinematic’, in the sense that they would describe the motion of any passive scalar advected by the ﬂow. Meanwhile, point (iii) is more ‘dynamical’, in the sense that it links the passive scalar advection back to the fully self-consistent ﬂow. First considering point (i), the concept of PV conservation persists in a broad range of ﬂuid and plasma systems (see e.g. Müller 1995; Gurcan & Diamond 2015). Many ﬂuid and plasma models of interest also conserve other scalar material invariants, temperature and density. such as speciﬁc entropy or related notions of potential 958 A28-19 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h N.M. Cao Furthermore, recent work has extended notions of Liouville integrability to generic dynamical systems (Bogoyavlenskij 1998; Zung 2016). This suggests the possibility of extending the Lagrangian integrability results in § 3 to more complex and realistic ﬂow models, such as the shallow water or Boussinesq equations. In the absence of forcing and dissipation, these systems typically conserve some form of PV, with the latter also typically conserving some form of buoyancy or potential density. Now considering point (ii), intuitively when a ﬂuid or plasma ﬂow conserves a scalar material invariant f , the level sets of f will generally form co-dimension one material surfaces (i.e. curves in two dimensions, surfaces in three dimensions) through which ﬂuid parcels do not pass transversally. If the dominant motions of these material surfaces are ‘reversible’, then the surfaces could correspond to invariant-torus-like LCSs. Recent extensions of KAM theory to measure-preserving dynamical systems (Xia 1992) may shed light on the conditions necessary for these transport barriers to survive in fully developed turbulent ﬂows. However for point (iii), in most ﬂuid and plasma systems the velocity ﬁeld is typically not determined by scalar material invariants like the PV or speciﬁc entropy alone. Together, these points suggest the ‘kinematic’ arguments of §§ 3 and 4 regarding the presence of invariant-torus-like LCSs in fully developed turbulent ﬂows may directly generalize to more complex ﬂow systems. However, the ‘dynamical’ arguments regarding the self-organization of ﬂows into waves do not appear to have a direct route for generalization. The development of these arguments in detail is left for future work. Finally, one other key aspect which is missing from this work is the presence of linear instabilities driven by gradients of quantities such as temperature or density, such as baroclinic instabilities or drift wave instabilities. In systems with marginally unstable gradients, found for example in fusion plasmas (Diamond & Hahm 1995), the interaction between zonal ﬂows and turbulent mixing of driving gradients can lead to non-trivial mesoscale behaviour such as avalanching (Dif-Pradalier et al. 2017). Future work can also explore how transitions to chaos in Lagrangian ﬂows interplay with linear instability dynamics. 6. Summary In summary, this work proposes that the dynamics of large-scale coherent ﬂows in QG turbulence can be understood as ‘nearly-integrable Rossby waves past the breaking point in zonally dominated turbulence’. The nearly steady zonal ﬂow background acts as a waveguide supporting Rossby wave eigenmodes which are standing in y and propagating in x. The largest-scale modes accumulate a signiﬁcant fraction of the ﬂow energy yet retain a strong linear coherent character. An integrability result on Lagrangian ﬂow in QG was given, and regular Rossby wave eigenmodes were characterized as minimally perturbing to this integrability. Chaos in the wave-induced Lagrangian ﬂow was identiﬁed using a Poincaré map, demonstrating that localization of chaos and subsequent ‘wave breaking’ was linked to the observed inhomogeneous PV mixing in the DNS. This inhomogeneous mixing was linked to a self-organization principle for the large-scale ﬂows, suggesting that surviving invariant tori organize the observed zonal jets plus Rossby waves into a single zonally and temporally varying laminar ﬂow. More broadly, this Lagrangian picture of turbulent ﬂow organization emphasizes the importance of considering the combined nonlinear effect of both zonal and non-zonal modes in the QG system. In the staircase regime when both zonal ﬂows and Rossby waves are strong, the effects of wave–mean ﬂow interactions were emphasized, whereas the roles 958 A28-20 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Rossby waves past the breaking point in turbulence of low-order wave–wave resonances were de-emphasized. Finally, it was discussed how there may be opportunities to extend the arguments in this work to more complex and realistic ﬂow systems. Acknowledgements. T. Grafke, J. Shatah and E. Vanden-Eijnden are thanked for fruitful discussions with the author, and several anonymous reviewers are thanked for their constructive commentary and insight. Funding. This research was supported by the Simons Collaboration Grant on Wave Turbulence, grant no. 617006. Declaration of interests. The author reports no conﬂict of interest. Data availability statement. The codes used in this study are openly available in GitHub at https://github. com/Maplenormandy/qg-edgeofchaos. Author ORCIDs. Norman M. Cao https://orcid.org/0000-0001-9745-0275. REFERENCES ARNOLD, V.I. 1989 Mathematical Methods of Classical Mechanics. Graduate Texts in Mathematics, vol. 60. Springer. BAKAS, N.A. & IOANNOU, P.J. 2014 A theory for the emergence of coherent structures in beta-plane turbulence. J. Fluid Mech. 740 (4), 312–341. BALASURIYA, S., JONES, C.K.R.T. & SANDSTEDE, B. 1998 Viscous perturbations of vorticity-conserving ﬂows and separatrix splitting. Nonlinearity 11 (1), 47–77. BALDWIN, M.P., RHINES, P.B., HUANG, H.-P. & MCINTYRE, M.E. 2007 The jet-stream conundrum. Science 315 (5811), 467–468. BEHRINGER, R.P., MEYERS, S.D. & SWINNEY, H.L. 1991 Chaos and mixing in a geostrophic ﬂow. Phys. Fluids A 3 (5), 1243–1249. BERKOOZ, G. 1993 The proper orthogonal decomposition in the analysis of turbulent ﬂows. Annu. Rev. Fluid Mech. 25 (1), 539–575. BERON-VERA, F.J., OLASCOAGA, M.J., BROWN, M.G., KOÇAK, H. & RYPINA, I.I. 2010 Invariant-tori-like Lagrangian coherent structures in geophysical ﬂows. Chaos 20 (1), 017514. BOFFETTA, G. & ECKE, R.E. 2012 Two-dimensional turbulence. Annu. Rev. Fluid Mech. 44 (1), 427–451. BOGOYAVLENSKIJ, O.I. 1998 Extended integrability and bi-Hamiltonian systems. Commun. Math. Phys. 196 (1), 19–51. BOUCHET, F., NARDINI, C. & TANGARIFE, T. 2013 Kinetic theory of jet dynamics in the stochastic barotropic and 2D Navier–Stokes equations. J. Stat. Phys. 153 (4), 572–625. BOUCHET, F., ROLLAND, J. & SIMONNET, E. 2019 Rare event algorithm links transitions in turbulent ﬂows with activated nucleations. Phys. Rev. Lett. 122 (7), 074502. BROWN, M.G. & SAMELSON, R.M. 1994 Particle motion in vorticity-conserving, two-dimensional incompressible ﬂows. Phys. Fluids 6 (9), 2875–2876. BURNS, K.J., VASIL, G.M., OISHI, J.S., LECOANET, D. & BROWN, B.P. 2020 Dedalus: a ﬂexible framework for numerical simulations with spectral methods. Phys. Rev. Res. 2 (2), 023068. CHIRIKOV, B.V. 1979 A universal instability of many-dimensional oscillator systems. Phys. Rep. 52 (5), 263–379. CONNAUGHTON, C., NAZARENKO, S. & QUINN, B. 2015 Rossby and drift wave turbulence and zonal ﬂows: the Charney–Hasegawa–Mima model and its extensions. Phys. Rep. 604, 1–71. CONSTANTINOU, N.C., FARRELL, B.F. & IOANNOU, P.J. 2016 Statistical state dynamics of jet–wave coexistence in barotropic beta-plane turbulence. J. Atmos. Sci. 73 (5), 2229–2253. DANILOV, S. & GRYANIK, V.M. 2004 Barotropic beta-plane turbulence in a regime with strong zonal jets revisited. J. Atmos. Sci. 61 (18), 2283–2295. DANILOV, S. & GURARIE, D. 2004 Scaling, spectra and zonal jets in beta-plane turbulence. Phys. Fluids 16 (7), 2592–2603. DEL-CASTILLO-NEGRETE, D. & MORRISON, P.J. 1993 Chaotic transport by Rossby waves in shear ﬂow. Phys. Fluids A 5 (4), 948–965. DELSHAMS, A. & DE LA LLAVE, R. 2000 KAM theory and a partial justiﬁcation of Greene’s criterion for nontwist maps. SIAM J. Math. Anal. 31 (6), 1235–1269. 958 A28-21 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h N.M. Cao DIAMOND, P.H. & HAHM, T.S. 1995 On the dynamics of turbulent transport near marginal stability. Phys. Plasmas 2 (10), 3640–3649. DIAMOND, P.H., ITOH, S.I., ITOH, K. & HAHM, T.S. 2005 Zonal ﬂows in plasma—a review. Plasma Phys. Control. Fusion 47 (5), R35–R161. DIF-PRADALIER, G., HORNUNG, G., GARBET, X., GHENDRIH, P., GRANDGIRARD, V., LATU, G. & SARAZIN, Y. 2017 The E × B staircase of magnetised plasmas. Nucl. Fusion 57 (6), 066026. DIF-PRADALIER, G., et al. 2015 Finding the elusive E × B staircase in magnetized plasmas. Phys. Rev. Lett. 114 (8), 085004. DRITSCHEL, D.G. 1988 The repeated ﬁlamentation of two-dimensional vorticity interfaces. J. Fluid Mech. 194 (-1), 511. DRITSCHEL, D.G. & MCINTYRE, M.E. 2008 Multiple jets as PV staircases: the Phillips effect and the resilience of eddy-transport barriers. J. Atmos. Sci. 65 (3), 855–874. FRISCH, U. 1995 Turbulence. Cambridge University Press. GALPERIN, B. & READ, P.L. (Ed.) 2019 Zonal Jets. Cambridge University Press. GARFINKEL, C.I., SHAMIR, O., FOUXON, I. & PALDOR, N. 2021 Tropical background and wave spectra: contribution of wave-wave interactions in a moderately nonlinear turbulent ﬂow. J. Atmos. Sci. 76 (6), 1773–1789. GÜRCAN, Ö.D. & DIAMOND, P.H. 2015 Zonal ﬂows and pattern formation. J. Phys. A 48 (29), 293001. HALLER, G. 2015 Lagrangian coherent structures. Annu. Rev. Fluid Mech. 47 (1), 137–162. HAYNES, P.H., POET, D.A. & SHUCKBURGH, E.F. 2007 Transport and mixing in kinematic and dynamically consistent ﬂows. J. Atmos. Sci. 64 (10), 3640–3651. HOLM, D.D., MARSDEN, J.E., RATIU, T. & WEINSTEIN, A. 1985 Nonlinear stability of ﬂuid and plasma equilibria. Phys. Rep. 123 (1–2), 1–116. KRAICHNAN, R.H. 1967 Inertial ranges in two-dimensional turbulence. Phys. Fluids 10 (7), 1417. KUO, H.-L. 1949 Dynamic instability of two-dimensional nondivergent ﬂow in a barotropic atmosphere. J. Meteorol. 6 (2), 105–122. LIEN, R.-C. & SANFORD, T.B. 2019 Small-scale potential vorticity in the upper-ocean thermocline. J. Phys. Oceanogr. 49 (7), 1845–1872. DE LA LLAVE, R. 2001 A tutorial on KAM theory. In Smooth Ergodic Theory and Its Applications (ed. A. Katok, R. de la Llave, Y. Pesin & H. Weiss), Proceedings of Symposia in Pure Mathematics, vol. 69, pp. 175–292. American Mathematical Society. L’VOV, V.S. & NAZARENKO, S. 2010 Discrete and mesoscopic regimes of ﬁnite-size wave turbulence. Phys. Rev. E 82 (5), 056322. MALHOTRA, N. & WIGGINS, S. 1998 Geometric structures, lobe dynamics, and Lagrangian transport in ﬂows with aperiodic time-dependence, with applications to Rossby wave ﬂow. J. Nonlinear Sci. 8 (4), 401–456. MARSTON, J.B., CHINI, G.P. & TOBIAS, S.M. 2016 Generalized quasilinear approximation: application to zonal jets. Phys. Rev. Lett. 116 (21), 214501. MEISS, J.D. 1992 Symplectic maps, variational principles, and transport. Rev. Mod. Phys. 64 (3), 795–848. MORRISON, P.J. 1998 Hamiltonian description of the ideal ﬂuid. Rev. Mod. Phys. 70 (2), 467–521. MÜLLER, P. 1995 Ertel’s potential vorticity theorem in physical oceanography. Rev. Geophys. 33 (1), 67. ORR, W.M.F. 1907 The stability or instability of the steady motions of a perfect liquid and of a viscous liquid. Part II. A viscous liquid. Proc. R. Irish Acad. A 27, 69–138. PIERREHUMBERT, R.T. 1991 Chaotic mixing of tracer and vorticity by modulated travelling Rossby waves. Geophys. Astrophys. Fluid Dyn. 58 (1–4), 285–319. POLVANI, L.M. & PLUMB, R.A. 1992 Rossby wave breaking, microbreaking, ﬁlamentation, and secondary vortex formation: the dynamics of a perturbed vortex. J. Atmos. Sci. 49 (6), 462–476. RHINES, P.B. 1975 Waves and turbulence on a beta-plane. J. Fluid Mech. 69 (3), 417–443. RICHMAN, J.G., ARBIC, B.K., SHRIVER, J.F., METZGER, E.J. & WALLCRAFT, A.J. 2012 Inferring dynamics from the wavenumber spectra of an eddying global ocean model with embedded tides. J. Geophys. Res. 117, C12012. ROCHA, C.B., CHERESKIN, T.K., GILLE, S.T. & MENEMENLIS, D. 2016 Mesoscale to submesoscale wavenumber spectra in drake passage. J. Phys. Oceanogr. 46 (2), 601–620. ROGERS, J.H 2009 The Giant Planet Jupiter. Cambridge University Press. ROSENBERG, E. 2020 Fractal Dimensions of Networks. Springer International Publishing. SALMON, R. 1988 Hamiltonian ﬂuid mechanics. Annu. Rev. Fluid Mech. 20 (1), 225–256. SALMON, R. 1998 Lectures on Geophysical Fluid Dynamics. Oxford University Press. SCHRÖTTLE, J., SUHAS, D.L., HARNIK, N. & SUKHATME, J. 2022 Turbulence and equatorial waves in moist and dry shallow-water ﬂow, excited through mesoscale stochastic forcing. Q. J. R. Meteorol. Soc. 148 (743), 599–619. 958 A28-22 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Rossby waves past the breaking point in turbulence SCOTT, R.K. & DRITSCHEL, D.G. 2012 The structure of zonal jets in geostrophic turbulence. J. Fluid Mech. 711, 576–598. SHEPHERD, T.G. 1987 Rossby waves and two-dimensional turbulence in a large-scale zonal jet. J. Fluid Mech. 183, 467–509. SIMONNET, E., ROLLAND, J. & BOUCHET, F. 2021 Multistability and rare spontaneous transitions in barotropic β-plane turbulence. J. Atmos. Sci., 78 (6), 1889–1911. SUHAS, D.L. & SUKHATME, J. 2015 Low frequency modulation of jets in quasigeostrophic turbulence. Phys. Fluids 27 (1), 016601. SUKORIANSKY, S., DIKOVSKAYA, N. & GALPERIN, B. 2008 Nonlinear waves in zonostrophic turbulence. Phys. Rev. Lett. 101 (17), 178501. THOMAS, J. & DANIEL, D. 2020 Turbulent exchanges between near-inertial waves and balanced ﬂows. J. Fluid Mech. 902, A7. THOMAS, J. & YAMADA, R. 2019 Geophysical turbulence dominated by inertia–gravity waves. J. Fluid Mech. 875, 71–100. TORRES, H.S., KLEIN, P., MENEMENLIS, D., QIU, B., SU, Z., WANG, J., CHEN, S. & FU, L.-L. 2018 Partitioning ocean motions into balanced motions and internal gravity waves: a modeling study in anticipation of future space missions. J. Geophys. Res. 123 (11), 8084–8105. VASAVADA, A.R. & SHOWMAN, A.P. 2005 Jovian atmospheric dynamics: an update after Galileo and Cassini. Rep. Prog. Phys. 68 (8), 1935–1996. WAGNER, F. 2007 A quarter-century of H-mode studies. Plasma Phys. Control. Fusion 49 (12B), B1–B33. WEBB, G.M. & ANCO, S.C. 2019 Conservation laws in magnetohydrodynamics and ﬂuid dynamics: Lagrangian approach. AIP Conf. Proc. 2153, 020024. WHEELER, M. & KILADIS, G.N. 1999 Convectively coupled equatorial waves: analysis of clouds and temperature in the wavenumber–frequency domain. J. Atmos. Sci. 56 (3), 374–399. WOILLEZ, E. & BOUCHET, F. 2019 Barotropic theory for the velocity proﬁle of Jupiter’s turbulent jets: an example for an exact turbulent closure. J. Fluid Mech. 860, 577–607. XIA, Z. 1992 Existence of invariant tori in volume-preserving diffeomorphisms. Ergod. Theory Dyn. Sys. 12 (3), 621–631. ZUNG, N.T. 2016 Geometry of Integrable non-Hamiltonian Systems, pp. 85–140. Springer International Publishing. 958 A28-23 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 0 9 3 2 0 2 m . f j / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h	null
10.1371_journal.ppat.1011346.pdf	Data Availability Statement: All relevant data are within the manuscript and its Supporting Information files.	All relevant data are within the manuscript and its Supporting Information files.	RESEARCH ARTICLE A modified Agrobacterium-mediated transformation for two oomycete pathogens Luyao Wang1,2, Fei Zhao2, Haohao Liu2, Han Chen2, Fan Zhang2, Suhua Li1, Tongjun Sun1, Vladimir NekrasovID 3, Sanwen Huang1, Suomeng DongID 2 a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Wang L, Zhao F, Liu H, Chen H, Zhang F, Li S, et al. (2023) A modified Agrobacterium- mediated transformation for two oomycete pathogens. PLoS Pathog 19(4): e1011346. https:// doi.org/10.1371/journal.ppat.1011346 Editor: Paul Birch, University of Dundee, UNITED KINGDOM Received: June 23, 2022 Accepted: April 6, 2023 Published: April 21, 2023 Peer Review History: PLOS recognizes the benefits of transparency in the peer review process; therefore, we enable the publication of all of the content of peer review and author responses alongside final, published articles. The editorial history of this article is available here: https://doi.org/10.1371/journal.ppat.1011346 Copyright: © 2023 Wang et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the manuscript and its Supporting Information files. Funding: In this research, S.W. and S.D. received funding support from Guangdong Major Project of 1 Shenzhen Branch, Guangdong Laboratory of Lingnan Modern Agriculture, Genome Analysis Laboratory of the Ministry of Agriculture and Rural Affairs, Agricultural Genomics Institute at Shenzhen, Chinese Academy of Agricultural Sciences, Shenzhen, China, 2 Department of Plant Pathology and Key Laboratory of Integrated Management of Crop Disease and Pests (Ministry of Education), Nanjing Agricultural University, Nanjing, China, 3 Plant Sciences and the Bioeconomy, Rothamsted Research, Harpenden, United Kingdom * [email protected] (SH); [email protected] (SD) Abstract Oomycetes are a group of filamentous microorganisms that include some of the biggest threats to food security and natural ecosystems. However, much of the molecular basis of the pathogenesis and the development in these organisms remains to be learned, largely due to shortage of efficient genetic manipulation methods. In this study, we developed modi- fied transformation methods for two important oomycete species, Phytophthora infestans and Plasmopara viticola, that bring destructive damage in agricultural production. As part of the study, we established an improved Agrobacterium-mediated transformation (AMT) method by prokaryotic expression in Agrobacterium tumefaciens of AtVIP1 (VirE2-interact- ing protein 1), an Arabidopsis bZIP gene required for AMT but absent in oomycetes genomes. Using the new method, we achieved an increment in transformation efficiency in two P. infestans strains. We further obtained a positive GFP transformant of P. viticola using the modified AMT method. By combining this method with the CRISPR/Cas12a genome editing system, we successfully performed targeted mutagenesis and generated loss-of- function mutations in two P. infestans genes. We edited a MADS-box transcription factor- encoding gene and found that a homozygous mutation in MADS-box results in poor sporula- tion and significantly reduced virulence. Meanwhile, a single-copy avirulence effector- encoding gene Avr8 in P. infestans was targeted and the edited transformants were virulent on potato carrying the cognate resistance gene R8, suggesting that loss of Avr8 led to suc- cessful evasion of the host immune response by the pathogen. In summary, this study reports on a modified genetic transformation and genome editing system, providing a poten- tial tool for accelerating molecular genetic studies not only in oomycetes, but also other microorganisms. Author summary Oomycetes are a unique group of animal and plant pathogens that include some of the biggest threats to food security and natural ecosystems. However, much of the PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 1 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Basic and Applied Basic Research (2021B0301030004); S.D. received funding support from National Natural Science Foundation of China (NSFC, 31721004) and China Agriculture Research System (CARS-09-P20); L.W was supported by NSFC (31900303) and the Agricultural Science and Technology Innovation Program (CAASZDRW202101); V.N. was supported by the Biotechnology and Biological Sciences Research Council (BBSRC) through the Designing Future Wheat (DFW) Institute Strategic Programme (grant number BB/P016855/1). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. pathogenesis and development in these organisms remains to be learned, largely due to shortage of efficient genetic manipulation methods for a long time. In this manuscript, we developed modified Agrobacterium-mediated genetic transformation strategies for Phy- tophthora infestans, a notorious oomycete species that caused Irish Famine in the 19th cen- tury, and Plasmopara viticola, the causal agent of grapevine downy mildew that is known as a highly destructive disease of grapevines in all grape-growing areas of the world. Using prokaryotic expression of a plant protein in A. tumefaciens, we achieved considerable increase in transformation rate in different P. infestans strains, and also acquired one posi- tive transformant of P. viticola, a oomycete species that is extremely hard-to-transform and cannot be grown as an axenic culture. Using our improved transformation protocol, combined with the CRISPR/Cas12a system, we performed genome editing and created loss-of-function alleles in P. infestans. In summary, our study reports on modified genetic transformation methods, for two important oomycete species, that have the potential to accelerate the molecular genetic study of many other microorganisms. Introduction Oomycetes are regarded as an important lineage of eukaryotic organisms that includes a con- siderable number of plant pathogens, which pose a threat to natural and arable ecosystems. Grapevine downy mildew, caused by Plasmopara viticola, is considered to be the most devas- tating disease of grapevines in climates with relatively warm and humid summers. P. viticola is an obligate biotrophic oomycete species that cannot be grown as an axenic culture and is very recalcitrant to genetic transformation as demonstrated in previous studies [1]. The genetic transformation obstacle in the case of P. viticola has severely hampered functional genomics research and studies on screening molecular drug targets. Thus, it is important to set up a workable transformation method for P. viticola and other hard-to-transform biotrophic microorganisms. Potato late blight is another one of the most well-known but not well-understood plant dis- eases in terms of molecular genetics of Phytophthora infestans, its causal pathogenic microor- ganism. P. infestans, the oomycete pathogen responsible for the devastating potato late blight, poses a worldwide threat [2], and plays an essential role in studying plant–microbe interac- tions, epidemiology, and field disease control [3]. Currently, reverse genetics studies in P. infes- tans are also hampered by inefficient genetic transformation methods. Up until now, functional genomic research in P. infestans has relied mostly on transient/stable gene overex- pression or target-gene silencing by RNAi [4, 5]. Although polyethylene glycol (PEG)-medi- ated transformation, Agrobacterium-mediated transformation (AMT), microprojectile bombardment and zoospore electroporation have been already set up in P. infestans for years, we need to be open to new methods as other transformation options are always encouraged [6–9]. Moreover, phenotypes of Phytophthora transformants often vary significantly between different pathogen lines, experiment operators, and individual experiments, possibly because of random gene integration sites and different silencing or overexpression qualities [10]. Despite of multiple attempts to apply CRISPR/Cas9 for the purpose of targeted mutagenesis in P. infestans, no genome editing events, generated using this system, have been reported so far [11]. In a recent study, the CRISPR/Cas12a (Cpf1) system was utilized to produce genome editing events in a P. infestans strain using the PEG-mediated protoplast transformation method [12]. To propel future research, it is of great interest to expand the range of available genetic transformation methods for genome editing in P. infestans. PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 2 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Agrobacterium tumefaciens, as a typical plant pathogenic bacterium, causes crown gall dis- ease in a wide range of plants [13], and was developed into an efficient genetic transformation tool since Binns and Thomashaw first demonstrated that A. tumefaciens can integrate exoge- nous gene segments into the plant nucleus [14]. Known as the first example of prokaryote to eukaryote horizontal gene transfer, the capacity of A. tumefaciens to transfer alien genes to host cells has become the basis of one of the most essential technologies in manifold research fields, and plant science in particular [15]. By now, transgenesis, achieved through A. tumefa- ciens-mediated transformation (AMT), has been established in many model plant species, including Arabidopsis, tobacco and rice [16–18]. In addition to plants, Agrobacterium is also able to infect a broad range of other non-plant hosts during the co-cultivation process, includ- ing microorganisms identified as plant pathogens, and AMT protocols have been developed as a transformation system for many fungi, including Aspergilus species, Beauveria species, Botry- tis cinerea, Candida species, Coccidiodes species, Colletotrichum species, Fusarium species, Tri- choderma species and Verticillium species [19–30]. Attempts to improve the efficiency of AMT have been undertaken since the method was first reported. However, in most cases, the efforts were focused on optimization of conditions, such as explant treatment, tissue culture medium composition, temperature or selectable marker, but not the core tool—A. tumefaciens [31]. One successful attempt to improve the plant transformation ability of A. tumefaciens involved introducing a compatible plasmid carrying a virG mutant (virGn45D), which constitutively induced vir gene expression during AMT [32]. The first AMT protocol for P. infestans was established in 2003 by Vijin and Govers; in this study, the A. tumefaciens LBA1100 strain was selected, and as many as 30 transformants per 107 zoospores could be obtained [7]. A recent study has also described an efficient AMT proto- col for another oomycete species, Phytophthora palmivora, which produces large amount of zoospores (2–5 × 106/mL) during in vitro preparation [33]. However, the amount of zoospores produced may vary dramatically in different Phytophthora strains, and some oomycete species produce hardly any zoospores under artificial conditions. Thus, the ability to produce suffi- cient amounts of zoospores under in vitro conditions is likely a serious limiting factor for applying zoospore-dependent transformation methods in oomycetes [34–36]. Previous studies illustrated that ectopic expression of several plant proteins significantly improved AMT rate in a range of plant species [37]. A set of proteins from the host are involved with Agrobacterium-mediated transformation of plant [37]. Overexpressing several of these host proteins, mostly from Arabidopsis, such as AtVIP1, AtRTNLB1, Ku80, histone H2A and SGA1, significantly increases AMT efficiency [37–42]. Arabidopsis VirE2-interact- ing-protein1 (AtVIP1) is known as an important plant protein that contributes to the AMT process [38, 43]. AtVIP1 has been proven to interact with Agrobacterium effector VirE2, and its ability to form homomultimeric protein complexes and interact with histone H2A in host cells is required for Agrobacterium-mediated pathogenicity or stable genetic transformation [44]. AtVIP1-overexpressing plant lines have shown significantly increased susceptibility to A. tumefaciens infection, and AtVIP1 seems to facilitate nuclear transport of VirE2 and the T-DNA complex [42]. It would be more practical if we could directly utilize AtVIP1 protein during the AMT process instead of constructing AtVIP1 transgenic material prior to transfor- mation of a gene of interest. Here we report on a modified AMT method utilizing AtVIP1 as an enhancer in the trans- formation of two important oomycete pathogens. In this study, an increment in the transfor- mation efficiency of Phytophthora infestans was achieved by prokaryotic expression of AtVIP1 in the A. tumefaciens EHA105 strain as compared to AMT without AtVIP1. We then extended application of this modified AMT procedure to Plasmopara viticola. By combining our modi- fied AMT method and the recently established Cas12a-based genome editing system, we PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 3 / 27 PLOS PATHOGENSModification of AMT for two oomycete species successfully generated genome-edited P. infestans transformants and observed the expected virulence phenotypes after inoculating potato leaves with these mutants. Our work is provid- ing a new direction for AMT improvement and a way to potentially accelerate molecular genetic studies of the two devastating plant pathogens and other microorganisms. Results Oomycetes lack homologues of specific plant proteins important for Agrobacterium-mediated transformation To better understand the low AMT efficiency in P. infestans and test whether introducing plant proteins into transformation system would significantly improve Phytophthora transfor- mation efficiency, we first performed a bioinformatic study of proteins with similarities to pre- viously identified plant proteins, required for AMT, in different oomycete species. We used target protein sequences from Arabidopsis to perform local BLAST using publicly available genomic data from each selected plant or oomycete species. Sequence similarity shown in Fig 1A is based on the ratio between the highest alignment score (bit-score) of local BLAST results from a target species and the one from Arabidopsis. In the selected monocotyledonous and dicotyledonous plants, the sequence similarities of the obtained proteins were all above 0.25, except that the sequence similarity of Rad50 in Solanum tuberosum and Vitis vinifera were merely 0.04 and 0.17, respectively (Fig 1A and S3 Table). In oomycete species, similar proteins with sequence similarity greater than 0.45 were detected for IMPA-4, Rab8, SGA1 and histone H2A. For AGP17, PP2C, Ku70, Rad50 and Ku80, sequence similarities of similar proteins were mostly above 0.15 (average values were 0.19, 0.19, 0.20, 0.18 and 0.15, respec- tively) in selected oomycete species. Interestingly, sequence similarity of AtVIP1 or its related Arabidopsis paralogues, including bZIP29, bZIP30, bZIP69, posF21, bZIP18 and bZIP52, were less than 0.08 in all selected oomycete species, indicating that absence of an AtVIP1 homo- logue might have a negative impact on the AMT rate in oomycetes. AtVIP1 enhances Agrobacterium-mediated genetic transformation To further test our hypothesis that introduction of plant proteins, required for AMT, into a transformation system increases AMT efficiency in oomycetes, we constructed a vector to deliver AtVIP1 from Agrobacterium to host cells to potentiate the transformation process. To make AtVIP1 translocatable by Agrobacterium T4SS, we fused the coding sequence of AtVIP1 to virFΔ42N (truncated virF, lacking the sequence encoding the 42 N-terminal amino acids, that functions as a C-terminal transport signal for VirB/D4-translocated proteins in Agrobac- terium [45]), and added green fluorescent protein-encoding sequence (GFP) to the N-terminus (S1A Fig). As shown in S1B Fig, GFP signal was observed in both cytoplasm and nucleus of infected Nicotiana benthaminana cells 3 days post leaf infiltration with A. tumefaciens EHA105 strain carrying p302b-gfp-AtVIP1-virFΔN42. These data suggested that the GFP-At- VIP1-virFΔ42N fusion protein was translocated into infected cells through T4SS during Agro- bacterium infection. To test whether our modified protocol enhances AMT efficiency, we first tested it in wheat, a cereal crop with low AMT efficiency [46]. We introduced the prokaryotic expression plasmid p302b-AtVIP1-virFΔN42 (pV1F) and the binary plasmid pWMB110-GUS carrying the β-glu- curonidase coding gene gus (contains maize adh1 intron) into Agrobacterium strain EHA105 (S2A Fig). By inoculating leaf segments of wheat cv. ‘Fielder’ with A. tumefaciens EHA105 strain containing pV1F and pWMB110-GUS, a significantly higher transient transformation efficiency in leaf tissue was observed as compared to the control treatment inoculated with PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 4 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Fig 1. (a) The protein similarity analysis of plant proteins, important for AMT of Arabidopsis, and their homologues in other plant and oomycete species. Representative plants with an established Agrobacterium-mediated transformation protocol and eight oomycete species were selected for phylogenetic analysis. The left phylogenetic tree was generated by the OrthoFinder software based on variations of single-copy orthologous genes in released genomic data of different species using IQ-tree with JTT + G4 model of evolution. The left bar indicates amino acid substitutions per site. The right heatmap indicates sequence similarity of known plant proteins envolved in Arabidopsis AMT process and their PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 5 / 27 PLOS PATHOGENSModification of AMT for two oomycete species homologues in selected plants and oomycete species. The right bar indicates sequence similarity. (b) Schematic representation of the main idea of our improved transformation strategy. The use of the helper plasmid pV1F results in expression of virFΔN42 tagged AtVIP1. The recombinant AtVIP1 is delivered into oomycete zoospores via VirB/VirD4 T4SS, and thereby facilitates the process of T-DNA complex (T-complex) targeting the nucleus of an infected zoospore or an encysted zoospore with a germ tube. https://doi.org/10.1371/journal.ppat.1011346.g001 EHA105 strain containing the vector lacking the AtVIP1 gene (pEV) (S2B Fig). In addition, gus expression in root segments increased from 1.8% to 11.6% when transformed with the vec- tor containing the AtVIP1 gene (S2C and S2D Fig). These results indicate that AtVIP1 is an effective enhancer of Agrobacterium-mediated genetic transformation in plant tissues. Expressing AtVIP1 in A. tumefaciens induces T-DNA integration in P. infestans To extend our AMT system to a wider range of eukaryotic species, characterized by low trans- formation efficiency, we tested it in one of the most important phytopathogenic oomycetes, P. infestans, in which AMT protocols have already been set up since years ago [7, 47]. A sche- matic illustration of our modified AMT method for P. infestans is presented in Fig 1B. To test if AtVIP1 also facilitates AMT in P. infestans, we first constructed the binary vector pLY40-gfp (Fig 2A), which carries a T-DNA region containing the gfp gene driven by Bremia lactucae Ham34 promoter, and the geneticin (G418) resistance gene nptII driven by Phytophthora sojae Rpl41 promoter. The pLY40-gfp was introduced into A. tumefaciens strain EHA105 together with either pV1F or pEV. A. tumefaciens strains carrying above-mentioned DNA construct combinations (Fig 2A) were co-cultivated with zoospores of P. infestans strains JH19 and HB1501, and subsequently geneticin-resistant transformants were obtained according to the methods presented in the flow diagram in S3 Fig. While G418 resistant colonies were obtained with both tested construct combinations, co-cultivation with A. tumefaciens containing the pV1F plasmid resulted in sig- nificantly more G418 resistant colonies (or colonies with observed GFP-signal) in both HB1501 and JH19 (Fig 2B–2C and Table 1). One of the representative transformants (JH19 background), named as T1, which was obtained with this modified AMT method, expressed a GFP-size protein and showed a strong GFP signal compared to the transformant T3 (JH19 background) that was obtained using the empty pLY40 as negative control (Fig 2D and 2E). The Southern blot analysis of representative transformants of the HB1501 strain background suggested that T-DNA fragments with the nptII gene were integrated into the genomic DNA of all seven of them (Fig 2F). These results indicate that prokaryotic-expressed AtVIP1 in A. tumefaciens considerably promotes T-DNA integration during AMT in P. infestans. Transformation of Plasmopara viticola using the modified AMT procedure In this study, we also extended application of our modified AMT protocol to P. viticola, a diffi- cult-to-transform oomycete that causes grapevine downy mildew. Based on the previously established AMT protocol for the biotrophic fungus Podosphaera xanthii [48], A. tumefaciens EHA105, carrying LY40-gfp and pV1F, was used for co-cultivation with released zoospores of P. viticola isolate BS5. After resistance selection, applied by rinsing P. viticola inoculated grape leaves with G418, we isolated one transformant called T1 (Figs 3 and S5B). We observed a strong GFP signal in both sporangia and mycelia of T1 during infection using confocal micros- copy, and the GFP signal was stable on grape leaves in three sub-inoculation generations (Fig 3A and 3C). Western blot analysis indicated that a GFP-size band could also be detected in protein extracts from the transformant T1 (Fig 3B), and Southern blot analysis, carried out using PCR-amplified nptII gene as a probe, showed that one T-DNA segment was successfully PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 6 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Fig 2. Generating P. infestans transformants expressing gfp using the modified AMT method. (a) Schematic representation of the constructs used in this experiment. A. tumefaciens strain EHA105 with pV1F and pLY40-gfp was used for P. infestans transformation. (b) Quantification of colonies that acquired G418 resistance as a result of transforming gfp into P. infestans strains JH19 and HB1501 using our modified AMT procedure. EHA105 with the helper plasmid pEV was used as control in this experiment. All data represent average values from three independent experiments with the indicated standard deviations. Statistical differences among the samples were analyzed with Sˇı´da´k’s multiple comparisons test (P< 0.0021: , P< 0.0001: *). (c) Two representative plates from the transformation experiment shown in (b). (d) Immunoblot of P. infestans JH19 transformant T1, expressing free gfp, probed with an anti-GFP antibody. Protein extracted from N. benthamiana leaves transiently expressing gfp driven by the CaMV35s promoter was used as positive control in lane 1. (e) The P. infestans JH19 transformant T1 expressing a detectable GFP signal was obtained by AMT using A. tumefaciens carrying constructs described in (a). Scale bars = 40 μm. T3, a randomly selected empty vector transformant was used as negative control. GFP expression in the transformant was analysed by PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 7 / 27 PLOS PATHOGENSconfocal microscopy. The protein blot was stained with Coomassie Blue to confirm equal loading. (f) Southern blot analysis of representative gfp transformants of P. infestans strain HB1501. Genomic DNA (4 μg) was digested with HindIII and all blots were probed with a fragment containing the nptII gene to detect the presence of T-DNA. Numbers on the left indicate the positions of molecular weight markers (kb). https://doi.org/10.1371/journal.ppat.1011346.g002 Modification of AMT for two oomycete species integrated into the genome of T1 (Fig 3D). Consequently, the extensibility of our strategy to P. viticola suggests that AtVIP1 could also be considered for optimizing AMT methods in the case of other oomycete species with poor efficiency of transformation. AtVIP1 facilitates Agrobacterium-mediated delivery of LbCas12a into P. infestans To test if our modified AMT method could also facilitate genome editing in oomycetes, we selected P. infestans for the first attempt as the Cas12a-based method had recently been set up in this species [12]. We first fused the coding sequences of the Phytophthora sojae NLS, human codon-optimized Lachnospiraceae bacterium Cas12a (LbCas12a) and GFP, and inserted the recombinant sequence into the T-DNA region of pLY40 to obtain pLY40-Cas12a-gfp (Fig 4A). A. tumefaciens strain EHA105 carrying pV1F and pLY40-Cas12a-gfp was used to transform zoospores of P. infestans strains JH19 and HB1501. As a result, up to 61 and 22 G418 resistant transformants were obtained for JH19 and HB1501, respectively, in three independent trans- formation experiments (Table 1). The localization of the GFP signal in the transformants T5 (JH19) and T6 (HB1501) was distinctly nuclear as visualized by confocal microscopy (Fig 4B). Table 1. Genetic transformations of P. infestans strains JH19 and HB1501 by A. tumefaciens EHA105 with or without the helper plasmid pV1F. Binary plasmid pLY40-gfp pLY40-gfp pLY40-gfp pLY40-gfp pLY40-Cas12a- gfp pLY40-Cas12a- gfp pLY40-Cas12a- gfp pLY40-Cas12a- gfp Binary plasmid pLY40-PiMADS- KO pLY40-PiMADS -KO pLY40-Avr8-KO pLY40-Avr8-KO Helper plasmida pEV pV1F pEV pV1F pEV pV1F pEV P. infestans strain JH19 JH19 HB1501 HB1501 JH19 JH19 HB1501 pV1F HB1501 Helper plasmid pEV pV1F pEV pV1F P. infestans strain JH19 JH19 HB1501 HB1501 GFP+/G418R colonies in attempt 1b 0/3 11/21 0/0 2/8 1/1 13/22 0/0 3/6 GFP+/G418R colonies in attempt 2 GFP+/G418R colonies in attempt 3 Average GFP/G418R colonies 1/2 10/29 0/1 7/12 0/7 3/19 0/0 3/7 1/3 13/24 0/0 5/12 0/2 6/20 0/0 7/9 0.67/2.67 11.33/24.67* 0/0.33 4.67/10.67 0.33/3.33 7.33/20.33 0/0 4.33/7.33 G418R colonies in attempt 1 G418R colonies in attempt 0 7 2 11 2 0 4 2 9 G418R colonies in attempt 3 Average G418R colonies 1 5 1 8 0.33 5.33* 1.67 9.33* a Helper plasmids used in this experiment are described in detail in S2 Table. b As determined with GFP-observation by confocal microscopy (GFP). As determined 14 days after acquired transformants transferred to another rye-sucrose medium with 5 or 10 μg/L geneticin (G418). The marked values indicate significantly different with the AMT process used same binary plasmid but pEV as the helper plasmid. Student’s t-test was used to determine the differences. https://doi.org/10.1371/journal.ppat.1011346.t001 PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 8 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Fig 3. Obtaining Plasmopara viticola BS5 stable transformant T1 using the modified AMT protocol. (a) P. viticola BS5 strain was used for transformation assay and was obtained by AMT using A. tumefaciens carrying constructs pLY40-gfp and pV1F. Scale bars = 40 μm. The confocal microscopy images were taken 7 days post inoculation with the zoospores from the third sub-generation of T1; wild type BS5 was used as negative control. (b) Immunoblot of P. viticola BS5 transformant T1 expressing free GFP, probed with an anti-GFP antibody. Protein extracted from N. benthamiana leaves transiently expressing gfp driven by the CaMV35s promoter was used as positive control in lane 2. (c) Transformant T1 of P. viticola BS5 expresses detectable GFP signal when infecting grapevine leaves (Zitian seedless, A17). Scale bars = 100 μm. Images were taken 7 days post inoculation and wild type BS5 was used as negative control. (d) Southern blot analysis of transformant T1 of P. viticola BS5. Genomic DNA (4 μg) was digested with HindIII and all blots were probed with a fragment containing the nptII gene to detect the presence of T-DNA. Numbers on the left indicate the positions of molecular weight markers (kb). https://doi.org/10.1371/journal.ppat.1011346.g003 As shown in Fig 4C, the hlbCas12a-GFP fusion protein was successfully detected by western blotting in both transformants T5 and T6 with hlbCas12a-gfp expressed in N. benthamiana used as a positive control. These data suggest successful delivery of the Cas12a gene into P. infestans using our modified AMT method, paving a way for an efficient genome editing method for this species. PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 9 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Fig 4. Integration of LbCas12a in P. infestans. (a) Schematic representation of the constructs used in the experiment. A. tumefaciens EHA105 carrying pLY40-Cas12a-gfp and either pV1F or pEV was used for P. infestans transformation. (b) Confocal micrograph of T5 and T6 transformants expressing PsNLS-lbCas12-GFP. The nuclear localization pattern of the fusion protein, indicated by white arrows, revealed mycelia and sporangia. T3, an empty vector pLY40 transformed line, is used as negative control. Bright field and GFP channels are presented. Scale bars = 40 μm. (c) Immunoblot of two representative P. infestans transformants expressing PsNLS-lbCas12-GFP. The gfp tagged hlbCas12a driven by CaMV35S promoter was expressed in N. benthamiana as positive control. The expected size of the protein is 176.6 kDa. The protein blot was stained with Coomassie Blue to confirm equal loading. (d) Quantification of positive P. infestans transformants expressing GFP tagged Cas12a (JH19 and HB1501 backgrounds) generated using the modified AMT procedure. A. tumefaciens EHA105 carrying the helper plasmid pEV was used as control in this experiment. Statistical differences among the samples were analyzed with Sˇı´da´k’s multiple comparisons test (P< 0.0021: , P< 0.0001: *). https://doi.org/10.1371/journal.ppat.1011346.g004 Editing a MADS-box-encoding gene in P. infestans To investigate the prospect of using the modified AMT method for CRISPR/Cas12a-mediated P. infestans genome editing, a single-copy gene encoding a MADS-box transcription factor (PITG_07059) was chosen as the first editing target in this study. The homologous MADS-box genes play an essential role in asexual reproduction and zoosporogenesis in both P. infestans and P. sojae [49, 50]. Two gRNAs, MADS-g1 and MADS-g2, targeting the 1332 nt MADS-box- PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 10 / 27 PLOS PATHOGENSModification of AMT for two oomycete species encoding sequence were designed using the procedure described by Ah-Fong et al. (Figs 5A and S6) [12]. In the binary construct pLY40-MADS-box-KO, MADS-g1 and MADS-g2 were flanked by 21-bp short direct repeats (DRs) and inserted in a tandem downstream of LbCas12a and the 73-nt poly-adenine sequence (pA), which was added to promote translation of Cas12a mRNA (S6C Fig). As a 3’ uridine-rich tail positively regulates CRISPR/Cas12a crRNA forma- tion, four thymines were added at the 3’ends of sequences encoding the gRNAs (S6C Fig) [12, 51]. The editing events within the MADS-box gene in the transformants were detected by per- forming PCR, with primers amplifying the full length of the gene, and sequencing the PCR amplicons. In 2 out of 16 G418 resistant transformants (T8 and T16), shorter PCR amplicons were observed suggesting presence of CRISPR/Cas12a-induced deletions (Fig 5B). Sequencing analysis confirmed that T8 and T16 carried identical 993 bp deletions, between MADS-g1 and MADS-g2 loci, resulting in a truncated MADS-box gene coding sequence (Figs 5C and S6E). The MEF2-like domain of the MADS-box protein contributes to the DNA binding activity, which is essential for a transcription factor. The MEF2-like domain was predicted to be located at the 2–72 aa sites by InterproScan (version 5.52–86.0) (Fig 5A). The sequencing data pre- sented in Figs 5C and S6E showed that both mutant MADS-box alleles, present in T8 and T16, lack the majority of the sequence encoding the MEF2-like domain, suggesting loss of function of the MADS-box proteins in these transformants. As shown in Fig 5D and 5F, MADS-box-edited JH19 strains T8 and T16 showed merely no sporangia output after 5 days of cultivation in PEA broth compared with untransformed wild type JH19, which yielded 49.22 sporangia in 10 μL PEA broth culture. Another transformant (T5) was selected as a non-edited control, in which no significant reduction of sporangia out- put was detected (Fig 5D and 5F). To check whether the phenotype observed in mutant T8 and T16 strains was specific to sporangia production, we decided to measure the vegetative mycelia growth rate in them. As a result, both T8 and T16 showed similar vegetative growth rates, comparable to the controls, on the rye-sucrose medium at 18˚C (S8A and S8B Figs). In addition, we performed an in planta test by inoculating potato leaves (Solanum tuberosum cv. De´sire´e) with T5, T8, T16 and wild type JH19. As MADS-box-edited JH19 strains yielded no sporangia, we selected mycelial discs instead of zoospores for inoculation assays in this experi- ment. T5 and wild type JH19 strains caused complete infectious lesions and developed obvious aerial mycelia. In contrast, T8 and T16 only showed partial watery lesions without formation of aerial mycelia (Figs 5E, 5G and S8C). We further examined the disease area by microscopy: both T5 and wild type JH19 strains produced sporangia in heavily diseased potato leaves, while T8 and T16 only developed sparse mycelia without detectable sporangia (S8C Fig). Editing Avr8-encoding gene in P. infestans HB1501 Phytophthora avirulence (Avr) genes are key determinant factors for gene-for-gene interaction with host plants such as potato and soybean (Dong et al., 2011). We selected Avr8 (PITG_07558), known as an avirulence gene recognized by potato late blight resistance gene R8 [52], as the second editing target. We selected P. infestans strain HB1501 for Avr8 editing since JH19 naturally overcomes the R8-mediated resistance, as was previously established using inoculation assays. As shown in S7A Fig, genome-wide single nucleotide polymorphism (SNP) analysis revealed that HB1501 is diploid. Avr8 encodes a 245 aa RxLR effector and was confirmed as a single-copy gene in strain HB1501 based on the read depth analysis (S7B Fig). The Avr8 protein includes 3 LWY motifs (63–107, 105–162, 165–218 aa), which might con- tribute to its novel activities as an RxLR effector during infection progress (Fig 6A) [53]. Cas12a guides RNAs Avr8-g1 and Avr8-g2 were designed to target Avr8 before the first LWY motif to increase chances of introducing a frame-shift mutation before this motif due to PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 11 / 27 PLOS PATHOGENSModification of AMT for two oomycete species PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 12 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Fig 5. Editing a MADS-box transcription factor coding gene in P. infestans strain JH19. (a) Target sites of two gRNAs in the MADS-box coding sequence. The MEF2-like domain was predicted by InterproScan (version 5.52–86.0). The MEF2-like domain locates at 2–77 aa sites. (b) Detecting of editing events in MADS-box in JH19 transformants. The PCR assay on T8 and T16 revealed two homozygous editing events. (c) Analysis of PCR amplicon sequencing results from (b). The PAM sequences are marked in red, the target sequences are marked in blue, and the stop codons are indicated by red arrowheads. Each dash line represents a deleted nucleotide. (d) Culture scrapings from 10 days Pea broth cultures of wild type JH19, T5, T8 and T16. Scale bars = 0.4 mm. (e) Infection phenotypes of wild type JH19 (WT), transformants T5, T8 and T16 on leaves of susceptible cultivar potato cv. De´sire´e. Detached potato leaves were inoculated with mycelia medium discs, and images were recorded 5 days post inoculation. Scale bars = 2 cm. (f) Quantification of sporangia numbers in 10 μL PEA broth culture of wild type JH19, T5, T8 and T16 in (d). (g) Quantification of lesion size in detached leaves of potato cv. De´sire´e inoculated with wild type JH19 (WT), T5, T8 and T16 in (e). All data represent average values from three independent experiments with the indicated standard deviations. Statistical differences among the samples were analyzed with Sˇı´da´k’s multiple comparisons test (P< 0.0021: , P< 0.0002: *). https://doi.org/10.1371/journal.ppat.1011346.g005 editing events (S7D Fig). The cartoon illustrating the design of the Avr8 CRISPR/Cas12a knockout construct is shown in S7C Fig, with Avr8-g1 and Avr8-g2 targets shown in red in S7D Fig. The transformants were PCR-genotyped for edits in Avr8 using primers amplifying the full-length gene, with PCR amplicons being subsequently sequenced. Out of 27 G418 resistant transformants, variant bands were observed in two of them, T3 and T10, suggesting both Avr8 alleles were altered in them (Fig 6B). Sequencing of the PCR amplicons from the shifted T3 and T10 bands showed deletions spanning both Avr8-g1 and Avr8-g2 target sites within Avr8. Unlike the editing events detected in MADS-box, edits in the Avr8 gene in T3 and T10 resulted in frame-shift mutations predicted to cause early termination of protein translation before the first LWY motif (Fig 6C). As a following step, we performed virulence assays by inoculating detached leaves of R8 transgenic potato (De´sire´e R8) with transformants T3, T10, T22 (unedited) and wild type HB1501 strain. We recorded the infection phenotypes 5 days post zoospore inoculation. Com- pared with unedited T22 and wild type HB1501, T3 and T10 caused lesions of significantly larger size upon infection of R8 transgenic potato leaves. Importantly, both T3 and T10 caused infection symptoms similar to T22 and wild type HB1501 when using detached leaves from susceptible wild type potato (De´sire´e WT) lacking the R8 gene (Fig 6D and 6E). We therefore reason that creating a loss-of-function Avr8 allele enables P. infestans to evade R8-mediated host resistance. Discussion As a widely used approach for transient and stable expression of exogenous genes, AMT has been set up for a broad spectrum of biological categories, including plants, microorganisms and even human cells [20, 54–56]. The AMT protocols for oomycete species, the majority of which are classified as destructive plant pathogens, have been reported, including Phy- tophthora palmivora, Phythium ultimum and Phytophthora infestans [57]. Although these AMT methods have been utilized for years, they require optimization due to generally low transformation efficiencies. Arabidopsis bZIP family transcription factor AtVIP1, known as a binding partner of A. tumefaciens effector VirE2, facilitates VirE2 nuclear import and A. tume- faciens infectivity [43, 58]. Moreover, significant improvement of transformation was observed in an A. tumefaciens transformation assay using AtVIP1-overexpressing tobacco plants [42]. Until now, utilization of AtVIP1 as a factor boosting the AMT efficiency required two steps: (1) generating a stable AtVIP1 transgenic line; (2) transforming a gene of interest into the AtVIP1 transgenic line. Overexpressing AtVIP1 in a stable transgenic line might cause an unpredictable pleiotropic phenotype as well as limit the range of selectable markers available for transformation of a gene of interest. In addition, when applying AtVIP1 the way described above, one is presented with a chicken-and-egg problem in the case of plant genotypes, which are not amenable to AMT to start with. A recent study in wheat reported that co- PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 13 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Fig 6. Editing an avirulence gene Avr8 in P. infestans strain HB1501. (a) Target sites of two gRNAs in the Avr8 coding sequence. The LWY motifs were predicted by InterproScan (version 5.52–86.0). The LWY1 domain locates at 63–107 aa sites, the LWY2 domain locates at 105–162 aa sites, and the LWY3 domain locates at 165–218 aa sites. (b) Detecting editing of Avr8 in HB1501 transformants. The PCR assay in T3 and T10 revealed two homozygous editing events. (c) Analysis of PCR amplicon sequencing results from (b). The PAM sequences are marked in red, the target sequences are marked in blue, and stop codons are indicated by red arrowheads. Each dash line represents a deleted nucleotide. (d) Infection phenotypes of wild type HB1501 (WT), T3, T10 and T22 on detached leaves of R8 transgenic potato. Detached potato leaves were inoculated with zoospores of selected strains, and images were taken 5 days post inoculation. Scale bars = 2 cm. (e) Quantification of lesion sizes in (d). All data represent average values from three independent experiments with the indicated standard deviations. Statistical differences among the samples were analyzed with Sˇı´da´k’s multiple comparisons test (P< 0.0002: ). https://doi.org/10.1371/journal.ppat.1011346.g006 PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 14 / 27 PLOS PATHOGENSModification of AMT for two oomycete species transformation of the wheat gene TaWOX5 from the WUSCHEL family with a gene of interest dramatically increases transformation efficiency, resulting in a lower genotype dependency, in 29 wheat varieties. The latter was similar to our original idea of modifying AMT for oomycete species [59]. Here, we present an alternative method utilizing AtVIP1 to optimize AMT for two oomycete species using prokaryotic expression of AtVIP1 fused with a sequence encoding a T4SS translocation tag in A. tumefaciens. In addition to utilizing AtVIP1, we have done additional modifications to the procedure while developing our modified AMT protocol for P. infestans (S3 Fig). Murashige and Skoog (MS) medium is commonly used for co-cultivation of A. tumefaciens and the explant material [60, 61]. Considering that A. tumefaciens shows a better growth rate in MS medium than in induction medium (IM), we used MS medium instead of IM for co-culturing A. tumefaciens and P. infestans zoospores. The germination of zoospores and transformation efficiency were not affected by this medium substitution. We used geneticin G418 as a selection agent in this study, while different P. infestans wild-type strains showed variation in G418 tolerance. For the JH19 strain, we used 5 μg/mL of G418 for selection of positive transformants, while for the HB1501 strain—10 μg/mL. In previously reported PEG-mediated P. infestans transformation or genome-editing proce- dures, several different situations might occur upon plasmid introduction: (a) the plasmid might integrate into an unstable site in the transformant’s genome; (b) the plasmid DNA might be later degraded by enzymes inside P. infestans cells [62]. The modified AMT proce- dure presented in this study has the advantages of the previously described AMT protocol, including no need to produce protoplasts, no need for a large amount of high-quality plasmid DNA, and, in addition, the integrated T-DNA fragment in the genomic DNA might be more stable than plasmid DNA [63]. The AMT method also comes with its issues e.g. during the T-DNA integration step: (a) some transgene instability might be observed, probably due to rearrangement of the T-DNA region, and/or due to homologous recombination between cop- ies of the transgene inserted into the DNA in the same nucleus; (b) only part of the T-DNA might integrate into genomic DNA [64]. Importantly, during PEG-mediated transformation, the plasmid DNA might not integrate straightaway in protoplasts but at a later stage, during cell differentiation. During the AMT process, T-DNA integration also potentially occurs in encysted zoospores that is followed by emergence of germ tubes. Both of the described above situations would result in chimeric mycelia in resulting transformants, although there are currently no data to indicate that this is the case with AMT [65]. In this study, each AMT experiment started with approximately 8 × 105 zoospores, and 9 of 18 experiments, conducted using our modified AMT method, produced at least 10 G418-resis- tant colonies; 5 of 18 attempts produced at least 20 G418-resistant colonies (Table 1). However, further characterization of acquired transformants revealed that part of the isolates, trans- formed with gfp, that showed resistance to G418, did not show a GFP signal during confocal microscopy analysis or expressed a GFP-size protein detectable by western blotting (Table 1 and S4 Fig). The Southern blot analysis of representative transformants of the HB1501 back- ground suggested that the T-DNA segments with the nptII gene were integrated into all six selected isolates, which were positive for the GFP signal, and one transformant (T1) with no GFP signal (Fig 2F). The latter could be due to the issue (b) of the AMT protocol that was men- tioned above. During the Southern blot analysis, we observed bands of similar small size (< 4.3 kb) in lanes 2, 3, 4, 7 and 8 (Fig 2F). These results might be caused by non-specific hybridization or, possibly, partial T-DNA integration events in selected HB1501 transformants. The above- mentioned observations are consistent with a similar phenomenon observed during plant PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 15 / 27 PLOS PATHOGENSModification of AMT for two oomycete species transformation [66]. To characterize better the integration events in transformants of oomy- cete species, we would like to perform next generation sequencing (NGS), that represents a highly sensitive approach to detect T-DNA insertions in transgenic isolates. There are few more differences between the PEG-mediated transformation and AMT methods, e.g. the different recipient cell types. Importantly, many protoplast cells used for PEG-mediated transformation are coenocytic, while most zoospores used for AMT are single- nucleated and wall-less [63, 67]. However, we have no sufficient evidence to suggest that AMT of P. infestans zoospores could eliminate the issue of potential heterokaryoteyotic transfor- mants, particularly because it was reported that a certain fraction of P. infestans zoospores had been observed to be multinucleate [68]. Additionally, the conditions used in our modified AMT method could not eliminate the possibility of encystment of the zoospores followed by emergence of a germ tube during co-cultivation of zoospores and Agrobacterium cells. To illus- trate such scenario, we included the potential situation of T-DNA integration in multinucleate cells in Fig 1B. We also performed a comparison between the published PEG-mediated trans- formation and AMT protocol, and our modified AMT protocol for P. infestans (S4 Table). As a result, we would define our modified AMT protocol as an option to be considered, rather than top choice, for P. infestans transformation, as transformation efficiencies associated with different protocols are difficult to compare due to different standards used. In addition to P. infestans, genetic transformation methods have not been set up for many biotrophic oomycetes. Martı´nez-Cruz et al. has reported an AMT method for Podosphaera xanthii, a biotrophic fungus that causes cucurbit powdery mildew [48]. Based on the AMT method for P. xanthii, we extended our modified AMT protocol to Plasmopara viticola and successfully obtained a positive transformant (Figs 3 and S5). There are a few more issues that need to be addressed when it comes to AMT of P. viticola, including (a) transformation effi- ciency might vary in different P. viticola isolates; (b) leaves from different grapevine varieties may contribute differently to screening of resistant transformants; (c) it would be difficult to recover P. vitcola transformants from infected grapevine leaves without aseptic conditions; (d) stable expression of an exogenous gene needs further validation in subsequent generations of obtained transformants [1]. Although we have acquired one positive stable transformant of P. viticola based on our modified AMT method with AtVIP1, we still need to do more transfor- mation attempts and analysis to draw a solid conclusion, in respect to the transformation effi- ciencies, in the future. Surely, we know the importance of acquiring a stable transformant of P. viticola, while our experimental procedure comes with a few shortcomings: (a) the transforma- tion efficiency is not high enough as compared to other oomycete species; (b) we have not per- formed a sufficient number of transformation assays with other P. viticola isolates. Consequently, the extensibility of our strategy to P. viticola suggests that AtVIP1 could be con- sidered for modifying AMT methods for other oomycete species, especially some hard-to- transform biotrophic oomycetes, such as Bremia lactucae. Unlike gene overexpression, integration of a CRISPR/Cas cassette in transformants does not guarantee successful editing of a gene target. Previous reports have demonstrated that edit- ing efficiency varies greatly, from 1% to 100%, among different eukaryotic pathogens [69]. The editing frequency of CRISPR/Cas12a in P. infestans reached 13% by using the PEG transfor- mation method [10, 12]. In this study, editing of the MADS-box gene resulted in two homozy- gous mutants out of 16 transformants, and editing of Avr8 genes resulted in two homozygous mutants out of 28 transformants (Table 1). Although introducing AtVIP1 does not seem to increase the frequency of CRISPR/Cas12a-mediated genome editing events in P. infestans, our modified AMT method still provides a valuable option for future related studies. We selected two target genes for CRISPR/Cas12a-mediated genome editing in this study based on two criteria. The first is avoiding multi-copy target genes to reduce the difficulty of PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 16 / 27 PLOS PATHOGENSModification of AMT for two oomycete species editing and transformant screening, and the second is that transformants with successful edit- ing events should present explicit phenotypes. The cycle of aerial asexual sporangia dispersal plays a crucial role in late blight development [70]. Thus, we selected MADS-box as our first target gene that is reported to express only in sporulating mycelia and spores of P. infestans. MADS-box (PITG_07059) was first identified by Leesutthiphonchai and Judelson in 2018 and MADS-box transcription factors play significant roles in eukaryotes [50]. We utilized modified AMT and CRISPR/Cas12a-mediated genome editing methods and obtained two edited MADS-box mutants (Fig 5). Consistently with the results obtained using the RNAi approach, our edited MADS-box mutants produced no sporangia and showed reduced ability to infect potato leaves (Fig 5). RNAi is triggered by siRNAs whose silencing efficiency is not guaranteed and varies widely in different transformants. In contrast to RNAi methods, genome editing would provide more stable phenotypic data for further research on P. infestans. Since during CRISPR/Cas12a-mediated genome editing, based on the AMT method, the Cas12a expression cassette stably integrates into the P. infestans genome, an additional self-fertilization step could be used to remove Cas12a from the original transformant, similarly to the situation in planta [71]. Disrupting the recognition of phytopathogen avirulence (AVR) genes by plant resistance (R) genes normally produces easily observable phenotypes. Thus, we selected Avr8 (also called AVRSmira2 and PITG_07558), a single-copy AVR gene recognized by potato late blight resis- tance gene R8, as our second genome editing target. We chose P. infestans strain HB1501 for Avr8 gene editing because seemingly the JH19 strain broke down the R8-mediated resistance in potato. Avr8 was identified by analysis of variance (ANOVA) using the average AUDPC val- ues of both its responses to the R gene and field trials [72]. The broad spectrum late blight resistance gene R8 that recognizes Avr8 was cloned from Solanum demissum, based on a previ- ously published coarse map position on the lower arm of chromosome IX, and the correlation between the expression levels of Avr8 and R8-mediated resistance had been proven in a previ- ous study [73]. However, an inoculation assay of an Avr8 knockout P. infestans mutant on potato carrying the R8 gene has not been reported yet. In our study, we produced two genome-edited Avr8 mutants of P. infestans (strain HB1501), and both mutants induced infec- tion lesions in leaves of the R8 transgenic potato line cv. De´sire´e (Fig 6D and 6E). Collectively, generating P. infestans transformants or genome edited mutants via our modi- fied AMT procedure is less laborious and results in an acceptable transformation rate as com- pared with the previously reported AMT protocol and PEG-mediated protoplast transformation method. Successfully acquiring a stable transformant of P. viticola gives a strong hint of gaining a potential advantage by utilizing proteins, such as AtVIP1, that play an important role in host cells during the AMT process. Although AMT has already been estab- lished in plenty of phytopathogen species, it would be particularly interesting to investigate whether our modified AMT protocol would increase the efficiency of transformation or that of CRISPR/Cas-mediated genome editing in them or even extend the effect to other oomycete or plant species. Materials and methods Growth conditions for P. infestans, P. viticola, bacteria and plants P. infestans strains were cultured on rye-sucrose medium (agar 15g/L) at 18˚C. P. infestans strain JH19 was isolated from infected tomato in San Diego Country, California in 1982 and kindly provided by Howard S. Judelson lab [12]. P. infestans strain HB1501 was isolated from Hebei Province in 2015 [74]. Plasmopara viticola isolate BS5 was kindly provided by Dr. Linfei Shangguan at Nanjing Agricultural University and was isolated in Nanjing, China in 2017. A. PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 17 / 27 PLOS PATHOGENSModification of AMT for two oomycete species tumefaciens strains and E. coli strains were grown on Luria–Bertani (LB) agar (NaCl 10g/L, Yeast extract 5g/L, Tryptone 5g/L, agar 15g/L) at 28˚C and 37˚C, respectively. Solanum tubero- sum L. cv. De´sire´e (2n = 4x = 48) and trangenic lines were grown in soil or in MS medium (MES 0.5 g/L, sucrose 20 g/L, agar 8 g/L, pH5.8), after seed surface sterilization, and main- tained in vitro. Vitis vinifera (Cultivar grape, Zitian Seedless, A17) was used for culturing of P. viticola isolate BS5. All plants were grown in environment-controlled growth chambers under long-day conditions (16 h light/8 h dark cycle at 140 μE sec-1m-2 light intensity) at 22˚C. Construction of plasmids Primer sequences used in these cloning procedures are described in S1 Table, and plasmids and cloning strategies are summarized in S2 Table. For AtVIP1 gene expression in A. tumefa- ciens, the coding sequences of AtVIP1 (AT1G43700) and virF (NC_003065.2) without the N terminal 42 amino acids (ΔN42virF) were PCR-amplified, using A. thaliana Col-0 cDNA library and A. tumefaciens C58 gDNA, respectively, as templates and cloned into p533BL. A pCB302B backbone construct with the virB1 promoter from A. tumefaciens C58 was used to drive acetosyringone induced gene expression in A. tumefaciens. To construct binary plasmid for P. infestans transformation, Bremia lactucae Ham34 pro- moter and terminator were PCR-amplified and cloned into I-CeuI site of pPZP-RCS2, while nptII expression cassette driven by Phytophthora sojae RPL41 promoter and B. lactucae Hsp70 terminator was inserted into AscI site of pPZP-RCS2, which resulted in pLY40. The coding sequence of gfp was PCR-amplified and cloned into I-CeuI site of pLY40 separately to obtain pLY40-gfp. The sequences of the RPL41 promoter from P. sojae, Ham34 promoter and termi- nator, Hsp70 terminator from B. lactucae were all PCR-amplified from pYF515, a plasmid con- struct used for genome editing in P. sojae [75]. To construct binary plasmids for CRISPR/Cas12a mediated P. infestans genome editing, the coding sequences of NLS derived from a P. sojae bZIP transcription factor [10], human codon-optimized LbCas12a from p33lb [76] and artificial synthetic polyA-crRNA-Ham34 ter- minator segment (S6C and S7C Figs) were fused and cloned into I-CeuI/PacI sites in pLY40, to obtain pLY40-MADS-box-KO and pLY40-Avr8-KO. The plasmid sequences of pLY40, pLY40-gfp and pLY40-Cas12a-gfp are presented in Appendix S1. Transient transformation assays in wheat tissue Transient expression assays in wheat tissue were performed as previously described for N. benthamiana with some modifications [77]. A. tumefaciens overnight culture was diluted in LB liquid medium without antibiotics, grown for 3–4 h and adjusted to OD600 = 0.5. Leaves and roots from 4-weeks old in vitro wheat plant were collected and divided into 5 mm seg- ments, then immersed for 10 min in Agrobacterium suspension, placed on MS medium at 22˚C for 3 days in growth chamber. Leaf and root segments were then rinsed by sterilized water and transferred into GUS staining solution and incubated at 37˚C overnight [78]. Agrobacterium meditated transformation of P. infestans zoospores AMT of P. infestans zoospores followed previous published method with few modifications and summarized in S3 Fig [7]. Briefly, P. infestans strains were ready for zoospore-induction after 14 days of growth on solid rye-sucrose medium (90mm petri dish plates). Approximately 5 mL of ice-cold sterilized water was used to flush and soak P. infestans culture and the culture plates were plated at 4˚C for 2 hours, and zoospores will be released (100 zoospores counts/μL for isolate JH19, 200 zoospores counts/μL for isolate HB1501). A. tumefaciens strains with PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 18 / 27 PLOS PATHOGENSModification of AMT for two oomycete species constructed plasmids for transformation was growing overnight with proper antibiotics (25 μg/mL rifampicin, 50 μg/mL kanamycin, 100 μg/mL spectinomycin), 1 mL of Agrobacter- ium culture was added into 50 mL LB liquid medium and grown for 3–4 h till OD600 reach 1.0. Agrobacterium cells were centrifuged at 4000 rpm for 10 min to collect cells and resuspended by MS liquid medium with 200 μM acetosyringone. Agrobacterium suspension was cultured at room temperature for 1 h and then mixed with harvested fresh P. infestans zoospores (about 8 × 105 zoospores for both isolate JH19 and HB1501). The mixture was cultured at room tem- perature for 45 min and zoospores were collected by centrifuge at 265 g for 5 min, each 200 μL of zoospore culture was spread at a 5 cm × 5 cm Nytran membrane upon MS solid medium plate (8 g/L agar), and the plates were cultured in dark at 22˚C for 4 d. The membrane was then moved upside down on Plich medium (0.5 g/L KH2PO4, 0.25 g/L MgSO4.7H2O, 1 g/L Asparagine, 1 mg/L Thiamine, 0.5 g/L Yeast extract, 10 mg/L β–sitosterol, 25 g/L Glucose, 15 g/L agar) with 1.5 mg/L G418 and 300 mg/L timentin, and plates were cultured at 18˚C for 4 d, the membranes were then removed and germinated mycelia were supposed to be observed, the plates were keep in 18˚C for 4 d. Melt rye-sucrose medium with 3 mg/L of G418 was then covered on Plich medium plates for further selection, the G418 concentration could be increased up to 5 mg/L at this stage. Agrobacterium meditated transformation of P. viticola zoospores AMT of P. viticola zoospores followed previous published method on cucurbit powdery mil- dew pathogen Podosphaera xanthii and summarized in S5A Fig [48]. Briefly, culture of A. tumefaciens strain EHA105 with pV1F and proper T-DNA construct was prepared following the same methods that described in AMT of P. infestans. Separately, the sporangia from P. viti- cola BS5 inoculated grape leaves were harvested by immersion of infected tissues in 30 mL of sterilized water with 0.01% Tween-20 and keep in room temperature for 1 h until zoospores released. About 10 mL (about 1 × 106 zoospores) P. viticola zoospores were gently mixed with same volume of A. tumefaciens suspension and co-cultivated for 1 h at room temperature in the dark in an orbital shaker at 65 rpm. Zoospores were then centrifuged for collection (265 g for 5 min), and zoospores in 1 mL of residual were deposited in young detached grape leaves (must be Zitian Seedless, A17 for the best inoculation rate). Two days after inoculation, the grape leaves were rinsed with 5 mg/L G418 and 300 mg/L Timentin to kill A. tumefaciens and select P. viticola transformants. To be noted, rinsing infection samples with G418 (5 mg/L) and Timentin (300 mg/L) should have no visible negative effect on grape leaves during the above-mentioned procedure. To acquire the stable transformant of P. viticola, the obtained isolate was sub-inoculated at zoospore stage on young grape leaves without G418 treatment. After at least three rounds of sub-inoculation, the P. viticola isolate that showed stable GFP sig- nal, was defined as a stable transformant. Virtualization of GFP CLSM (Leica AF6000 modular microsystems) was used to take pictures of P. infestans. A 489-nm line from an argon ion laser were used to excite green fluorescent protein (GFP). For each assay, six independent leaves were observed for each experiment and each experiment has at least 3 repeats. sgRNA design and cloning CRISPR/Cas12a targets for MADS-box (PITG_07059) and Avr8 (PITG_07558) were designed with overall consideration based on output data from EuPaGDT, CRISPOR and Deep-Cpf1 [79–81]. The used crRNAs were carefully inspected with sequencing data of P. infestans JH19 PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 19 / 27 PLOS PATHOGENSModification of AMT for two oomycete species and 1501 to avoid off-target events. DNA oligonucleotides contains crRNAs and direct repeats (DR) were artificial synthesized and cloned into pLY40 based strategy described above and in S6C and S7C Figs. Detection of target gene editing Genomic DNA (gDNA) of P. infestans strain or relative transformants was first isolated from liquid cultural mycelia. Specifically, 5 mycelia discs of each P. infestans strain from rye-sucrose medium plates were cut and placed into 10 mL of PEA broth, and mycelia for gDNA extrac- tion were harvested after 5 days growth in dark at 18˚C. Mycelia were then blot up by filter paper and gDNA was extracted with Omega E.Z.N.A. Plant DNA Kit (HP). To confirm edit- ing, primer pair F: 5’-ATGGGCCGCAAGAAGATCCAG-3’ and R: 5’-TCAAACAGCCACA CGTTGACGCTTG-3’ were for checking MADS-box editing in P. infestans JH19 derived transformants, and primer pair F: 5’-ATGCGCTCAATCCAACTTCTG-3’ and R: 5’-TTAC GATGTTTTCGCTTCTTTAAAAAG-3’ was for checking Avr8 editing in P. infestans 1501 derived transformants. Protein analysis P. infestans protein assay referred to previously described method [12]. Briefly, mycelia were collected from PEA broth culture and total protein was extracted with Beyotime RIPA P0013B Lysis Buffer. Immunoblots were performed as described [78]. Total protein was eluted in sodium dodecylsulfate (SDS) sampling buffer and proceed to western blot analysis. LbCas12a was detected by immunoblotting with anti-LbCas12a (Cpf1) antibody (Sigma/SAB4200777, dilution 1:4000), followed by a secondary antibody conjugated to FITC (ThermoFisher Scien- tific, dilution 1:5000). Southern blot analysis For T-DNA integration analysis, we performed southern blot with genomic DNA obtained from P. infestans transformants or wild-type (WT) isolates. Four μg genomic DNA was digested with restriction enzyme HindIII (Takara), and separated on DNA agarose gel (1%) and then blotted on positively charged nylon membranes. A nptII gene fragment was used as probe. Further preparations of probes DIG-labeling, hybridization and chemiluminescent detection were conducted according to the operation protocol of DIG-High Prime DNA label- ing and Detection Starter Kit (ROCHE/11585614910). Leaf inoculation with P. infestans For zoospores inoculation assays, zoospores of P. infestans strains were collected from plates followed by the same method described above. 10 μL drops with 200,000 to 400,000 zoospores per mL were inoculated on detached leaves from 4 to 6 weeks old potato plants of wild type De´sire´e and R8 trangenic potato lines. For mycelial disc inoculation assays, detached potato leaves were inoculated with mycelial discs (5 mm diameter) taken from the edge of 7 days-old rye-sucrose medium culture of P. infestans. Inoculated leaves were kept in a plastic tray (30 × 44 × 8 cm) covered with polypropylene film at 22 ± 3˚C. Results were recorded 5 days post inoculation. For each inoculation assay, 8 independent leaves were used for each experi- ment and each experiment has at least 3 repeats. PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 20 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Supporting information S1 Table. Primers used in this study. (XLSX) S2 Table. Plasmids used in this study. (XLSX) S3 Table. Sequence similarity data used in Fig 1A. (XLSX) S4 Table. Comparisons of modified AMT protocol and previous reported transformation protocols for P. infestans. (XLSX) S1 Fig. Translocating AtVIP1 from A. tumefaciens to host cells. (a) The expression cassette used for translocating AtVIP1 fused with GFP (b) Confocal microscopy observation of N. benthamiana leaves infiltrated with A. tumefaciens EHA105 carrying the construct described in (a). Images were taken at 3 days post infiltration. Images are single confocal sections and are representative of images obtained in three independent experiments. White arrows indicate observed nuclei in tobacco cells. Scale bars = 40 μm. Three independent experiments were per- formed for each assay with similar results. (TIF) S2 Fig. AtVIP1 enhances transient AMT efficiency in wheat tissues. (a) Schematic represen- tation of plasmid constructs used in this experiment. The pWMB110-gus construct contains a β-glucuronidase expression cassette, carrying the maize adh1 intron, in the T-DNA region. (b-c) Transient transformation on wheat leaf and root segments. Dissected wheat tissue seg- ments were inoculated with A. tumefaciens EHA105 carrying the binary plasmid pWMB110- gus and either pV1F or the control plasmid pEV. At 3 days post inoculation, GUS activity was analyzed by histochemical staining. Scale bars = 2 mm. At least 50 leaf or root segments were recorded in each experiment. Each experiment was repeated 3 times and representative results were presented. (d) Quantification of root segments that expressed the gus gene in (c). Statisti- cal differences among the samples were analyzed with Sˇı´da´k’s multiple comparisons test (P< 0.0001: ***). (TIF) S3 Fig. Schematic outline of modified AMT for P. infestans. (TIF) S4 Fig. Western blot characterization of P. infestans HB1501 transformants obtained using pLY40-gfp and pV1F. Total protein samples were purified from 14 transformants with a GFP signal (up) and 14 transformants without a GFP signal (down). All blots were probed with an anti-GFP antibody. The protein blot was stained with Ponceau S to confirm equal loading. (TIF) S5 Fig. Obtaining Plasmopara viticola BS5 transformant T1, expressing gfp, using the opti- mized AMT method. (a) Schematic outline of the optimized AMT method for P. viticola BS5. (b) AMT with only pLY40-gfp produced no G418 resistant transformants of P. viticola BS5 (left), while AMT with pLY40-gfp and pV1F produced the transformant T1 (white arrow) that is resistant to G418 (right). Scale bars = 2 mm. (TIF) PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 21 / 27 PLOS PATHOGENSModification of AMT for two oomycete species S6 Fig. The gRNA target regions used for MADS-box genome editing in P. infestans. (a) Genome-wide allele ratio analysis of P. infestans JH19. (b) Copy number of PITG_07059 (MADS-box) relative to single-copy control gene (= 1.0), determined based on read depth in DNA library of JH19 strain. (c) Schematic representation of the constructs used in this experi- ment. The pLY40-MADS-box-KO with either pV1F or pEV were used for P. infestans transfor- mation in this experiment. Two gRNAs for MADS-box editing are named as g1 and g2. (d) Gene sequence of PITG_07059 (MADS-box). Sequences marked in red are targeted by g1 and g2. (e) Sequencing chromatograms of MADS-box in wild type JH19, T8 and T16. Both T8 and T16 showed single peaks in either the g1 (left) or g2 (right) target sites. The wild type sequences with gRNA targets are shown at the top of the panel; black arrows indicate the 5’ border of the detected deletion. (TIF) S7 Fig. The gRNA regions used for Avr8 editing in P. infestans. (a) Genome-wide allele ratio analysis of P. infestans HB1501. (b) Copy number of PITG_07558 (Avr8) relative to single- copy control gene (= 1.0), determined based on read depth in DNA library of HB1501 strain. (c) Schematic representation of the constructs used in this experiment. (d) Gene sequence of PITG_07558 (Avr8). Two selected gRNA target regions are marked in red. (e) Sequencing chromatograms of Avr8 in T3, T10 and T22 of HB1501. Both T3 and T22 showed single peaks in both g1 and g2 target sites. The wild type sequences with gRNA targets are shown at the top of the panel; black arrows indicate the 5’ border of the detected deletion. (TIF) S8 Fig. Vegetative growth of P. infestans strains. (a) Mycelia cultured on the rye-sucrose medium were photographed at 5 days post inoculation. (b) Quantification of mycelium diame- ter of P. infestans strains in (a). All data represent average values from three independent experiments with the indicated standard deviations. (c) Microscopy images of the opposite side of the inoculated region of detached potato leaves in Fig 5E show the details of mycelia generated during infection. Images were taken at 5 days post inoculation. White arrowheads indicate the observed sporangia. Scale bars = 1 mm. (TIF) S1 Appendix. Sequences of pLY40, pLY40-gfp, pLY40-Cas12a-gfp. (DOCX) Acknowledgments We thank Dr. Howard Judelson (UCR, USA) and Prof. Yuanchao Wang (NAU) for the discus- sion on Phytophthora genome editing. Potato R8 transgenic line was kindly provided by Dr. Jack Vossen (WUR, Netherlands). The Plasmopara viticola BS5 was a kind gift from Prof. Lin- fei Shangguan (NAU). We thank Dr. Vitaly Citovsky from Stonybrook University for support- ive discussion on the original project design. We thank Ms. Ying Zheng (NAU) for confocal microscopy. We thank Ms. Zhao Hu, Dr. Xinyu Liu and Dr. Changling Mou (NAU) for their help with Southern blot analysis and constructive comments on this study. Author Contributions Conceptualization: Luyao Wang, Sanwen Huang, Suomeng Dong. Data curation: Luyao Wang, Fei Zhao, Haohao Liu, Suomeng Dong. Formal analysis: Luyao Wang, Fei Zhao, Haohao Liu, Suomeng Dong. PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 22 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Funding acquisition: Luyao Wang, Sanwen Huang, Suomeng Dong. Investigation: Luyao Wang, Fei Zhao, Haohao Liu, Suomeng Dong. Methodology: Luyao Wang, Fei Zhao, Haohao Liu, Han Chen, Fan Zhang, Vladimir Nekrasov, Suomeng Dong. Project administration: Luyao Wang, Sanwen Huang, Suomeng Dong. Resources: Luyao Wang, Sanwen Huang, Suomeng Dong. Software: Luyao Wang, Fan Zhang, Sanwen Huang, Suomeng Dong. Supervision: Luyao Wang, Sanwen Huang, Suomeng Dong. Validation: Luyao Wang, Sanwen Huang, Suomeng Dong. Visualization: Luyao Wang, Suomeng Dong. Writing – original draft: Luyao Wang, Sanwen Huang, Suomeng Dong. Writing – review & editing: Luyao Wang, Han Chen, Fan Zhang, Suhua Li, Tongjun Sun, Vladimir Nekrasov, Sanwen Huang, Suomeng Dong. References 1. Dubresson R, Kravchuk Z, Neuhaus J-M, Mauch-Mani B. Optimisation and comparison of transient expression methods to express the green fluorescent protein in the obligate biotrophic oomycete Plas- mopara viticola. Vitis. 2008; 47(4):235–40. 2. Fry W. Phytophthora infestans: the plant (and R gene) destroyer. Molecular plant pathology. 2008; 9 (3):385–402. https://doi.org/10.1111/j.1364-3703.2007.00465.x PMID: 18705878 3. Chen H, Shu H, Wang L, Zhang F, Li X, Ochola SO, et al. Phytophthora methylomes are modulated by 6mA methyltransferases and associated with adaptive genome regions. Genome biology. 2018; 19 (1):1–16. 4. Sabbadin F, Urresti S, Henrissat B, Avrova AO, Welsh LR, Lindley PJ, et al. Secreted pectin monooxy- genases drive plant infection by pathogenic oomycetes. Science. 2021; 373(6556):774–9. https://doi. org/10.1126/science.abj1342 PMID: 34385392 5. Luo X, Tian T, Feng L, Yang X, Li L, Tan X, et al. Pathogenesis-related protein 1 suppresses oomycete pathogen by targeting against AMPK kinase complex. Journal of Advanced Research. 2022. https://doi. org/10.1016/j.jare.2022.02.002 PMID: 36585103 6. Ah-Fong AM, Bormann-Chung CA, Judelson HS. Optimization of transgene-mediated silencing in Phy- tophthora infestans and its association with small-interfering RNAs. Fungal Genetics and Biology. 2008; 45(8):1197–205. https://doi.org/10.1016/j.fgb.2008.05.009 PMID: 18599326 7. Vijn I, Govers F. Agrobacterium tumefaciens mediated transformation of the oomycete plant pathogen Phytophthora infestans. Molecular Plant Pathology. 2003; 4(6):459–67. https://doi.org/10.1046/j.1364- 3703.2003.00191.x PMID: 20569405 8. Cvitanich C, Judelson HS. Stable transformation of the oomycete, Phytophthora infestans, using micro- projectile bombardment. Current genetics. 2003; 42(4):228–35. https://doi.org/10.1007/s00294-002- 0354-3 PMID: 12589474 9. 10. 11. van West P, de Jong AJ, Judelson HS, Emons AMC, Govers F. TheipiO Gene ofPhytophthora infestan- sIs Highly Expressed in Invading Hyphae during Infection. Fungal Genetics and Biology. 1998; 23 (2):126–38. https://doi.org/10.1006/fgbi.1998.1036 PMID: 9578626 Fang Y, Tyler BM. Efficient disruption and replacement of an effector gene in the oomycete P hytophthora sojae using CRISPR/C as9. Molecular plant pathology. 2016; 17(1):127–39. https://doi. org/10.1111/mpp.12318 PMID: 26507366 van den Hoogen J, Govers F. Attempts to implement CRISPR/Cas9 for genome editing in the oomycete Phytophthora infestans. bioRxiv. 2018:274829. 12. Ah-Fong AM, Boyd AM, Matson ME, Judelson HS. A Cas12a-based gene editing system for Phy- tophthora infestans reveals monoallelic expression of an elicitor. Molecular plant pathology. 2021. https://doi.org/10.1111/mpp.13051 PMID: 33724663 PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 23 / 27 PLOS PATHOGENSModification of AMT for two oomycete species 13. Mansfield J, Genin S, Magori S, Citovsky V, Sriariyanum M, Ronald P, et al. Top 10 plant pathogenic bacteria in molecular plant pathology. Molecular plant pathology. 2012; 13(6):614–29. https://doi.org/ 10.1111/j.1364-3703.2012.00804.x PMID: 22672649 14. Binns AN, Thomashow MF. Cell biology of Agrobacterium infection and transformation of plants. Annual Reviews in Microbiology. 1988; 42(1):575–606. 15. Gelvin SB. Integration of Agrobacterium T-DNA into the plant genome. Annual review of genetics. 2017; 51:195–217. https://doi.org/10.1146/annurev-genet-120215-035320 PMID: 28853920 16. Zhang X, Henriques R, Lin S-S, Niu Q-W, Chua N-H. Agrobacterium-mediated transformation of Arabi- dopsis thaliana using the floral dip method. Nature protocols. 2006; 1(2):641–6. https://doi.org/10.1038/ nprot.2006.97 PMID: 17406292 17. Gallois P, Marinho P. Leaf disk transformation using Agrobacterium tumefaciens-expression of heterol- ogous genes in tobacco. Plant gene transfer and expression protocols. 1995:39–48. https://doi.org/10. 1385/0-89603-321-X:39 PMID: 8563823 18. Hiei Y, Komari T, Kubo T. Transformation of rice mediated by Agrobacterium tumefaciens. Plant molec- ular biology. 1997; 35(1):205–18. PMID: 9291974 19. D’spain S, Andrade PI, Brockman NE, Fu J, Wickes BL. Agrobacterium tumefaciens-Mediated Trans- formation of Candida glabrata. Journal of Fungi. 2022; 8(6):596. https://doi.org/10.3390/jof8060596 PMID: 35736079 20. Soltani J, van Heusden GPH, Hooykaas PJ. Agrobacterium-mediated transformation of non-plant organisms. Agrobacterium: from biology to biotechnology. 2008:649–75. 21. Michielse CB J Hooykaas PJ, JJ van den Hondel CA, J Ram AF. Agrobacterium-mediated transforma- tion of the filamentous fungus Aspergillus awamori. Nature protocols. 2008; 3(10):1671–8. https://doi. org/10.1038/nprot.2008.154 PMID: 18833205 22. Cardoza RE, Vizcaino JA, Hermosa MR, Monte E, Gutieˊrrez S. A comparison of the phenotypic and genetic stability of recombinant Trichoderma spp. generated by protoplast-and Agrobacterium-medi- ated transformation. Journal of Microbiology. 2006; 44(4):383–95. PMID: 16953173 23. Michielse CB, Hooykaas PJ, van den Hondel CA, Ram AF. Agrobacterium-mediated transformation as a tool for functional genomics in fungi. Current genetics. 2005; 48:1–17. https://doi.org/10.1007/ s00294-005-0578-0 PMID: 15889258 24. 25. Takahara H, Tsuji G, Kubo Y, Yamamoto M, Toyoda K, Inagaki Y, et al. Agrobacterium tumefaciens- mediated transformation as a tool for random mutagenesis of Colletotrichum trifolii. Journal of General Plant Pathology. 2004; 70:93–6. dos Reis MCl Fungaro MHP, Duarte RTD, Furlaneto L, Furlaneto MC. Agrobacterium tumefaciens- mediated genetic transformation of the entomopathogenic fungus Beauveria bassiana. Journal of microbiological methods. 2004; 58(2):197–202. https://doi.org/10.1016/j.mimet.2004.03.012 PMID: 15234517 26. Dobinson KF, Grant SJ, Kang S. Cloning and targeted disruption, via Agrobacterium tumefaciens-medi- ated transformation, of a trypsin protease gene from the vascular wilt fungus Verticillium dahliae. Cur- rent genetics. 2004; 45:104–10. https://doi.org/10.1007/s00294-003-0464-6 PMID: 14618375 27. Rolland S, Jobic C, Fèvre M, Bruel C. Agrobacterium-mediated transformation of Botrytis cinerea, sim- ple purification of monokaryotic transformants and rapid conidia-based identification of the transfer- DNA host genomic DNA flanking sequences. Current genetics. 2003; 44:164–71. https://doi.org/10. 1007/s00294-003-0438-8 PMID: 12937946 28. Sarah F, Kapoor P, Lee M-h, Briley A, Nairn CJ. Agrobacterium tumefaciens-mediated transformation of Fusarium circinatum. Mycological Research. 2001; 105(3):259–64. 29. Mullins ED, Chen X, Romaine P, Raina R, Geiser D, Kang S. Agrobacterium-mediated transformation of Fusarium oxysporum: an efficient tool for insertional mutagenesis and gene transfer. Phytopathology. 2001; 91(2):173–80. https://doi.org/10.1094/PHYTO.2001.91.2.173 PMID: 18944391 30. Abuodeh RO, Orbach MJ, Mandel MA, Das A, Galgiani JN. Genetic transformation of Coccidioides immi- tis facilitated by Agvobactevium tumefaciens. Journal of Infectious Diseases. 2000; 181(6):2106–10. 31. Opabode JT. Agrobacterium-mediated transformation of plants: emerging factors that influence effi- ciency. Biotechnology and Molecular Biology Reviews. 2006; 1(1):12–20. 32. van der Fits L, Deakin EA, Hoge JHC, Memelink J. The ternary transformation system: constitutive virG on a compatible plasmid dramatically increases Agrobacterium-mediated plant transformation. Plant molecular biology. 2000; 43(4):495–502. https://doi.org/10.1023/a:1006440221718 PMID: 11052201 33. Wu D, Navet N, Liu Y, Uchida J, Tian M. Establishment of a simple and efficient Agrobacterium-medi- ated transformation system for Phytophthora palmivora. BMC microbiology. 2016; 16(1):1–8. https:// doi.org/10.1186/s12866-016-0825-1 PMID: 27599726 PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 24 / 27 PLOS PATHOGENSModification of AMT for two oomycete species 34. Walker CA, van West P. Zoospore development in the oomycetes. Fungal biology reviews. 2007; 21 (1):10–8. 35. Ahonsi MO, Banko TJ, Hong C. A simple in-vitro ‘wet-plate’method for mass production of Phytophthora nicotianae zoospores and factors influencing zoospore production. Journal of microbiological methods. 2007; 70(3):557–60. https://doi.org/10.1016/j.mimet.2007.06.012 PMID: 17683817 36. Eye LL, Sneh B, Lockwood J. Factors affecting zoospore production by Phytophthora megasperma var. sojae. Phytopathology. 1978; 68(12):1766. 37. Gelvin SB. Plant proteins involved in Agrobacterium-mediated genetic transformation. Annual review of phytopathology. 2010; 48:45–68. https://doi.org/10.1146/annurev-phyto-080508-081852 PMID: 20337518 38. Djamei A, Pitzschke A, Nakagami H, Rajh I, Hirt H. Trojan horse strategy in Agrobacterium transforma- tion: abusing MAPK defense signaling. Science. 2007; 318(5849):453–6. https://doi.org/10.1126/ science.1148110 PMID: 17947581 39. Hwang H-H, Gelvin SB. Plant proteins that interact with VirB2, the Agrobacterium tumefaciens pilin pro- tein, mediate plant transformation. The Plant Cell. 2004; 16(11):3148–67. https://doi.org/10.1105/tpc. 104.026476 PMID: 15494553 40. 41. 42. 43. 44. Li J, Vaidya M, White C, Vainstein A, Citovsky V, Tzfira T. Involvement of KU80 in T-DNA integration in plant cells. Proceedings of the National Academy of Sciences. 2005; 102(52):19231–6. https://doi.org/ 10.1073/pnas.0506437103 PMID: 16380432 Tenea GN, Spantzel J, Lee L-Y, Zhu Y, Lin K, Johnson SJ, et al. Overexpression of several Arabidopsis histone genes increases Agrobacterium-mediated transformation and transgene expression in plants. The Plant Cell. 2009; 21(10):3350–67. https://doi.org/10.1105/tpc.109.070607 PMID: 19820187 Tzfira T, Vaidya M, Citovsky V. Increasing plant susceptibility to Agrobacterium infection by overexpres- sion of the Arabidopsis nuclear protein VIP1. Proceedings of the National Academy of Sciences. 2002; 99(16):10435–40. https://doi.org/10.1073/pnas.162304099 PMID: 12124400 Tzfira T, Vaidya M, Citovsky V. VIP1, an Arabidopsis protein that interacts with Agrobacterium VirE2, is involved in VirE2 nuclear import and Agrobacterium infectivity. The EMBO Journal. 2001; 20(13):3596– 607. https://doi.org/10.1093/emboj/20.13.3596 PMID: 11432846 Li J, Krichevsky A, Vaidya M, Tzfira T, Citovsky V. Uncoupling of the functions of the Arabidopsis VIP1 protein in transient and stable plant genetic transformation by Agrobacterium. Proceedings of the National Academy of Sciences. 2005; 102(16):5733–8. 45. Vergunst AC, Schrammeijer B, den Dulk-Ras A, de Vlaam CM, Regensburg-Tuïnk TJ, Hooykaas PJ. VirB/D4-dependent protein translocation from Agrobacterium into plant cells. Science. 2000; 290 (5493):979–82. https://doi.org/10.1126/science.290.5493.979 PMID: 11062129 46. Uauy C. Wheat genomics comes of age. Current opinion in plant biology. 2017; 36:142–8. https://doi. org/10.1016/j.pbi.2017.01.007 PMID: 28346895 47. Judelson HS, Michelmore RW. Transient expression of genes in the oomycete Phytophthora infestans using Bremia lactucae regulatory sequences. Current Genetics. 1991; 19(6):453–9. 48. Martı´nez-Cruz J, Romero D, de Vicente A, Pe´ rez-Garcı´a A. Transformation of the cucurbit powdery mil- dew pathogen Podosphaera xanthii by Agrobacterium tumefaciens. New Phytologist. 2017; 213 (4):1961–73. https://doi.org/10.1111/nph.14297 PMID: 27864969 49. 50. Lin L, Ye W, Wu J, Xuan M, Li Y, Gao J, et al. The MADS-box transcription factor PsMAD1 is involved in zoosporogenesis and pathogenesis of Phytophthora sojae. Frontiers in Microbiology. 2018; 9:2259. https://doi.org/10.3389/fmicb.2018.02259 PMID: 30319576 Leesutthiphonchai W, Judelson HS. A MADS-box transcription factor regulates a central step in sporu- lation of the oomycete Phytophthora infestans. Molecular Microbiology. 2018; 110(4):562–75. https:// doi.org/10.1111/mmi.14114 PMID: 30194883 51. Bin Moon S, Lee JM, Kang JG, Lee N-E, Ha D-I, Kim DY, et al. Highly efficient genome editing by CRISPR-Cpf1 using CRISPR RNA with a uridinylate-rich 30-overhang. Nature communications. 2018; 9 (1):1–11. 52. Jiang R, Li J, Tian Z, Du J, Armstrong M, Baker K, et al. Potato late blight field resistance from QTL dPI09c is conferred by the NB-LRR gene R8. Journal of experimental botany. 2018; 69(7):1545–55. https://doi.org/10.1093/jxb/ery021 PMID: 29385612 53. Dong S, Ma W. How to win a tug-of-war: the adaptive evolution of Phytophthora effectors. Current Opin- ion in Plant Biology. 2021; 62:102027. https://doi.org/10.1016/j.pbi.2021.102027 PMID: 33684881 54. Hooykaas PJ, van Heusden GPH, Niu X, Reza Roushan M, Soltani J, Zhang X, et al. Agrobacterium- mediated transformation of yeast and fungi. Agrobacterium Biology: From Basic Science to Biotechnol- ogy. 2018:349–74. https://doi.org/10.1007/82_2018_90 PMID: 29770864 PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 25 / 27 PLOS PATHOGENSModification of AMT for two oomycete species 55. Gelvin SB. Agrobacterium-mediated plant transformation: the biology behind the “gene-jockeying” tool. Microbiology and molecular biology reviews. 2003; 67(1):16–37. 56. Kunik T, Tzfira T, Kapulnik Y, Gafni Y, Dingwall C, Citovsky V. Genetic transformation of HeLa cells by Agrobacterium. Proceedings of the National Academy of Sciences. 2001; 98(4):1871–6. https://doi.org/ 10.1073/pnas.98.4.1871 PMID: 11172043 57. Judelson HS, Ah-Fong AM. Progress and challenges in oomycete transformation. Oomycete Genetics and Genomics: Diversity, Interactions, and Research Tools. 2009:435–53. 58. Citovsky V, Kapelnikov A, Oliel S, Zakai N, Rojas MR, Gilbertson RL, et al. Protein interactions involved in nuclear import of the Agrobacterium VirE2 protein in vivo and in vitro. Journal of Biological Chemistry. 2004; 279(28):29528–33. https://doi.org/10.1074/jbc.M403159200 PMID: 15123622 59. Wang K, Shi L, Liang X, Zhao P, Wang W, Liu J, et al. The gene TaWOX5 overcomes genotype depen- dency in wheat genetic transformation. Nature Plants. 2022:1–8. 60. Krishnamurthy K, Suhasini K, Sagare A, Meixner M, De Kathen A, Pickardt T, et al. Agrobacterium mediated transformation of chickpea (Cicer arietinum L.) embryo axes. Plant Cell Reports. 2000; 19 (3):235–40. https://doi.org/10.1007/s002990050005 PMID: 30754901 61. Choi PS, Soh WY, Kim YS, Yoo OJ, Liu JR. Genetic transformation and plant regeneration of water- melon using Agrobacterium tumefaciens. Plant cell reports. 1994; 13(6):344–8. https://doi.org/10.1007/ BF00232634 PMID: 24193834 62. 63. Zhang M, Chen S, Yu X, Vikesland P, Pruden A. Degradation of extracellular genomic, plasmid DNA and specific antibiotic resistance genes by chlorination. Frontiers of Environmental Science & Engineer- ing. 2019; 13(3):1–12. Judelson HS, Tyler BM, Michelmore RW. Transformation of the oomycete pathogen, Phytophthora infestans. Mol Plant-Microbe Interact. 1991; 4:602–7. https://doi.org/10.1094/mpmi-4-602 PMID: 1804404 64. Kohli A, Miro B, Twyman RM. Transgene integration, expression and stability in plants: strategies for improvements. Transgenic crop plants: Springer; 2010. p. 201–37. 65. Fitzgerald AM, Mudge AM, Gleave AP, Plummer KM. Agrobacterium and PEG-mediated transforma- tion of the phytopathogen Venturia inaequalis. Mycological research. 2003; 107(7):803–10. https://doi. org/10.1017/s0953756203008086 PMID: 12967207 66. Schouten HJ, vande Geest H, Papadimitriou S, Bemer M, Schaart JG, Smulders MJ, et al. Re-sequenc- ing transgenic plants revealed rearrangements at T-DNA inserts, and integration of a short T-DNA frag- ment, but no increase of small mutations elsewhere. Plant Cell Reports. 2017; 36:493–504. https://doi. org/10.1007/s00299-017-2098-z PMID: 28155116 67. Savory AI, Grenville-Briggs LJ, Wawra S, van West P, Davidson FA. Auto-aggregation in zoospores of Phytophthora infestans: the cooperative roles of bioconvection and chemotaxis. Journal of the Royal Society Interface. 2014; 11(94):20140017. https://doi.org/10.1098/rsif.2014.0017 PMID: 24598206 68. Oertel B, Jelke E. Formation of multinucleate zoospores inPhytophthora infestans (Mont.) de Bary (Oomycetes, Pythiaceae). Protoplasma. 1986; 135(2):173–9. 69. Shi T-Q, Liu G-N, Ji R-Y, Shi K, Song P, Ren L-J, et al. CRISPR/Cas9-based genome editing of the fila- mentous fungi: the state of the art. Applied microbiology and biotechnology. 2017; 101(20):7435–43. https://doi.org/10.1007/s00253-017-8497-9 PMID: 28887634 70. Skelsey P, Rossing W, Kessel G, Van der Werf W. Scenario approach for assessing the utility of dis- persal information in decision support for aerially spread plant pathogens, applied to Phytophthora infestans. Phytopathology. 2009; 99(7):887–95. https://doi.org/10.1094/PHYTO-99-7-0887 PMID: 19522587 71. Schaeffer SM, Nakata PA. CRISPR/Cas9-mediated genome editing and gene replacement in plants: transitioning from lab to field. Plant Science. 2015; 240:130–42. https://doi.org/10.1016/j.plantsci.2015. 09.011 PMID: 26475194 72. Rietman H, Bijsterbosch G, Cano LM, Lee H-R, Vossen JH, Jacobsen E, et al. Qualitative and quantita- tive late blight resistance in the potato cultivar Sarpo Mira is determined by the perception of five distinct RXLR effectors. Molecular Plant-Microbe Interactions. 2012; 25(7):910–9. https://doi.org/10.1094/ MPMI-01-12-0010-R PMID: 22414442 73. Vossen JH, van Arkel G, Bergervoet M, Jo K-R, Jacobsen E, Visser RG. The Solanum demissum R8 late blight resistance gene is an Sw-5 homologue that has been deployed worldwide in late blight resis- tant varieties. Theoretical and Applied Genetics. 2016; 129(9):1785–96. https://doi.org/10.1007/ s00122-016-2740-0 PMID: 27314264 74. Zhang F, Chen H, Zhang X, Gao C, Huang J, Lu¨ L, et al. Genome analysis of two newly emerged potato late blight isolates sheds light on pathogen adaptation and provides tools for disease management. PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 26 / 27 PLOS PATHOGENSModification of AMT for two oomycete species Phytopathology. 2021; 111(1):96–107. https://doi.org/10.1094/PHYTO-05-20-0208-FI PMID: 33026300 75. Cao J, Qiu M, Ye W, Wang Y. Phytophthora sojae Transformation Based on the CRISPR/Cas9 System. Bio-protocol. 2022; 12(6). https://doi.org/10.21769/BioProtoc.4352 PMID: 35434189 76. Ding D, Chen K, Chen Y, Li H, Xie K. Engineering introns to express RNA guides for Cas9-and Cpf1- mediated multiplex genome editing. Molecular plant. 2018; 11(4):542–52. https://doi.org/10.1016/j. molp.2018.02.005 PMID: 29462720 77. Lacroix B, Citovsky V. Extracellular VirB5 enhances T-DNA transfer from Agrobacterium to the host plant. PLoS One. 2011; 6(10):e25578. https://doi.org/10.1371/journal.pone.0025578 PMID: 22028781 78. Wang L, Lacroix B, Guo J, Citovsky V. The Agrobacterium VirE2 effector interacts with multiple mem- bers of the Arabidopsis VIP1 protein family. Molecular plant pathology. 2018; 19(5):1172–83. https:// doi.org/10.1111/mpp.12595 PMID: 28802023 79. Concordet J-P, Haeussler M. CRISPOR: intuitive guide selection for CRISPR/Cas9 genome editing experiments and screens. Nucleic acids research. 2018; 46(W1):W242–W5. https://doi.org/10.1093/ nar/gky354 PMID: 29762716 80. Kim HK, Min S, Song M, Jung S, Choi JW, Kim Y, et al. Deep learning improves prediction of CRISPR– Cpf1 guide RNA activity. Nature biotechnology. 2018; 36(3):239–41. https://doi.org/10.1038/nbt.4061 PMID: 29431740 81. Peng D, Tarleton R. EuPaGDT: a web tool tailored to design CRISPR guide RNAs for eukaryotic patho- gens. Microbial genomics. 2015; 1(4). https://doi.org/10.1099/mgen.0.000033 PMID: 28348817 PLOS Pathogens \| https://doi.org/10.1371/journal.ppat.1011346 April 21, 2023 27 / 27 PLOS PATHOGENS	null
10.1371_journal.pone.0269555.pdf	Data Availability Statement: All relevant data are within the paper and its Supporting Information files.	All relevant data are within the paper and its Supporting Information files.	RESEARCH ARTICLE Grapevine trunk diseases of cold-hardy varieties grown in Northern Midwest vineyards coincide with canker fungi and winter injury David H. DeKreyID 1 1, Annie E. Klodd2, Matthew D. Clark3, Robert A. BlanchetteID 1 Department of Plant Pathology, University of Minnesota, St. Paul, Minnesota, United States of America, 2 University of Minnesota Extension, Farmington, Minnesota, United States of America, 3 Department of Horticultural Science, University of Minnesota, St. Paul, Minnesota, United States of America [email protected] Abstract Grapevine trunk diseases make up a disease complex associated with several vascular fun- gal pathogenic species. Surveys to characterize the composition of grapevine trunk dis- eases have been conducted for most major grape growing regions of the world. This study presents a similar survey characterizing the fungi associated with grapevine trunk diseases of cold-hardy interspecific hybrid grape varieties grown nearly exclusively in the atypical harsh winter climate of Northern Midwestern United states vineyards. From the 172 samples collected in 2019, 640 isolates obtained by culturing were identified by ITS sequencing and represent 420 sample-unique taxa. From the 420 representative taxa, opportunistic fungi of the order Diaporthales including species of Cytospora and Diaporthe were most frequently identified. Species of Phaeoacremonium, Paraconiothyrium, and Cadophora were also prevalent. In other milder Mediterranean growing climates, species of Xylariales and Botryo- sphaeriales are often frequently isolated but in this study they were isolated in small num- bers. No Phaeomoniellales taxa were isolated. We discuss the possible compounding effects of winter injury, the pathogens isolated, and management strategies. Additionally, difficulties in researching and understanding the grapevine trunk disease complex are discussed. Introduction Grapevine trunk diseases (GTDs) make up a disease complex most often associated with sev- eral wood-inhabiting fungal species [1] and more recently possibly some bacterial species [2]. Sub-groups of these diseases are frequently categorized by symptomology and or taxonomic designation of causal fungal agents. Common names given to GTDs include Esca [3], folletage or berry shrivel [4], Petri disease, young esca, young vine decline [5], hoja de malvo´n [6], Botryosphaeria dieback, bot canker, black goo [7], slow stroke [8], eutypiosis, Eutypa dieback [9], black dead arm, dying arm, dead arm [10], swelling arm [11], grapevine leaf stripe disease a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: DeKrey DH, Klodd AE, Clark MD, Blanchette RA (2022) Grapevine trunk diseases of cold-hardy varieties grown in Northern Midwest vineyards coincide with canker fungi and winter injury. PLoS ONE 17(6): e0269555. https://doi.org/ 10.1371/journal.pone.0269555 Editor: Hernaˆni Gero´s, Universidade do Minho, PORTUGAL Received: January 24, 2022 Accepted: May 23, 2022 Published: June 3, 2022 Copyright: © 2022 DeKrey et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the paper and its Supporting Information files. Funding: MDC, RAB, AEK, DHD. This research was funded by the Minnesota Agricultural Rapid Response Fund and USDA Hatch Project MIN-22- 081 and MIN-22-089. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 1 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest [12], Phomopsis dieback, black spot [13], black measles [14], and black foot disease [15]. These diseases can be difficult to diagnose due to their sporadic symptom display and similarity of external and internal symptoms. Such symptoms may include interveinal foliar chlorosis and necrosis or tiger striping, generalized dieback, apoplexy or sudden death, gummosis, vascular streaking, wedge- or V-shaped vascular discoloration, cankers, and wood decay (Fig 1) [16]. GTDs cause serious grapevine health and economic problems and can be found in all grape growing regions of the world [17–19]. In the Northern Midwest United States (NMW), grow- ers often struggle with unproductive cordon sections commonly referred to as “skips in the cordon” or “blind wood” (Fig 1A and 1B). Cordon skips come at the cost of vineyard managers with lower yields and require retraining new cordons. In the past, chemicals such as sodium arsenate was used to control GTDs but health [20] and environmental concerns [21] have eliminated its widespread use. Often the dramatic increase in the incidence of GTDs in the last two decades is associated with the 2003 ban on sodium arsenate [22]. However, the increasing incidence of GTDs in countries which have never used sodium arsenate points to other factors being involved [22]. To date, very few chemical options are available to growers but recent research in the use of biological control agents has shown some promise for controlling spe- cific fungal GTD pathogens [23]. In most situations, development of best practices for GTD management remains the best option for growers. Management strategies can include prac- tices such as variety selection, rootstock selection, training system, pruning timing, double- pruning, wound-protection, multi-trunking, trunk renewal, trunk surgery, debris removal, tool sterilization, and other practices [24]. Management options of GTD pathogens tend to be region specific with considerations to climate, weather, cultural practices, and varieties grown. In the NMW, wine grape growing is a relatively new industry that is increasing at a consider- able pace. According to the 2016 University of Minnesota Extension vineyards and grapes sta- tus report, planted cold-hardy grapevine varieties increased from 5900 acres to 7580 acres from 2011 to 2015 [25]. However, Tuck et al. also reported an average decrease in yield of 3.5 to 3.2 tons per acre from 2011 to 2015 which indicates a need for better-informed, variety and region-specific GTD management practices. To accomplish this, it is important to identify the GTDs responsible for the problems. Traditional European Vitis vinifera cultivars are not often grown in the NMW due to diffi- culties brought on by harsh winters and a short growing season. Instead, own-rooted cold- hardy interspecific hybrid grape (CIHG) varieties are widely and often exclusively grown in the region. The genetic contribution of the native riverbank grape (V. riparia) provide CIHG varieties developed in Minnesota their cold-hardiness (rated down to -30˚C) [26, 27] and some resistance to endemic diseases and insect pests like phylloxera [28, 29]. Over the past four decades, the University of Minnesota has become a leader in the development of several CIHG wine and table grape varieties. The varieties most produced in the region include Mar- quette, Frontenac blanc, Frontenac, La Crescent, Petite Pearl, Brianna, and Frontenac Gris [30]. As many NMW vineyards are now reaching a decade in age since their first vines were planted, the characteristic cordon skip (Fig 1A and 1B) and dieback (Fig 1D) symptoms of GTDs have begun to appear. In addition, the compounding effect that GTDs and winter injury have on vines is becoming a major concern (Fig 1L). In many other parts of the world where grapes are grown, surveys have been conducted to characterize the region-specific composi- tion of GTD pathogens. In Europe and nearby Mediterranean countries where GTDs were first reported, major causal agents include fungal species of the genera Eutypa, Diplodia, Botryosphaeria, and Phaeomoniella [19]. Similar fungal species have also been identified as major causal agents of GTDs in Australia, New Zealand [31], South Africa [32], China [33– 35], and Chile [36, 37] as well as southern US and west coast US [38, 39]. Species of PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 2 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest Fig 1. Symptoms of grapevine trunk diseases in Northern Midwest vineyards. Cankers were often associated with skips in the cordons but had rarely wedge-shaped discoloration (A). Cankers more often had irregular shaped xylem reactions (B). Pycnidia were sometimes observed fruiting from cankers (B) and bleached canes (C). Dieback symptoms are common and the result of successive skips starting from tips of cordons (D). Pruning wounds were associated with minor (E), moderate (F), and severe (G) vascular streaking symptoms. Near completely healthy vascular tissue observed in wild Vitis riparia vines (H). Infrequent shallow cracks (I), several shallow cracks (J), and deep cracks (K and L) were associated with minor to moderate (I), moderate to severe (J), and severe (K) vascular symptoms. Winter injury often results in deep cracks on the trunk (L). Observations included cankers (ck), skips in the cordons (sk), bleached canes (bc), dieback (db), pruning wounds (pw), shallow cracks (sc), deep cracks (dc), black spotting (bs), black lines (bl), brown-red wood streaking (br), brown to black necrotic streaking (bn), discolored xylem (dx), sometimes healthy tissue (h), and white rot (wr). Bars = 1 cm. https://doi.org/10.1371/journal.pone.0269555.g001 Fomitiporia are often the main white-rot pathogen found in older vines in most of these regions as well [19]. Species of Phaeoacremonium have been identified in grapevines and other woody hosts in several countries around the world [40]. Species of Cadophora are on occasion identified as well notably found in Canada [41]. Species of Diaporthe and Cytospora have also PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 3 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest been identified in most of these regions though often to a lesser extent and usually in more humid growing regions [32]. However, no surveys have been conducted in the NMW or exclusively on CIHG cultivars. The objective of this study was to identify the major GTD species throughout the grape-grow- ing regions of Minnesota and Wisconsin. Three hypotheses were explored in this study. First, NMW GTDs will have a regionally distinct composition compared to other grape-growing regions of the world given the harsh growing climate and CIHG varieties grown. Second, die- back symptoms and internal vascular streaking can be associated with pruning wounds and winter injury. Third, isolation frequency of fungal genera will significantly differ compared to sample variety, variety berry color, sample section type, and sample county origin. Methods Sample collection Our sample collection was targeted towards symptomatic grapevines showing skips in the cor- dons, generalized dieback, reduced productivity, vascular discoloration, vascular decay, or apoplexy (Fig 1). A few externally asymptomatic vines were also collected for comparison. In 2019, a total of 172 samples were collected and brought to the laboratory. Samples were col- lected throughout both the dormant and growing season of 2019. Some samples were shipped by priority mail. Most samples collected were woody sections of grapevines, especially of cor- dons and trunks. It is important to note that regular re-trunking is frequently practiced in NMW vineyards and therefore main woody trunks of vines rarely, if ever, exceed ten years in age. Samples were stored at -20˚C until processed. Samples were acquired from 34 vineyards in Minnesota and Wisconsin from a total of 21 counties (Fig 2). However, data reported in this study is down to the county level to conserve anonymity of contributing vineyards. Primarily named CIHG varieties were collected as well as a few wild vines and genetically unique breed- ing lines. Sample processing Large diameter vine samples were cut and 3–5 mm3 chips excised from the margins of discol- ored or decayed internal vascular wood tissue or from the edge of cankers. For smaller diame- ter vine samples, the bark was peeled off and 3–5 mm thick discs were cut. Some disks were kept whole while others cut in half or in fourths depending on diameter of the sample. Excised chips where surface sterilized for 30 sec in an aqueous 10% sodium hypochlorite solution, fol- lowed by two washes in sterile distilled H2O, and one wash in 70% EtOH then left to dry in a clean air cabinet prior to plating. Between 3–5 chips per plate were then semi-embedded into 3 different culturing medias including malt extract agar (MEA; 15 g of Difco Bacto-agar, 15 g of Difco Bacto malt extract, and 1 L of deionized water with 0.1 g streptomycin sulphate dissolved in a small amount of 95% EtOH added post-autoclaving once cooled to 50˚C), basidiomycete semi-selective agar (BSA; same as MEA recipe plus 2 g of Difco yeast extract and 0.06 g Aldrich benomyl dissolved in a small amount of 95% EtOH added pre-autoclaving with 2 mL 85% lac- tic acid added post-autoclaving once cooled to 50˚C, adapted from Worrall, 1991) [42], and sabouraud dextrose agar (SDA; same as MEA recipe with 0.1 g Aldrich cycloheximide dis- solved in a small amount of deionized water added post-autoclaving once cooled to 50˚C, adapted from Harrington, 1981) [43]. Plates were left to incubate at 20–23˚C in darkness and checked daily. Emerging fungi were transferred onto fresh MEA. All cultures were maintained and stored in plastic bins at 20–23˚C. PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 4 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest Fig 2. Map of counties sampled. Grapevine wood samples collected from Minnesota and Wisconsin counties traced in yellow included Blue Earth (BE), Carver (C), Crow Wing (CW), Douglas (D), Fillmore (F), Goodhue (G), Jackson (J), Kanabec (K), Lac Qui Parle (LQP), Le Sueur (LS), Meeker (Mk), Mower (Mw), Murray (Mr), Pine (Pn), Polk (Pl), Trempealeau (T), Vernon (V), Wabasha (Wb), Walworth (Wl), Washington (W), and Winona (Wn). Color scale indicates elevation in meters. Pink points denote locations of the University of Minnesota (UMN) St. Paul campus and the UMN Horticultural Research Center (HRC) where the grape breeding program is located and where several samples were collected. Map constructed in R with the public domain map collections Natural Earth (https://www.naturalearthdata.com/) and Terrain Tiles (https://registry.opendata.aws/terrain-tiles/). https://doi.org/10.1371/journal.pone.0269555.g002 Isolate selection Fungal isolates for each sample were selected by culture macro-morphology on MEA and genetic identification by sequencing the internal transcribed spacer (ITS) genomic region. The primary macro-morphological characteristics considered included isolate color, growth rate, hyphal branching, hyphal depth, hyphal extension, hyphal margin, fruiting, sporulation, and metabolite staining of media. At the start of this study all unidentified cultures with unique morphologies isolated from a single sample were selected for sequencing. Isolates were later selected by macro-morphology in a more targeted manner as the study progressed by choosing unique cultures or cultures similar to known pathogens we previously identified in the isolate collection. Any isolates with questionable, non-descript, or similar culture macro-morphology were sequenced to be sure of their identity. DNA extraction and amplification The DNA of select isolates was extracted using the NaOH protocol according to Osmundson et al. (2013) [44]. Hyphae were scraped using a sterile scalpel from cultures of select isolates on MEA that had grown out larger than 2.5 cm in diameter. Hyphal tissue was transferred to a 1.7 mL microcentrifuge tube with 300 μL of 5 mM NaOH and 3 to 5 3.5 mm glass beads. The sam- ples where then vortexed for 1 to 5 min and centrifuged for 30 sec at 10,000 rpm. Then 5 μL of supernatant was transferred to new tubes containing 495 μL Tris-HCL 5 mM, pH 8.0. PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 5 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest The ITS region of the isolated DNA was targeted for PCR amplification using the ITS1F/4 primer pair [45] according to Blanchette et al. (2016) [46]. Each PCR had a final volume of 25.5 μL consisting of 12.5 μL GoTaq1 Green Master Mix, 9.5 μL molecular grade water, 1 μL of each primer at 10 μM, and 0.5 μL bovine serum albumin. The ITS locus was amplified using a Bio-Rad T100™ Thermal Cycler following a program of 94˚C for 5 min, 35 cycles of 94˚C for 1 min, 50˚C for 1 min, and 72˚C for 1 min, followed by a final extension step of 72˚C for 5 min. Locus amplification was confirmed by gel electrophoresis of SYBR stained PCR products prior to sequencing. Crude PCR products were Sanger sequenced by ABI 3730xl DNA sequences, Applied Biosystems, Foster City, CA. Molecular identification Sequences were processed using Geneious v9.0. The processed sequences where then identified with the basic local alignment search tool algorithm program for nucleotide sequences (BLASTn) initially against the TrunkDiseaseID.org [47] database and also against the standard complete NCBI GenBank. Best sequence identity match was selected for by consideration of highest score for published data as denoted in GenBank at the time of BLASTn searches. Iden- tity of isolates were matched to published sequences from taxonomic studies and identified to the species level whenever possible. Isolates with greater than 97% sequences identity match were considered homologous. Pathogenicity of identified fungal species on grapevines were denoted according initially to TrunkDiseaseID.org [47] and then confirmed and expanded by an assortment of grapevine pathogenicity trials found in published literature. However, most pathogenicity trials for these fungi were conducted on traditional V. vinifera grapevine culti- vars. Samples were scored as GTD+ upon sequence confirmation of at least one known patho- genic species. Additional isolation and sequencing was discontinued once a sample was designated GTD+. Data analysis Data were analyzed using the R statistical programming language in the RStudio integrated development environment using an assortment of packages but most notably the collection of Tidyverse Packages (v1.3.0) [48], the iNEXT package (v2.0.20) [49] to analyze sample coverage, and the vcd package (v1.4–9) for multivariant analysis. The vcd package was used to explore potential differences in isolation frequencies of genera-level taxa compared to a few variables of interest that are descriptive of the 168 collected woody samples. In brief, Hill numbers are used in the iNEXT package to estimate and then visualized sample completeness [50]. Addi- tionally, diversity Pearson residuals statistics were used to analyze the measure of discrepancy between observed and expected values within the vcd package. For each statistical comparison, a p-value is returned from a corresponding Chi-square test and a residual shaded mosaic plot was produced. Mosaic plots are graphs used for visualizing the comparison of multi-categori- cal data where both the x- and y-axis are sized proportionally to the input data, i.e. the sum area of the blocks represent 100% of the data and individual blocks are size proportionally to the frequency with which the categories are observed. Results Internal symptoms of GTDs following the terminology of Mugnai et al. 1999 [16] included brown-red wood streaking in a clearly defined wedge-shape from the cambium to pith which is indicative of canker fungi were observed in few of our samples. Cankers more often occurred in irregular forms and were associated with skips in the cordons (Fig 1A and 1B). Centrally diffuse brown-red wood as well as brown to black necrotic streaking originating PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 6 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest from the pith was often associated with pruning wounds (Fig 1E–1G) and cracks (Fig 1I–1L). All samples collected had discolored xylem to some extent and nearly completely healthy vas- cular tissues were only observed in some cross-sections of wild riverbank grapevines in forest and urban environments not included in this survey (Fig 1H). Cross-sections near pruning spurs often showed discolored wood symptoms without being preceded by diffuse brown-red wood or brown to black streaking (Fig 1E). Concentric black spotting, the result of longitudi- nal streaking, was also frequently observed, and in some cases, black spotting would begin to coalesce into shorter black lines (Fig 1F). Most samples had severe mottled expression of vas- cular symptoms especially for brown-red wood streaking, brown to black necrotic streaking, and discolored xylem (Fig 1G). Severe symptoms were sometimes associated with several points of origin from shallow cracking from winter injury or hail damage (Fig 1J). Rarely, if ever, have GTD foliar symptoms been observed in the NMW which possibly may be the result of our overall young vineyards or different climate. Foliar symptoms are more often observed in older vines under particular seasonal conditions [51, 52]. Moreover, lack of foliar symptoms may also be the result of regular re-trunking, a common cultural practice in the NMW. It is uncommon in the NMW for grapevine trunk wood to exceed ten years in age. All wood samples collected had some degree of internal vascular symptoms including exter- nally healthy samples (Fig 1). From 172 samples with various symptoms that included cankers and vascular discoloration, dieback as well as pruning wounds and cracks from cold injury or other environmental stresses yielded 640 isolates. These isolates represented 420 species-level taxa unique to individual samples. Rarefication using the 420 representative taxa estimate a sample coverage of 83% that reached to 90% by doubling the number of representative taxa (Fig 3). We found 32 of the 34 sampled vineyard locations, 20 of 21 counties, in this survey to have at least one GTD+ sample. Of the 172 samples we collected, 142 (83%) had taxa reported as pathogens associated with GTDs. Most samples were cordon sections and of the Marquette variety (Tables 1 and 2). Most taxa are of the phylum Ascomycota (398 isolates, 94.76%) which encompass 19 different orders, 38 genera, and potentially 89 species (Table 3). The most frequently isolated genera obtained in this study that were known to be associated with GTDs from previous reports included Cytospora, Phaeoacremonium, Diaporthe, Fig 3. Rarefaction sample coverage curve. Observed sample coverage reaches 83% for the 420 sample representative taxa. Extrapolated sample coverage reaches 90% by doubling the number of sample representative taxa. https://doi.org/10.1371/journal.pone.0269555.g003 PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 7 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest Table 1. Type of samples collected from the Northern Midwest in 2019. The majority of the 172 samples included woody sections of cordon, trunk, root, sucker, shoots, and unknown. Type cordon trunk root sucker shoot unknown slime flux bark basidiocarp Sample no. 113 32 11 5 4 3 2 1 1 Percent 66% 19% 6% 3% 2% 2% 1% 1% 1% https://doi.org/10.1371/journal.pone.0269555.t001 Cadophora, Pestalotiopsis, Diatrypella, Diplodia, and Botryosphaeria, respectively (Fig 4). Of these genera the most frequent species level sequence matches associated with GTDs included Cy. viticola, Ph. fraxinopennsylvanicum, Ph. minimum, Dpr. ampelina, Cd. luteo-olivacea, Ps. neglecta, Dt. verruciformis, Dpl. seriata, and Bt. dothidea (Table 3). Table 2. Varieties of the 172 samples collected in the Northern Midwest in 2019. Variety Marquette La Crescent Frontenac St. Pepin Frontenac Blanc Brianna Frontenac Gris Edelweiss Itasca Marechal Foch unknown Petite Pearl Prairie Star Valiant MN1069 MN1016 Sabrevois slime flux St. Croix MN43765 basidocarp Millot MN1005 Osceola Muscat Riverbank Grape Sauvignon Virginia Creeper Sample no. 42 28 20 Percent 24% 16% 12% 9 8 7 7 6 6 6 6 3 3 3 2 2 2 2 2 1 1 1 1 1 1 1 1 5% 5% 4% 4% 3% 3% 3% 3% 2% 2% 2% 1% 1% 1% 1% 1% 1% 1% 1% 1% 1% 1% 1% 1% https://doi.org/10.1371/journal.pone.0269555.t002 PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 8 / 25 PLOS ONETable 3. Taxonomy, isolation frequency, and pathogenicity of fungi identified by ITS with greater than 97% homology match. Grapevine trunk diseases of the Northern Midwest Family (n) Genus (n) Isolate Host variety Origin GenBank Phylum (n) Order (n) Ascomycota (398) Diaporthales (98) Valsaceae (61) Diaporthaceae (37) Togniniales (40) Togniniaceae (40) Pleosporales (99) Didymosphaeriaceae (31) Didymellaceae (28) Pleosporaceae (26) Astrosphaeriellaceae (3) Coniothyriaceae (3) Phaeosphaeriaceae (3) undefined family (3) Cucurbitariaceae (1) Species (n) [pathogenicity studies] Cytospora (60) Cy. viticola (57) [53] Cy. piceae (2) Cy. mali (1) GI-174 Frontenac Gris Blue Earth, MN OM307727 GI-847 Virginia Creeper Carver, MN GI-89 Marquette Meeker, MN OM307728 OM307729 Frontenac Crow Wing, MN OM307730 Frontenac Gris Douglas, MN OM307731 GI-31 GI-75 GI-384 GI-856 GI-413 GI-347 GI-741 GI-212 Valsa (3) Vl. sordida (2) Vl. salicina (1) Diaporthe (38) Dpr. ampelina (31) [54–56] Dpr. eres (7) [54] Phaeoacremonium (41) Ph. fraxinopennsylvanicum (22) [41] Ph. minimum (15) [3, 38] Ph. amstelodamense (1) Ph. angustius (1) [57] Ph. canadense (1) [41] Ph. hungaricum (1) Paraconiothyrium brasiliense (30) [58, 59] GI-516 Paraphaeosphaeria sporulosa (1) GI-157 Didymella (16) Dd. pinodella (11) Dd. glomerata (2) Dd. pomorum (2) Dd. bellidis (1) Epicoccum (10) Ep. nigrum (9) Ep. sorghinum (1) Nothophoma spiraeae (2) Alternaria (26) Al. alternata (19) Al. tenuissima (5) Al. arborescens (1) Al. infectoria (1) Pithomyces chartarum (3) Coniothyrium palmicola (3) GI-162 GI-389 GI-885 GI-879 GI-190 GI-837 GI-882 GI-878 Frontenac Carver, MN GI-236 Marquette GI-422 Marquette St. Pepin Edelweiss Goodhue, MN Goodhue, MN Douglas, MN Goodhue, MN La Crescent Carver, MN St. Pepin Pine, MN Osceola Muscat Wabasha, MN Frontenac Edelweiss Douglas, MN Wabasha, MN GI-257 Marquette GI-116 MN43765 GI-787 MN1016 Wabasha, MN Carver, MN Carver, MN La Crescent Walworth, WI La Crescent Carver, MN GI-223 Marquette Blue Earth, MN Riverbank Grape Carver, MN La Crescent La Crescent Frontenac GI-829 Marquette Carver, MN Blue Earth, MN Wabasha, MN Blue Earth, MN Frontenac Gris Blue Earth, MN La Crescent Blue Earth, MN OM307732 OM307733 OM307734 OM307735 OM307736 OM307737 OM307738 OM307739 OM307740 OM307741 OM307742 OM307743 OM307744 OM307745 OM307746 OM307747 OM307748 OM307749 OM307750 OM307751 OM307752 OM307753 OM307754 OM307755 OM307756 Sclerostagonospora (2) Sc. cycadis (1) Sc. lathyri (1) Neosetophoma cerealis (1) Microsphaeropsis olivacea (3) [60] Neocucurbitaria quercina (1) GI-141 Marquette Meeker, MN GI-140 Marquette Blue Earth, MN GI-480 Frontenac Blanc Crow Wing, MN OM307757 GI-505 Marquette Lac Qui Parle, MN OM307758 GI-37 Marquette Meeker, MN OM307759 (Continued ) PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 9 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest Family (n) Genus (n) Isolate Host variety Origin GenBank Table 3. (Continued) Phylum (n) Order (n) Hypocreales (49) Nectriaceae (29) Bionectriaceae (17) Hypocreaceae (2) Xylariales (25) Sporocadaceae (11) Diatrypaceae (9) Hypoxylaceae (2) Apiosporaceae (1) Xylariaceae (2) Helotiales (21) undefined family (18) Dermateaceae (1) Porodiplodiaceae (1) Sclerotiniaceae (1) Dothideales (14) Saccotheciaceae (14) Botryosphaeriales (10) Botryosphaeriaceae (10) Species (n) [pathogenicity studies] Fusarium (26) Fs. acuminatum (5) Fs. equiseti (3) Fs. solani (3) Fs. culmorum (1) Ilyonectria liriodendri (1) [61–63] Scolecofusarium ciliatum (1) Thyronectria austroamericana (1) Clonostachys (18) Cln. rosea (16) [60] Cln. byssicola (1) Trichoderma (2) Trc. atroviride (1) Trc. deliquescens (1) Pestalotiopsis (10) Ps. neglecta (6) Ps. uvicola (2) [38, 56] Ps. brassicae (1) Ps. chamaeropis (1) Neopestalotiopsis mesopotamica (1) Seimatosporium lichenicola (1) Seiridium rosarum (1) Diatrypella Dt. verruciformis (8) [64, 65] Dt. pulvinata (1) Diatrype stigma (1) [65] Hypomontagnella submonticulosa (1) Hypoxylon invadens (1) Arthrinium arundinis (1) Rosellinia corticium (1) Cadophora (16) Cd. luteo-olivacea (13) [41, 66–68] Cd. melinii (2) [68] Cd. ferruginea (1) Discohainesia oenotherae (1) Porodiplodia vitis (1) Botrytis cinerea (1) Aureobasidium pullulans (14) Diplodia (7) Dpl. seriata (6) [38, 56, 58, 64, 69] Dpl. corticola (1) [38, 56] Botryosphaeria dothidea (2) [38, 56, 70] Phaeobotryon negundinis (1) GI-820 Frontenac GI-95 MN1005 Fillmore, MN Carver, MN OM307760 OM307761 GI-376 Frontenac Crow Wing, MN OM307762 GI-151 La Crescent Carver, MN GI-322 Marechal Foch Goodhue, MN GI-149 La Crescent GI-796 slime flux Carver, MN Carver, MN OM307763 OM307764 OM307765 OM307766 GI-870 La Crescent Polk, MN OM307767 GI-874 Marquette Trempealeau, WI OM307768 GI-795 slime flux Carver, MN GI-351 Frontenac Blanc Fillmore, MN OM307769 OM307770 GI-491 Edelweiss Trempealeau, WI OM307771 GI-738 Edelweiss Trempealeau, WI OM307772 GI-403 Marquette Lac Qui Parle, MN OM307773 GI-231 St. Pepin Goodhue, MN GI-220 Frontenac Gris Blue Earth, MN GI-99 Marquette GI-352 La Crescent Carver, MN Murray, MN OM307774 OM307775 OM307776 OM307777 GI-464 Marquette GI-416 Valiant GI-895 Valiant Carver, MN Douglas, MN OM307778 OM307779 Crow Wing, MN OM307780 GI-817 Frontenac Blanc Blue Earth, MN GI-350 La Crescent Blue Earth, MN GI-70 GI-62 Frontenac St. Croix Carver, MN Wabasha, MN GI-370 La Crescent Mower, MN GI-316 Marechal Foch Goodhue, MN GI-328 La Crescent Goodhue, MN GI-442 Frontenac Gris Winona, MN GI-269 Frontenac Blue Earth, MN GI-386 Frontenac Walworth, WI GI-886 La Crescent Blue Earth, MN OM307781 OM307782 OM307783 OM307784 OM307785 OM307786 OM307787 OM307788 OM307789 OM307790 OM307791 GI-408 La Crescent Blue Earth, MN OM307792 GI-373 Petite Pearl Crow Wing, MN OM307793 GI-225 Marquette Blue Earth, MN GI-131 MN1005 Carver, MN OM307794 OM307795 (Continued ) PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 10 / 25 PLOS ONETable 3. (Continued) Phylum (n) Order (n) Family (n) Genus (n) Isolate Host variety Origin GenBank Species (n) [pathogenicity studies] Grapevine trunk diseases of the Northern Midwest Cladosporiales (11) Cladosporiaceae (11) Eurotiales (8) Aspergillaceae (4) Trichocomaceae (4) Sordariales (6) Chaetomiaceae (4) Sordariaceae (1) Trichosphaeriales (3) Trichosphaeriaceae (3) Chaetomellales (1) Chaetomellaceae (1) Chaetothyriales (1) Herpotrichiellaceae (1) Coniochaetales (1) Coniochaetaceae (1) Glomerellales (1) Glomerellaceae (1) Saccharomycetales (1) Dipodascaceae (1) Thelebolales (1) Thelebolaceae (1) Valsariales (1) Valsariaceae (1) Basidiomycota (15) Polyporales (9) Phanerochaetaceae (3) Irpicaceae (2) Polyporaceae (2) Cerrenaceae (1) Meruliaceae (1) Russulales (2) Peniophoraceae (1) Stereaceae (1) Agaricales (2) Physalacriaceae (1) Schizophyllaceae (1) Hymenochaetales (1) Hymenochaetaceae (1) Cystofilobasidales (1) Mrakiaceae (1) Mucoromycota (7) Mucorales (7) Mucoraceae (7) Cladosporium (11) Cld. cladosporioides (9) Cld. anthropophilum (1) Cld. westerdijkiae (1) Talaromyces amestolkiae (4) Penicillium (5) Pn. pulvillorum (2) Pn. raistrickii (1) Pn. simplicissimum (1) Pn. sumatraense (1) Ovatospora (3) Ov. brasiliensis (2) Ov. mollicella (1) Chaetomium concavisporum (1) Sordaria fimicola (1) Nigrospora oryzae (3) Chaetomella raphigera (1) Rhinocladiella quercus (1) Coniochaeta velutina (1) Colletotrichum acutatum (1) Geotrichum candidum (1) Thelebolus microsporus (1) Valsaria spartii (1) Bjerkandera adusta (2) Hyphodermella rosae (1) Irpex lacteus (2) Trametes versicolor (2) Cerrena unicolor (1) Phlebia radiata (1) Peniophora cinerea (1) Stereum complicatum (1) Cylindrobasidium laeve (1) Chondrostereum purpureum (1) Phellinus conchatus (1) Tausonia pullulans (1) Mucor (7) Mucr. circinelloides (6) Mucr. moelleri (1) GI-866 Marquette GI-209 Marquette GI-194 Marquette GI-801 Vitis spp. Blue Earth, MN Blue Earth, MN Blue Earth, MN Carver, MN GI-337 Marquette Blue Earth, MN GI-309 St. Pepin Goodhue, MN GI-210 Marquette Blue Earth, MN GI-36 Marquette Meeker, MN GI-828 Marquette Blue Earth, MN GI-881 La Crescent Blue Earth, MN GI-865 Marquette Blue Earth, MN GI-848 Virginia Creeper Carver, MN GI-846 Edelweiss GI-217 Marquette Winona, MN Goodhue, MN GI-486 Marquette Blue Earth, MN GI-472 Brianna GI-287 La Crescent GI-164 Edelweiss Winona, MN Goodhue, MN Goodhue, MN GI-88 Frontenac Blanc Pine, MN GI-43 Marquette Meeker, MN OM307796 OM307797 OM307798 OM307799 OM307800 OM307801 OM307802 OM307803 OM307804 OM307805 OM307806 OM307807 OM307808 OM307809 OM307810 OM307811 OM307812 OM307813 OM307814 OM307815 GI-417 Marquette Trempealeau, WI OM307816 GI-823 Frontenac GI-806 Vitis spp. GI-68 La Crescent GI-198 Prairie Star GI-798 Vitis spp. GI-342 La Crescent GI-200 Marquette GI-263 Itasca GI-444 Itasca GI-805 Vitis spp. slime flux GI-61 Fillmore, MN Carver, MN Wabasha, MN Carver, MN Carver, MN Blue Earth, MN Goodhue, MN Wabasha, MN Le Sueur, MN Carver, MN Carver, MN OM307817 OM307818 OM307819 OM307820 OM307821 OM307822 OM307823 OM307824 OM307825 OM307826 OM307827 GI-71 La Crescent GI-365 Marquette Vernon, WI Mower, MN OM307828 OM307829 Taxonomic rankings from order to species are denoted followed by isolation frequency in parenthesis. The isolation frequency is the count of samples each taxa was isolated from a possible 172 samples. Representative isolates deposited to GenBank are listed for each species along with the sample variety and county origin of that isolate. Pathogenicity studies conducted for each species are listed in brackets following species. See references for complete citations. Highlighted species have associated pathogenicity studies. Highlighted isolates pictured in Fig 4. https://doi.org/10.1371/journal.pone.0269555.t003 PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 11 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest Fig 4. Fungal genera isolated from more than one grapevine sample. Individual culture image areas are relatively proportional to the number of samples each fungal genus was isolated from. All pictured fungal isolates grown on MEA. Cytospora (Cy = 60, GI-174, 80 dpi); Phaeoacremonium (Ph = 41, GI-422, 94 dpi); Diaporthe (Dpr = 38, GI-449, 47 dpi); Paraconiothyrium (Pr = 30, GI-516, 47 dpi); Alternaria (Al = 27, GI-879, 36 dpi); Fusarium (Fs = 27, GI-151, 17 dpi); Clonostachys (Cln = 18, GI-870, 27 dpi); Cadophora (Cd = 16, GI-370, 94 dpi); Didymella (Dd = 16, GI-257, 21 dpi); Aureobasidium (Ar = 14, GI-886, 21 dpi); Cladosporium (Cld = 11, GI-866, 21 dpi); Epicoccum (Ep = 10, GI-837, 19 dpi); Pestalotiopsis (Ps = 10, GI-491, 40 dpi); Diatrypella (Dt = 9, GI-464, 15 dpi); Diplodia (Dpl = 7, GI-408, 20 dpi); Mucor (Mucr = 7, GI-71, 20 dpi); Talaromyces (Tl = 6, GI-801, 55 dpi); Penicillium (Pn = 5, GI-337, 14 dpi); Coniothyrium (Cn = 3, GI-882, 21 dpi); Microsphaeropsis (Mcrs = 3, GI-505, 20 dpi); Nigrospora (Ng = 3, GI-846, 14 dpi); Ovatospora (Ov = 3, GI-828, 19 dpi); Pithomyces (Pt = 3, GI-162, 28 dpi); Valsa (Vl = 3, GI-384, 35 dpi); Bjerkandera (Bj = 2, GI-417, 28 dpi); Botryosphaeria (Bt = 2, GI-225, 35 dpi); Irpex (Ir = 2, GI-806, 55 dpi); Nothophoma (Nt = 2, GI-878, 31 dpi); Sclerostagonospora (Sc = 2, GI-140, 17 dpi); Trametes (Trm = 2, GI-68, 9 dpi); and Trichoderma (Trc = 2, GI-795, 14 dpi). https://doi.org/10.1371/journal.pone.0269555.g004 There were 15 taxa of Basidiomycota (3.57%) which encompass 5 orders, 11 genera, and 12 species. Bjerkandera adusta, Irpex lacteus, and Trametes versicolor where the most frequently identified Basidiomycota isolated. These fungi were present in 5 samples but were found in PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 12 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest counties not adjacent to one another. Very few taxa of Mucoromycota (7 isolates, 1.68%) were identified. Mucor circinelloides, not considered associated with GTD, was isolated from 6 sam- ples which originated from different vineyards in nonadjacent counties. There were no significant differences in isolation frequencies of genera based on sample berry color (S1 Fig). There is some indication of significant isolation frequency differences of genera by sample variety (S2 Fig) or sample county origin (S3 Fig). For most genera not enough sample representative taxa were obtained to be sure these isolation frequency differ- ences between counties are truly significant. However, Cadophora obtained in this study nota- bly had significant differences in isolation frequency between sample section types with a p- value of 0.013. Cadophora spp. were less often isolated from cordons with a Pearson residual of -2.9 and more frequently isolated from root, trunk, and sucker sections with Pearson residuals between 2.2 to 3.9 (S4 Fig). Discussion The vast majority of the fungal taxa isolated in this study are of the phylum Ascomycota, sev- eral of which are considered pathogenic to grapevines (references cited in Table 3). Cytospora spp. and Diaporthe spp. of the order Diaporthales as well as Phaeoacremonium spp. make up the largest majority of isolates identified and are known to be pathogenic on grapevines. Fre- quent species identification of Phaeoacremonium in this study aligns similarly with most other GTD surveys. However, the other major results confirm our first hypothesis that the composi- tion of GTDs for the NMW is different in comparison to most other studied grape growing regions. These fungi in the Diaporthales typically are considered a minor or secondary group of causal agents in other regions were GTD surveys were completed [55, 71–74] but in the study presented here they are the most commonly found GTD fungi in the NMW. The GTD species genera of Diaporthe was previously named Phomopsis by some other investigators [55, 75]. Species of Cytospora have been reported previously as prevalent in cold climate regions [74] and areas of high humidity [32]. The etiology of these Diaporthales pathogens on grapevines has been previously well char- acterized [76]. In addition to symptoms of internal wood discoloration, pathogenic species of Cytospora and Diaporthe can induce symptoms of cane bleaching (Fig 1C). They also produce asexual fruiting bodies known as pycnidia which can be found on all affected tissues (Figs 1B, 1C and 4). On succulent green tissue, pycnidia may be surrounded by a halo or half halo of chlorotic tissues or may reside hidden just under the bark of infected vines. Conidio-spores that ooze out from pycnidia serve as a major source of inocula that can re-infect the same host or infect other nearby hosts. Conidia are most often disseminated by rain or irrigation splash but also by contaminated tools and more rarely by wind alone. These species overwinter in col- onized wood of canes, spurs, pruning debris, and dormant buds [77]. However, symptoms appear to differ from locations of sample collection and isolation particularly for Diaporthe spp. [11, 55, 78, 79]. Symptom differences may be explained by genetic differences of local fun- gal populations due to the result of horizontal gene transfer of transposable elements for the acquisition or loss of pathogenicity [80, 81]. However, horizontal gene transfer has never been studied in fungal GTD pathogens. Many of the Diaporthales are assumed opportunistic patho- gens, causing disease only in stressed or weakened hosts or may live endophytically without causing disease [82]. Since these fungi can colonize wounds, the prevalence of these fungi found may be a result of wounds caused by cold injury. In January of 2019, an atypical polar vortex occurred in the NMW. In Minnesota on Janu- ary 30, 2019, the temperature dropped to -33˚C (-48˚C with wind chill) for the Minneapolis- St. Paul area while the lowest temperature recorded in the state was -39˚C (-53˚C with wind PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 13 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest chill) [83]. The polar vortex temperatures were well below the lowest ratings for most of the CHIG varieties in many counties. Winter injury was most apparent on Marquette variety grapevines and in vineyards with little wind protection. Exposure of grapevines to these extreme weather conditions resulted in frost cracks of woody tissues and damage to dormant buds. However, winter injury of grapevines more typically occurs by means of sun exposure. Injury occurs when both direct and snow reflected sunlight warms trunks to above freezing during the day followed by a rapid decrease in temperature to below freezing at night. The sud- den drop in temperature ruptures just the outer most layer of phloem cells for mild cases while more serious cases kill cambial cells and damage xylem tissues. When this occurs on trees in the NMW it is often referred to as sunscald. On grapevines this could be considered “winter sunscald”, not to be confused with sunscald of grape berries in the summer. Like extreme cold weather exposure, winter sunscald can also result in both shallow and deep frost cracks depending on severity (Fig 1I–1L). Additionally, winter sunscald of grapevines results in a blackened appearance of the bark on the south to southwest facing side of the vine (Fig 1L). In either case of winter injury often the roots and the lower trunk of vines are protected by the insulating snow covering. Thus, trunk replacement by sucker is a viable and common manage- ment strategy in the NMW [84]. Regardless, associated observations of winter injury, vascular discoloration, and identification of fungi suggests our second hypothesis is true and that the polar vortex likely predisposed grapevines to these opportunistic canker pathogens. Wounds, perhaps from winter injury or mechanical pruning, serve as portals for infection by GTD pathogens under conducive weather conditions such as cool spring or fall rains [32, 85–87]. Fungal spores then colonize and spread through the vascular tissue either by hyphal growth or additional sporulation. Many canker pathogens secrete cell wall degrading enzymes or other compounds to spread laterally through xylem tissues eventually circumnavigating and killing the entire cambium. However, the grapevine host produces tyloses, gels, phenolics, and suberin to compartmentalize the damaged tissues and invading microorganisms [88]. How- ever, restricted balanced production of these defensive structures and compounds is essential. Overproduction of occlusions in response to pathogenic infections can lead to extensive hydraulic failure resulting in external foliar symptoms and often vine death [89, 90]. In cross sections the defense response of the grapevine host is seen as a continuum of brown-red wood to brown-black necrotic tissue (Fig 1). Lighter vascular discoloration indicates more recent responding tissues and likely the front of pathogen spread. Darker vascular discoloration indi- cates long responding tissues and the probable point of pathogen entry [91]. Alternatively, pri- marily pectinolytic active pathogens degrade gels in xylem vessels and spread longitudinally by spores through the small spaces between tyloses partially occluding xylem conduits [92]. Lon- gitudinal spread of these pathogens is seen in cross section by the host defense response as black spotting and black lines (Fig 1F) [16]. Genomes of Cytospora spp. and Diaporthe spp. reveal these fungi employ an abundance of cell wall degrading enzymes [93, 94]. Xylem vessel anatomy likely influences host resistance, pathogen spread, and environmen- tal resilience. In the Dutch elm disease pathosystem, smaller diameter xylem vessels appeared to confer some level of resistance to the causal fungal agents Ophiostoma ulmi and O. novo- ulmi [95]. Reduced vessel diameter permits a more energetically conserved faster occlusion of tissues adjacent to damaged or infected xylem tissues. Pouzoulet et al. (2014; 2017; 2020) con- ducted histological and pathogenicity studies comparing a few V. vinifera grapevine cultivars that had varying susceptibility to GTDs and showed cultivars with smaller diameter xylem ves- sels may likely confer some resistance [88, 96, 97]. Unfortunately, histological pathogenicity studies of grapevines against vascular pathogens are few and completely lacking for hybrid varieties. However, hybrid varieties as well as traditional cultivars have been studied and show links between vessel anatomy and environmental resilience against freezing and drought PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 14 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest conditions, though more research is needed [98, 99]. Interestingly, developmental histological studies of grapevine xylem tissues have shown plasticity of vessel diameter even within a vari- ety or individual based on early season precipitation [100]. Therefore, abundant early irriga- tion influence vines to develop larger xylem vessels that can allow for vigorously growing higher yielding grapevines but may also render vines more susceptible to biotic vascular patho- gens and abiotic environmental stresses. Moreover, environmental stresses, including extreme weather events, will be more frequent due to climate change which may provide more oppor- tunities for some vascular pathogens [101]. However, the effects of climate change are likely dependent on the pathogens, cultivars, and environments in question. For example, intensive drought conditions have shown both positive [102] or negative [103] effects for grapevines suf- fering from GTDs. Species of Phaeoacremonium are found in many grape growing countries and often associ- ated with GTDs of young vines. Phaeoacremonium spp. are often found to spread through wounds, nursery propagation and grafting [104–112], see review by Gramaje et al. (2015) [40]. Ph. fraxinopennsylvanicum is widespread throughout the world on other hosts and has been found in many other investigations in the Midwestern United States [113] but Ph. minimum appears to be the most widespread Phaeoacremonium spp. throughout grape growing countries. Cd. luteo-olivacea has been isolated from many substrates including soils [114], decaying wood [115–117], and grapevines [68] as well as from various grafting tools [118] and pruning shears [119]. Cd. luteo-olivacea is often referred to as a weak pathogen and this fungus was not recognized as pathogenic on grapevine until extended grapevine inoculation studies were con- ducted (see Table 3). Interestingly, Cadophora was the only genus in this study that showed some differences in isolation frequency in comparison with four tested criteria of hypothesis three. No significant differences were observed for any fungal genera isolated compared to variety berry color (S1 Fig). Some researchers have indicated suspicions that red cultivars are more susceptible to GTD pathogens, although much more research is needed [120]. Some sig- nificant differences were observed for a few genera compared to sample variety (S2 Fig) or sample county origin (S3 Fig). However, the residuals were only slight for the variety and county origin comparisons and more research is needed to be sure of these correlations. Yet, Cadophora spp. were found to be significantly less isolated from cordon sample sections and significantly more isolated from woody sections of trunk, roots, and suckers (S4 Fig). Increased isolation of Cadophora spp. from the more central main trunk of the vine may be an indication that infection occurs from the soil or possibly the infection was acquired prior to planting. Additional research of vineyard soils and nursery stock materials would better eluci- date the origin of Cadophora spp. in NMW grapevines. Additionally, Ilyonectria liriodendri, isolated just once in this study, is another weak pathogen often associated with GTDs of roots and often found in nurseries [121, 122]. Fungi considered nonpathogenic to grapevines according to TrunkDiseaseID.org [47] included Penicillium, Alternaria, Didymella, Epicoccum, and Paraconiothyrium which were some of the more frequently isolated genera in this study. However, few GTD pathogenicity studies have tested Pleosporales fungi. Paraconiothyrium spp. have been demonstrated to be pathogenic on fruit trees and other woody species [105] and potentially pathogenic on grape- vines [59]. In the NMW, Paraconiothyrium spp. could be a potential pathogen. This fungus was recently found associated with the emerald ash borer and found to cause small cankers on healthy ash trees [113, 123]. Further investigation of the pathogenesis of Pleosporales may prove interesting considering the persistence in isolation and sequencing of these fungi. Basidiomycota have been also found to play a role in GTDs [124–126]. Often the presence of these wood decay fungi are found mainly in older vines following the colonization of faster PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 15 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest growing host detoxifying pioneering ascomycota [16]. In many parts of the grape growing regions of the world, the primary Basidiomycota associated with GTDs are Fomitporia medi- terriana [127] and Stereum hirsutum [128]. However, in our study no species of Fomitporia were isolated but a different species of Stereum was one of the more frequently isolated Basi- diomycota. Additionally, Trametes versicolor and Cerrena unicolor were also isolated. These fungi are commonly found on forest and shade trees locally (personal observations). Fruiting bodies of these fungi have been found on trunks of grapevines that had advanced stages of GTD symptoms. These Basidiomycota have not been tested for pathogenicity on grapevines but Trametes versicolor causes cankers and decay on fruit trees [129] and Cerrena unicolor is well characterized on hardwoods where it is an aggressive canker rot pathogen [130]. GTD fungi in the Xylariales, Botryosphaeriales, and Phaeomoniellales are of concern in many grape growing countries and this includes fungi in the families Diatrypaceae, Botryo- sphaeriaceae, and Phaeomoniellaceae [73]. In our study, only 10 Diatrypaceae, and 10 Botryo- sphaeriaceae isolates were identified. No Phaeomoniellaceae isolates were identified. No species of Eutypella or Eutypa (both Diatrypaceae), common in some other grape growing regions, were isolated. However, a few isolates not included in analysis closely matched with Eutypa but with less than 97% and may prove to be new species following additional detailed taxonomic studies (See S1 Table). Pestalotiopsis spp., Neopestalotiopsis spp., Diatrypella verru- ciformis, and Diatrype stigma of the Xylariales as well as Diplodia seriata, Botryosphaeria dothi- dea, Diplodia corticola, and Phaeobotryon negundinis of the Botryosphaeriales represented a minority of the GTD pathogens isolated in our study (Table 3). At the start of our research project, many grape growers expressed concerns about GTDs in their young vineyards. There was also considerable concern about Botryosphaeriaceae GTD often called “Bot-rot” or otherwise known as Bot canker. Grape growers in the NMW regularly associate any wedge-shaped discoloration of cross-sections of grapevines as Bot-rot. Based on this survey, Botryosphaeriaceae GTD is rare in the NMW. Confusion and concerns of growers, viticultural professionals, and even fellow researchers is understandable given the complexity of GTDs in addition to the many various names used in an attempt to sub-categorize GTDs. Many of these sub-categorized GTDs have been associated with irregular generalized symp- toms of grapevines influenced by the cultivar or variety, climate, and environmental condi- tions [32, 118, 131]. Moreover, many GTD designations rapidly become obsolete with each taxonomic recategorization of fungal species. Fungal taxonomy will likely continue to change as more genetic information is gathered into databases and mycologist strive to dissolve the two-name system for fungi [132]. The isolation frequency differences of these typically important GTD groups in other grape growing regions is especially notable. Such differences could possibly be correlated with the different climate of the NMW as compared to the many other grape growing regions which typically have more seasonally mild, often Mediterranean climates. Several spore trapping studies from various countries have attempted to characterize sporulation events of various GTD pathogens in correlation with varying weather measurements [87, 133–135]. Given the drastically different GTDs composition of the NMW, additional studies using spore trapping would prove insightful to obtain a better understanding of the GTD pathogens in the NMW. Notably, culturing methods may present bias as some faster growing fungal species such as those of Cytospora, Diaporthe, Diplodia, and Botryosphaeria may outgrow slower growing spe- cies such as those of Phaeoacremonium, Phaeomoniella, and many basidiomycota. This bias of culture-based studies would benefit being paired with modern metagenomic techniques to characterize all potential microorganismal species present in a substrate. However, many metagenomic techniques also present bias such as detection of non-viable organisms or unin- tended preferential over-identification. Thus, metagenomic techniques also benefit being PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 16 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest paired with classical culture-based techniques. Additionally, inclusion of culture-based meth- ods allows for the curation of live microbial collections for use in future pathogenicity or char- acterization studies. Therefore, future studies which use both classical culturing and metagenomic techniques would better elucidate the NMW GTD complex. This combination of classical and modern techniques has been demonstrated effective in recent local studies of Heterobasidion Root Rot [136]. Regardless, in the study we present here the sample coverage curve (Fig 3) was observed to have reached a plateau providing confidence all major fungal species were identified in this study. Moreover, the use of three types of media allowed for fre- quent isolation of slow growing fungal species such as those of Phaeoacremonium as well as infrequent isolation of fast growing fungal species such as those of the Botryosphaeriales. Therefore, confidence is assured that the isolation frequency of these fungi identified is repre- sentative for grapevines of the NMW. In this study we revealed a large diversity of fungal species associated with cold-hardy hybrid grapevines in the NMW. A handful of these isolates (those with less than 97% sequence match provided in S1 Table) could potentially be revealed as new species following additional detailed taxonomic studies. However, the majority of fungal species we identified show Dia- porthales predominate GTDs in the NMW. Diaporthales GTD species, Cytospora and Dia- porthe, are generally opportunistic fungi and largely spread to new hosts within short distances by asexual conidia via rain splash or contaminated tools. Basic understanding of these oppor- tunistic pathogens lifecycles emphasizes the benefit growers would gain from more intentional phytosanitary practices such as prompt removal and destruction of pruning debris as well as the regular sanitization of tools. Pruning debris and diseased canes left unpruned have recently been shown to be a major source of Diaporthe GTD Inoculum [135]. Our current recommen- dation for grape growers in the NMW is to prune in the dormant winter season during a period of cold and dry weather. Recommendations on pruning timing could be fine-tuned by epidemiological spore trapping studies in NMW vineyards and may possibly allow for some degree of GTDs forecasting. Vineyard spore trapping could also provide the opportunity for broader biosurveillance of invasive pathogenic microbial species of forest, shade, and orchard trees. Knowing the prevalence of GTDs in the NMW provides insight for the development of improved management practices. Similar studies of GTD pathogens spread from nurseries of cold-hardy grapevine hybrid varieties would also provide insight to improved propagation practices as well as yield less stressed, higher quality, and more vigorous growing nursery stock plants for growers. Assessment of these hybrid varieties against a panel of GTD pathogens may reveal novel evidence of resistance or susceptibility that would be useful for grape breeders. The development of cost-effective rapid molecular assays for the most prevalent GTDs in the NMW would be a useful tool to measure the effectiveness of practices or the variability of vari- ety susceptibility to GTDs. Supporting information S1 Table. Isolates with less than 97% homology match. Taxonomic rankings from order to species are denoted followed by isolation frequency in parenthesis. The isolation frequency is the count of samples each taxa was isolated from a possible 172 samples. Isolates deposited to GenBank are listed for each species along with the sample variety and county origin of that iso- late. Pathogenicity studies conducted for each species are listed in brackets following species. See references for complete citations. Highlighted species have associated pathogenicity stud- ies. (DOCX) PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 17 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest S1 Fig. Mosaic plot of genus level taxa isolated from sample varieties. (TIF) S2 Fig. Mosaic plot of genus level taxa isolated from sample variety berry color. (TIF) S3 Fig. Mosaic plot of genus level taxa isolated from sample section types. (TIF) S4 Fig. Mosaic plot of genus level taxa isolated from counties. (TIF) Acknowledgments We thank Owen Geier and Brian Prior for technical support and the staff and management of all the contributing Northern Midwestern vineyards that provided samples for this study. Author Contributions Conceptualization: David H. DeKrey. Data curation: David H. DeKrey. Formal analysis: David H. DeKrey. Funding acquisition: Matthew D. Clark. Investigation: David H. DeKrey. Methodology: David H. DeKrey. Project administration: David H. DeKrey. Resources: David H. DeKrey, Robert A. Blanchette. Software: David H. DeKrey. Supervision: Matthew D. Clark, Robert A. Blanchette. Validation: David H. DeKrey. Visualization: David H. DeKrey. Writing – original draft: David H. DeKrey. Writing – review & editing: David H. DeKrey, Annie E. Klodd, Matthew D. Clark, Robert A. Blanchette. References 1. Bertsch C, Ramı´rez-Suero M, Magnin-Robert M, Larignon P, Chong J, Abou-Mansour E, et al. Grape- vine trunk diseases: complex and still poorly understood. Plant Pathology. 2013; 62: 243–265. https:// doi.org/10.1111/j.1365-3059.2012.02674.x 2. Haidar R, Yacoub A, Vallance J, Compant S, Antonielli L, Saad A, et al. Bacteria associated with wood tissues of Esca-diseased grapevines: functional diversity and synergy with Fomitiporia mediterranea to degrade wood components. Environmental Microbiology. 2021; 23: 6104–6121. https://doi.org/10. 1111/1462-2920.15676 PMID: 34288352 3. Larignon P, Dubos B. Fungi associated with esca disease in grapevine. European Journal of Plant Pathology. 1997; 103: 147–157. https://doi.org/10.1023/A:1008638409410 4. Galloway BT, Fairchild DG. Diseases of the Grape in Western New York. The Journal of Mycology. 1891; 6: 95–99. https://doi.org/10.2307/3752563 PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 18 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest 5. Scheck H, Vasquez S, Fogle D, Gubler WD. Grape growers report losses to black-foot and grapevine decline. Cal Ag. 1998; 52: 19–23. https://doi.org/10.3733/ca.v052n04p19 6. The Gatica M. “Hoja de Malvo´ n” grape disease in Argentina. Phytopathologia Mediterranea. 2000; 1– 5. https://doi.org/10.1400/57809 7. Pascoe IG. Grapevine trunk diseases in Australia: diagnostics and taxonomy. Symptoms and occur- rence of grape declines. Fort Valley, VA: International Ampelography Society and International Coun- cil on Grapevine Trunk Diseases; 1998. pp. 56–77. 8. Fontaine F, Gramaje D, Armengol J, Smart R, Nagy ZA, Borgo M, et al. Grapevine trunk diseases. A review. OIV Publications; 2016. 9. Moller WJ, Kasimatis AN. Dieback of grapevines caused by Eutypa armeniacae. Plant Disease Reporter. 1978; 62: 254–258. 10. Lehoczky J. Black dead-arm disease of grapevine caused by Botryosphaeria stevensii infection. Acta Phytopathologica Academiae Scientiarum Hungaricae. 1974; 9: 319–327. 11. Kajitani Y, Kanematsu S. Diaporthe kyushuensis sp. nov., the teleomorph of the causal fungus of grapevine swelling arm in Japan, and its anamorph Phomopsis vitimegaspora. Mycoscience. 2000; 41: 111–114. https://doi.org/10.1007/BF02464318 12. Fontaine F, Pinto C, Vallet J, Cle´ ment C, Gomes AC, Spagnolo A. The effects of grapevine trunk dis- eases (GTDs) on vine physiology. Eur J Plant Pathol. 2016; 144: 707–721. https://doi.org/10.1007/ s10658-015-0770-0 13. Hewitt WB. Phomopsis cane and leaf spot. Compendium of grape diseases. APS Press; 1988. pp. 17–18. 14. Chiarappa L. Wood decay of the grapevine and its relationship with black measles disease. Phytopa- thology. 1959;49. 15. Halleen F, Schroers HJ, Groenewald JZ, Crous PW. Novel species of Cylindrocarpon (Neonectria) and Campylocarpon gen. nov. associated with black foot disease of grapevines (Vitis spp.). Studies in Mycology. 2004; 50: 431–455. 16. Mugnai L, Graniti A, Surico G. Esca (black measles) and brown wood-streaking: two old and elusive diseases of grapevines. Plant Disease. 1999; 83: 404–418. https://doi.org/10.1094/PDIS.1999.83.5. 404 PMID: 30845530 17. Kaplan J, Travadon R, Cooper M, Hillis V, Lubell M, Baumgartner K. Identifying economic hurdles to early adoption of preventative practices: The case of trunk diseases in California winegrape vineyards. Wine Economics and Policy. 2016; 5: 127–141. https://doi.org/10.1016/j.wep.2016.11.001 18. Baumgartner K, Hillis V, Lubell M, Norton M, Kaplan J. Managing grapevine trunk diseases in Califor- nia’s southern San Joaquin Valley. Am J Enol Vitic. 2019; 70: 267–276. https://doi.org/10.5344/ajev. 2019.18075 19. Guerin-Dubrana L, Fontaine F, Mugnai L. Grapevine trunk disease in European and Mediterranean vineyards: occurrence, distribution and associated disease-affecting cultural factors. Phytopathologia Mediterranea. 2019; 58: 49–71. https://doi.org/10.13128/Phytopathol_Mediterr-25153 20. 21. Lee T-C, Ho I-C. Modulation of cellular antioxidant defense activities by sodium arsenite in human fibroblasts. Arch Toxicol. 1995; 69: 498–504. https://doi.org/10.1007/s002040050204 PMID: 8526746 Tanner CC, Clayton JS. Persistence of arsenic 24 years after sodium arsenite herbicide application to Lake Rotoroa, Hamilton, New Zealand. New Zealand Journal of Marine and Freshwater Research. 1990; 24: 173–179. https://doi.org/10.1080/00288330.1990.9516412 22. Rezgui A, Vallance J, Ben Ghnaya-Chakroun A, Bruez E, Dridi M, Demasse RD, et al. Study of Lasi- diodiplodia pseudotheobromae, Neofusicoccum parvum and Schizophyllum commune, three patho- genic fungi associated with the grapevine trunk diseases in the North of Tunisia. Eur J Plant Pathol. 2018; 152: 127–142. https://doi.org/10.1007/s10658-018-1458-z 23. Van Jaarsveld WJ, Halleen F, Mostert L. In vitro screening of Trichoderma isolates for biocontrol of black foot disease pathogens. Phytopathologia Mediterranea. 2020; 59: 465–471. 24. Mondello V, Songy A, Battiston E, Pinto C, Coppin C, Trotel-Aziz P, et al. Grapevine trunk diseases: a review of fifteen years of trials for their control with chemicals and biocontrol agents. Plant Disease. 2018; 102: 1189–1217. https://doi.org/10.1094/PDIS-08-17-1181-FE PMID: 30673583 25. Tuck B, Gartner W, Appiah G. Vineyards and grapes of the North. University of Minnesota; 2016 Dec. Available: http://conservancy.umn.edu/handle/11299/197814. 26. Hemstad PR, Luby JJ. Utilization of Vitis riparia for the development of new wine varieties with resis- tance to disease and extreme cold. VII International Symposium on Grapevine Genetics and Breeding 528. 2000. pp. 487–496. https://doi.org/10.17660/ActaHortic.2000.528.70 PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 19 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest 27. Hemstad PR, Luby JL. La Crescent, a new cold hardy, high quality, white wine variety. VIII Interna- tional Conference on Grape Genetics and Breeding 603. 2003. pp. 719–722. https://doi.org/10. 17660/ActaHortic.2003.603.100 28. Clark MD. Development of Cold Climate Grapes in the Upper Midwestern U.S. Plant Breeding Reviews. John Wiley & Sons, Ltd; 2019. pp. 31–60. https://doi.org/10.1002/9781119616801.ch2 29. Yin L, Karn A, Cadle-Davidson L, Zou C, Underhill A, Atkins P, et al. Fine mapping of leaf trichome density revealed a 747-kb region on chromosome 1 in cold-hardy hybrid wine grape populations. Fron- tiers in Plant Science. 2021;12. https://doi.org/10.3389/fpls.2021.587640 PMID: 33746993 30. Clark M, Tuck B. Minnesota grape production statistics: 2020—estimates for the yield, production, and pricing data of the Minnesota grape industry. 2021. Available: http://conservancy.umn.edu/handle/ 11299/224882. 31. Billones-Baaijens R, Savocchia S. A review of Botryosphaeriaceae species associated with grapevine trunk diseases in Australia and New Zealand. Australasian Plant Pathology. 2019; 48: 3–18. https:// doi.org/10.1007/s13313-018-0585-5 32. Van Niekerk JM, Bester W, Halleen F, Crous PW, Fourie PH. The distribution and symptomatology of grapevine trunk disease pathogens are influenced by climate. Phytopathologia Mediterranea. 2011; 50: S98–S111. 33. Ye Q, Jia J, Manawasinghe IS, Li X, Zhang W, Mugnai L, et al. Fomitiporia punicata and Phaeoacre- monium minimum associated with Esca complex of grapevine in China. Phytopathology Research. 2021; 3: 11. https://doi.org/10.1186/s42483-021-00087-w 34. Yan J-Y, Xie Y, Zhang W, Wang Y, Liu J-K, Hyde KD, et al. Species of Botryosphaeriaceae involved in grapevine dieback in China. Fungal Diversity. 2013; 61: 221–236. https://doi.org/10.1007/s13225- 013-0251-8 35. Li H, Li RY, Wang H. New Disease for Wine-making Grape-Eutypa Dieback. Liquor Making Science and Technology. 2007; 5: 48. 36. Dı´az GA, Latorre BA. Infection Caused by Phaeomoniella chlamydospora Associated with Esca-like Symptoms in Grapevine in Chile. Plant Disease. 2014; 98: 351–360. https://doi.org/10.1094/PDIS-12- 12-1180-RE PMID: 30708428 37. Dı´az GA, Auger J, Besoain X, Bordeu E, Latorre BA. Prevalence and pathogenicity of fungi associated with grapevine trunk diseases in Chilean vineyards. Ciencia e investigacio´ n agraria. 2013; 40: 327– 339. https://doi.org/10.4067/S0718-16202013000200008 38. Urbez-Torres JR, Peduto F, Striegler RK, Urrea-Romero KE, Rupe JC, Cartwright RD, et al. Charac- terization of fungal pathogens associated with grapevine trunk diseases in Arkansas and Missouri. Fungal Diversity. 2012; 52: 169–189. https://doi.org/10.1007/s13225-011-0110-4 39. Gubler W, Rolshausen P, Trouillas F, U´ rbez-Torres JR, Voegel T, Leavitt GM, et al. Grapevine trunk diseases in California: Research update. Practical Winery & Vineyards. 2005; 1–9. 40. Gramaje D, Mostert L, Groenewald JZ, Crous PW. Phaeoacremonium: From esca disease to phaeo- hyphomycosis. Fungal Biology. 2015; 119: 759–783. https://doi.org/10.1016/j.funbio.2015.06.004 PMID: 26321726 41. U´ rbez-Torres JR, Haag P, Bowen P, O’Gorman DT. Grapevine trunk diseases in British Columbia: incidence and characterization of the fungal pathogens associated with esca and petri diseases of grapevine. Plant Disease. 2014; 98: 469–482. https://doi.org/10.1094/PDIS-05-13-0523-RE PMID: 30708729 42. Worrall JJ. Media for selective isolation of hymenomycetes. Mycologia. 1991; 83: 296–302. https://doi. org/10.1080/00275514.1991.12026013 43. Harrington TC. Cycloheximide sensitivity as a taxonomic character in ceratocystis. Mycologia. 1981; 73: 1123–1129. https://doi.org/10.1080/00275514.1981.12021447 44. Osmundson TW, Eyre CA, Hayden KM, Dhillon J, Garbelotto MM. Back to basics: an evaluation of NaOH and alternative rapid DNA extraction protocols for DNA barcoding, genotyping, and disease diagnostics from fungal and oomycete samples. Molecular Ecology Resources. 2013; 13: 66–74. https://doi.org/10.1111/1755-0998.12031 PMID: 23121735 45. Gardes M, Bruns TD. ITS primers with enhanced specificity for basidiomycetes—application to the identification of mycorrhizae and rusts. Molecular Ecology. 1993; 2: 113–118. https://doi.org/10.1111/ j.1365-294x.1993.tb00005.x PMID: 8180733 46. Blanchette RA, Held BW, Hellmann L, Millman L, Bu¨ ntgen U. Arctic driftwood reveals unexpectedly rich fungal diversity. Fungal Ecology. 2016; 23: 58–65. https://doi.org/10.1016/j.funeco.2016.06.001 47. Lawrence DP, Travadon R, Nita M, Baumgartner K. TrunkDiseaseID.org: a molecular database for fast and accurate identification of fungi commonly isolated from grapevine wood. Crop Protection. 2017; 102: 110–117. https://doi.org/10.1016/j.cropro.2017.08.017 PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 20 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest 48. Wickham H, Averick M, Bryan J, Chang W, McGowan LD, Franc¸ois R, et al. Welcome to the Tidy- verse. Journal of Open Source Software. 2019; 4: 1686. https://doi.org/10.21105/joss.01686 49. Hsieh TC, Ma KH, Chao A. iNEXT: an R package for rarefaction and extrapolation of species diversity (Hill numbers). Methods in Ecology and Evolution. 2016; 7: 1451–1456. https://doi.org/10.1111/2041- 210X.12613 50. Chao A, Gotelli NJ, Hsieh TC, Sander EL, Ma KH, Colwell RK, et al. Rarefaction and extrapolation with Hill numbers: a framework for sampling and estimation in species diversity studies. Ecological Mono- graphs. 2014; 84: 45–67. https://doi.org/10.1890/13-0133.1 51. Sosnowski MR, Shtienberg D, Creaser ML, Wicks TJ, Lardner R, Scott ES. The Influence of Climate on Foliar Symptoms of Eutypa Dieback in Grapevines. Phytopathology®. 2007; 97: 1284–1289. https://doi.org/10.1094/PHYTO-97-10-1284 PMID: 18943686 52. Marchi G, Peduto F, Mugnai L, Di Marco S, Calzarano F, Surico G. Some observations on the relation- ship of manifest and hidden esca to rainfall. Phytopathologia Mediterranea. 2006; 45: S117–S126. 53. 54. Lawrence DP, Travadon R, Pouzoulet J, Rolshausen PE, Wilcox WF, Baumgartner K. Characteriza- tion of Cytospora isolates from wood cankers of declining grapevine in North America, with the descriptions of two new Cytospora species. Plant Pathology. 2017; 66: 713–725. https://doi.org/10. 1111/ppa.12621 Lawrence D, Travadon R, Baumgartner K. Diversity of Diaporthe species associated with wood can- kers of fruit and nut crops in northern California. Mycologia. 2015; 107: 926–940. https://doi.org/10. 3852/14-353 PMID: 26240309 55. U´ rbez-Torres JR, Peduto F, Smith RJ, Gubler WD. Phomopsis dieback: a grapevine trunk disease caused by Phomopsis viticola in California. Plant Disease. 2013; 97: 1571–1579. https://doi.org/10. 1094/PDIS-11-12-1072-RE PMID: 30716818 56. U´ rbez-Torres JR, Adams P, Kamas J, Gubler WD. Identification, incidence, and pathogenicity of fun- gal species associated with grapevine dieback in texas. Am J Enol Vitic. 2009; 60: 497–507. 57. Aroca A, Raposo R. Pathogenicity of Phaeoacremonium Species on Grapevines. Journal of Phytopa- thology. 2009; 157: 413–419. https://doi.org/10.1111/j.1439-0434.2008.01513.x 58. Cloete M, Fourie PH, Damm U, Crous PW, Mostert L. Fungi associated with die-back symptoms of apple and pear trees, a possible inoculum source of grapevine trunk disease pathogens. Phytopatho- logia Mediterranea. 2011; 50: S176–S190. 59. Elena G, Bruez E, Rey P, Luque J. Microbiota of grapevine woody tissues with or without esca-foliar symptoms in northeast Spain. Phytopathologia Mediterranea. 2018. https://doi.org/10.14601/ Phytopathol_Mediterr-23337 60. Bahmani Z, Abdollahzadeh J, Amini J, Evidente A. Biscogniauxia rosacearum the charcoal canker agent as a pathogen associated with grapevine trunk diseases in Zagros region of Iran. Sci Rep. 2021; 11: 14098. https://doi.org/10.1038/s41598-021-93630-w PMID: 34239007 61. Berlanas C, Ojeda S, Lo´ pez-Manzanares B, Andre´ s-Sodupe M, Bujanda R, del Pilar Martı´nez-Diz M, et al. Occurrence and diversity of black-foot disease fungi in symptomless grapevine nursery stock in Spain. Plant Disease. 2020; 104: 94–104. https://doi.org/10.1094/PDIS-03-19-0484-RE PMID: 31738690 62. Cabral A, Cecı´lia R, CROUS P, Oliveira H. Virulence and cross-infection potential of Ilyonectria spp. to grapevine. Phytopathologia Mediterranea. 2012; 51: 340–354. 63. Probst CM, Ridgway HJ, Jaspers MV, Eirian Jones E. Pathogenicity of Ilyonectria liriodendri and Dac- tylonectria macrodidyma propagules in grapevines. Eur J Plant Pathol. 2019; 154: 405–421. https:// doi.org/10.1007/s10658-018-01664-0 64. 65. Lawrence DP, Travadon R, Baumgartner K. Novel Seimatosporium species from grapevine in north- ern california and their interactions with fungal pathogens involved in the trunk-disease complex. Plant Disease. 2018; 102: 1081–1092. https://doi.org/10.1094/PDIS-08-17-1247-RE PMID: 30673434 Trouillas FP, Gubler WD. Pathogenicity of diatrypaceae species in grapevines in California. Plant Dis- ease. 2010; 94: 867–872. https://doi.org/10.1094/PDIS-94-7-0867 PMID: 30743554 66. Halleen F, Mostert L, Crous PW. Pathogenicity testing of lesser-known vascular fungi of grapevines. Australasian Plant Pathology. 2007; 36: 277–285. https://doi.org/10.1071/AP07019 67. Raimondo ML, Carlucci A, Ciccarone C, SADDALLAH A, Francesco L. Identification and pathogenicity of lignicolous fungi associated with grapevine trunk diseases in southern Italy. Phytopathologia Medi- terranea. 2019; 58: 639–662. 68. Travadon R, Lawrence DP, Rooney-Latham S, Gubler WD, Wilcox WF, Rolshausen PE, et al. Cado- phora species associated with wood-decay of grapevine in North America. Fungal biology. 2015; 119: 53–66. https://doi.org/10.1016/j.funbio.2014.11.002 PMID: 25601149 PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 21 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest 69. U´ rbez-Torres JR, Leavitt GM, Guerrero JC, Guevara J, Gubler WD. Identification and pathogenicity of Lasiodiplodia theobromae and Diplodia seriata, the causal agents of bot canker disease of grapevines in Mexico. Plant Disease. 2008; 92: 519–529. https://doi.org/10.1094/PDIS-92-4-0519 PMID: 30769642 70. Marsberg A, Kemler M, Jami F, Nagel JH, Postma-Smidt A, Naidoo S, et al. Botryosphaeria dothidea: a latent pathogen of global importance to woody plant health. Molecular Plant Pathology. 2017; 18: 477–488. https://doi.org/10.1111/mpp.12495 PMID: 27682468 71. Baranek M, Armengol J, Holleinova´ V, Pečenka J, Calzarano F, Peňa´zova´ E, et al. Incidence of symp- toms and fungal pathogens associated with grapevine trunk diseases in Czech vineyards: first exam- ple from a north-eastern European grape-growing region. Phytopathologia Mediterranea. 2018; 57: 449–458. https://doi.org/10.14601/Phytopathol_Mediterr-22460 72. Mundy DC, Brown A, Jacobo F, Tennakoon K, Woolley RH, Vanga B, et al. Pathogenic fungi isolated in association with grapevine trunk diseases in New Zealand. New Zealand Journal of Crop and Horti- cultural Science. 2020; 48: 84–96. https://doi.org/10.1080/01140671.2020.1716813 73. Kaliterna J, Miličević T, Cvjetković B. Grapevine trunk diseases associated with fungi from the dia- porthaceae family in Croatian vineyards. Arhiv za higijenu rada i toksikologiju. 2012; 63: 471–478. https://doi.org/10.2478/10004-1254-63-2012-2226 PMID: 23334042 74. Rolshausen PE, Wilcox W, Baumgartner K. Distribution and occurrence of fungi associated with grapevine trunk diseases in Northeastern American vineyards. Phytopathologia Mediterranea. 2010; 49: 105. 75. van Niekerk JM, Groenewald JZ, Farr DF, Fourie PH, Halleen F, Crous PW. Reassessment of Pho- mopsis species on grapevines. Austral Plant Pathol. 2005; 34: 27. https://doi.org/10.1071/AP04072 76. Mostert L, Crous PW, Petrini O. Endophytic fungi associated with shoots and leaves of Vitis vinifera, with specific reference to the Phomopsis viticola complex. Sydowia. 2000; 52: 46–58. 77. Anderson HW, Colby AS. Dead arm disease of grape in Illinois. The Plant Disease Reporter. 1943; 27: 245–246. 78. Baumgartner K, Fujiyoshi PT, Travadon R, Castlebury LA, Wilcox WF, Rolshausen PE. Characteriza- tion of species of Diaporthe from wood cankers of grape in eastern North American vineyards. Plant Disease. 2013; 97: 912–920. https://doi.org/10.1094/PDIS-04-12-0357-RE PMID: 30722541 79. Dissanayake AJ, Liu M, Zhang W, Chen Z, Udayanga D, Chukeatirote E, et al. Morphological and molecular characterisation of Diaporthe species associated with grapevine trunk disease in China. Fungal Biology. 2015; 119: 283–294. https://doi.org/10.1016/j.funbio.2014.11.003 PMID: 25937058 80. Fitzpatrick DA. Horizontal gene transfer in fungi. FEMS Microbiology Letters. 2012; 329: 1–8. https:// doi.org/10.1111/j.1574-6968.2011.02465.x PMID: 22112233 81. Casacuberta E, Gonza´ lez J. The impact of transposable elements in environmental adaptation. Molec- ular Ecology. 2013; 22: 1503–1517. https://doi.org/10.1111/mec.12170 PMID: 23293987 82. Christensen CM. Studies on the biology agric. Valsa sordida and Cytospora chrysosperma. Phytopa- thology. 1940;30. Available: https://www.cabdirect.org/cabdirect/abstract/19401101375. 83. Cold Outbreak: January 27–31, 2019. In: Minnesota Department of Natural Resources [Internet]. [cited 19 Dec 2021]. Available: https://www.dnr.state.mn.us/climate/journal/cold-outbreak-january-27- 31-2019.html. 84. Svyantek A, Ko¨se B, Stenger J, Auwarter C, Hatterman-Valenti H. Cold-hardy grape cultivar winter injury and trunk re-establishment following severe weather events in North Dakota. Horticulturae. 2020; 6: 75. https://doi.org/10.3390/horticulturae6040075 85. van Niekerk JM, Calitz FJ, Halleen F, Fourie PH. Temporal spore dispersal patterns of grapevine trunk pathogens in South Africa. Eur J Plant Pathol. 2010; 127: 375–390. https://doi.org/10.1007/s10658- 010-9604-2 86. Ayres M, Billones -Baaijens Regina, Savocchia S, Scott E, Sosnowski M. Trunk diseases: susceptibil- ity of pruning wounds to grapevine trunk disease pathogens. Wine & Viticulture Journal. 31: 48–50. https://doi.org/10.3316/informit.523883064121148 87. Kraus C, Voegele RT, Fischer M. Temporal development of the culturable, endophytic fungal commu- nity in healthy grapevine branches and occurrence of GTD-associated fungi. Microb Ecol. 2019; 77: 866–876. https://doi.org/10.1007/s00248-018-1280-3 PMID: 30397796 88. Pouzoulet J, Scudiero E, Schiavon M, Rolshausen PE. Xylem vessel diameter affects the compart- mentalization of the vascular pathogen Phaeomoniella chlamydospora in grapevine. Frontiers in Plant Science. 2017; 8: 1442. https://doi.org/10.3389/fpls.2017.01442 PMID: 28871268 89. Bortolami G, Gambetta GA, Delzon S, Lamarque LJ, Pouzoulet J, Badel E, et al. Exploring the Hydrau- lic Failure Hypothesis of Esca Leaf Symptom Formation. Plant Physiology. 2019; 181: 1163–1174. https://doi.org/10.1104/pp.19.00591 PMID: 31455632 PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 22 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest 90. Bortolami G, Farolfi E, Badel E, Burlett R, Cochard H, Ferrer N, et al. Seasonal and long-term conse- quences of esca grapevine disease on stem xylem integrity. Journal of Experimental Botany. 2021; 72: 3914–3928. https://doi.org/10.1093/jxb/erab117 PMID: 33718947 91. Agrios GN. Plant Pathology. fifth. Elsevier Inc.; 2005. 92. Valtaud C, Larignon P, Roblin G, Fleurat-Lessard P. Developmental and ultrastructural features of Phaeomoniella chlamydospora and Phaeoacremonium aleophilum in relation to xylem degradation in esca disease of the grapevine. Journal of Plant Pathology. 2009; 91: 37–51. 93. Yin Z, Liu H, Li Z, Ke X, Dou D, Gao X, et al. Genome sequence of Valsa canker pathogens uncovers a potential adaptation of colonization of woody bark. New Phytologist. 2015; 208: 1202–1216. https:// doi.org/10.1111/nph.13544 PMID: 26137988 94. Morales-Cruz A, Amrine KCH, Blanco-Ulate B, Lawrence DP, Travadon R, Rolshausen PE, et al. Dis- tinctive expansion of gene families associated with plant cell wall degradation, secondary metabolism, and nutrient uptake in the genomes of grapevine trunk pathogens. BMC Genomics. 2015; 16: 469. https://doi.org/10.1186/s12864-015-1624-z PMID: 26084502 95. Blanchette RA. Degradation of the lignocellulose complex in wood. Can J Bot. 1995; 73: 999–1010. https://doi.org/10.1139/b95-350 96. Pouzoulet J, Pivovaroff AL, Santiago LS, Rolshausen PE. Can vessel dimension explain tolerance toward fungal vascular wilt diseases in woody plants? Lessons from Dutch elm disease and esca dis- ease in grapevine. Frontiers in Plant Science. 2014; 5: 253. https://doi.org/10.3389/fpls.2014.00253 PMID: 24971084 97. Pouzoulet J, Rolshausen PE, Charbois R, Chen J, Guillaumie S, Ollat N, et al. Behind the curtain of the compartmentalization process: Exploring how xylem vessel diameter impacts vascular pathogen resistance. Plant, Cell & Environment. 2020; 43: 2782–2796. https://doi.org/10.1111/pce.13848 PMID: 32681569 98. Knipfer T, Eustis A, Brodersen C, Walker AM, Mcelrone AJ. Grapevine species from varied native hab- itats exhibit differences in embolism formation/repair associated with leaf gas exchange and root pres- sure. Plant, Cell & Environment. 2015; 38: 1503–1513. https://doi.org/10.1111/pce.12497 PMID: 25495925 99. Smith MS, Centinari M. Young grapevines exhibit interspecific differences in hydraulic response to freeze stress but not in recovery. Planta. 2019; 250: 495–505. https://doi.org/10.1007/s00425-019- 03183-6 PMID: 31089803 100. Lovisolo C, Schubert A. Effects of water stress on vessel size and xylem hydraulic conductivity in Vitis vinifera L. Journal of Experimental Botany. 1998; 49: 693–700. https://doi.org/10.1093/jxb/49.321.693 101. Garrett KA, Nita M, De Wolf ED, Esker PD, Gomez-Montano L, Sparks AH. Chapter 24—Plant patho- gens as indicators of climate change. In: Letcher TM, editor. Climate Change (Third Edition). Elsevier; 2021. pp. 499–513. https://doi.org/10.1016/B978-0-12-821575-3.00024–4 102. Bortolami G, Gambetta GA, Cassan C, Dayer S, Farolfi E, Ferrer N, et al. Grapevines under drought do not express esca leaf symptoms. Proceedings of the National Academy of Sciences. 2021;118. https://doi.org/10.1073/pnas.2112825118 PMID: 34675082 103. Galarneau ERA, Lawrence DP, Travadon R, Baumgartner K. Drought Exacerbates Botryosphaeria Dieback Symptoms in Grapevines and Confounds Host-based Molecular Markers of Infection by Neo- fusicoccum parvum. Plant Disease. 2019; 103: 1738–1745. https://doi.org/10.1094/PDIS-09-18-1549- RE PMID: 31082329 104. Whiteman SA, Stewart A, Ridgway HJ, Jaspers MV. Infection of rootstock mother-vines by Phaeomo- niella chlamydospora results in infected young grapevines. Australasian Plant Pathology. 2007; 36: 198–203. https://doi.org/10.1071/AP07012 105. Damm U, Verkley GJM, Crous PW, Fourie PH, Haegi A, Riccioni L. Novel Paraconiothyrium species on stone fruit trees and other woody hosts. Persoonia. 2008; 20: 9–17. https://doi.org/10.3767/ 003158508X286842 PMID: 20467483 106. Groenewald M, Kang J-C, Crous PW, Gams W. ITS and β-tubulin phylogeny of Phaeoacremonium and Phaeomoniella species. Mycological Research. 2001; 105: 651–657. https://doi.org/10.1017/ S0953756201004282 107. Gramaje D, Alaniz S, Pe´rez-Sierra A, Abad-Campos P, Garcı´a-Jime´nez J, Armengol J. First report of Phaeoacremonium mortoniae causing Petri disease of grapevine in Spain. Plant Disease. 2007; 91: 1206–1206. https://doi.org/10.1094/PDIS-91-9-1206A PMID: 30780685 108. Mohammadi H. First report of Phaeoacremonium mortoniae associated with grapevine decline in Iran. Plant Disease. 2011; 95: 1034–1034. https://doi.org/10.1094/PDIS-05-11-0361 PMID: 30732074 PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 23 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest 109. Crous PW, Gams W, Wingfield MJ, van Wyk PS. Phaeoacremonium gen. nov. associated with wilt and decline diseases of woody hosts and human infections. Mycologia. 1996; 88: 786–796. https://doi. org/10.1080/00275514.1996.12026716 110. Gatica M, Dupont J, Magnin S, Ce´ sari C. Phaeoacremonium species and Phaeomoniella chlamydos- pora in vines showing “hoja de malvo´ n” and young vine decline symptoms in Argentina. Phytopatholo- gia Mediterranea. 2001; 1000–1008. https://doi.org/10.1400/14644 111. Cottral E, Pascoe I. Developments in grapevine trunk diseases research in Australia. Phytopathologia Mediterranea. 2000; 39: 68–75. https://doi.org/10.1400/57813 112. Essakhi S, Mugnai L, Crous PW, Groenewald JZ, Surico G. Molecular and phenotypic characterisation of novel Phaeoacremonium species isolated from esca diseased grapevines. Persoonia. 2008; 21: 119–134. https://doi.org/10.3767/003158508X374385 PMID: 20396582 113. Held BW, Simeto S, Rajtar NN, Cotton AJ, Showalter DN, Bushley KE, et al. Fungi associated with gal- leries of the emerald ash borer. Fungal Biology. 2021; 125: 551–559. https://doi.org/10.1016/j.funbio. 2021.02.004 PMID: 34140151 114. Arenz BE, Held BW, Jurgens JA, Farrell RL, Blanchette RA. Fungal diversity in soils and historic wood from the Ross Sea Region of Antarctica. Soil Biology and Biochemistry. 2006; 38: 3057–3064. https:// doi.org/10.1016/j.soilbio.2006.01.016 115. Blanchette RA, Held BW, Jurgens JA, McNew DL, Harrington TC, Duncan SM, et al. Wood-destroying soft rot fungi in the historic expedition huts of Antarctica. Applied and Environmental Microbiology. 2004 [cited 21 Jan 2022]. https://doi.org/10.1128/AEM.70.3.1328-1335.2004 PMID: 15006750 116. Blanchette RA, Held BW, Jurgens J, Stear A, Dupont C. Fungi attacking historic wood of Fort Conger and the Peary Huts in the High Arctic. PLOS ONE. 2021; 16: e0246049. https://doi.org/10.1371/ journal.pone.0246049 PMID: 33497418 117. Rusman Y, Held BW, Blanchette RA, He Y, Salomon CE. Cadopherone and colomitide polyketides from Cadophora wood-rot fungi associated with historic expedition huts in Antarctica. Phytochemistry. 2018; 148: 1–10. https://doi.org/10.1016/j.phytochem.2017.12.019 PMID: 29366851 118. Gramaje D, Mostert L, Armengol J. Characterization of Cadophora luteo-olivacea and C. melinii iso- lates obtained from grapevines and environmental samples from grapevine nurseries in Spain. Phyto- pathologia Mediterranea. 2011; 50: S112–S126. 119. Agustı´-Brisach C, Leo´ n M, Garcı´a-Jime´nez J, Armengol J. Detection of grapevine fungal trunk patho- gens on pruning shears and evaluation of their potential for spread of infection. Plant Disease. 2015; 99: 976–981. https://doi.org/10.1094/PDIS-12-14-1283-RE PMID: 30690978 120. Murolo S, Romanazzi G. Effects of grapevine cultivar, rootstock and clone on esca disease. Austral- asian Plant Pathology. 2014; 43: 215–221. https://doi.org/10.1007/s13313-014-0276-9 121. Agustı´-Brisach C, Mostert L, Armengol J. Detection and quantification of Ilyonectria spp. associated with black-foot disease of grapevine in nursery soils using multiplex nested PCR and quantitative PCR. Plant Pathology. 2014; 63: 316–322. https://doi.org/10.1111/ppa.12093 122. del Pilar Martı´nez-Diz M, Andre´ s-Sodupe M, Berbegal M, Bujanda R, Dı´az-Losada E, Gramaje D. Droplet digital PCR technology for detection of Ilyonectria liriodendri from grapevine environmental samples. Plant Disease. 2020; 104: 1144–1150. https://doi.org/10.1094/PDIS-03-19-0529-RE PMID: 32053059 123. Rajtar NN, Held BW, Blanchette RA. Fungi from galleries of the emerald ash borer produce cankers in ash trees. Forests. 2021; 12: 1509. https://doi.org/10.3390/f12111509 124. Biodiversity Fischer M. and geographic distribution of basidiomycetes causing esca-associates white rot in grapevine: a worldwide perspective. Phytopathologia Mediterranea. 2006; 1–13. 125. Fischer M, Garcı´a G. V. An annotated checklist of European basidiomycetes related to white rot of grapevine. Phytopathologia Mediterranea. 2015; 54: 281–298. 126. Brown AA, Lawrence DP, Baumgartner K. Role of basidiomycete fungi in the grapevine trunk disease esca. Plant Pathology. 2020; 69: 205–220. https://doi.org/10.1111/ppa.13116 127. 128. Fischer M, Edwards J, Cunnington JH, Pascoe IG. Basidiomycetous pathogens on grapevine: a new species from Australia-Fomitiporia australiensis. Mycotaxon. 2005; 92: 85–96. Tomoiaga LL, Chedea VS. Grapevine trunk diseases, anagement in vineyards from central Transylva- nia. BUASVMCN-HORT. 2020; 77: 117. https://doi.org/10.15835/buasvmcn-hort:2019.0041 129. Dilley MA, Covey RP. Survey of wood decay and associated hymenomycetes in central Washington apple orchards. Plant Disease. 1980; 64: 561. 130. Enebak SA, Blanchette RA. Canker formation and decay in sugar maple and paper birch infected by Cerrena unicolor. Canadian Journal of Forest Research. 2011 [cited 21 Jan 2022]. https://doi.org/10. 1139/x89-031 PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 24 / 25 PLOS ONEGrapevine trunk diseases of the Northern Midwest 131. Claverie M, Notaro M, Fontaine F, Wery J. Current knowledge on grapevine trunk diseases with com- plex etiology: a systemic approach. Phytopathologia Mediterranea. 2020; 59: 29–53. 132. Wingfield MJ, De Beer ZW, Slippers B, Wingfield BD, Groenewald JZ, Lombard L, et al. One fungus, one name promotes progressive plant pathology. Molecular Plant Pathology. 2012; 13: 604–613. https://doi.org/10.1111/j.1364-3703.2011.00768.x PMID: 22146077 133. Silva F, Santos K, Tamiris R, Armengol J, Rossi V, Michereff S, et al. Temporal conidial dispersal pat- tern of Botryosphaeriaceae species in table-grape vineyards in Northeastern Brazil. Phytopathologia Mediterranea. 2018; 57: 547–556. https://doi.org/10.14601/Phytopathol_Mediterr-24240 134. Urbez-Torres JR, Gispert C, Trouillas FP. Spore dispersal of Eutypella species under desert grape- growing conditions of southern California. Phytopathologia Mediterranea. 2020; 59: 485–501. https:// doi.org/10.14601/Phyto-11807 135. Gonza´lez-Domı´nguez E, Caffi T, Languasco L, Latinovic N, Latinovic J, Rossi V. Dynamics of Dia- porthe ampelina conidia released from grape canes that overwintered in the vineyard. Plant Disease. 2021; 105: 3092–3100. https://doi.org/10.1094/PDIS-12-20-2639-RE PMID: 33755509 136. Otto EC, Held BW, Gould TJ, Blanchette RA. Fungal Diversity in Multiple Post-harvest Aged Red Pine Stumps and Their Potential Influence on Heterobasidion Root Rot in Managed Stands Across Minne- sota. Frontiers in Fungal Biology. 2021;2. Available: https://www.frontiersin.org/article/10.3389/ffunb. 2021.782181. PLOS ONE \| https://doi.org/10.1371/journal.pone.0269555 June 3, 2022 25 / 25 PLOS ONE	null
10.1038/s41467-022-29297-2	Data availability The microbial growth data generated in this study are provided in the Supplementary Table S2. Code availability The model code is deposited on GitHub (https://github.com/LevineLab/POMmodel) and citable using https://doi.org/10.5281/zenodo.6015020.	Data availability The microbial growth data generated in this study are provided in the Supplementary Table S2 . Code availability The model code is deposited on GitHub ( https://github.com/LevineLab/POMmodel ) and citable using https://doi.org/10.5281/zenodo.6015020 .	ARTICLE https://doi.org/10.1038/s41467-022-29297-2 OPEN Microbes contribute to setting the ocean carbon ﬂux by altering the fate of sinking particulates Trang T. H. Nguyen Kapil Amarnath Otto X. Cordero 3, Uria Alcolombri 2 & Naomi M. Levine 1✉ 1, Emily J. Zakem1, Ali Ebrahimi 2, Julia Schwartzman 2, Tolga Caglar3, 4, François J. Peaudecerf 4, Terence Hwa 3, Roman Stocker 4, ; , : ) ( 0 9 8 7 6 5 4 3 2 1 Sinking particulate organic carbon out of the surface ocean sequesters carbon on decadal to millennial timescales. Predicting the particulate carbon ﬂux is therefore critical for under- standing both global carbon cycling and the future climate. Microbes play a crucial role in particulate organic carbon degradation, but the impact of depth-dependent microbial dynamics on ocean-scale particulate carbon ﬂuxes is poorly understood. Here we scale-up essential features of particle-associated microbial dynamics to understand the large-scale vertical carbon ﬂux in the ocean. Our model provides mechanistic insight into the microbial contribution to the particulate organic carbon ﬂux proﬁle. We show that the enhanced transfer of carbon to depth can result from populations struggling to establish colonies on sinking particles due to diffusive nutrient loss, cell detachment, and mortality. These dynamics are controlled by the interaction between multiple biotic and abiotic factors. Accurately capturing particle-microbe interactions is essential for predicting variability in large-scale carbon cycling. 1 Department of Biological Sciences, University of Southern California, Los Angeles, CA 90089, USA. 2 Ralph M. Parsons Laboratory for Environmental Science and Engineering, Department of Civil and Environmental Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA. 3 Department of Physics, University of California at San Diego, La Jolla, CA 92093, USA. 4 Institute of Environmental Engineering, Department of Civil, Environmental and Geomatic Engineering, ETH Zurich, 8093 Zurich, Switzerland. email: [email protected] ✉ NATURE COMMUNICATIONS \| (2022) 13:1657 \| https://doi.org/10.1038/s41467-022-29297-2 \| www.nature.com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS \| https://doi.org/10.1038/s41467-022-29297-2 The vertical ﬂux of particulate organic carbon (POC) in the ocean has been the subject of numerous ﬁeld campaigns, laboratory experiments, and modeling studies over the past four decadese.g.,1–25. Ultimately, the vertical ﬂux of POC is determined by the rate of POC production in the surface ocean where de novo particle production takes place, the sinking speed of particles, and the rate of POC consumption in the subsurface ocean. Previous carbon cycle modeling studies aimed at under- standing basin-scale ﬂuxes out of the upper ocean have focused primarily on physical and chemical processes5,7,8,26 and have shown that differences in particle size spectra5,9,16,17, lability6, and temperature8,27 play important roles in determining the efﬁciency of POC transfer, consistent with observations4,28,29. However, these large-scale models have not explicitly included microbial ecological dynamics. This is problematic as many of the aforementioned physical and chemical mechanisms shown to impact the vertical ﬂux of POC are controlled by microbial and grazer dynamics22–25,30. For example, lability as represented by these models is inherently related to microbial activity (as we demonstrate below). Even the size of particles and thus their sinking speed is strongly inﬂuenced by the rate of degradation by microbes and consumption by zooplankton31–33. Organisms alter the POC ﬂux through a variety of processes. Heterotrophic microbes directly consume organic carbon within particles with an estimated contribution of 70-92% of POC remineralization25. Grazers such as zooplankton con- tribute to the aggregation, disaggregation, and consumption of particles23,25,30,34–40 with small particles (0.7-53 μm) acting as a primary food source for zooplankton in the mesopelagic ocean through ﬂux feeding36. However, rather than explicitly capturing these dynamics, current biogeochemical models rely on low and invariant rates of POC consumption (typically using bulk remi- neralization terms) tuned in order to match observed POC ﬂux proﬁles5,6,8. Here we focus on the mechanisms behind the per- sistence of particulate organic carbon in the deep ocean and the role that heterotrophic microbes play in decreasing the vertical ﬂux of particulate carbon out of the surface ocean. Additional work is needed to understand the coupled impact of microbial and grazer dynamics on the POC ﬂux and the relative contribution of these processes to the shape of the POC ﬂux proﬁle (see Discus- sion below). Particle-associated heterotrophic microbial communities are incredibly dynamic41–44 and able to achieve relatively rapid growth rates (order of 1 to 10 day −1)32,42,43,45,46 even on seemingly ‘recalcitrant’ organic carbon compounds41,44. As a result, microbes are capable of consuming particles on the timescales of days32,41–44. This contrasts with observations of ‘labile’ particles in the bath- ypelagic and epibenthic zones21,47. Previous work suggests that temperature and pressure limitation on microbial activity might allow for the persistence of particles in the deep ocean16,48–50. In addition, microbial dynamics such as enzyme production, attach- ment, detachment, and mortality have been shown to play a key role in the rate of POC degradation31,41–43,46,51–53. Laboratory studies suggests that, for successful colonization of a particle to occur, particle-associated microbes must surpass a critical popula- tion size46,54. This critical population size or density of cells (cells per surface area) is necessary for countering the diffusive losses of both extracellular enzymes that are used to break down polymers (the main component of POC) and the resulting low-molecular- weight degradation products42,46,54–56. These micro-scale observa- tions suggest that the rate of POC consumption in the ocean is highly variable and can vary as a function of microbial processes, in contrast representation in carbon cycle models45,46,53,57,58. Therefore, to mechanistically understand the vertical ﬂux of carbon in the ocean and generate robust predictions of future changes, we must account for particle-associated microbial to the conventional behavior22,24,59, in addition to other dynamics such as particle size distribution and zooplankton activity. is Here we present a water-column model that explicitly accounts for micro-scale observations, reconciles rapid microbial growth rates with slow POC remineralization timescales (order of 10−3 to 10−1 day−1) in the upper ocean (<2,500 m)3,4,29,60, and determines the impact of shifts in microbial dynamics on the rate of POC ﬂux attenuation. Speciﬁcally, we identify key aspects of particle-associated microbial community dynamics that contribute to setting the shape of the POC ﬂux attenuation proﬁle. We also predict that changes in microbial community dynamics can rapidly shift rates of POC remineralization. This work demonstrates that the assumption of low constant POC incorrect5,61,62, and that microbial consumption rates dynamics alone can generate signiﬁcant variability in the POC ﬂux. Our results challenge the classic idea of particles being inherently labile or recalcitrant and propose that lability is an emergent ecosystem property and a function of the microbial community, organic matter chemistry, and environmental conditions. One of the difﬁculties in studying the vertical POC ﬂux in the ocean is that POC ﬂux observations provide a poor constraint on models13—in fact only 2 free parameters are needed to represent the classic POC ﬂux curve while mechanistic carbon-cycle models rely on many more parameters. Thus, capturing observed POC ﬂux proﬁles is not in of itself sufﬁcient validation of a proposed mechanism. Our model allows for targeted hypotheses that can be tested in the ﬁeld—the results of which will provide enhanced constraints on global carbon-cycle models allowing for more robust predictions of future changes in carbon export out of the surface ocean. Results Scaling-up micro-scale dynamics to the water-column. Our mechanistic model captures the colonization of particles by free- living microbes and conversion from polymeric to low- molecular-weight organic matter (LMWOM) compounds as the particles sink (Methods and Supplementary Material 1). To represent the complexity of the interactions between diverse particle-associated microbial communities and a chemically diverse organic carbon pool, we use lability as an ecosystem property that speciﬁes the conversion rate of polymeric material to LMWOM of a speciﬁc POC pool by a speciﬁc microbial group63. As lability is poorly constrained by observations, we test a wide range of POC lability values assuming a log-normal distribution63, though the modeled POC ﬂuxes at the ocean-scale are found to be independent of the speciﬁc form of the dis- tribution chosen (Supplementary Fig. 1). We represent the organic carbon content of each model particle using a single, stochastically assigned, lability value and represent the particle- associated microbial community using a single microbial group per particle (in all we resolve 18 microbial types with different enzyme kinetics and growth rates). Sensitivity tests with complex particles, multiple microbial groups per particle, and greater numbers of microbial groups show that these simpliﬁcations do not impact the overall POC ﬂux proﬁles (Supplementary Figs. 2 and 3). The model is initialized using a particle export depth between 50 and 100 m (with a default value of 100 m) with particle size distributions spanning the observed range (power law with exponent s = −2, −3, or −4; Supplementary Table 1). For simplicity, we do not include any additional particle formation below the export depth. Each particle is stochastically assigned an initial radius at the export depth and type of microbial degrader. The microbial groups are deﬁned by their enzyme kinetics (i.e., POC degradation rate), 2 NATURE COMMUNICATIONS \| (2022) 13:1657 \| https://doi.org/10.1038/s41467-022-29297-2 \| www.nature.com/naturecommunications NATURE COMMUNICATIONS \| https://doi.org/10.1038/s41467-022-29297-2 ARTICLE Fig. 1 Micro-scale model dynamics. a Illustration showing micro-scale model dynamics occurring on sinking particles. Primary degraders (red microbes) convert polymeric organic matter (dark blue sphere) into low molecular weight organic matter (LMWOM, light blue) using extracellular enzymes (yellow). The particle-associated community experiences loss due to mortality (gray microbes) and detachment (purple microbes). b Illustration of water column model dynamics with an emphasis on a single particle (blue sphere) falling through the water column. Each particle is stochastically assigned an initial radius, lability, and set of biological parameter values at the depth of formation (see Methods). The particle-associated microbial dynamics then evolve prognostically for each particle as it sinks through the water column and is consumed by microbial activity. The total particulate organic carbon ﬂux throughout the water column is obtained by summing across all sinking particles. maximum growth rate, and abundance in the free-living microbial pool, which is set to decrease exponentially with depth64 (Fig. 1, see Methods and Supplementary Material 1 for more details). The particle is initialized with a stochastically assigned initial microbial biomass and is continually colonized by the free-living microbial pool as it sinks through the water column. The sinking speed of size and speciﬁc each particle is calculated based on its gravity5,16,28,65 (Supplementary Material 1 eqs. s13–s15). As each particle sinks through the water column, microbial growth on the particle evolves prognostically, and organic carbon is consumed or lost to the surrounding water column due to diffusion and advection55. The microbially-catalyzed degradation of the particle causes the radius to shrink and the sinking speed to decrease. Temperature dependence of microbial growth rates is applied using a typical water column temperature proﬁle66 (Supplementary Figs. 4 and 5). The model captures the observed density-dependent growth of particle-associated microbial populations, where the population density (cells per surface area) must surpass a critical threshold in order for successful particle colonization to occur46 (Fig. 2). This critical threshold is a consequence of mortality and population- size dependent per capita growth rate (the ‘Allee effect’54); see Supplementary Figs. 6–9, Supplementary Material 3. Speciﬁcally, microbial populations must overcome multiple sources of loss including mortality (e.g., viral lysis or bacterivory), detachment, and the diffusive and advective loss of LMWOM away from the loss particle processes). Model populations below a critical threshold cannot establish a colony on the particle due to these loss processes, consistent with laboratory observations46 (Fig. 2 and Supple- mentary Fig. 10). Using a simpliﬁed mathematical model of POC degradation, we show that the population-dependent growth rate arises naturally as an interplay of simple microbial dynamics (e.g., saturating (Monod) growth kinetics, uptake kinetics, and yield) and particle chemical and physical properties (e.g., particle size and monomer diffusivity leading to nutrient loss) (Supplementary Material 4, Supplementary Fig. 11). referred to collectively as (hereafter surface Ebrahimi et al 2019 Our Model 2.5 2 1.5 1 0.5 s s a m o B n i i e g n a h C d o F l 0 100 102 Initial population (cells/particle) 104 Fig. 2 Density-dependent growth validation. The model captures observed density-dependent growth of particle-associated communities. Fold change in microbial population biomass after 10 hours of growth is shown for the model simulations (red asterisks) and experimental data (open circles, Ebrahimi et al. (2019)). Error bars on the open circles represent standard error of measurement from at least three measurements of particle colonization density from Ebrahimi et al. (2019). Role of microbes in setting water-column POC ﬂuxes. Multiple biotic and abiotic factors control the rate of particle-associated microbial growth and therefore the microbial consumption of POC. Critically, the emergent particle consumption rates that result from interactions between these factors (e.g., particle labi- lity and particle-microbe encounter rate) are not predictable from one factor alone (Fig. 3b). For example, with an initial radius of −1 day−1, half of 500 μm and lability of 32 mmol CPOC mmol Ccell the time the particle is consumed in the upper 1,000 m (50% of particles) and half of the time the particle will persist into the NATURE COMMUNICATIONS \| (2022) 13:1657 \| https://doi.org/10.1038/s41467-022-29297-2 \| www.nature.com/naturecommunications 3 ARTICLE NATURE COMMUNICATIONS \| https://doi.org/10.1038/s41467-022-29297-2 a) 3 2.5 2 1.5 1 0.5 2 - m m s l l e c 104 b) 100 80 60 40 i s u d a r i i g n n a m e r 0 50 100 Time (hour) 150 % 20 0 0 50 100 Time (hour) 150 c) 500 ) m ( h t p e D 1000 1500 2000 0 100 50 % remaining of initial radius d) % particle persisting to 1,000 m 0.005 17 17 17 17 0 0.01 50 34 17 17 17 0.02 100 59 25 17 17 0.05 100 84 50 25 17 0 0 9 9 0.1 100 92 59 34 25 12 0.2 100 100 75 42 25 12 ) m c ( i s u d a r l e c i t r a P 0 0 7 9 9 9 0.4 100 100 92 67 41 21 10 0 0 0 0 5 5 6 0 0 0 0 0 0 2 102 101 100 (A) struggling community: low lability (B) struggling community: slow growth (C) rescued community (D) exponential growth: long lag (E) exponential growth: small particle (F) exponential growth: large particle (G) exponential growth: high lability 10 18 32 56 100 180 315 Lability (mmol C mmol C POC 560 1000 -1 day-1) cell Fig. 3 Particle-associated microbial growth dynamics. Example degradation dynamics are shown for rapidly growing particle-associated microbial populations (curves E, F, G), ‘rescued’ populations (C, D), and ‘struggling’ populations (A, B). Parameter values are given in Supplementary Fig. 14. For each simulation, the change in microbial density on the particle surface over time (a), percentage changes in the particle radius over time (b) and depth (c) are shown. To quantify the impact of microbial growth dynamics on POC persistence at depth, the fraction of particles within a given lability class and radius at formation that persist below 1,000 m is calculated (d). The colors and values in each grid correspond to the percentage of model parameter combinations for a given lability and initial radius (n = 72) that persist past 1,000 m depth (see Supplementary Material 1). deep ocean (>2,000 m). The timescale for particle degradation by the microbial community varies from ~5 days to >200 days, with most particles lasting for a few weeks below the export depth (~30 days, lognormally distributed), consistent with observations67,68. The degradation timescale for each particle is determined by the duration of the lag phase for the particle- associated microbial populations: once populations reach expo- nential growth, the particle is consumed rapidly. When popula- tion density at particle formation (100 m in the default simulations) is high, encounter rates are high, and/or lability is high, the population quickly overcomes loss processes and reaches exponential growth (Fig. 3, E-G curves). This results in the complete consumption of particles within the upper 500 m of the water column. When the population density at formation is low, encounter rates are low, and/or the lability of the particle is low, loss rates can exceed the particle-associated microbial growth rates. This results in a particle-associated community that is unable to successfully colonize the particle and reach exponential growth (here termed a “struggling” community). This in turn results in slow microbially mediated POC remineralization rates, and a higher transfer of carbon into the deep ocean (Fig. 3, A-B curves). A struggling particle-associated population can be “res- cued” by recruitment from the free-living microbial pool as the particle sinks if the encounter rates are sufﬁciently high, resulting in shallower particle consumption (Fig. 3, e.g., A vs. C curve). The depth over which a particle sinks before it is completely con- sumed by the particle-associated microbial community thus depends on multiple biotic and abiotic factors that determine microbial population behavior (Fig. 3d, see Supplementary Material 3): the conversion rate of POC to LMWOM, particle size (which determines diffusive loss and sinking speed), microbial biomass at particle formation, temperature, and exchange between the particle and free-living community (encounter rate). The dichotomy between particle-associated populations that successfully colonize particles and populations that struggle to colonize particles provides a mechanistic explanation for the commonly used approximation of two labilities of POC6 or the double exponential representation of the POC ﬂux10. Here we show that this dichotomy emerges mechanistically as a result of lability cannot be deﬁned microbial dynamics. Furthermore, solely as a chemical property of the particle but must be considered an ecosystem property63,69–71 deﬁned by both the organic carbon composition and the enzyme systems of the colonizing microbial community. The result of these dynamics is that particle remineralization rates for a single particle type (e.g., radius and lability) can vary by several orders of magnitude across particles with different microbial dynamics and over depth (Supplementary Figs. 12 and 13). This dichotomy also emerges when a modiﬁed version of the model is used where the particle- associated populations are able to directly consume POC and so are not subjected to diffusive loss (Supplementary Material 3 and Supplementary Fig. 15). In this direct-uptake formulation, the rate of particle degradation is limited by the rate at which the microbes can incorporate carbon into their biomass. We demonstrate that both models produce similar dynamics and rates of particle remineralization (Supplementary Fig. 16). Our 4 NATURE COMMUNICATIONS \| (2022) 13:1657 \| https://doi.org/10.1038/s41467-022-29297-2 \| www.nature.com/naturecommunications NATURE COMMUNICATIONS \| https://doi.org/10.1038/s41467-022-29297-2 ARTICLE model mechanistically captures these particle-associated micro- bial dynamics and their impacts on particle consumption. This provides insight into how micro-scale processes might drive geographic variations in POC ﬂux attenuation and allows for the generation of testable hypotheses for experimentalists. The model predicts a shift in particle-associated microbial community composition with depth. Fast growing microbes associated with high-lability particles are abundant in the upper water column, whereas slower-growing microbes associated with low-lability particles are relatively more abundant at depth (Supplementary Fig. 17). This pattern emerges because popula- tions growing on labile particles and populations with faster maximum growth rates can more easily surpass the critical threshold necessary for successful colonization and so thrive in the upper water column. In contrast, slower-growing commu- nities take longer to build up biomass on particles and so are more abundant at depth. Since slow growing populations are also present in the surface ocean, the model predicts a more diverse particle-associated microbial community in the surface waters compared to the deeper ocean (Supplementary Fig. 17). These ﬁndings are consistent with documented changes in particle- communities with depth18,57,58,72,73: associated prokaryotic particle associated communities tend to be more homogenous at depth and community richness is positively correlated with the rate of POC remineralization. The ocean is associated with strong vertical gradients in temperature which have been hypothesized to play an important role in POC ﬂux attenuation8,60,74. We investigate the role that temperature plays in our modeled microbial dynamics using a suite of model simulations with varied temperature proﬁles and two temperature limitation functions, which span the observed temperature-growth rate relationship for marine microbial heterotrophs (Supplementary Fig. 5). Temperature does play an important role in the loss processes described above as decreased growth rates due to temperature limitation makes it more difﬁcult to overcome loss processes and for microbial populations successfully colonize particles. However, our results suggest that temperature limitation is not the primary driver of the observed dynamics (Supplementary Material 3, Supplementary Figs. 18–20). Speciﬁcally, if a population is able to grow exponentially, the impact of temperature becomes secondary. Temperature becomes important for the struggling communities which are associated with slow microbial growth rates and low particle consumption rates, consistent with previous work48,50. The global relationships observed between POC ﬂux attenuation rates and temperature may be partially explained by co-varying factors such as shifts in POC lability and microbial community function4,64,75,76. (cid:1) Variability in microbially mediated POC ﬂux attenuation. The explicit representation of micro-scale dynamics on particles in our model generates water-column-scale estimates of bacterially mediated POC ﬂuxes (Fig. 4). To compare across large numbers of model simulations and with observed POC ﬂux proﬁles, we quantify the rate of POC ﬂux attenuation using the commonly (cid:3)(cid:2)b; where b is the attenua- used power law function F ¼ F100 tion exponent, which is inversely related to the vertical carbon transfer rate, F100 is the ﬂux at 100 m, and z is depth in meters1–3. The model reproduces known relationships. For example, higher b values in the model output are associated with higher abundance of small particles (ﬂatter particle size spectra) and higher lability. Thus, the model predicts that oligotrophic regions with ﬂatter particle size spectra would coincide with higher b values, consistent with pre- vious studies5–9,16 In addition, the model results suggest that shifts in the depth of particle formation within the euphotic zone play z 100 a key role in the attenuation rates observed in the upper water column (<500 m) (Supplementary Fig. 21). The observed range in the POC ﬂux across different oceanic regions is large (i.e., 1–30% of the initial ﬂux remains at 1,000 m)1,3. We demonstrate that variable microbial dynamics is sufﬁcient to generate the observed mean, range, and distribution of observed POC ﬂux proﬁles from across the global oceans (n = 897)3 (Fig. 4). While other processes not included in the model (e.g., zooplankton dynamics) are also critical for setting the POC ﬂux attenuation proﬁle, our results suggest that variations in microbial POC consumption rates may play a signiﬁcant role in determining spatial and temporal changes in the POC ﬂux proﬁle. For example, we show that particle-associated microbial community dynamics such as shifts in maximum growth rate or particle lability can alter the POC ﬂux to the same extent as a change in particle size spectra (Fig. 4). These model dynamics (from rapidly growing populations to those struggling to survive) emerge as a result of stochastic interactions between biological, chemical, and physical controls on microbial growth. Discussion Our results suggest that the stochastic assembly of communities on particles frequently results in communities struggling to overcome losses, thereby generating the long tail of persistent POC at depth (>1000 m). Furthermore, for particles that persist at depth due to sub-optimal growth of the particle-associated communities, rela- tively small changes in microbial dynamics can rescue these communities, allowing them to rapidly reach exponential growth and generating large changes in POC ﬂux. For example, while a given particle type (e.g., size and lability) might persist below 1000 m in one region of the ocean due to low initial biomass or low microbial encounter rates, that same particle might be consumed rapidly in a different region where microbial encounter rates are higher (Fig. 3). Our model provides an alternative hypothesis for pulsed export of carbon from oligotrophic regions77–79. In addi- tion to a shift in particle size associated with these export eventse.g.,77,80,81, the microbial community in oligotrophic regions is typically associated with picoplankton dominated communities and so may not be adapted to consume particles generated by larger phytoplankton groups. These differences in microbial dynamics, represented as lower biomass at formation and lower encounter rates in our model, will yield greater transfer efﬁciency of carbon to depth than would occur if the same type and quantity of particles were released in more productive regions where the microbial community is primed to consume POC generated by larger phytoplankton groups. This work suggests that shifts in microbial communities both in surface waters and at depth can result in signiﬁcantly dif- ferent POC ﬂuxes and that the magnitude of microbial driven variations in the POC ﬂux is similar to other previously pro- posed processes (e.g., particle size spectrum). This is not to say that other processes do not also impact the POC ﬂux. Dynamics not explicitly represented in the model such as particle aggre- gation and disaggregation32,35,39, zooplankton grazing on particles30,36,37,40,82, phytoplankton dynamics83–85, and the formation of new particles within the water column83–85 also play an important role. An exciting avenue of future work is to investigate the extent to which complex ecological interactions between the microbial communities and zooplankton dynamics impact the POC ﬂux, the relative importance of these different processes for determining the rate of organic carbon export, and how these dynamics may vary geographically. Our model provides a unique framework for understanding the carbon cycle impact of zooplankton bactivory on particle-associated communities42, and the direct consumption of sinking particles NATURE COMMUNICATIONS \| (2022) 13:1657 \| https://doi.org/10.1038/s41467-022-29297-2 \| www.nature.com/naturecommunications 5 ARTICLE NATURE COMMUNICATIONS \| https://doi.org/10.1038/s41467-022-29297-2 a) Size spectra b) Lability e) Attenuation ) m ( h t p e D 500 1000 1500 2000 ) m ( h t p e D 500 1000 1500 2000 500 1000 1500 2000 s=-4 s=-3 s=-2 1.5 =50 =200 =500 avg avg avg 0 0.5 1 1.5 0 0.5 1 c) Max growth d) Encounter 500 1000 1500 2000 High Enc Low Enc Martin 1987 0 0.5 1 1.5 =7.2 d -1 =1.2 d -1 max max 0 0.5 1 1.5 POC flux (mol C m -2 year-1) Global UVP Our model Avg. b 1.5 1 0.5 y c n e u q e r f d e z i l a m r o N 0 0 1 2 Attenuation exponent (b) Fig. 4 Microbial contribution to large-scale particulate organic carbon (POC) ﬂuxes. Shifts in POC ﬂuxes over depth are shown as a function of varying particle and microbial dynamics. Each gray line is the integrated ﬂux over 2,256 to 763,615 stochastically initialized particles (depending on the particle size spectra). For each parameter set, 1,000 stochastic simulations were conducted (1,000 gray lines of different shades). The average for each parameter set is shown with the thick colored line. The data from Martin et al (1987) (in open circles) are also shown. POC transfer efﬁciency decreases (larger attenuation −1 day−1) (b), higher exponent b) with more negative particle size spectra power law exponent s (a), higher mean particle lability (β maximal growth rate (μmax, day−1) (c), and higher particle-microbe encounter rate (Enc) (d). The distribution of attenuation exponent (b) for all 10,000 stochastic simulations from panels a-d (gray bars) are compared against observed attenuation coefﬁcients from 897 global Underwater Video Proﬁler (UVP) measurements compiled by Guidi et al. (2015) (open bars) (e). avg, mmol Cpoc mmol Ccell by zooplankton (e.g., exploring preferences for different types of POC and particle sizes). Direct inclusion of dynamics such as zooplankton disaggregation of particles and zooplankton vertical migration would allow for further mechanistic con- straints on spatial and temporal variability of organic carbon export. This work takes a ﬁrst step towards explicitly integrating micro- scale dynamics into large-scale models to generate predictions of organic carbon ﬂuxes in the ocean. We show that microbial growth dynamics can generate temporal and spatial variability in POC consumption rates, suggesting that current parametrizations for POC degradation are inadequate (Supplementary Mate- rial 1–3). Our model generates hypotheses as to the relative importance of particle-associated microbial dynamics throughout the water column that can be tested by targeted ﬁeld and laboratory studies. These studies will in turn improve the model parameterizations and generate more robust estimates of the POC vertical ﬂux. For example, incubation experiments could test whether a signiﬁcant fraction of particle-associated populations at depth are below the critical density threshold. Many aspects of particle-associated microbial dynamics are currently poorly con- strained (e.g., encounter rates, bacterial growth efﬁciencies). Our results highlight the need for better in situ measurements of these key biological processes such as loss processes for microbial communities, microbial abundance on particles, enzyme activities, growth rates on particles, and encounter and detachment rates for dominant particle-associated marine species. Ultimately, robust predictions of future shifts in carbon cycle dynamics require accurate, mechanistic representation of the primary processes in global climate models. Here we demonstrate that particle- associated microbial dynamics are one of these processes. Methods This model captures key micro-scale dynamics occurring on particulate organic carbon (POC) in a manner scalable to the water column. For the results presented in the main text, we represent particle-associated microbial diversity using 18 groups of heterotrophic microbes, deﬁned based on their enzyme kinetics and growth rates. Supplementary Material 2 presents a sensitivity test with a continuum of microbial classes and demonstrates that this discretization does not impact our results. We also make the simplifying assumption that each particle type consists of a single lability of organic carbon colonized by a single type of microbial primary degrader, though the conclusions of this work are not dependent on this assumption (Supplementary Fig. 2 and Supplementary Material 2). Here we track each unique particle type i, which is deﬁned based on radius at formation and lability (β i). We include enzymatic degradation of polymer into low molecular weight organic matter (Clmwom), density dependent growth of the particle- associated microbial community (Bi), and the attachment (Ei;z) and detachment (Li) of heterotrophic microbes to/from the particles. This model can be coupled to a full ecosystem model such that the generation of each particle type can be calculated prognostically. However, here we focus on the degradation of POC below the export depth and so simply include a source term to represent net particle formation above the export depth (default 100 m, see Supplementary Material 2 for simulations with alternative formation depths). For simplicity, we also do not allow for aggregation or disaggregation to occur within the water column. An extended model description is provided in Supplementary Material 1. The change in the carbon content of particle i (Cpoc;i, mmol CPOC particle−1) over time is deﬁned as: iBi ¼ (cid:2)β i (mmol CPOC mmol Ccell dCpoc;i dt −1 day−1) represents the polymer degradation where β rate of Cpoc;i by microbial group Bi (mmol Ccell particle−1) similar to59. Speciﬁcally, β i captures differences in ‘lability’ of particles, which is a function the organic carbon itself, the microbial enzymes speciﬁc to group Bi, and production rate of = 0), the particle- those enzymes by Bi. When the particle is fully consumed (Cpoc associated microbial community detaches and so consumption stops. ð1Þ The enzymatic degradation of POC results in the production of low molecular weight organic matter (LMWOM) (Clmwom, mmol C m−3 particle−1) which sup- ports microbial growth. We assume that there is no loss of carbon during the enzymatic cleavage from POC to LMWOM such that 1 mmol Cpoc degraded = 1 6 NATURE COMMUNICATIONS \| (2022) 13:1657 \| https://doi.org/10.1038/s41467-022-29297-2 \| www.nature.com/naturecommunications NATURE COMMUNICATIONS \| https://doi.org/10.1038/s41467-022-29297-2 ARTICLE mmol Clmwom produced. There is however diffusive loss of LMWOM away from the particle as described in Eq. 2. Speciﬁcally, the LMWOM concentration is cal- culated assuming steady state dynamics as: Clmwom (cid:4) ¼ β (cid:5) iBi (cid:2) μ iBi ylmwom =dloss ð2Þ i is the growth rate of microbial group Bi on the particle (day−1), ylmwom is −1), and dloss is where μ the aerobic microbial growth efﬁciency (mmol Ccell mmol Clmwom the diffusion loss rate of the LMWOM (m3 day−1) (Supplementary Material 1 eq. s3 and Supplementary Material 4 for full calculation). Microbial dynamics on each particle are deﬁned as: (cid:6) ¼ μ i dBi dt (cid:7) (cid:2) Li (cid:2) mlin;i Bi þ Ei;z ð3Þ where mlin;i is the microbial mortality rate (day−1) and Li is the detachment rate (day−1). The microbial encounter rate (Ei;z, mmol Ccell day−1 particle−1) repre- sents the rate of colonization of the particle by the free-living microbial pool. Ei;zvaries with depth based on particle size, sinking speed, and the abundance of free-living, motile microbes in group i51,64 (eq. s11 in Supplementary Material 1). The free-living abundance is assumed to decrease exponentially with depth64. Microbial growth rate, μ the particle surface and is represented with the Monod equation: i (day−1) is dependent on the LMWOM concentration at μ i ¼ V max;i Clmwom Clmwom þ km;i γ T;z ð4Þ ; i;z ω Fpoc;z where V max;i, km;i, and γ T;z represent the maximum LMWOM uptake rate, the half saturation of LMWOM uptake by microbial group Bi, and the temperature lim- itation at depth z (eq. s11 in Supplementary Material 1), respectively. Model parameter values are given in Supplementary Table 1 and sensitivity tests are described in Supplementary Material 2. The total POC ﬂux at a certain depth z (Fpoc z in mmol CPOC m−2 day−1), is calculated as the sum of the vertical ﬂuxes of each individual particle as they sink ; through the water column, where Fpoc z is: ¼ ∑Cpoc;i;zN i;z ð5Þ where N i;z is the number of particles of type i per m3 water column at depth z. To test the impact of particle-associated microbial dynamics on the POC vertical ﬂux through the water column, we perform a set of stochastic simulations in which para- meters are randomly chosen from within a reasonable range. Here we simulate a single 2.24 m ´ 2.24 m water column initialized between 50 and 100 m with an observed particle size distribution yielding a total ﬂux of 1.5 mol m−2 yr−1 1,5,13. Speciﬁcally, we simulate 2,256 particles (particle size spectra with power law exponent s = −2), 69,000 particles (s = −3), or 763,615 particles (s = −4) ranging from 50 μm to 0.4 cm in radius as they fall through the water column. For each particle, the following model para- meters are stochastically assigned within a reasonable range (Supplementary Table 1) using a uniform distribution, except for lability for which a log-normal distribution is used: maximum growth rate V max (1.2 or 7.2 day−1), particle lability β (10–1,000 mmol −1 day−1), initial cell density (400–2,800 cell mm−2), and density of CPOC mmol Ccell free-living community of microbes in group i (Fi) (10–285 cell mm−3). The stochastic simulations are conducted 1,000 times for each particle size distribution. Simulations are run for 600 days which is sufﬁcient for all particles to be fully consumed or exported to >4,000 m. The attenuation exponent b for the modeled POC ﬂux is calculated using a least square ﬁt of the power law function. Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The microbial growth data generated in this study are provided in the Supplementary Table S2. Code availability The model code is deposited on GitHub (https://github.com/LevineLab/POMmodel) and citable using https://doi.org/10.5281/zenodo.6015020. Received: 29 June 2021; Accepted: 2 March 2022; References 1. Martin, J. H., Knauer, G. A., Karl, D. M. & Broenkow, W. W. VERTEX: carbon cycling in the northeast Paciﬁc. Deep Sea Res. Part A. Oceanographic Res. Pap. 34, 267–285 (1987). 2. Gloege, L., McKinley, G. A., Mouw, C. B. & Ciochetto, A. B. Global evaluation of particulate organic carbon ﬂux parameterizations and implications for atmospheric pCO2. Glob. Biogeochemical Cycles 31, 1192–1215 (2017). 3. Guidi, L. et al. A new look at ocean carbon remineralization for estimating deepwater sequestration. Glob. Biogeochemical Cycles 29, 1044–1059 (2015). 4. Marsay, C. M. et al. Attenuation of sinking particulate organic carbon ﬂux through the mesopelagic ocean. PNAS 112, 1089–1094 (2015). 5. Omand, M. M., Govindarajan, R., He, J. & Mahadevan, A. Sinking ﬂux of particulate organic matter in the oceans: Sensitivity to particle characteristics. Sci. Rep. 10, 5582 (2020). 6. Aumont, O. et al. Variable reactivity of particulate organic matter in a global ocean biogeochemical model. Biogeosciences 14, 2321–2341 (2017). 7. DeVries, T., Liang, J.-H. & Deutsch, C. A mechanistic particle ﬂux model applied to the oceanic phosphorus cycle. Biogeosciences 11, 5381–5398 (2014). 8. DeVries, T. & Weber, T. The export and fate of organic matter in the ocean: new constraints from combining satellite and oceanographic tracer observations. Glob. Biogeochemical Cycles 31, 535–555 (2017). 9. Kriest, I. & Oschlies, A. On the treatment of particulate organic matter sinking in large-scale models of marine biogeochemical cycles. Biogeosciences 5, 55–72 (2008). 10. Lutz, M., Dunbar, R. & Caldeira, K. Regional variability in the vertical ﬂux of particulate organic carbon in the ocean interior. Glob. Biogeochemical Cycles 16, 11-1–11-18 (2002). 11. Pavia, F. J. et al. Shallow particulate organic carbon regeneration in the South Paciﬁc Ocean. PNAS 116, 9753–9758 (2019). 12. Weber, T., Cram, J. A., Leung, S. W., DeVries, T. & Deutsch, C. Deep ocean nutrients imply large latitudinal variation in particle transfer efﬁciency. Proc. Natl Acad. Sci. USA 113, 8606–8611 (2016). 13. Cael, B. B. & Bisson, K. Particle ﬂux parameterizations: quantitative and mechanistic similarities and differences. Front. Mar. Sci. 5, (2018). 14. Cael, B. B. & White, A. E. Sinking versus suspended particle size distributions in the North Paciﬁc Subtropical Gyre. Geophys. Res. Lett. 47, e2020GL087825 (2020). 15. Lam, P. J., Doney, S. C. & Bishop, J. K. B. The dynamic ocean biological pump: Insights from a global compilation of particulate organic carbon, CaCO3, and opal concentration proﬁles from the mesopelagic. Global Biogeochemical Cycles 25, (2011). 16. Cram, J. A. et al. The role of particle size, ballast, temperature, and oxygen in the sinking ﬂux to the deep sea. Glob. Biogeochemical Cycles 32, 858–876 (2018). 17. Boyd, P. W., Claustre, H., Levy, M., Siegel, D. A. & Weber, T. Multi-faceted particle pumps drive carbon sequestration in the ocean. Nature 568, 327–335 (2019). 18. Boeuf, D. et al. Biological composition and microbial dynamics of sinking particulate organic matter at abyssal depths in the oligotrophic open ocean. PNAS 116, 11824–11832 (2019). 19. Grabowski, E., Letelier, R. M., Laws, E. A. & Karl, D. M. Coupling carbon and energy ﬂuxes in the North Paciﬁc Subtropical Gyre. Nat. Commun. 10, 1895 (2019). 20. Karl, D. M., Knauer, G. A. & Martin, J. H. Downward ﬂux of particulate organic matter in the ocean: a particle decomposition paradox. Nature 332, 438–441 (1988). 21. Karl, D. M., Church, M. J., Dore, J. E., Letelier, R. M. & Mahaffey, C. Predictable and efﬁcient carbon sequestration in the North Paciﬁc Ocean supported by symbiotic nitrogen ﬁxation. PNAS 109, 1842–1849 (2012). 22. Church, M. J. et al. Production and diversity of microorganisms associated with sinking particles in the subtropical North Paciﬁc Ocean. Limnol. Oceanogr. 66, 3255–3270 (2021). 23. Briggs, N., Dall’Olmo, G. & Claustre, H. Major role of particle fragmentation in regulating biological sequestration of CO2 by the oceans. Science 367, 791–793 (2020). 24. Cho, B. C. & Azam, F. Major role of bacteria in biogeochemical ﬂuxes in the ocean’s interior. Nature 332, 441–443 (1988). 25. Giering, S. L. C. et al. Reconciliation of the carbon budget in the ocean’s twilight zone. Nature 507, 480–483 (2014). 26. Bianchi, D., Weber, T. S., Kiko, R. & Deutsch, C. Global niche of marine anaerobic metabolisms expanded by particle microenvironments. Nat. Geosci. 11, 263–268 (2018). 27. Cavan, E. L., Henson, S. A. & Boyd, P. W. The sensitivity of subsurface microbes to ocean warming accentuates future declines in particulate carbon export. Front. Ecol. Evol. 6, (2019). 28. McDonnell, A. M. P. & Buesseler, K. O. Variability in the average sinking velocity of marine particles. Limnol. Oceanogr. 55, 2085–2096 (2010). 29. Bendtsen, J., Hilligsøe, K. M., Hansen, J. L. S. & Richardson, K. Analysis of remineralisation, lability, temperature sensitivity and structural composition of organic matter from the upper ocean. Prog. Oceanogr. 130, 125–145 (2015). 30. Steinberg, D. K. et al. Bacterial vs. zooplankton control of sinking particle ﬂux in the ocean’s twilight zone. Limnol. Oceanogr. 53, 1327–1338 (2008). 31. Alcolombri, U. et al. Sinking enhances the degradation of organic particles by marine bacteria. Nat. Geosci. 1–6 https://doi.org/10.1038/s41561-021-00817-x (2021). NATURE COMMUNICATIONS \| (2022) 13:1657 \| https://doi.org/10.1038/s41467-022-29297-2 \| www.nature.com/naturecommunications 7 ARTICLE NATURE COMMUNICATIONS \| https://doi.org/10.1038/s41467-022-29297-2 32. Biddanda, B. & Pomeroy, L. Microbial aggregation and degradation of 60. phytoplankton-derived detritus in seawater. I. Microbial succession. Mar. Ecol. Prog. Ser. 42, 79–88 (1988). 33. Dilling, L. & Alldredge, A. L. Fragmentation of marine snow by swimming macrozooplankton: a new process impacting carbon cycling in the sea. Deep Sea Res. Part I: Oceanographic Res. 47, 1227–1245 (2000). 61. 34. Buesseler, K. O. & Boyd, P. W. Shedding light on processes that control particle export and ﬂux attenuation in the twilight zone of the open ocean. Limnol. Oceanogr. 54, 1210–1232 (2009). 35. Burd, A. B. & Jackson, G. A. Particle aggregation. Annu. Rev. Mar. Sci. 1, 65–90 (2009). 36. Romero‐Romero, S. et al. Deep zooplankton rely on small particles when particle ﬂuxes are low. Limnol. Oceanogr. Lett. 5, 410–416 (2020). 37. Maas, A. E. et al. Migratory zooplankton excreta and its inﬂuence on Iversen, M. H., Nowald, N., Ploug, H., Jackson, G. A. & Fischer, G. High resolution proﬁles of vertical particulate organic matter export off Cape Blanc, Mauritania: Degradation processes and ballasting effects. Deep Sea Res. Part I: Oceanographic Res. Pap. 57, 771–784 (2010). Ilyina, T. et al. Global ocean biogeochemistry model HAMOCC: Model architecture and performance as component of the MPI-Earth system model in different CMIP5 experimental realizations. J. Adv. Modeling Earth Syst. 5, 287–315 (2013). 62. Garber, J. H. Laboratory study of nitrogen and phosphorus remineralization during the decomposition of coastal plankton and seston. Estuar., Coast. Shelf Sci. 18, 685–702 (1984). 63. Zakem, E. J., Cael, B. B. & Levine, N. M. A uniﬁed theory for organic matter accumulation. PNAS https://doi.org/10.1101/2020.09.25.314021 (2021). 64. Sunagawa, S. et al. Structure and function of the global ocean microbiome. prokaryotic communities. Front. Mar. Sci. 0, (2020). Science 348, (2015). 38. Möller, K. O. et al. Marine snow, zooplankton and thin layers: indications of a 65. Alldredge, A. The carbon, nitrogen and mass content of marine snow as a trophic link from small-scale sampling with the Video Plankton Recorder. Mar. Ecol. Prog. Ser. 468, 57–69 (2012). 39. Karakaş, G. et al. Impact of particle aggregation on vertical ﬂuxes of organic matter. Prog. Oceanogr. 83, 331–341 (2009). function of aggregate size. Deep Sea Res. Part I: Oceanographic Res. Pap. 45, 529–541 (1998). 66. Zakem, E. J. et al. Ecological control of nitrite in the upper ocean. Nat. Commun. 9, 1206 (2018). 40. Cavan, E. L., Trimmer, M., Shelley, F. & Sanders, R. Remineralization of particulate organic carbon in an ocean oxygen minimum zone. Nat. Commun. 8, 1–9 (2017). 67. Boyd, P. W. et al. Transformations of biogenic particulates from the pelagic to the deep ocean realm. Deep Sea Res. Part II: Topical Stud. Oceanogr. 46, 2761–2792 (1999). 41. Datta, M. S., Sliwerska, E., Gore, J., Polz, M. F. & Cordero, O. X. Microbial interactions lead to rapid micro-scale successions on model marine particles. Nat. Commun. 7, 11965 (2016). 68. Schmidt, S., Chou, L. & Hall, I. R. Particle residence times in surface waters over the north-western Iberian Margin: comparison of pre-upwelling and winter periods. J. Mar. Syst. 32, 3–11 (2002). 42. Kiørboe, T., Tang, K., Grossart, H.-P. & Ploug, H. Dynamics of microbial communities on marine snow aggregates: colonization, growth, detachment, and grazing mortality of attached bacteria. Appl. Environ. Microbiol. 69, 3036–3047 (2003). 43. Grossart, H.-P., Kiørboe, T., Tang, K. & Ploug, H. Bacterial colonization of particles: growth and interactions. Appl Environ. Microbiol 69, 3500–3509 (2003). 44. Enke, T. N., Leventhal, G. E., Metzger, M., Saavedra, J. T. & Cordero, O. X. Microscale ecology regulates particulate organic matter turnover in model marine microbial communities. Nat. Commun. 9, 2743 (2018). 45. Kirchman, D. L. Growth Rates of Microbes in the Oceans. Annu. Rev. Mar. Sci. 8, 285–309 (2016). 46. Ebrahimi, A., Schwartzman, J. & Cordero, O. X. Cooperation and spatial self- organization determine rate and efﬁciency of particulate organic matter degradation in marine bacteria. PNAS https://doi.org/10.1073/ pnas.1908512116 (2019). 47. Agusti, S. et al. Ubiquitous healthy diatoms in the deep sea conﬁrm deep carbon injection by the biological pump. Nat. Commun. 6, 1–8 (2015). 48. Tamburini, C. et al. Effects of hydrostatic pressure on microbial alteration of sinking fecal pellets. Deep Sea Res. Part II: Topical Stud. Oceanogr. 56, 1533–1546 (2009). 49. Tamburini, C., Garcin, J., Ragot, M. & Bianchi, A. Biopolymer hydrolysis and bacterial production under ambient hydrostatic pressure through a 2000 m water column in the NW Mediterranean. Deep Sea Res. Part II Topical Stud. Oceanogr. 49, 2109–2123 (2002). 50. Tamburini, C., Boutrif, M., Garel, M., Colwell, R. R. & Deming, J. W. Prokaryotic responses to hydrostatic pressure in the ocean – a review. Environ. Microbiol. 15, 1262–1274 (2013). 51. Lambert, B. S., Fernandez, V. I. & Stocker, R. Motility drives bacterial encounter with particles responsible for carbon export throughout the ocean. Limnol. Oceanogr. Lett. 4, 113–118 (2019). 52. Ploug, H. & Grossart, H.-P. Bacterial growth and grazing on diatom aggregates: respiratory carbon turnover as a function of aggregate size and sinking velocity. Limnol. Oceanogr. 45, 1467–1475 (2000). 53. Enke, T. N. et al. Modular assembly of polysaccharide-degrading marine microbial communities. Curr. Biol. 29, 1528–1535.e6 (2019). 54. Kaul, R. B., Kramer, A. M., Dobbs, F. C. & Drake, J. M. Experimental demonstration of an Allee effect in microbial populations. Biol. Lett. 12, 20160070 (2016). 55. Kiørboe, T., Ploug, H. & Thygesen, U. H. Fluid motion and solute distribution around sinking aggregates I: Small-scale ﬂuxes and heterogeneity of nutrients in the pelagic environment. Mar. Ecol. - Prog. Ser. 211, 1–13 (2001). 56. Kiørboe, T. & Jackson, G. A. Marine snow, organic solute plumes, and optimal chemosensory behavior of bacteria. Limnol. Oceanogr. 46, 1309–1318 (2001). 57. Baumas, C. M. J. et al. Mesopelagic microbial carbon production correlates with diversity across different marine particle fractions. The ISME Journal 1–14 https://doi.org/10.1038/s41396-020-00880-z (2021). 58. Mestre, M. et al. Spatial variability of marine bacterial and archaeal communities along the particulate matter continuum. Mol. Ecol. 26, 6827–6840 (2017). 59. Mislan, K. A. S., Stock, C. A., Dunne, J. P. & Sarmiento, J. L. Group behavior among model bacteria inﬂuences particulate carbon remineralization depths. J. Mar. Res. 72, 183–218(36) (2014). 69. Schmidt, M. W. I. et al. Persistence of soil organic matter as an ecosystem property. Nature 478, 49–56 (2011). 70. Lehmann, J. et al. Persistence of soil organic carbon caused by functional complexity. Nat. Geosci. 13, 529–534 (2020). 71. Dittmar, T. et al. Enigmatic persistence of dissolved organic matter in the ocean. Nat. Rev. Earth Environ. 2, 570–583 (2021). 72. Poff, K. E., Leu, A. O., Eppley, J. M., Karl, D. M. & DeLong, E. F. Microbial dynamics of elevated carbon ﬂux in the open ocean’s abyss. PNAS 118, (2021). 73. Pelve, E. A., Fontanez, K. M. & DeLong, E. F. Bacterial Succession on Sinking Particles in the Ocean’s Interior. Front. Microbiol. 8, (2017). 74. Boscolo-Galazzo, F. et al. Temperature controls carbon cycling and biological evolution in the ocean twilight zone. Science 371, 1148–1152 (2021). 75. Riley, J. S. et al. The relative contribution of fast and slow sinking particles to ocean carbon export. Global Biogeochemical Cycles 26, (2012). 76. Guidi, L. et al. Plankton networks driving carbon export in the oligotrophic ocean. Nature 532, 465–470 (2016). 77. Steinberg, D. K. et al. Overview of the US JGOFS Bermuda Atlantic Time-series Study (BATS): a decade-scale look at ocean biology and biogeochemistry. Deep Sea Res. Part II: Topical Stud. Oceanogr. 48, 1405–1447 (2001). 78. Conte, M. H., Dickey, T. D., Weber, J. C., Johnson, R. J. & Knap, A. H. Transient physical forcing of pulsed export of bioreactive material to the deep Sargasso Sea. Deep Sea Res. Part I: Oceanographic Res. Pap. 50, 1157–1187 (2003). 79. Smith, K. L., Ruhl, H. A., Huffard, C. L., Messié, M. & Kahru, M. Episodic organic carbon ﬂuxes from surface ocean to abyssal depths during long-term monitoring in NE Paciﬁc. Proc. Natl Acad. Sci. USA 115, 12235–12240 (2018). 80. Alkire, M. B. et al. Estimates of net community production and export using high-resolution, Lagrangian measurements of O2, NO3−, and POC through the evolution of a spring diatom bloom in the North Atlantic. Deep Sea Res. Part I: Oceanographic Res. Pap. 64, 157–174 (2012). 81. Briggs, N. et al. High-resolution observations of aggregate ﬂux during a sub- polar North Atlantic spring bloom. Deep Sea Res. Part I: Oceanographic Res. Pap. 58, 1031–1039 (2011). 82. Talmy, D. et al. An empirical model of carbon ﬂow through marine viruses and microzooplankton grazers. Environ. Microbiol. 21, 2171–2181 (2019). 83. Kostadinov, T. S., Milutinović, S., Marinov, I. & Cabré, A. Carbon-based 84. phytoplankton size classes retrieved via ocean color estimates of the particle size distribution. Ocean Sci. 12, 561–575 (2016). Jin, X., Gruber, N., Dunne, J. P., Sarmiento, J. L. & Armstrong, R. A. Diagnosing the contribution of phytoplankton functional groups to the production and export of particulate organic carbon, CaCO3, and opal from global nutrient and alkalinity distributions. Global Biogeochemical Cycles 20, (2006). 85. Mouw, C. B., Barnett, A., McKinley, G. A., Gloege, L. & Pilcher, D. Phytoplankton size impact on export ﬂux in the global ocean. Glob. Biogeochemical Cycles 30, 1542–1562 (2016). Acknowledgements This work was supported by a grant from the Simons Foundation (542387 to NML, 542395 to RS and OC, 542389 to TH). We thank N. Norris and E. Lee for assistance with model development. 8 NATURE COMMUNICATIONS \| (2022) 13:1657 \| https://doi.org/10.1038/s41467-022-29297-2 \| www.nature.com/naturecommunications NATURE COMMUNICATIONS \| https://doi.org/10.1038/s41467-022-29297-2 ARTICLE Author contributions TTHN and NML designed the study, developed the full model, and conducted the numerical analysis; TC and TH developed and analyzed the mathematical model in Supplementary Material 4; KA and TC performed measurements of bacterial parameters for Vibrio sp. 1A01; EJZ, OXC, AE, JS, UA, FJP, TH, and RS contributed data and to model development; all authors contributed to the writing. Competing interests The authors declare no competing interests. Additional information Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-022-29297-2. Correspondence and requests for materials should be addressed to Naomi M. Levine. Peer review information Nature Communications thanks Jacob Cram, Philip Boyd and the other, anonymous, reviewer for their contribution to the peer review of this work. Peer reviewer reports are available. Reprints and permission information is available at http://www.nature.com/reprints Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional afﬁliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/. © The Author(s) 2022 NATURE COMMUNICATIONS \| (2022) 13:1657 \| https://doi.org/10.1038/s41467-022-29297-2 \| www.nature.com/naturecommunications 9	null
10.1073_pnas.1913292117.pdf	Data Availability Statement. All data discussed in the paper are available to readers through Harvard Dataverse, https://doi.org/10.7910/DVN/RFGT4S.	Data Availability Statement. All data discussed in the paper are available to readers through Harvard Dataverse, https://doi.org/10.7910/DVN/RFGT4S .	B cells migrate into remote brain areas and support neurogenesis and functional recovery after focal stroke in mice Sterling B. Ortegaa,b,c,1, Vanessa O. Torresb,c,1, Sarah E. Latchneyd,e, Cody W. Whooleryd, Ibrahim Z. Noorbhaib,c, Katie Poinsatteb,c, Uma M. Selvarajb,c, Monica A. Bensonb,c, Anouk J. M. Meeuwissenb,c, Erik J. Plautzb,c, Xiangmei Kongb,c, Denise M. Ramirezb,c, Apoorva D. Ajayb,c, Julian P. Meeksb,c,f, Mark P. Goldbergb,c, Nancy L. Monsonb,c, Amelia J. Eischd,g,h, and Ann M. Stoweb,c,i,2 aDepartment of Pathology, University of Iowa, Iowa City, IA 52242; bDepartment of Neurology and Neurotherapeutics, UT Southwestern Medical Center, Dallas, TX 75390; cPeter O’Donnell Jr. Brain Institute, UT Southwestern Medical Center, Dallas, TX 75390; dDepartment of Psychiatry, UT Southwestern Medical Center, Dallas, TX 75390; eDepartment of Biology, St. Mary’s College of Maryland, St. Mary’s City, MD 20686; fDepartment of Neuroscience, UT Southwestern Medical Center, Dallas, TX 75390; gDepartment of Neuroscience, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA 19104; hDepartment of Anesthesiology and Critical Care, Children’s Hospital of Philadelphia, Philadelphia, PA 19104; and iDepartment of Neurology, University of Kentucky, Lexington, KY 40506 Edited by Lawrence Steinman, Stanford University School of Medicine, Stanford, CA, and approved January 16, 2020 (received for review August 1, 2019) Lymphocytes infiltrate the stroke core and penumbra and often exacerbate cellular injury. B cells, however, are lymphocytes that do not contribute to acute pathology but can support recovery. B cell adoptive transfer to mice reduced infarct volumes 3 and 7 d after transient middle cerebral artery occlusion (tMCAo), in- dependent of changing immune populations in recipient mice. Testing a direct neurotrophic effect, B cells cocultured with mixed cortical cells protected neurons and maintained dendritic arbori- zation after oxygen-glucose deprivation. Whole-brain volumetric serial two-photon tomography (STPT) and a custom-developed image analysis pipeline visualized and quantified poststroke B cell diapedesis throughout the brain, including remote areas support- ing functional recovery. Stroke induced significant bilateral B cell diapedesis into remote brain regions regulating motor and cogni- tive functions and neurogenesis (e.g., dentate gyrus, hypothalamus, olfactory areas, cerebellum) in the whole-brain datasets. To confirm a mechanistic role for B cells in functional recovery, rituximab was given to human CD20+ (hCD20+) transgenic mice to continuously deplete hCD20+-expressing B cells following tMCAo. These mice experienced delayed motor recovery, impaired spatial memory, and increased anxiety through 8 wk poststroke compared to wild type (WT) littermates also receiving rituximab. B cell depletion re- duced stroke-induced hippocampal neurogenesis and cell survival. Thus, B cell diapedesis occurred in areas remote to the infarct that mediated motor and cognitive recovery. Understanding the role of B cells in neuronal health and disease-based plasticity is critical for developing effective immune-based therapies for protection against diseases that involve recruitment of peripheral immune cells into the injured brain. B lymphocytes \| focal stroke \| serial two-photon tomography \| adaptive immunity \| neurogenesis Stroke leads to central nervous system (CNS) damage, which results in functional deficits (1) and is exacerbated by an inflammatory immune response derived from both the innate and adaptive immune systems (2). Mechanistic studies using murine stroke models show a significant infiltration of innate immune cells, including monocytes, macrophages, and neutrophils, pre- dominantly in the area of ischemic injury (i.e., infarct, periinfarct regions) (3). The role of the adaptive immune system is also pivotal to stroke recovery, as it can both exacerbate and amelio- rate long-term neuropathology, depending on the lymphocyte population, location, and timing of activation (4–10). Location and timing are particularly relevant, as recovery of lost function in stroke patients depends on functional plasticity in areas outside of the infarct (i.e., remote cortices) to subsume lost function (11). Neurons in remote cortical areas that are interconnected to the infarct up-regulate growth factors and plasticity-related genes af- ter stroke (12, 13), but it is unknown if neuroinflammation is concomitant with remote genetic and proteomic responses to is- chemic injury and subsequent reorganization. B cells, critical effector cells for antibody production and an- tigen presentation, are one adaptive immune cell subset with the capacity to also produce neurotrophins to support neuronal survival and plasticity (14–16). Prior work found that the loss of B cells increases stroke-induced infarct volumes, functional deficits, and mortality in immunodeficient μMT mice, which lack B cells (17). Polyclonal activated and nonactivated B cell −/− E C N E I C S O R U E N Significance Neuroinflammation occurs immediately after stroke onset in the ischemic infarct, but whether neuroinflammation occurs in remote regions supporting plasticity and functional recovery remains unknown. We used advanced imaging to quantify whole-brain diapedesis of B cells, an immune cell capable of producing neurotrophins. We identify bilateral B cell diapede- sis into remote regions, outside of the injury, that support motor and cognitive recovery in young male mice. Poststroke depletion of B cells confirms a positive role in neurogenesis, neuronal survival, and recovery of motor coordination, spatial learning, and anxiety. More than 80% of stroke survivors have long-term disability uniquely affected by age and lifestyle factors. Thus, identifying beneficial neuroinflammation during long-term recovery increases the opportunity of therapeutic interventions to support functional recovery. Author contributions: S.B.O., S.E.L., C.W.W., K.P., M.P.G., N.L.M., A.J.E., and A.M.S. de- signed research; S.B.O., V.O.T., S.E.L., C.W.W., I.Z.N., K.P., U.M.S., M.A.B., A.J.M.M., E.J.P., X.K., D.M.R., A.D.A., J.P.M., M.P.G., A.J.E., and A.M.S. performed research; N.L.M. contrib- uted new reagents/analytic tools; S.B.O., V.O.T., S.E.L., C.W.W., D.M.R., A.D.A., J.P.M., and A.M.S. analyzed data; and S.B.O., V.O.T., S.E.L., C.W.W., N.L.M., A.J.E., and A.M.S. wrote the paper. The authors declare no competing interest. This article is a PNAS Direct Submission. This open access article is distributed under Creative Commons Attribution-NonCommercial- NoDerivatives License 4.0 (CC BY-NC-ND). Data deposition: All raw data have been uploaded to Harvard Dataverse, https://doi.org/ 10.7910/DVN/RFGT4S. 1S.B.O. and V.O.T. contributed equally to this work. 2To whom correspondence may be addressed. Email: [email protected]. This article contains supporting information online at https://www.pnas.org/lookup/suppl/ doi:10.1073/pnas.1913292117/-/DCSupplemental. First published February 12, 2020. www.pnas.org/cgi/doi/10.1073/pnas.1913292117 PNAS \| March 3, 2020 \| vol. 117 \| no. 9 \| 4983–4993 −/− transfer into syngeneic mice also ameliorates stroke pathology and functional deficits within the first 96 h after stroke via in- terleukin (IL)-10–mediated up-regulation (9, 10). Reconstitution of Rag (i.e., B- and T cell-knockout) mice with B cells does not exacerbate infarct volume, confirming a nonpathological role in acute stroke (18, 19). Others confirmed a nonpathological role for B cells after stroke in several other strains (20). However, additional studies do not extend the neuroprotective and/or nonpathological role beyond 4 d after stroke onset, with the only investigation into the long-term role of B cells in stroke recovery finding that B cells contribute to cognitive decline weeks after stroke in multiple strains of mice across various stroke models (21). These data, coupled with human data associating self- reactive antibodies with cognitive decline (8), underscore the importance of additional studies to clarify the contribution of B cells to poststroke injury and repair. This includes the impor- tance of timing, location, and function of B cells recruited into the injured brain in defining the role in either injury or repair. The goal of this series of experiments is to elucidate the en- dogenous role of B cells, including localization of B cell diape- desis after stroke using advanced volumetric imaging strategies (e.g., serial two-photon tomography), and the behavioral rele- vance of B cells using refined neurological motor and cognitive assessments. In this work, we show that poststroke B cell re- pletion reduces neuropathology via an IL-10–dependent mech- anism that is independent of immune modulation. Second, we show that B cells exert a direct neuroprotective effect on neurons and preserve neuronal dendritic arborization following oxygen- glucose deprivation in vitro. We find that B cells diapedese bi- laterally into remote areas of the brain that associate with long- term motor and cognitive function following a transient middle cerebral artery occlusion (tMCAo), while long-term depletion of B cells after stroke exacerbates motor and cognitive deficits and reduces poststroke neurogenesis. In summary, we show that B cells play a prominent role in stroke recovery beyond the first 4 d after stroke by directly protecting neurons and promoting recovery within the injured and reorganizing brain. Results B Cells Use IL-10 to Reduce Stroke-Induced Neuropathology. Pre- viously published studies found that B cells exhibit conflicting roles in stroke recovery dependent on the state of activation: acute in vitro activation benefits (i.e., 1 to 4 d) versus long-term in vivo activation detriments (i.e., months) (10, 21). Further- more, the use of antibody-mediated B cell depletion in wild type and transgenic mice revealed no change in the neuropathology or functional deficits at the acute phase of stroke (days 1 and 3 post induction) (20). Thus, to further dissect the role of B cells in subacute to long-term stroke recovery, we investigated the en- dogenous function of B cells by adoptive transfer of naïve, wild type B cells into stroke-induced recipient mice, measuring neuropathology and motor coordination at time points up to 2 wk following stroke induction. We purified unstimulated B cells from naïve mice and injected them i.v. into recipient mice 7 h after stroke onset (60-min tMCAo; Fig. 1A). On all mice, suc- cessful tMCAo induction was qualified as a 20% reduction for 60 min followed by reperfusion to at least 50% of baseline (pretMCAo induction) cerebral blood flow. We measured motor function using rotarod (days 2, 6, and 14 poststroke), forearm strength (force grip), neuropathology via T2-weighted MRI (days 3 and 7), and cytotoxicity via standard histology (day 15). Rela- tive to PBS controls, B cell recipients exhibited reduced infarc- tions identified by serial MRI at both 3 d (P < 0.01) and 7 d (P < 0.05; Fig. 1 B and C) post tMCAo. Interestingly, the rotarod and force grip assays did not detect differences between the two cohorts, even though all groups initiated acute deficits con- firming similar degree of initial stroke-induced neuropathology (SI Appendix, Fig. S1 A and B). We sought to determine if naïve B cells, without in vitro activation, caused reduced infarct vol- umes via an IL-10–mediated mechanism. Adoptive transfer of wild type B cells from naïve mice reduced infarct volume, and this protection was lost with IL-10–deficient B cells (Fig. 1D). As protection was absent in both IL-10–deficient (Fig. 1D) and “preconditioned” (22) B cells (SI Appendix, Fig. S1 B and C), these data show that naïve B cells provide an endogenous neuroprotective function (22). In contrast to other studies (9), Fig. 1. B cell adoptive transfer reduces infarct volumes via IL-10 but independently of immunomodulation. (A) Experimental timeline for B cell adoptive transfer 7 h after tMCAo. (B) Infarct volumes quantified from (C) serial MR images show that naïve B cells (green upward triangles) significantly decrease infarct volumes 3 and 7 d poststroke as compared to PBS control (black circles). (D) B cells isolated from wild type (WT) B6 mice (solid green triangles, white bar) and IL-10 KO mice (open green triangles, gray bar) show that only WT B cells reduced infarct volumes 3 d after tMCAo compared to PBS controls (black T cells in the spleen, nor did they (F) induce a circles; determined by one-way ANOVA). (E) WT B cell treatment did not differentially activate CD4 regulatory T cell population in the spleen. Significance determined by two-way repeated-measure ANOVA or Student’s t test (P < 0.05, P < 0.01 vs. PBS control unless indicated by bracket). CD25 + + 4984 \| www.pnas.org/cgi/doi/10.1073/pnas.1913292117 Ortega et al. + CD25 protection by naïve B cells was indicated neither by peripheral + T cells (Fig. 1E), nor by suppression of activated CD4 the induction of splenic regulatory T cell population (Fig. 1F). Together with prior studies showing no induction of regula- tory T cells in the CNS by B cell adoptive transfer (10), our data suggest a previously unreported and distinct mechanism employed by the adoptive transfer of naive B cells during stroke recovery. B Cells Directly Promote Neuronal Viability and Dendritic Arborization In Vitro. One regulatory immune cell-independent mechanism of poststroke neuroprotection that could be induced by B cells is a direct neurotrophic effect of B cells on neurons at risk for cell death (14). Using an in vitro approach to determine a direct role in neuronal protection, mixed cortical cells were subjected to 2 h of oxygen-glucose deprivation (OGD) followed immedi- ately by addition of naïve B cells to the culture for 4 d (Fig. 2 A and B). In the mixed cortical culture condition without B cells, + OGD caused loss of microtubule-associated protein (MAP)2 neurons (P < 0.001) and loss of arborization (P < 0.01; Fig. 2 C + and D). Addition of naïve B cells to cortical cells (0.1:1.0 ratio) reduced both neuronal death (P < 0.05) and loss of dendritic arborization (P < 0.05). In the absence of ischemic injury, the addition of naïve B cells increased the number of surviving neurons (P < 0.05; Fig. 2C) and increased dendritic MAP2 arborization (P < 0.05, 0.1:1.0 ratio; P < 0.01, 1:1; Fig. 2D) compared to cultures without B cells. Replication of these ex- periments with naïve IL-10–deficient B cells showed no dose– response effect on neuronal cell loss either with or without OGD (Fig. 2E). In contrast, IL-10–deficient B cells still preserved dendritic arborization after OGD, albeit at a higher concen- tration (P < 0.05; Fig. 2F and SI Appendix, Fig. S2). Increased dendritic arborization also occurred in the absence of OGD with IL-10–deficient B cells (P < 0.05), suggesting that IL-10 is a re- dundant mechanism that supports maintenance of mature neu- rons with dendrites. In summary, these data confirm the capability of naïve B cells for direct neuronal protection in vitro by an IL-10 mechanism within the context of neuronal ischemic injury. E C N E I C S O R U E N Fig. 2. B cells induce a neurotrophic effect in mixed cortical cultures. (A) Experimental timeline for the placement of naïve B cells on culture after oxygen- + glucose deprivation (OGD) and prior to immunocytochemistry (ICC). (B) Images at 10× of microtubule-associated protein (MAP2) neurons (red) for treatment (txn) groups; higher-magnification image is also shown (Inset). (Scale bar, 50 μm.) (C) Increasing ratio of B cells to mixed cortical (MC) cells increased the number of MAP2+ neurons, as well as (D) the number of MAP2+ neurons with dendrites, in the non-OGD experiments. OGD decreased overall numbers, with only a 0.1:1.0 B cell:mixed cortical cell ratio preserving neuronal survival and dendritic arborization. (E and F) Experiments were replicated with IL-10–deficient B cells (images shown in SI Appendix, Fig. S2). IL-10–deficient B cells (E) did not preserve cell survival after OGD, but (F) a 1:1 ratio did preserve dendritic arborization after OGD. Replicate, independent experiments (each with interexperimental replicates) are shown by red circles for total n. Nonparametric one- way ANOVA (Kruskal–Wallis) determined within-group significance, and t test was used for untreated non-OGD:OGD comparisons. Data graphed as mean ± SD, and significance determined by nonparametric one-way ANOVA or Student’s t test (P < 0.05, P < 0.01, P < 0.001 vs. untreated control unless otherwise indicated by brackets). Ortega et al. PNAS \| March 3, 2020 \| vol. 117 \| no. 9 \| 4985 B Cells Are Present within the Parenchyma of the Poststroke Brain. At the whole-brain level, B cells are an abundant leukocyte pop- ulation within the stroke-injured brain 48 h after stroke onset (22, 23). In fact, the cortical and subcortical brain vascular en- dothelium exhibits an up-regulation of CXCL13, a B cell homing chemokine (22). Recent advances in volumetric whole-brain imaging methods including serial two-photon tomography (STPT) have enabled automated, unbiased methods to visualize and computationally detect signals of interest, including fluorescently labeled cell populations as well as neuronal substructures in the entire brain (24–26). We established a custom pipeline including STPT, supervised machine learning-based pixel classification, and image registration to visualize and quantify adoptively transferred immune cells labeled with the fluorophore e450 throughout the whole mouse brain. We verified this methodology to accurately + T cell diapedesis into the whole brain after quantify labeled CD8 tMCAo in Poinsatte et al., 2019 (27). The output of our machine learning-based pixel classification step is visualized as a probability map of pixels automatically detected by the trained algorithm as B cells (Fig. 3A) (28). The brighter the area in the probability map, the more likely the trained algorithm identified the fluorescence as a B cell. To quantify whole brain neuroinflammation, we used two mice or B cell donor littermate controls, with all cohorts of mice: (i) B cell-depleted recipient hCD20 wild type (WT) hCD20 + − + − + recipient mice receiving rituximab prior to tMCAo to target B cells for depletion (29) (Fig. 3B); and (ii) B cell- hCD20 sufficient C57BL/J6 mice with unaltered B cell populations. In order to account for endogenous poststroke autofluorescence secondary to ischemic injury and cell death, we included PBS + recipient cohorts. STPT detected signal from fluorescent e450 donor hCD20 B cells in both the ipsilesional and contralesional hemispheres (Fig. 3C) of B cell-depleted recipient mice (P < 0.05 for both hemispheres vs. PBS; Fig. 3D). We also observed bi- lateral e450 B cell signal in the B cell-sufficient cohort, albeit with larger group variation (Fig. 3E). B cell recipients in both B cells identified by flow cohorts of mice exhibited e450 cytometry in both the cervical lymph nodes and spleen 4 d after tMCAo (SI Appendix, Fig. S3), confirming the presence of live B cells following adoptive transfer. The surprising increase of B cells in the contralesional hemisphere indicates that B cell di- apedesis is not only localized to infarct and periinfarct areas with high cytokine and chemokine up-regulation, but there is also an active recruitment from the periphery in brain regions remote from the ischemic core. + Adoptively Transferred B Cells Localize to CNS Motor Nuclei and Areas of Neurogenesis. We previously observed B cell infiltration into the brain following stroke (22, 30). Now, understanding that B cells could exert in vitro neuroprotection (Fig. 2), we wanted to Fig. 3. Establishing whole-brain quantification of neuroinflammation using STPT. (A) Representative coronal section image acquired via STPT shown as a multichannel raw fluorescence merged image. (Magnification: Left, 16×.) (Inset) e450 B cells present in the hippocampus (blue). (Right) Corresponding 2D probability map from the same coronal section in which the trained machine-learning algorithm classified pixels likely to represent brain parenchyma (gray) and e450+ B cells (magenta). A 3D rendering of the B cell and parenchyma probability maps for each section in the whole brain are shown (Right). (B) Timeline for B cell depletion immediately prior to a transient middle cerebral artery occlusion (tMCAo), with adoptive transfer (AT) of e450+ B cells. (C) Whole-brain images show a representative brain from the PBS- (Left) and B cell-treated (Right) cohorts. The probability maps of e450+ B cells are shown in green and fluorescence overlaid onto the average template from the atlas shown in grayscale. Ipsilesional (ipsi) hemisphere is identified. (D) Whole-hemisphere e450 (in pixels per cubic millimeter) is shown for hCD20 -depleted mice that received PBS (black squares) or B cells (red squares). Dashed lines show mean PBS fluorescence that is used for normalization. (E) Similar to D, showing fluorescence in C57BL/6J mice that received PBS or B cells but without depletion of endogenous B cells. Significance determined by two-way ANOVA (P < 0.05 vs. respective PBS-treated control). + + + 4986 \| www.pnas.org/cgi/doi/10.1073/pnas.1913292117 Ortega et al. determine if B cells were infiltrating bilateral brain areas favor- able to poststroke functional recovery. To achieve specific lo- calization of neuroinflammation, raw fluorescent STPT images of every coronal section were registered with the corresponding sections in the CCFv3.0 atlas (Fig. 4A) in order to assess cellular + B cells across brain regions of interest (Fig. 4B). Quantifi- e450 cation of e450-labeled B cells in whole-brain STPT datasets revealed brain regions with significant B cell diapedesis, including ipsilesional cerebral cortex, olfactory areas, and hypothalamus (Table 1). Contralesional areas with significant B cell diapedesis included olfactory areas, dentate gyrus, and hypothalamus, creat- ing a bilateral diapedesis of B cells into brain regions typically spared in tMCAo-induced injury and that mediate motor and cognitive recovery, as well as active neurogenesis. To determine if endogenous B cells have an effect on adoptively transferred B + cells, we also adoptively transferred naive WT e450 B cells into + donor B poststroke WT mice that were not B cell-depleted. e450 cells were again localized to the spleen and CLN after stroke (SI Appendix, Fig. S3E). STPT detected e450 B cells throughout + B cell signal was not the poststroke brain, even though e450 significantly elevated over PBS controls due to more within-group fluorescent variation. + B cell diapedesis was elevated in five brain regions associated with motor function (i.e., cerebral cortex, cortical subplate, midbrain [sensory-related], cerebellum, and substantia nigra [reticular]) in B cell-depleted mice, which are highlighted in 3D surface renderings created by the Allen Institute for Brain Science Brain Explorer application (Fig. 4C). All areas demonstrated a + B cell signal (100- to 200-fold higher similar magnitude of e450 than PBS fluorescence), with ipsilesional signals significant in E C N E I C S O R U E N Fig. 4. B cells show diapedesis into remote motor areas after stroke. (A) Serial two-photon tomography images are coregistered to the Allen Institute CCF3.0 + to allow identification of region-specific fluorescence in 3D whole-brain probability maps. (Magnification: A, 16×.) (B) Representative areas classified as e450 B cells via machine learning (punctate magenta dots [Insets], identified by white arrows) that show B cell diapedesis in cortex and hippocampus (enlarged areas indicated by white squares). (C) Three-dimensional surface renderings created with the Brain Explorer application (Allen Institute for Brain Science) show regions with bilateral B cell diapedesis after stroke, with areas labeled to the letter corresponding to the data graphs. (D–H) Quantification of e450+ B cell fluorescence in both B cell-depleted mice (red squares) and mice with endogenous B cells at time of stroke (black circles). Ipsilesional (ipsi, white bars) and contralesional (contra, hatched bars) brain regions show predominantly ipsilesional B cell diapedesis for (D) cerebral cortex, (E) cortical subplate, (F) midbrain, (G) cerebellum, and (H) substantia nigra with percent change (y-axes, pixels per cubic millimeter). Significance determined by paired t tests (P < 0.05, P < 0.01 vs. PBS controls unless indicated by brackets). Ortega et al. PNAS \| March 3, 2020 \| vol. 117 \| no. 9 \| 4987 Table 1. STPT data for e450 + pixels for each brain region Brain region Gray matter Cerebral cortex Olfactory areas Dentate gyrus Cerebral nuclei Striatum Striatum ventral region Cortical subplate Hypothalamus Midbrain, sensory related Cerebellum White matter Corpus callosum Ipsilesional, mm3 Contralesional, mm3 Allen Brain Atlas abbreviation PBS control (n = 3) B cells (n = 4) P value PBS control (n = 3) B cells (n = 4) P value Laterality index CTX OLF DG CNU STR STRv CTXsp HY MBsen CB CC 8.6 8.05 26.55 4.25 3.88 5.26 2.5 11.85 28.09 10.32 1.29 22.44 33.5 67.45 10.62 9.27 13.52 5.46 32.21 75.47 24.14 0.025 0.019* 0.073 0.061 0.06 0.06 0.066 0.039* 0.056 0.061 11.64 10.1 30.2 8.09 8.16 12.18 6.45 12.35 35.38 12.96 26.62 32.46 55 18.78 18.34 21.09 9.15 36.1 85.27 22.32 0.155 0.041* 0.028* 0.202 0.227 0.17 0.514 0.020* 0.136 0.128 4.41 0.313 1.98 6.44 0.242 0.9 1.1 1.2 0.7 0.6 0.6 0.8 0.9 0.9 1.1 0.8 STPT, serial two-photon tomography. Significance between PBS and B cell recipients per hemisphere was analyzed by unpaired, parametric. Student’s t test: P < 0.05; bolded text, P ≤ 0.06. + cerebral cortex (P < 0.05). Ipsilesional B cell diapedesis was also significantly higher compared to the contralesional hemisphere in the cerebellum (P < 0.05; Fig. 4G). In contrast, B cell-sufficient mice exhibited higher laterality for ipsilesional diapedesis, with B cell signal localized to the ipsilesional cortical more e450 subplate (P < 0.01; Fig. 4E) and midbrain (P < 0.05; Fig. 4F). Interestingly, one motor region, the substantia nigra pars retic- ulata, had much higher B cell diapedesis (1,000- to 2,000-fold vs. PBS) in the mice with endogenous B cells versus B cell-depleted mice (Fig. 4H), with highest values in the contralesional hemi- sphere. In summary, these data show that, in a mouse devoid of B cells, adoptively transferred B cells migrate to the motor areas to a similar magnitude in both the ipsi- and contralesional hemi- spheres, possibly exerting a supportive role in functional recovery. B Cells Contribute to Long-Term Motor Function Recovery. As several motor areas remote from the tMCAo-mediated infarction exhibi- ted B cell diapedesis, we hypothesized that long-term depletion of poststroke B cells could impede motor recovery. We tested this mice or WT with prestroke depletion of B cells in hCD20 − littermate controls receiving rituximab, with continuous hCD20 depletion for 2 wk after tMCAo (Fig. 5A and SI Appendix, Fig. S4A). While an effect of B cell depletion on infarct volumes between cohorts (Fig. 5B) was observed, there was no geno- type effect on motor skill acquisition during rotarod training (SI Appendix, Fig. S4B). WT mice exhibited a motor deficit at 2 d after tMCAo (P < 0.001; Fig. 5C) that recovered by 4 d relative to prestroke baseline. In contrast, B cell-depleted mice exhibited a significant loss in poststroke motor function at 2 d (P < 0.01; + Fig. 5. B cells support motor recovery after stroke. (A) Experimental timeline for motor testing after transient middle cerebral artery occlusion (tMCAo). Rituximab treatment in wild type (WT; black circles) and B cell-depleted (red squares) mice show (B) infarct volumes that were similar between cohorts. Cresyl violet images for animals closest to the mean are shown. (Magnification: B, 20×, whole brain images.) (C) WT mice exhibited a motor deficit on a rotarod 2 d poststroke compared to prestroke baseline (horizontal gray bar) that (D) correlated to infarct volume. (E) B cell-depleted mice exhibited a loss in motor performance 2, 4, and 14 d poststroke compared to prestroke baseline (red horizontal bar), an observation (F) independent of infarct volume. Significance determined by Student’s t test, one-way repeated-measure ANOVA, or linear correlation (P < 0.05, P < 0.01, P < 0.001 vs. prestroke baseline). Dotted lines indicate 95% confidence interval. 4988 \| www.pnas.org/cgi/doi/10.1073/pnas.1913292117 Ortega et al. Fig. 5E), 4 d (P < 0.01), and 14 d (P < 0.05) relative to prestroke baseline. There were no significant between-group differences. These data suggest that the absence of B cells after stroke could potentially impede plasticity in the motor network(s), located outside of the area of infarction, that support recovery of motor coordination. B cell depletion, however, did not affect the muscle strength, as analyzed by force grip analysis (SI Appendix, Fig. S4C). Infarct volumes only in WT mice negatively correlated with motor performance (Fig. 5D), with larger infarct volumes associated with poorer motor performance but lost with B cell depletion (Fig. 5F). With the observed bilateral B cell migration into brain regions associated with motor function, combined with the behavioral deficits in B cell-depleted mice, these data suggest that the pres- ence of B cells may affect remote brain regions required for motor recovery (11). Poststroke B Cell Depletion Increases General Anxiety and Spatial B Memory Deficits. Initially, we observed significant e450 cell signal in several brain regions (e.g., olfactory areas, dentate gyrus, substantia nigra, and hypothalamus) also asso- ciated with cognitive function (Fig. 6A) (31, 32). These areas, with the exception of the substantia nigra, show high bilateral e450 B cell signal. In the olfactory areas and the hypothalamus, + + + diapedesis of adoptively transferred B cells in B cell-depleted mice was significant (P < 0.05) for both the ipsi- and con- tralesional hemispheres (both P < 0.05 vs. PBS; Fig. 6 B and C). The dentate gyrus is unique in that e450 B cell signal was sig- nificant in the contralesional hemisphere (P < 0.05; Fig. 6D). Both the olfactory areas and the hypothalamus also exhibit in- creased signal in ipsi- vs. contralesional hemispheres, though the latter, surprisingly, is in the B cell-sufficient cohort. Thus, we sought to determine any associated poststroke cognitive deficits (e.g., locomotor activity, anxiety, learning, and memory) with B cell depletion extended through 8 wk after tMCAo (Fig. 6E). Interestingly, while rituximab maintained B cell depletion for 8 cohorts (SI Appendix, wk in both uninjured and tMCAo hCD20 Fig. S5A), WT poststroke mice also exhibited reductions in splenic B cell representation compared to uninjured WT mice (P < 0.05). + As with the 2-wk motor recovery cohort shown in Fig. 5, there was no effect of long-term B cell depletion on infarct volumes (Fig. 6F). For the cognitive tests, we first measured lo- comotor activity using open field and found that the total distance traveled decreased between day 1 and day 2 for both poststroke WT (P < 0.01) and B cell-depleted (P < 0.05) cohorts (Fig. 6G). E C N E I C S O R U E N Fig. 6. B cell depletion increases general anxiety and spatial memory deficits independent of infarct volume. (A) Brain Explorer surface rendering shows brain regions with B cell diapedesis related to cognitive function, including (B) olfactory areas, (C ) hypothalamus, and (D) dentate gyrus, as indicated by letter in the figure. Both B cell-depleted mice (red squares) and mice with endogenous B cells at time of stroke (black circles) show B cell diapedesis in ipsilesional (ipsi, white bars) and contralesional (contra, hatched bars) brain regions. Fold change (y-axes) and significance vs. PBS controls shown unless indicated by brackets. (E) Experimental timeline for behavior testing in uninjured mice (white bars) or after transient middle cerebral artery occlusion (tMCAo; gray bars). (F ) B cell depletion did not affect infarct volumes or (G) total movement in the open field. (H) Both uninjured and poststroke B cell- depleted mice spent more time in the periphery (“P”) of the open field compared to the center (“C”) of the field. (I) On the first day of cued and contextual fear-conditioning, poststroke B cell-depleted mice exhibited prolonged freezing during the training tones/shocks. (J) The next day, these same mice exhibited a worse spatial memory. (K) On the last day, B cell-depleted mice again exhibited more freezing during cued memory. Data graphed as mean ± SD. Significance determined by paired t test, Student’s t test, or repeated-measures ANOVA (P < 0.05, *P < 0.01 vs. nondepleted control or D1 unless indicated by brackets). Ortega et al. PNAS \| March 3, 2020 \| vol. 117 \| no. 9 \| 4989 Interestingly, though all cohorts traveled similar distances, B cell-depleted mice, regardless of injury, exhibited a preference to travel in the periphery of the open field compared to WT uninjured (P < 0.05) or stroke-injured mice (P < 0.01; Fig. 6H). Subsequent nonaversive cognitive assessment by novel object and novel location (33) failed to show differences between cohorts based on either genotype or stroke (SI Appendix, Fig. S5 C–F). Finally, we assessed hippocampal-dependent and -independent forms of learning and memory using a contextual and cued fear-conditioning paradigm, respectively. During the training phase, there was a significant effect of genotype for poststroke mice (P < 0.001), as B cell-depleted mice froze for longer durations during the second, third (both P < 0.05; Fig. 6I), and final (P < 0.05; SI Appendix, Fig. S5G) tones. The next day, evaluation of contextual, hippocampal-dependent memory showed that B cell-depleted poststroke mice had worse recall compared to WT mice (P < 0.05; Fig. 6J). In contrast, hippocampal- independent (i.e., amygdala-mediated) cued memory recall was exaggerated in B cell-depleted mice (P < 0.01; Fig. 6K). These data suggest that long-term B cell depletion independently in- fluences both hippocampal- and amygdala-dependent cognitive functions after stroke. B Cell Support of Poststroke Neurogenesis. Adult neurogenesis occurs in the subventricular zone (SVZ) and the subgranular zone (SGZ) of the dentate gyrus, within the hippocampus (34). Stroke is also a potent inducer of neurogenesis (35, 36), al- though the mechanism is still poorly understood. Interestingly, of the 40 brain regions assessed for B cell migration in the poststroke brain, the two associated with neurogenesis were significant: olfactory areas, including the lateral olfactory tract, and the dentate gyrus (Fig. 6 B and D). In fact, a greater mag- nitude of e450 B cell signal over PBS controls occurred in the dentate gyrus of mice with endogenous B cell populations, again with an almost equivalent bilateral signal similar to the B cell- depleted cohort. These data demonstrate that B cells exhibit sig- nificant migration patterns into areas associated with neurogenesis yet outside of the ischemic injury. We thus quantified hippocampal + neurogenesis 2 wk after tMCAo. Doublecortin-expressing (DCX ) cells are immature neuroblasts generated in the inner granular cell + layer (GCL) of the hippocampal dentate gyrus (37), and DCX cell bodies are an index of neurogenesis. Two weeks of B cell depletion + + cell numbers (Fig. 7B) and GCL volume (SI did not affect DCX Appendix, Fig. S6A) in otherwise healthy mice. Two weeks after + tMCAo, however, WT mice had more DCX neuroblasts in the dentate gyrus of the ipsilesional hemisphere compared to the contralesional hemisphere (P < 0.01; Fig. 7C), consistent with other studies (35, 38). In contrast, the B cell-depleted mice did not exhibit increased poststroke neurogenesis (P < 0.001 vs. WT). + Changes (or lack thereof) in DCX cell number were not asso- ciated with changes in GCL volume (Fig. 7D). In combination with the in vivo data (Fig. 6), these data suggest that B cells contribute to both stroke-induced neurogenesis and poststroke cognitive recovery, although removal of B cells does not affect basal (i.e., homeostatic) neurogenesis. Next, we quantified hip- pocampal cell survival via bromodeoxyuridine (BrdU) injection prior to the first rituximab treatment in the cognitive cohorts shown in Fig. 6. BrdU birthdates new neurons, and we determined long-term (i.e., 8 wk) cell survival in the hippocampus following continued poststroke B cell depletion. Basal cell survival in the uninjured mice was unaffected by long-term rituximab treatment + (SI Appendix, Fig. S6B), similar to the DCX population data. + Stroke induced an increase in the ipsilateral BrdU neurons of adult-generated hippocampal cells in the ischemic hemisphere of WT mice 8 wk after tMCAo (P < 0.05 vs. contralesional hemi- sphere), but not B cell-depleted mice (Fig. 7 E and F). As all mice received rituximab to control for off-target drug effects, this con- firms a role for B cells in sustaining neuronal cell survival in areas outside of the ischemic injury. Discussion Under physiological conditions, neuroplasticity occurs through- out life in cortical, subcortical, and cerebellar brain regions [e.g., olfactory areas (34), motor cortex (39, 40), hippocampus (34, 41), and cerebellum (42, 43)]. These areas are responsible for brain development, neurogenesis, learning, and memory forma- tion. Following stroke, the spinal cord and brain, including the brain regions regulating motor and cognitive function, undergo significant neuroplasticity to reinnervate injured areas and re- organize neural networks supporting recovery (44–49). While neuroplasticity occurring within the ipsilesional hemisphere can improve functional outcome (46, 50–52), bihemispheric plasticity in the contralesional hemisphere also supports recovery (53–57). Fig. 7. B cells support poststroke neurogenesis. (A) Brain Explorer surface rendering shows neurogenic regions with bilateral B cell diapedesis after stroke. (B) cells) B cell depletion did not alter basal neurogenesis in the dentate gyrus in the absence of stroke. (C) Stroke induced neurogenesis (i.e., doublecortin [DCX] in the ipsilesional (ipsi) hemisphere of wild type (WT; black circles) mice but not B cell-depleted (red squares) mice (D) independent of changes in granule cell layer (GCL) volume. (E) Stroke increased bromodeoxyuridine (BrdU) cells in the dentate gyrus over contralesional (contra) numbers only in WT mice, (F) as shown by immunohistochemical detection of BrdU-positive cells, that was lost with long-term B cell depletion of healthy mice. Significance determined by one-way ANOVA (P < 0.05; P < 0.01; *P < 0.001 vs. ipsilesional hemisphere unless indicated by brackets). + + 4990 \| www.pnas.org/cgi/doi/10.1073/pnas.1913292117 Ortega et al. While our understanding of whole-brain neuroplasticity following stroke is ever-evolving, few studies investigate neuroinflammation in reorganizing brain regions. We previously found a high diape- desis of B cells into the poststroke brain, with a concomitant up- regulation of CXCL13 in cortical and subcortical brain areas (22, 23). Here, our advanced whole-brain imaging and analysis meth- ods highlight significant bilateral migration of B cells to select brain regions, including the aforementioned regions undergoing poststroke plasticity, that are outside of the area of infarction and thus distant to the epicenter of poststroke neuroinflammation. B cells are part of the adaptive immune system, and are re- sponsible for many immunological functions, including antibody production, modulation of T cell responses, and antigen pre- sentation (14). However, what is becoming increasingly apparent in animal studies of stroke is that not only the location, but also the timing and type, of B cell activation is critical in determining the contribution to stroke injury and/or repair (10, 21). B cells lack an acute (day 1 to 3 post stroke) physiological role in stroke pathology, indicating that their effector function is irrelevant to stroke etiology or innate immune modulation (20). B cells can produce several neurotrophins capable of supporting plasticity and neurogenesis, including brain-derived neurotrophic factor (BDNF) and nerve growth factor (NGF), in addition to IL-10 (15, 58, 59). Our in vitro data confirm a neurotrophic effect of B cells on the survival and health of neurons, as well as dendritic arborization, in mixed cortical cultures. Interestingly, treatment of mixed cortical cultures with IL-10–deficient B cells was not neuroprotective for total neuronal counts, but a 1:1 ratio of IL- 10–deficient B cells to mixed cortical cells still protected arbor- ization. This suggests that IL-10 is not the only neuroprotective effector produced by B cells following oxygen-glucose depriva- tion that should be addressed in future studies. The neurotrophic effect of B cells is also reflected in our long- term hippocampal neurogenesis and cell survival data, wherein increases in both cellular processes require the active presence of B cells after stroke. The adult brain has the capacity to induce neurogenesis within the SGZ of the dentate gyrus and the SVZ of the lateral ventricle (60). Ischemic insult significantly in- creases neurogenesis in the SGZ and SVZ (61), with several studies also suggesting that neuroblasts migrate from the SVZ and SGZ to areas of ischemia, with the potential to adapt a neuronal phenotype and integrate into existing circuitry (35, 62). Others found that Copaxone (i.e., glatiramer acetate), which promotes the development of regulatory T (63) and B (Breg) cells (64, 65), induced both ipsi- and contralesional neurogenesis in a murine model of stroke concomitant with improved recovery (66), even though a specific B cell response was not investigated. However, we did not see an effect of B cell depletion on basal levels of neurogenesis or neuronal survival in otherwise healthy young mice. Other forms of neuroinflammation, particularly the microglial response (67), can modulate induction of poststroke neurogenesis. However, microglia do not play a role in the age- related decline of basal levels of neurogenesis in rats (68). Thus, future studies should investigate whether a natural, age-related decline in B cell lymphopoiesis, and thus declining circulating B cells and their neurotrophic capacity (69), contributes to lower basal neurogenesis with age. Of course, our study does not exclude other neuronal regeneration mechanisms, and future studies could use additional methods such as electrophysiology in order to more fully elucidate the neurotrophic function of B cells. Our data suggest several future directions for study. While skilled motor rehabilitation improves motor function concomi- tant with neurogenesis (70), we used several additional cognitive tests to assess hippocampal-related recovery. Long-term B cell depletion increased anxiety with a concomitant decrease in con- textual fear memory, suggesting both hippocampal and amygdala- related deficits, though these deficits were only present under stressful testing conditions and not during nonaversive testing. Additional tests should augment our cognitive testing, especially in light of prior conflicting studies showing an active role of B cells in the development of poststroke cognitive decline in mice (21). We also previously identified that preconditioning down-regulated B cell antigen presentation and antibody production (22) and that poststroke B cell diapedesis in preconditioned mice correlated with fewer CD4 T cells and macrophages in the ipsilesional hemisphere (30). It was therefore understandably surprising to find that the adoptive transfer of this novel B cell subset did not reduce infarct volumes and, in fact, inhibited long-term motor recovery. B cell adoptive transfer into Rag mice did not show evidence of poststroke neuroprotection either, indicating that B cell neuroprotection may be dependent on multiple lymphocyte subset synergism (18). It must be noted that our study used mature adult mice, and further understanding would necessitate the in- corporation of female and aged (>18 mo) mice. This interaction with resident cells or, more intriguingly, the possibly detrimental inability of interaction from anergic B cells suggests a complicated poststroke immune response that potentially may be manipulated for a much longer window of time to induce neuroprotection. −/− Several clinical studies confirm a role for the B cell immune response in stroke, evident by the presence of immunoglobulins in the CSF of stroke patients (71–73) and the presence of anti- body deposits in the brain of individuals with poststroke dementia (21). However, whether these are epiphenomena or represent an active role of B cells in neuropathology and/or functional recovery remains to be determined. Recovery occurs for months following stroke onset in patients, with better recovery associated with an- atomical and functional plasticity (74, 75) and angiogenesis (76) in the ipsilesional hemisphere. Even though neuroplasticity occurs rapidly and robustly within an acute time window after stroke, neuroplasticity can last for years after stroke and support the long- term improvement of functional outcomes (11, 77). As stroke remains a leading cause of adult disability (1), it is vital to identify mechanisms that can contribute to neuroplasticity in neural networks. This includes new studies into long-term neu- roinflammatory mechanisms that could be either detrimental or beneficial to stroke recovery depending on the timing of activa- tion, the responding leukocyte subset, and now also the remote brain region(s) in which the responding immune cells migrate. Materials and Methods Mice. All experiments used male mice that were 2 to 4 mo old at the start of experimentation and maintained in accordance to NIH guidelines for the care and use of laboratory animals. UT Southwestern Medical Center approved all procedures according to AAALAC accreditation and current PHS Animal Wel- fare Assurance requirements. Transgenic human CD20+ expressing (hCD20+/−) mice were obtained from Mark Shlomchik (University of Pittsburgh, Pittsburgh, PA). C57BL/6J mice and B6.Cg-Tg(Thy1-YFP)16Jrs/J mice were purchased from Jackson Laboratory, and Swiss Webster mice were purchased from Harlan. Mice were group-housed in cages of 2 to 4 mice in standard animal housing with cob bedding, a 12/12-h light cycle with lights on at 6:00 AM, and food and water ad libitum. A total of 125 mice were used in experiments, with 84 un- dergoing tMCAo surgery. A total of 23 were excluded (73% success), including 6 for failure to meet blood flow criteria, 16 for death during or after surgery, and 1 for a lack of B cell depletion. E C N E I C S O R U E N −/− +/+ and hCD20 B Cell Depletion and Repletion. Rituximab (Micromedex) was given to (littermate controls) at a standard dose of 100 μg i.p. hCD20 for three consecutive days before tMCAo (29). Each mouse was administered a weekly additional dose of rituximab thereafter to target the turnover of B cells. Successful and sustained B cell depletion was confirmed by isolating lymphocytes from spleen 2 or 8 wk post tMCAo that were stained with a general antibody panel and quantified by flow cytometry (78–80). B cells were purified from the splenocyte population using magnetic bead sepa- ration (Stem Cell Technologies). Purity was verified by flow cytometry. A total of 5 × 106 B cells were transferred i.v. in 0.2 mL 0.1 M PBS. B cells for imaging were labeled with eFluor 450 (e450) proliferation dye (eBioscience 65–0842-85). Hemocytometer counts were not collected for one cohort of B Ortega et al. PNAS \| March 3, 2020 \| vol. 117 \| no. 9 \| 4991 cell-depleted mice in the long-term cognitive studies, so percent represen- tation of CD45 cells is shown SI Appendix, Fig. S4A. + Statistical Analysis. Power analysis based on previous results and published data determined the approximate number of animals given an expected 30% stroke-induced mortality rate (80). Statistical differences were analyzed using unpaired parametric two-sample Student’s t test, ratio paired Student’s t test, one-way repeated-measures ANOVA, or two-way ANOVA where appropriate and as indicated in the text (Graph Pad Prism 6.0). Values of P < 0.05 were considered significant. All experimenters were blinded to condition, and all animals were randomly assigned to group. All detailed protocols for tMCAo induction, imaging (Tissuecyte, MRI), histology, stereology, flow cytometry, motor and cognitive assays, and in vitro culture assays are available in SI Appendix, Methods. Data Availability Statement. All data discussed in the paper are available to readers through Harvard Dataverse, https://doi.org/10.7910/DVN/RFGT4S. 1. A. S. Go et al.; American Heart Association Statistics Committee and Stroke Statistics Subcommittee, Executive summary: Heart disease and stroke statistics–2014 update: A report from the American Heart Association. Circulation 129, 399–410 (2014). 2. Á. Chamorro et al., The immunology of acute stroke. Nat. Rev. Neurol. 8, 401–410 (2012). 3. M. Gelderblom et al., Temporal and spatial dynamics of cerebral immune cell accu- mulation in stroke. Stroke 40, 1849–1857 (2009). 4. J. T. Walsh et al., MHCII-independent CD4+ T cells protect injured CNS neurons via IL-4. J. Clin. Invest. 125, 2547 (2015). 5. J. Wang et al., Activated regulatory T cell regulates neural stem cell proliferation in the subventricular zone of normal and ischemic mouse brain through interleukin 10. Front. Cell. Neurosci. 9, 361 (2015). 6. A. Liesz et al., Inhibition of lymphocyte trafficking shields the brain against delete- rious neuroinflammation after stroke. Brain 134, 704–720 (2011). 7. T. Shichita et al., Pivotal role of cerebral interleukin-17-producing gammadeltaT cells in the delayed phase of ischemic brain injury. Nat. Med. 15, 946–950 (2009). 8. K. J. Becker, P. Tanzi, D. Zierath, M. S. Buckwalter, Antibodies to myelin basic protein are associated with cognitive decline after stroke. J. Neuroimmunol. 295-296, 9–11 (2016). 9. S. Bodhankar, Y. Chen, A. A. Vandenbark, S. J. Murphy, H. Offner, IL-10-producing B-cells limit CNS inflammation and infarct volume in experimental stroke. Metab. Brain Dis. 28, 375–386 (2013). 10. S. Bodhankar, Y. Chen, A. A. Vandenbark, S. J. Murphy, H. Offner, Treatment of ex- perimental stroke with IL-10-producing B-cells reduces infarct size and peripheral and CNS inflammation in wild-type B-cell-sufficient mice. Metab. Brain Dis. 29, 59–73 (2014). 11. N. Dancause, R. J. Nudo, Shaping plasticity to enhance recovery after injury. Prog. Brain Res. 192, 273–295 (2011). 12. E. T. Urban, 3rd et al., Gene expression changes of interconnected spared cortical neurons 7 days after ischemic infarct of the primary motor cortex in the rat. Mol. Cell. Biochem. 369, 267–286 (2012). 13. A. M. Stowe et al., VEGF protein associates to neurons in remote regions following cortical infarct. J. Cereb. Blood Flow Metab. 27, 76–85 (2007). 14. U. M. Selvaraj, K. Poinsatte, V. Torres, S. B. Ortega, A. M. Stowe, Heterogeneity of B cell functions in stroke-related risk, prevention, injury, and repair. Neurotherapeutics 13, 729–747 (2016). 15. A. L. Fauchais et al., Role of endogenous brain-derived neurotrophic factor and sortilin in B cell survival. J. Immunol. 181, 3027–3038 (2008). 16. R. Tabakman, S. Lecht, S. Sephanova, H. Arien-Zakay, P. Lazarovici, Interactions between the cells of the immune and nervous system: Neurotrophins as neuroprotection medi- ators in CNS injury. Prog. Brain Res. 146, 387–401 (2004). 17. X. Ren et al., Regulatory B cells limit CNS inflammation and neurologic deficits in murine experimental stroke. J. Neurosci. 31, 8556–8563 (2011). 18. G. Yilmaz, T. V. Arumugam, K. Y. Stokes, D. N. Granger, Role of T lymphocytes and interferon-gamma in ischemic stroke. Circulation 113, 2105–2112 (2006). 19. C. Kleinschnitz et al., Early detrimental T-cell effects in experimental cerebral ischemia are neither related to adaptive immunity nor thrombus formation. Blood 115, 3835– 3842 (2010). 20. M. K. Schuhmann, F. Langhauser, P. Kraft, C. Kleinschnitz, B cells do not have a major pathophysiologic role in acute ischemic stroke in mice. J. Neuroinflammation 14, 112 (2017). 21. K. P. Doyle et al., B-lymphocyte-mediated delayed cognitive impairment following stroke. J. Neurosci. 35, 2133–2145 (2015). 22. N. L. Monson et al., Repetitive hypoxic preconditioning induces an immunosuppressed B cell phenotype during endogenous protection from stroke. J. Neuroinflammation 11, 22 (2014). 23. U. M. Selvaraj et al., Preconditioning-induced CXCL12 upregulation minimizes leu- kocyte infiltration after stroke in ischemia-tolerant mice. J. Cereb. Blood Flow Metab. 37, 801–813 (2017). ACKNOWLEDGMENTS. The authors acknowledge Drs. Abhijit Budge, Shanrong Zhang, and Shari Birnbaum for their technical expertise for imaging and behavioral analyses and data acquisition. The authors also acknowl- edge UT Southwestern’s Flow Cytometry Core, the Moody Foundation Flow Cytometry Facility at the Children’s Research Institution, the Advanced Imaging Research Center (Shared Instrumental Grant NIH1S10OD023552-01), Rodent Be- havior Core Facility, the Whole Brain Microscopy Facility, and the Neuro-Models Facility with Laura Ingle. This study was funded by grants to A.M.S. from the American Heart Association (14SDG18410020), NIH/NINDS (NS088555), the Dana Foundation David Mahoney Neuroimaging Program, and The Haggerty Center for Brain Injury and Repair (UTSW); to S.B.O. from the American Heart Associa- tion (14POST20480373) and NIH/NINDS (3R01NS088555-03S1); to V.O.T. from the NIH/NIAID (5T32AI005284-40) and NIH/NINDS (3R01NS088555-02S1); to U.M.S. from the American Heart Association (17PRE33660147); and to A.J.E. from the NIH (DA023701, DA023555, MH107945) and the US National Aeronautics and Space Administration (NNX15AE09G). S.E.L. and C.W.W. were funded by an NIH institutional training grant (DA007290, Basic Science Training Program in the Drug Abuse Research, Principal Investigator A.J.E.). The UT Southwestern Whole Brain Microscopy Facility and the Neuro-Models Facility are supported by the Texas Institute for Brain Injury and Repair (TIBIR). 25. Y. Kim et al., Mapping social behavior-induced brain activation at cellular resolution in the mouse. Cell Rep. 10, 292–305 (2015). 26. D. A. Vousden et al., Whole-brain mapping of behaviourally induced neural activa- tion in mice. Brain Struct. Funct. 220, 2043–2057 (2015). 27. K. Poinsatte et al., Visualization and quantification of post-stroke neural connectivity and neuroinflammation using serial two-photon tomography in the whole mouse brain. Front. Neurosci. 13, 1055 (2019). 28. D. M. O. Ramirez, D. A. Ajay, M. P. Goldberg, J. P. Meeks, Serial Multiphoton Tomogra- phy and Analysis of Volumetric Images of the Mouse Brain. Multiphoton Micros- copy, Neuromethods series (Springer Nature, 2019). 29. N. L. Monson et al., Rituximab therapy reduces organ-specific T cell responses and ameliorates experimental autoimmune encephalomyelitis. PLoS One 6, e17103 (2011). 30. U. M. Selvaraj et al., Preconditioning-induced CXCL12 upregulation minimizes leu- kocyte infiltration after stroke in ischemia-tolerant mice. J. Cereb. Blood Flow Metab. (Nihongoban) 37, 801–813 (2017). 31. E. R. Duval, A. Javanbakht, I. Liberzon, Neural circuits in anxiety and stress disorders: A focused review. Ther. Clin. Risk Manag. 11, 115–126 (2015). 32. L. M. Shin, I. Liberzon, The neurocircuitry of fear, stress, and anxiety disorders. Neu- ropsychopharmacology 35, 169–191 (2010). 33. K. O. Cho et al., Aberrant hippocampal neurogenesis contributes to epilepsy and associated cognitive decline. Nat. Commun. 6, 6606 (2015). 34. G. L. Ming, H. Song, Adult neurogenesis in the mammalian brain: Significant answers and significant questions. Neuron 70, 687–702 (2011). 35. W. Jiang, W. Gu, T. Brännström, R. Rosqvist, P. Wester, Cortical neurogenesis in adult rats after transient middle cerebral artery occlusion. Stroke 32, 1201–1207 (2001). 36. J. M. Parent, Z. S. Vexler, C. Gong, N. Derugin, D. M. Ferriero, Rat forebrain neuro- genesis and striatal neuron replacement after focal stroke. Ann. Neurol. 52, 802–813 (2002). 37. G. Kempermann, S. Jessberger, B. Steiner, G. Kronenberg, Milestones of neuronal development in the adult hippocampus. Trends Neurosci. 27, 447–452 (2004). 38. E. Gould, B. S. McEwen, P. Tanapat, L. A. Galea, E. Fuchs, Neurogenesis in the dentate gyrus of the adult tree shrew is regulated by psychosocial stress and NMDA receptor activation. J. Neurosci. 17, 2492–2498 (1997). 39. J. N. Sanes, J. P. Donoghue, Plasticity and primary motor cortex. Annu. Rev. Neurosci. 23, 393–415 (2000). 40. T. A. Clark, M. Fu, A. K. Dunn, Y. Zuo, T. A. Jones, Preferential stabilization of newly formed dendritic spines in motor cortex during manual skill learning predicts per- formance gains, but not memory endurance. Neurobiol. Learn. Mem. 152, 50–60 (2018). 41. Y. Gu, S. Janoschka, S. Ge, Neurogenesis and hippocampal plasticity in adult brain. Curr. Top. Behav. Neurosci. 15, 31–48 (2013). 42. J. Carrillo, S. Y. Cheng, K. W. Ko, T. A. Jones, H. Nishiyama, The long-term structural plasticity of cerebellar parallel fiber axons and its modulation by motor learning. J. Neurosci. 33, 8301–8307 (2013). 43. E. D’Angelo, The organization of plasticity in the cerebellar cortex: From synapses to control. Prog. Brain Res. 210, 31–58 (2014). 44. D. L. Adkins, J. E. Hsu, T. A. Jones, Motor cortical stimulation promotes synaptic plasticity and behavioral improvements following sensorimotor cortex lesions. Exp. Neurol. 212, 14–28 (2008). 45. L. I. Benowitz, S. T. Carmichael, Promoting axonal rewiring to improve outcome after stroke. Neurobiol. Dis. 37, 259–266 (2010). 46. J. Biernaskie, A. Szymanska, V. Windle, D. Corbett, Bi-hemispheric contribution to functional motor recovery of the affected forelimb following focal ischemic brain injury in rats. Eur. J. Neurosci. 21, 989–999 (2005). 47. J. D. Riley et al., Anatomy of stroke injury predicts gains from therapy. Stroke 42, 421– 426 (2011). 48. J. V. Perederiy, G. L. Westbrook, Structural plasticity in the dentate gyrus- revisiting a classic injury model. Front. Neural Circuits 7, 17 (2013). 24. T. Ragan et al., Serial two-photon tomography for automated ex vivo mouse brain 49. A. R. Luft et al., Brain activation of lower extremity movement in chronically impaired imaging. Nat. Methods 9, 255–258 (2012). stroke survivors. Neuroimage 26, 184–194 (2005). 4992 \| www.pnas.org/cgi/doi/10.1073/pnas.1913292117 Ortega et al. 50. M. L. Starkey et al., Back seat driving: Hindlimb corticospinal neurons assume forelimb control following ischaemic stroke. Brain 135, 3265–3281 (2012). 51. L. Wang, J. M. Conner, A. H. Nagahara, M. H. Tuszynski, Rehabilitation drives en- hancement of neuronal structure in functionally relevant neuronal subsets. Proc. Natl. Acad. Sci. U.S.A. 113, 2750–2755 (2016). 52. T. A. Jones, D. L. Adkins, Motor system reorganization after stroke: Stimulating and training toward perfection. Physiology (Bethesda) 30, 358–370 (2015). 53. N. T. Lindau et al., Rewiring of the corticospinal tract in the adult rat after unilateral stroke and anti-Nogo-A therapy. Brain 137, 739–756 (2014). 54. J. E. Hsu, T. A. Jones, Contralesional neural plasticity and functional changes in the less-affected forelimb after large and small cortical infarcts in rats. Exp. Neurol. 201, 479–494 (2006). 65. P. Putheti, M. Soderstrom, H. Link, Y. M. Huang, Effect of glatiramer acetate (Copaxone) on CD4+CD25high T regulatory cells and their IL-10 production in multiple sclerosis. J. Neuroimmunol. 144, 125–131 (2003). 66. Y. Cruz et al., Copolymer-1 promotes neurogenesis and improves functional recovery after acute ischemic stroke in rats. PLoS One 10, e0121854 (2015). 67. M. K. Tobin et al., Neurogenesis and inflammation after ischemic stroke: What is known and where we go from here. J. Cereb. Blood Flow Metab. 34, 1573–1584 (2014). 68. V. Darsalia, U. Heldmann, O. Lindvall, Z. Kokaia, Stroke-induced neurogenesis in aged brain. Stroke 36, 1790–1795 (2005). 69. Z. Keren et al., B-cell depletion reactivates B lymphopoiesis in the BM and rejuvenates the B lineage in aging. Blood 117, 3104–3112 (2011). 55. C. M. Lapash Daniels, K. L. Ayers, A. M. Finley, J. P. Culver, M. P. Goldberg, Axon sprouting in adult mouse spinal cord after motor cortex stroke. Neurosci. Lett. 450, 191–195 (2009). 70. F. Wurm, S. Keiner, A. Kunze, O. W. Witte, C. Redecker, Effects of skilled forelimb training on hippocampal neurogenesis and spatial learning after focal cortical infarcts in the adult rat brain. Stroke 38, 2833–2840 (2007). 56. Z. Liu, R. L. Zhang, Y. Li, Y. Cui, M. Chopp, Remodeling of the corticospinal innervation and spontaneous behavioral recovery after ischemic stroke in adult mice. Stroke 40, 2546–2551 (2009). 57. T. H. McNeill, S. A. Brown, E. Hogg, H. W. Cheng, C. K. Meshul, Synapse replacement in the striatum of the adult rat following unilateral cortex ablation. J. Comp. Neurol. 467, 32–43 (2003). 58. A. E. Edling, T. Nanavati, J. M. Johnson, V. K. Tuohy, Human and murine lymphocyte neurotrophin expression is confined to B cells. J. Neurosci. Res. 77, 709–717 (2004). 59. T. Kiyota et al., AAV serotype 2/1-mediated gene delivery of anti-inflammatory interleukin-10 enhances neurogenesis and cognitive function in APP+PS1 mice. Gene Ther. 19, 724–733 (2012). 71. H. Prüss et al., Evidence of intrathecal immunoglobulin synthesis in stroke: A cohort study. Arch. Neurol. 69, 714–717 (2012). 72. B. Roström, B. Link, Oligoclonal immunoglobulins in cerebrospinal fluid in acute ce- rebrovascular disease. Neurology 31, 590–596 (1981). 73. S. A. Tsementzis, S. W. Chao, E. R. Hitchcock, J. S. Gill, D. G. Beevers, Oligoclonal immunoglobulin G in acute subarachnoid hemorrhage and stroke. Neurology 36, 395–397 (1986). 74. K. R. Crafton, A. N. Mark, S. C. Cramer, Improved understanding of cortical injury by incorporating measures of functional anatomy. Brain 126, 1650–1659 (2003). 75. C. Braun et al., Crossed cortico-spinal motor control after capsular stroke. Eur. J. Neurosci. 25, 2935–2945 (2007). 60. X. Duan, E. Kang, C. Y. Liu, G. L. Ming, H. Song, Development of neural stem cell in the 76. J. Krupinski, J. Kaluza, P. Kumar, S. Kumar, J. M. Wang, Role of angiogenesis in patients adult brain. Curr. Opin. Neurobiol. 18, 108–115 (2008). with cerebral ischemic stroke. Stroke 25, 1794–1798 (1994). 61. K. Jin et al., Neurogenesis in dentate subgranular zone and rostral subventricular zone after focal cerebral ischemia in the rat. Proc. Natl. Acad. Sci. U.S.A. 98, 4710– 4715 (2001). 77. P. W. Duncan, L. B. Goldstein, D. Matchar, G. W. Divine, J. Feussner, Measurement of motor recovery after stroke. Outcome assessment and sample size requirements. Stroke 23, 1084–1089 (1992). 62. A. Arvidsson, Z. Kokaia, O. Lindvall, N-methyl-D-aspartate receptor-mediated increase of neurogenesis in adult rat dentate gyrus following stroke. Eur. J. Neurosci. 14, 10–18 (2001). 63. D. K. Tennakoon et al., Therapeutic induction of regulatory, cytotoxic CD8+ T cells in multiple sclerosis. J. Immunol. 176, 7119–7129 (2006). 78. S. B. Ortega et al., Stroke induces a rapid adaptive autoimmune response to novel neuronal antigens. Discov. Med. 19, 381–392 (2015). 79. U. M. Selvaraj et al., Selective nonnuclear estrogen receptor activation decreases stroke severity and promotes functional recovery in female mice. Endocrinology 159, 3848–3859 (2018). 64. P. H. Lalive et al., Glatiramer acetate in the treatment of multiple sclerosis: Emerging 80. C. Lane-Donovan et al., Physiologic Reelin does not play a strong role in protection concepts regarding its mechanism of action. CNS Drugs 25, 401–414 (2011). against acute stroke. J. Cereb. Blood Flow Metab. 36, 1295–1303 (2016). E C N E I C S O R U E N Ortega et al. PNAS \| March 3, 2020 \| vol. 117 \| no. 9 \| 4993	null
10.1371_journal.pntd.0010231.pdf	Data Availability Statement: De-identified data is available online via the COR-NTD Research Dataverse: https://dataverse.unc.edu/dataset. xhtml?persistentId=doi:10.15139/S3/WOYRUW.	De-identified data is available online via the COR-NTD Research Dataverse: https://dataverse.unc.edu/dataset. xhtml?persistentId=doi:10.15139/S3/WOYRUW .	RESEARCH ARTICLE Positive-case follow up for lymphatic filariasis after a transmission assessment survey in Haiti 1☯, Alain JavelID Marisa A. HastID 1, Tara A. BrantID Keri RobinsonID Aurèle Telfort5, Christine Dubray1 2☯, Eurica Denis2, Kira BarbreID 1, Ryan WiegandID 1, Katherine GassID 3, Jonas RigodonID 3, Marc 4, a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Hast MA, Javel A, Denis E, Barbre K, Rigodon J, Robinson K, et al. (2022) Positive-case follow up for lymphatic filariasis after a transmission assessment survey in Haiti. PLoS Negl Trop Dis 16(2): e0010231. https://doi.org/ 10.1371/journal.pntd.0010231 Editor: Peter Steinmann, Swiss Tropical and Public Health Institute, SWITZERLAND Received: October 21, 2021 Accepted: February 1, 2022 Published: February 25, 2022 Copyright: This is an open access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication. Data Availability Statement: De-identified data is available online via the COR-NTD Research Dataverse: https://dataverse.unc.edu/dataset. xhtml?persistentId=doi:10.15139/S3/WOYRUW. Funding: This work received financial support from the Coalition for Operational Research on Neglected Tropical Diseases, which is funded at The Task Force for Global Health primarily by the Bill & Melinda Gates Foundation, by the United States Agency for International Development through its Neglected Tropical Diseases Program, 1 Centers for Disease Control and Prevention, Atlanta, Georgia, United States of America, 2 IMA World Heath, Port-au-Prince, Haiti, 3 Neglected Tropical Diseases Support Center (NTD-SC), Task Force for Global Health, Atlanta, Georgia, United States of America, 4 Centers for Disease Control and Prevention Country Office, Port-au-Prince, Haiti, 5 Ministry of Public Health and Population (MSPP), Port-au-Prince, Haiti ☯ These authors contributed equally to this work. [email protected] Abstract Background Lymphatic filariasis (LF) has been targeted for global elimination as a public health problem since 1997. The primary strategy to interrupt transmission is annual mass drug administra- tion (MDA) for �5 years. The transmission assessment survey (TAS) was developed as a decision-making tool to measure LF antigenemia in children to determine when MDA in a region can be stopped. The objective of this study was to investigate potential sampling strategies for follow-up of LF-positive children identified in TAS to detect evidence of ongo- ing transmission. Methodology/Principle findings Nippes Department in Haiti passed TAS 1 with 2 positive cases and stopped MDA in 2015; however, 8 positive children were found during TAS 2 in 2017, which prompted a more thor- ough assessment of ongoing transmission. Purposive sampling was used to select the clos- est 50 households to each index case household, and systematic random sampling was used to select 20 households from each index case census enumeration area. All consent- ing household members aged �2 years were surveyed and tested for circulating filarial anti- gen (CFA) using the rapid filarial test strip and for Wb123-specific antibodies using the Filaria Detect IgG4 ELISA. Among 1,927 participants, 1.5% were CFA-positive and 4.5% were seropositive. CFA-positive individuals were identified for 6 of 8 index cases. Positivity ranged from 0.4–2.4%, with highest positivity in the urban commune Miragoane. Purposive sampling found the highest number of CFA-positives (17 vs. 9), and random sampling found a higher percent positive (2.4% vs. 1.4%). Conclusions/Significance Overall, both purposive and random sampling methods were reasonable and achievable methods of TAS follow-up in resource-limited settings. Both methods identified additional PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 1 / 18 PLOS NEGLECTED TROPICAL DISEASESand with UK aid from the British people. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. The authors received no specific funding or salary for this work. Competing interests: The authors have declared that no competing interests exist. Lymphatic filariasis follow-up in Haiti CFA-positives in close geographic proximity to LF-positive children found by TAS, and both identified strong signs of ongoing transmission in the large urban commune of Miragoane. These findings will help inform standardized guidelines for post-TAS surveillance. Author summary Lymphatic filariasis (LF) is a debilitating parasitic disease that has been targeted for global elimination. The transmission assessment survey (TAS) is a tool used to determine if LF transmission has reached low enough levels that prevention activities can be stopped. This study aimed to identify methods to investigate positive LF cases found during TAS. The investigation was conducted in Nippes Department, Haiti, where 8 positive cases were found in TAS in 2017. Participants were recruited through two methods: purposive selec- tion of the closest 50 households to the positive case, and random selection of 20 house- holds in the census enumeration area of the case. Participants completed a survey and were tested for LF antigen, indicative of current infection, and parasite-specific antibody, indicative of current or past infection. A total of 1,927 people participated in the study; 1.5% of these were antigen-positive, and 4.5% were antibody-positive. Purposive sampling found a higher number of antigen-positive individuals, and random sampling found a higher percent positive. Both sampling methods were feasible to use in this setting, and both methods identified signs of ongoing transmission in a large urban area. Additional research is needed to help standardize guidance for post-TAS surveillance to best identify ongoing transmission. Introduction Lymphatic filariasis (LF) is a mosquito-borne parasitic disease caused by the filarial worms Wuchereria bancrofti, Brugia Malayi, and B. timori, and is endemic to tropical areas in 72 countries [1]. The debilitating clinical disease caused by LF can result in fluid accumulation in the extremities, resulting in lymphedema, hydrocoele, and acute adenolymphangitis. LF is one of the leading causes of chronic disability worldwide, and it is estimated that LF was responsi- ble for over 5 million disability-adjusted life years prior to the implementation of disease con- trol programs [2]. Due to this high global burden, LF has been targeted for global elimination as a public health problem since 1997 [3]. The primary strategy for LF control and elimination is the use of annual mass drug admin- istration (MDA) for the entire population at risk for at least five consecutive years or until local transmission is interrupted [4]. In order to define when this criteria has been reached and MDA can be stopped in a region, the transmission assessment survey (TAS) was devel- oped as a decision-making tool to determine when prevalence of LF has reached low enough levels that transmission cannot be sustained, even in the absence of active control measures [5]. The current World Health Organization (WHO) criteria states that TAS can be initiated once pre-TAS sentinel site and spot checks show <2% LF antigenemia among the population over 5 years old in a given evaluation unit (EU), which is typically a district or a combination of districts [4]. As part of TAS, antigenemia in children aged 6–7 is systematically measured, typically using a cluster survey of at least 30 schools, although some variations exist globally. Measuring infection in children this age is relevant for LF since they are of sufficient age to PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 2 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti develop a mature parasite but young enough for a current infection to indicate recent trans- mission. WHO guidelines then state that MDA can be stopped in the EU if prevalence of LF antigenemia among children this age in the first TAS (TAS 1) is below 2% in regions where Anopheles or Culex mosquitoes are the main vectors or below 1% where Aedes mosquito spe- cies are the main vectors [4]. Once antigenemia falls below the appropriate threshold and MDA is stopped, the EU must then pass 2 additional TAS (TAS 2 and TAS 3) over a period of 4–6 years. Once all EUs in a country successfully pass three TAS, the country is eligible to sub- mit a dossier to the WHO to receive formal acknowledgement of validation of the elimination of LF as a public health problem, indicating that prevalence in the population is below the threshold to support onward transmission [6]. Since its inception in 2011, TAS has been widely integrated into national elimination pro- grams worldwide, with over 1,000 EUs having already passed TAS 1 [7]. However, a continu- ing challenge for LF programs is how to interpret and respond to antigen-positive children identified in a TAS 2 or 3 survey that passes the overall threshold. Although a passing result suggests that transmission has been interrupted, any positive result in young children without travel to other regions is cause for concern regarding ongoing transmission, particularly if those cases are clustered geographically or if the number of cases increases between surveys [8]. Furthermore, there is some concern that a cluster-based survey might not be sensitive enough to detect all hotspots of transmission, particularly in large or environmentally hetero- geneous EUs [9,10]. Due to these concerns, the WHO currently encourages program managers to conduct fol- low-up surveys in communities where antigen-positive children are detected [4]. However, lit- tle guidance exists on how follow-up should be conducted, what threshold should trigger a programmatic response, or what this response should entail. Given the recent evidence of residual LF transmission or resurgence in Sri Lanka and American Samoa following multiple passed TAS, additional research is needed to help guide these programmatic decisions and bet- ter determine the utility of TAS for measuring interruption of transmission [11–13]. Addi- tional research is also needed to determine the role of emerging tools, including the use of antibody testing which may detect both current and past infection, and how these complement antigen detection in the context of identifying ongoing transmission. Haiti is one of only four countries in the Americas where transmission of W. bancrofti still occurs [7]. In 2001, it was determined that LF endemicity was widespread throughout the country, and the decision was made to conduct MDA nationwide. By 2012, MDA was con- ducted in all 140 communes (equivalent to districts) [14]. As of 2020, 122 of these communes had reached the WHO criteria to stop MDA, many of which had low or moderate LF transmis- sion at baseline. However, the remaining 18 communes have had persistent LF transmission despite more than 10 years of MDA, and other communes have had an increase in antigen- positive children despite continuing to pass their TAS. In the context of potential continuing transmission, it is critical to follow-up positive cases from TAS 2 and TAS 3 to identify any areas where targeted MDA may be warranted to sustain the gains made towards LF elimina- tion. The objective of this study was to identify potential sampling strategies for positive case follow-up after a TAS using antigen and antibody detection that optimizes the chances of cor- rectly identifying evidence of ongoing transmission in Haiti, while saving program resources. Methods Ethics statement This study was approved by the National Bioethical Committee in Haiti (1718–84) and the Sci- entific Internal Review Board at the CDC (2018–430). Formal informed consent was obtained PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 3 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti by all participants or parents/guardians of children <18 years old. All study participants were read a consent form in Haitian Kreyol by the survey team and provided verbal informed con- sent or parental consent, and children aged 7–17 years provided verbal informed assent. A written copy of the consent form in Haitian Kreyol and the study investigator’s phone number was left with each household. Consent for collecting geographic coordinates was obtained from the head of household. All data were stored on secure servers and password-protected computers. Results of biological testing were kept confidential and only were shared with the study participant or their guardian. Individuals with positive test results were offered treat- ment with the standard of care in Haiti. Study site The site for this study was the department of Nippes, Haiti, which has a population of approxi- mately 342,000 and is comprised of 11 communes [15]. The most populous commune is Mira- goane, which has more than 62,000 residents and contains the department capital. Due to a low baseline prevalence of <5%, all communes were combined into one EU for the LF elimi- nation program, so the whole department is evaluated for TAS together. The department suc- cessfully received five consecutive rounds of MDA from 2009–2013 with sufficient coverage, defined as �65% of the population. TAS 1 was conducted in 2015 using a cluster survey of 30 primary schools, and two children were identified as antigen-positive using the rapid filariasis test strip (FTS) (Alere, Scarborough, ME), which detects circulating filarial antigenemia (CFA) for adult worms. This was below the critical cutoff of 2% for regions with Culex vectors, and MDA in the department was stopped. Nippes underwent TAS 2 in 2017 and passed again, however the number of CFA-positive children had increased to eight. Of these, four came from the commune of Miragoane, with two residing across the street from each other in a dense urban area. The four other CFA-positive children were located in the communes of Anse-a-Veau, L’Asile, Petit-Trou de Nippes, and Plaisance du Sud. Study design and sampling Each CFA-positive child identified in TAS 2 was considered an index case. The residential locations of the index cases were plotted in ArcGIS Version 10.7 (ESRI, Redland, CA), and each case was mapped to an Enumeration Area (EA) for follow-up (Fig 1). EA boundaries were previously determined by the Haitian Ministry of Health and Ministry of Statistics for use with the census and the Demographic Health Survey and range from 0.02–30.8 square kilometers. The index case in Anse-a-Veau had moved recently from a more rural area, so both the case’s current location and previous residence were mapped to their respective EAs for follow-up. Purposive and random sampling methods were used to select participants for this study. Purposive sampling For each index case, the location of the index case household was identified by the field team using global position system (GPS) coordinates. The index household and the 50 closest house- holds by straight-line distance were selected for inclusion in the sample. Fifty was chosen since it was the maximum deemed to be feasible and sustainable using this method. In all selected households, head of household consent was obtained, and household coordinates were col- lected on electronic mobile devices. If any of the 50 nearest households declined, it was replaced with the next nearest household until 50 had been enrolled. To identify the 50 nearest households to each index case, structures surrounding index case households were enumerated using high-definition satellite images (DigitalGlobe, Denver, PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 4 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti Fig 1. Locations of antigen-positive lymphatic filariasis cases identified in TAS 2, Nippes Department, Haiti 2017. Exact locations are jittered for participant confidentiality. Shapefiles for administrative boundaries are from the Centre National d’Information Geo-Spatiale and are available at https://data.humdata.org/dataset/hti-polbndl-adm1-cnigs-zip. https://doi.org/10.1371/journal.pntd.0010231.g001 CO), and the proportion of inhabited structures in that EA was calculated using geo-linked census data. Using a projected non-response/refusal rate of 10%, an approximate radius was calculated around each index household to encompass the 50 households anticipated to partic- ipate. A buffer of the appropriate radius was developed in ArcGIS for each index case, and these shapefiles were loaded onto Garmin GPSMAP 64 handheld GPS devices (Olathe, KS) and Locus Map software (Prague, Czech Republic) for use by the field team. Once the index household was identified, field teams started at the index house and proceeded outwards in concentric circles until 50 households had been enrolled, using the buffer shapefile as a guide. Due to their close proximity, the two Miragoane index cases in the same EA were treated as a unit and a purposive circle was drawn around the midpoint between the two houses. Random sampling: Index EA In each index case EA, field teams conducted a comprehensive census in the EA to identify and number all occupied households. The boundaries of the EA were determined using maps and GPS tools, and every structure in the EA was categorized as a household or non-house- hold. Each occupied household was given a number, and 20 households were selected for inclusion using systematic random sampling, which was the maximum deemed to be feasible PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 5 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti and sustainable using this method. A sampling interval was determined by dividing the total number of households by 20 and rounding down to the nearest whole number. A random start was selected between 1 and the sampling interval using random number generating soft- ware. A list of selected households was developed by serially adding the sampling interval to the random start until 20 households had been selected. At each selected house, head of house- hold consent was obtained, and household coordinates were collected on electronic mobile devices. If any of the selected households declined to participate, it was not replaced. Random sampling: Neighbor EA If any participant in a selected household from either purposive or random sampling tested CFA-positive for LF (methods to follow), additional random sampling of 20 households was conducted in each of the two neighboring EAs nearest to the positive case. For these neighbor- ing EAs, sampling methods were identical to random sampling conducted in index case EAs, including conducting the census, household numbering, identification of a sampling interval and random start, listing of selected households, and head of household consent. Field methods and sample collection Field data collection occurred from July to August 2019. In selected households, all household members aged 2 years and older were invited to participate in the sample. Adults provided informed consent or parental permission for their children under 18 years, and children aged 7–17 years provided informed assent. Exclusion criteria included age under 2 years, having another primary residence, and inability to provide informed consent due to physical or men- tal incapacity. All consenting household members were given a unique barcode and adminis- tered an electronic questionnaire using Secure Data Kit software (Atlanta, GA). The questionnaire collected information on participant demographics, time living in the EA, his- tory of travel, participation in the most recent MDA, and use of bed nets. All consenting participants provided approximately 250μl of blood, which was collected by finger stick into heparinized collection tubes. Collection tubes were marked with a matching barcode to the participant and were not otherwise marked with any personally identifying information. Antigen testing was done at point-of-collection with FTS using standardized methods [16]. In brief, 75 μL of blood were pipetted from the heparinized tubes onto the test sample pad and allowed to flow through the strip. After exactly 10 minutes, the result window was read for a positive, negative or invalid CFA result. Any positive or invalid FTS were repeated with a sec- ond confirmatory FTS. Individuals with two positive FTS were counted as CFA-positive, and individuals with one or more negative FTS were counted as CFA-negative. Tubes with the remainder of the blood samples were then transported to the field laboratory in a portable cooler, where dried blood spots (DBS) were prepared for each participant. Ten μL of heparin- ized blood were pipetted onto each of six extensions of Trop Bio filter paper (Cellabs, Sydney, Australia) for a total sample of 60 μL. Filter paper was labeled with the unique barcode for the participant, allowed to dry, and packed with desiccant in labelled bags until transport to the National Laboratory of Public Health in Port au Prince, where they were stored at -20˚C. Individuals with two positive FTS were treated with the standard combination of albenda- zole and diethylcarbamazine (DEC), with dosage calculated based on the age of the participant following standard procedure in Haiti. Pregnant women were eligible for participation in the study; however, due to the unknown effects of these medications in utero, pregnant partici- pants who tested CFA-positive were advised to seek treatment for LF a week after delivery. PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 6 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti Laboratory analysis Preserved DBS were transported to the Centers for Disease Control and Prevention (CDC) in Atlanta, GA where they were tested for IgG4 antibodies against the recombinant Wb123 anti- gen using the Filaria Detect IgG4 ELISA kit (InBios, Seattle, WA), a direct enzyme immunoas- say. The assay was performed according to the standard operating procedure provided by the manufacturer with minor modifications, as described previously [17]. In brief, blood spot extensions of DBS and positive and negative controls were diluted 1:50 in kit-provided sample buffer and stored overnight at 4˚C. Samples and controls were added to plate wells, sealed, and incubated at 37˚C for 30 minutes, washed using an automated plate washer and kit-provided wash buffer, incubated again with mouse anti-human IgG4 conjugated with horseradish per- oxidase for another 30 minutes, and washed again. Plates were developed at room temperature in the dark for 13 minutes with an added 100 μL of tetramethylbenzidine substrate in each well, stopped using kit-provided stop solution, and incubated for one minute. Plates were read on a microplate reader at 450 nm. To compare optical density (OD) values across plates the OD values were normalized by dividing the mean OD of the sample by the mean OD of the H3-positive control from the same plate. Additional details are described in Supplementary Materials S1 Text. Data analysis Cutoff determination for ELISA OD values. In order to analyze the ELISA OD values as a dichotomous variable, a fixed finite mixture model (FMM) [18] was fit to the serology OD dataset to determine a cutoff using the flexmix, mixtools, mixsmsn and sn packages in R version 4.0.2 (R Core Team, Vienna, Austria). These models use maximum likelihood estimation to fit a two-component FMM to the data to estimate the parameters of “seropositive” and “seroneg- ative” distributions [19,20]. Each component was fitted with a normal or skew-normal distri- bution to log-transformed OD values and assumed to be independent of age. The ELISA value at which the probability of positive was greater than 0.5 was used as the cutoff, which is also the point where the two distributions intersect [21]. Values above the cutoff were determined to be seropositive and values below were considered seronegative. Statistical analysis. All analyses were done in Stata 13.1 (Stata Corporation, College Sta- tion, TX) and R version 4.0.2 statistical software. Frequencies of demographic and behavioral characteristics of the study population were calculated. Both CFA and serology results are described by participant characteristics and sampling method, and bivariate comparisons were conducted using chi-squared and Fisher’s exact tests. For purposively sampled households, the proportion and number CFA positive were compared by increasing number of sampled households from the index case house (e.g., closest 10, closest 20, etc.). To better visualize their distribution, the natural log of serology OD values were plotted against participant characteris- tics in scatterplots. Exploratory analyses were conducted both by index case and by geographic zone. Index cases were defined by number as cases 1–8. The two index cases in the same EA in Mira- goane, designated as 1a and 1b, were analyzed together as index case 1 due to their close proximity and resulting inability to differentiate in sampling. Geographic zones were defined as Miragoane (containing index cases 1a, 1b, 2, and 3), L’Asile (index case 4), Plaisance du Sud (index case 5), Petit-Trou de Nippes (index case 6), Anse-a-Veau rural (index case 7), and Anse-a-Veau urban (index case 8). As described above, index cases 7 and 8 were the same individual but were analyzed separately since he or she resided in the two zones at dif- ferent times. PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 7 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti Results Study population A total of 1,927 participants from 786 unique households completed the survey and FTS test- ing. Of these, nearly 40% were male and 18% were under age 10 years (Table 1). Nearly half of participants resided in the commune of Miragoane, with 8–14% residing in each of the other 5 geographic zones. Forty percent of participants were students, 20% listed commerce as their primary occupation, and 15% did not have a stated occupation. Approximately one fifth of participants reported sleeping under a bed net the previous night, and a quarter reported travel outside their commune in the past year. Among participants aged 12 and above who would have been eligible to take MDA during the last administration, 65% reported ever taking medi- cations for LF as part of MDA in the past. By mode of sampling, 63% of participants were selected by purposive sampling, 20% were selected by random sampling in the index case EA, and 26% were selected by random sam- pling in a neighboring EA. Nine percent of participants (175) were selected for both purposive and either index case EA or neighbor EA random sampling, and thus these numbers add to more than 100%. Antigen results (CFA) Of the 1,927 participants who completed FTS testing, 29 were CFA-positive, for a total positivity rate of 1.5% (Table 2). Of these, 21 (72%) were from Miragoane, and each index case in this commune led to the identification of CFA-positive participants, with the number found per index case ranging from 6 to 8 (Fig 2). An additional 3 and 4 CFA-positive participants were found in Anse-a-Veau rural and urban zones, respectively, 1 was found in L’Asile, and no CFA- positive participants were found in Plaisance du Sud or Petit-Trou de Nippes (S1 Table). Over- all, 6 of the 8 original index cases (75%) were linked to at least one CFA-positive participant. Miragoane also had the highest rate of CFA positivity, with 2.4% of participants testing positive compared to the range of 0.0%–1.8% in other geographic zones (Chi squared P value = 0.03). Two CFA positives were found among participants aged <10 years (positivity 0.8%), both of which resided in Miragoane and were between the ages of 4 and 6 years. CFA positivity rates were slightly higher among participants aged 20–44 years (2.4%) compared to other age bands; however, this difference was not statistically significant (P = 0.08). Positivity was slightly higher among participants who did not sleep under a bed net (1.6% vs. 1.0%) or did not take MDA in past (2.3% vs. 1.5%) compared to those who did; however, these differences also were not statistically significant (P = 0.4 and 0.3, respectively). Rates of positivity were similar by participant sex and history of travel. Purposive sampling identified 17 CFA-positive participants for a positivity rate of 1.4%, ran- dom sampling in the index case EA identified 9 positive participants for a positivity rate of 2.4%, and random sampling in a neighboring EA identified only 3 positive participants for a positivity rate of 0.6% (Table 2). Thereby, purposive sampling identified the highest number of CFA-positive participants, but random sampling in the index case EA returned the highest per- cent positive. Among the 6 index cases that yielded any CFA-positive cases in this investigation, purposive sampling found positives for 5 index cases in 3 geographic zones (Miragoane, rural and urban Anse-a-Veau), and random sampling in the index EA found positives for 4 index cases in 3 geographic zones (Miragoane, L’asile, urban Anse-a-Veau) (S1 Table). Both of these sampling methods missed cases in one zone each. Positivity rate by geographic zone ranged from 1.3–2.7% for purposive sampling and 2.0–9.6% for random sampling. Random sampling in a neighboring EA found one additional CFA-positive participant for 3 of the 6 index cases. PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 8 / 18 PLOS NEGLECTED TROPICAL DISEASESTable 1. Demographics and participant characteristics among survey respondents in Nippes Department, Haiti, July-August 2019. N = 1,927. Lymphatic filariasis follow-up in Haiti Sex Male Female Age in years <10 10–19 20–44 �45 Commune of residence Miragoane L’Asile Plaisance du Sud Petit-Trou de Nippes Anse-a-Veau rural Anse-a-Veau urban Occupation Student Commerce Agriculture1 Manual labor Public sector Private sector None Other Missing Slept under bed net last night Yes No Traveled outside commune past year Yes No Have taken MDA in past2 Yes No Don’t know Sampling method3 Purposive Random (index) Random (neighbor) Total n 742 1,185 351 461 702 413 861 268 158 152 268 220 776 378 157 71 32 19 288 86 120 392 1,535 500 1,427 943 488 26 1,221 381 500 1927 % 38.5% 61.5% 18.2% 23.9% 36.4% 21.5% 44.7% 13.9% 8.2% 7.9% 13.9% 11.4% 40.3% 19.6% 8.1% 3.7% 1.7% 1.0% 15.0% 4.5% 6.2% 20.3% 79.7% 26.0% 74.0% 64.7% 33.5% 1.8% 63.4% 19.7% 25.9% 100% 1 Includes farming or fishing 2among participants aged 12 years and above 3 adds to more than 100% because 175 participants were from households selected for both purposive and random sampling; MDA = mass drug administration https://doi.org/10.1371/journal.pntd.0010231.t001 PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 9 / 18 PLOS NEGLECTED TROPICAL DISEASESTable 2. Filariasis test strip (FTS) and serology results by participant characteristic, Nippes Department, Haiti, July-August 2019. FTS N = 1927; Serology N = 1,914. CFA Positive n (%) P value� Serology Positive n (%) P value� Lymphatic filariasis follow-up in Haiti Sex Males Females Age in years <10 10–19 20–44 �45 Slept under bed net last night Bed net No bed net Travel outside commune past year Yes No Have taken MDA in past Yes No Geographic zone Miragoane L’Asile Plaisance du Sud Petit-Trou de Nippes Anse-a-Veau rural Anse-a-Veau urban Sampling method Purposive sampling Random (index) Random (neighbor) Total 0.7 0.08 0.4 0.8 0.3 0.03 0.08 12 (1.6%) 17 (1.4%) 2 (0.8%) 5 (1.1%) 17 (2.4%) 5 (1.2%) 4 (1.0%) 25 (1.6%) 8 (1.6%) 21 (1.5%) 14 (1.5%) 11 (2.3%) 21 (2.4%) 1 (0.4%) 0 (0.0%) 0 (0.0%) 3 (1.1%) 4 (1.8%) 17 (1.4%) 9 (2.4%) 3 (0.6%) 29 (1.5%) 43 (5.8%) 44 (3.7%) 11 (3.1%) 14 (3.0%) 35 (5.0%) 27 (6.6%) 9 (2.3%) 78 (5.1%) 27 (5.4%) 59 (4.2%) 42 (4.5%) 22 (4.6%) 43 (5.1%) 9 (3.4%) 6 (3.8%) 9 (6.0%) 13 (4.8%) 6 (2.7%) 62 (5.1%) 15 (4.0%) 16 (3.2%) 87 (4.5%) 0.02 0.06 0.02 0.2 0.9 0.5 0.2 �P values were Chi squared tests or Fisher’s Exact tests as appropriate; CFA = circulating filarial antigen https://doi.org/10.1371/journal.pntd.0010231.t002 Among purposively sampled households, average CFA positivity was highest in the closest 10 households to the index case at 1.6%, declined to 1.4% in the 20 closest households, and then remained relatively stable at 1.3%, 1.3%, and 1.4% in the closest 30, 40, and 50 households respectively. However, these differences were not statistically significant, and there was sub- stantial variation by index case and geographic zone (Figs 3 and S2). For 4 of 5 index cases with purposively identified CFA-positive participants, the majority of these were found within the 20 closest households to the index household, and prevalence in these households was higher than random sampling for 3 of 5 index cases (S1 Table and S1 Fig). The primary excep- tion to this was for Miragoane index case 2, which had >9% positivity among randomly sam- pled participants compared to 2.4% positivity among purposively sampled households. Serology results (ELISA OD values) Of the 1,927 participants who received FTS testing, 1,914 (99.3%) also provided DBS for serol- ogy analysis. Normalized mean OD values from ELISA testing ranged from 0.04 to 4.5, with a PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 10 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti Fig 2. Locations of individuals positive for circulating filarial antigen (CFA) identified through purposive or random sampling in relation to their index case, Nippes Department, Haiti July-August 2019. Exact locations are jittered for participant confidentiality. Shapefiles for administrative boundaries are from the Centre National d’Information Geo-Spatiale and are available at https://data. humdata.org/dataset/hti-polbndl-adm1-cnigs-zip. https://doi.org/10.1371/journal.pntd.0010231.g002 mean of 0.11. The cutoff for serology positivity determined by FMM methods was 0.18. Using this cutoff, there were 87 serology-positive participants, for a seropositivity rate of 4.5%. Con- cordance with CFA antigen results was 95%, with 6 participants positive by both methods, 1,813 negative by both methods, 23 CFA-positive only, and 81 seropositive only. Seropositivity was higher than CFA positivity across every category, however the size of the difference varied (Table 2). Seropositive participants were found for all 8 index cases, with pos- itivity rates ranging from 2.7–6.0% (S2 Table). The distribution of log-transformed OD values across participant characteristics can be seen in Fig 4. The highest number of seropositive par- ticipants was found in Miragoane (43, 5.1% positive), but the highest percent seropositive was found in Petit-Trou de Nippes (6.0% positive), which had no CFA positives. There were no sig- nificant differences in seropositivity by index case or geographic zone (Table 2 and Fig 4). Eleven seropositive results were found in participants <10 years (3.1%), and seropositivity increased by age to 6.6% in participants aged �45 years, although this increase was not statisti- cally significant (Table 2). Seropositive children were aged 4–9 years and were located in Mira- goane, rural Anse-a-Veau, and Petit-Trou de Nippes. Seropositivity was higher among male (5.8%) than female (3.7%) participants (P = 0.02) and was higher among participants who did not sleep under a bed net (5.1%) than those who did (2.3%) (P = 0.02) (Table 2 and Figs 4 and PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 11 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti Fig 3. Circulating filarial antigen (CFA) prevalence by geographic zone, index case, and number of purposively sampled households by increasing distance from index case household, Nippes Department, Haiti, July-August 2019. https://doi.org/10.1371/journal.pntd.0010231.g003 S3). Seropositivity was similar among participants who did or did not participate in MDA for LF in the past. By sampling method, seropositivity was 5.1% by purposive sampling, 4.0% by random sampling in the index case EA, and was 3.2% by random sampling in neighboring EAs. Purposive sampling found at least one seropositive participant for all 8 index cases, ran- dom sampling in the index EA found seropositive participants for 7 of 8, and random sam- pling in neighboring EAs found seropositive participants for 4 of 8 index cases. Discussion This investigation describes findings for several methods of follow-up for CFA-positive chil- dren identified in TAS 2 in the Department of Nippes, Haiti. From 8 initial index cases in 6 geographic zones, a total of 29 additional CFA-positive individuals were identified out of 1,927 tested, and CFA-positives were found during follow-up for 6 of the 8 index cases. High rates of CFA positivity were found in the commune of Miragoane (2.4%). Overall, both purposive and random sampling methods detected additional CFA-positive persons in the vicinity of most index cases, and both methods identified a high positivity rate in Miragoane. Purposive sampling tested, found, and treated the highest number of CFA-posi- tives, and this method found follow-up CFA-positive participants for a higher number of index cases. Correspondingly, random sampling in the index case EA returned the highest per- cent positive and was therefore most efficient at identifying additional cases. Random sam- pling in neighboring EAs returned both a low number and percent positivity, indicating that this additional element may be of limited usefulness in the future. Considering use of these sampling methods in the resource-limited context of Haiti, purpo- sive sampling tested a higher number of households (50 per index case compared to 20) and therefore was more resource-intensive for both field staff hours and study materials, but PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 12 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti Fig 4. Log normalized mean optical density (OD) values for IgG4 antibodies against the recombinant Wb123 antigen by participant characteristics. Red line indicates cutoff for seropositivity. https://doi.org/10.1371/journal.pntd.0010231.g004 random sampling was more time-consuming overall due to the need for a census of the index EA in order to select a random sample, which took 2.2 days per EA on average. Interestingly, purposive testing of only the closest 20 households to the index case yielded a similar percent positivity as the closest 50 households in sub-analyses, which corresponds to previous findings in Haiti that additional LF cases may be clustered close to the index case [22]. Purposive sam- pling of fewer households may therefore be an option in the future to save resources, although more research is needed to confirm this across settings and species of LF. Random sampling might also not be as time intensive if countries have already performed the appropriate census in advance, if satellite mapping could be used to enumerate households, or if alternatives to random sampling such as segmentation could be used. As a whole, both random and purpo- sive sampling had merits and continue to be potential methods for TAS follow-up. In this setting, having a higher number of index cases was also correlated with a higher per- cent positivity. Four of the original index cases were found in Miragoane, and percent positiv- ity in this geographic zone was 2.4% compared to 0.0–1.8% in the other 4 zones where only one index case was found. Despite this result, sample sizes were too small to draw definitive conclusions, and percent positivity approached the 2% cutoff in the urban zone of Anse-a- PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 13 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti Veau despite having only one index case. These observations indicate that, although having multiple index cases in an area might suggest heightened chance of ongoing transmission, at this time the conservative choice still would be to follow up every positive case found in TAS. Of importance for the LF elimination effort in Haiti, the CFA prevalence of 2.4% in Mira- goane was higher than the 2% cutoff typically needed in pre-TAS sentinel checks to initiate the TAS process and stop MDA in regions with Culex vectors [4]. The two CFA-positive children under 7 years were also found in Miragoane, which is concerning since MDA was stopped in Nippes eight years prior and might be a further indication of ongoing transmission. This is a change from historical data, as Miragoane had no positive LF cases during initial mapping; how- ever, the dramatic urbanization of the commune and influx of people from rural areas might have resulted in a higher concentration of cases and thereby allowed low levels of transmission despite the pressure of MDA. As shown in previous research, urban areas can have unique barriers to LF elimination, including high community mobility and lower compliance with MDA [23,24]. As a result of this study’s findings, the Haiti National Program to Eliminate LF has decided to adjust the practice of combining all the Nippes Department communes for the upcoming TAS 3 and will evaluate Miragoane separately from the rest of the department. This will likely provide a more sensitive result and will identify whether the commune needs to resume MDA. This study was also enriched by antibody testing, which provide a more thorough picture of LF in Nippes. As seen in other studies [25,26], a higher proportion of participants were sero- positive than CFA-positive since antibody responses may be seen in both past and present LF infection [27]. Seropositivity was not significantly different by geographic zone and was not correlated with number of index cases, but was higher among males, older adults, and among participants who reported not sleeping under a bed net. This might indicate higher prevalence of lifetime infection in these groups, consistent with other studies [28]. Of note, seropositivity in this population was approximately 3% in both children under 10 years and children aged 10–19 years. Seropositivity in the youngest age group could indicate persistent or recurrent LF transmission since infection was likely acquired after MDA was stopped eight years prior to the study and it takes approximately three years to develop antibodies after initial infection [29]. Antibody response to Wb123 has also been correlated with molecular xenomonitoring results in prior studies, further suggesting that ongoing transmission may be occurring [25]. However, this could also indicate residual seropositivity following interruption of transmission in the older children. Continuing vigilance in this population is warranted to determine if anti- body responses are indicative of ongoing transmission. Similar to previous investigations [30–32], individual-level concordance of Wb123-specific antibody and CFA antigen was poor in this investigation. Eighty-one participants (4%) were antibody positive only, which can be explained by past infection and persistent antibody response. However, 23 participants (1%) were CFA-positive only, which constituted a high proportion of the total 29 CFA-positives. This could potentially be due to recent infection or a less durable antibody response in some individuals [27,29], but further research is needed to characterize the relationship between LF antibody and antigen response to better understand the utility of these markers during post-MDA surveillance. Therefore, antibody results appear to primarily be useful at the population level to describe patterns and trends in transmission but may not be appropriate for individual diagnosis, while antigen detection remains the gold standard for both diagnosis and surveillance. This study had several limitations. Individuals in sampled households could have been missed if they were away during household visits, and responses to the survey could have been inaccurate if respondents did not sufficiently recall their history or provided answers they thought interviewers preferred. Due to the small sample size of positive participants and high number of zero cells, there was insufficient power to conduct multivariate analyses, including PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 14 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti more complex analyses comparing sampling methods or antigen vs. antibody results. Further- more, due to varying household density in urban vs. rural zones, the distance of the 50 closest household to the index case varied greatly during purposive sampling. In urban areas, the diameter of the purposive circle was as small as 135 m, but it ranged up to 2 km in rural areas, so an ideal minimum distance for sampling could not be established. This issue was also seen to a lesser extent in random sampling, since the size of the census EA also varied by household density and was as small as 0.02 square kilometers in urban settings. Future studies could sam- ple all households within a set diameter in order to determine a minimum sampling distance. Another limitation of this study is the possibility that the cutoff values for the ELISA OD values were inaccurate. Due to the absence of well-characterized panels for neglected tropical diseases, it is often challenging to determine cutoffs for serological assays. The method used in this study represents fitting two distributions to the OD data with the assumption that there is limited overlap between the results for true positive and negative samples. Although this method is becoming standard in neglected tropical disease research, it is unknown to what extent the assumptions are met, particularly given the unknown duration of antibody responses to these pathogens. Future studies and meta-analyses can better confirm the appro- priateness of this method. Despite these limitations, this study provides some of the first systematic analyses on fol- low-up of LF CFA-positive children identified by TAS. Two sampling methods were demon- strated to be achievable in resource-limited settings, and both identified strong signs of ongoing transmission in the large urban commune of Miragoane. This area of potential trans- mission after the cessation of MDA has the potential to disrupt the local program to eliminate LF in Haiti and further demonstrates the importance of TAS follow-up for the Global Program to Eliminate Lymphatic Filariasis. While this study will help inform standardized guidelines for post-TAS surveillance, more research is needed, and additional results including cost-effi- cacy analyses from forthcoming post-TAS surveillance analyses from the Philippines, Burkina Faso, and Nepal will continue to explore best methods for TAS follow up. Supporting information S1 Text. Detailed methods for ELISA antibody testing. (DOCX) S1 Table. Number and percent of participants who tested positive for circulating filarial antigen by index case and sampling method, Nippes Department, Haiti, July-August 2019. Purposive sampling N = 1,221; random sampling index case enumeration area (EA) N = 381; random sampling neighbor EA N = 500. (DOCX) S2 Table. Number and percent of participants who tested serology positive by index case and sampling method, Nippes Department, Haiti, July-August 2019. N = 1,914. (DOCX) S1 Fig. Circulating filarial antigen (CFA) prevalence by geographic zone, index case, and number of purposively sampled households by increasing distance in meters from index case household in comparison to households random sampled in the index case enumera- tion area (EA) where available, Nippes Department, Haiti, July-August 2019. (TIF) S2 Fig. Circulating filarial antigen (CFA) prevalence by geographic zone and index case by increasing 10-house band of increasing distance from index case household, Nippes PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 15 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti Department, Haiti, July-August 2019. Each band comprised only the 10 households in that distance band (e.g. 1–10, 11–20, 21–30, etc.) (TIF) S3 Fig. Log normalized mean optical density (OD) values for IgG4 antibodies against the recombinant Wb123 antigen by participant characteristics and circulating filarial antigen (CFA) results. Navy color indicates CFA-negative participants, and pink color indicates CFA- positive participants. Red line indicates cutoff for seropositivity. (TIF) Acknowledgments The findings and conclusions in the article are those of the authors and do not necessarily rep- resent the views of the US Centers for Disease Control and Prevention. The authors would like to thank the field study teams for their hard work and dedication to this project and would especially like to thank the participants in Nippes Department of Haiti for their time and cooperation with the study. We would also like to thank Amber Dismer, Kim Won, Paul Cantey, and Mary Kamb at CDC, and Patrick Lammie at the Task Force for Global Health for their expertise and advice. Author Contributions Conceptualization: Marisa A. Hast, Alain Javel, Kira Barbre, Tara A. Brant, Katherine Gass, Marc Aurèle Telfort, Christine Dubray. Data curation: Marisa A. Hast, Keri Robinson. Formal analysis: Marisa A. Hast, Ryan Wiegand. Funding acquisition: Kira Barbre, Katherine Gass, Christine Dubray. Investigation: Marisa A. Hast, Alain Javel, Eurica Denis, Keri Robinson, Christine Dubray. Methodology: Marisa A. Hast, Alain Javel, Eurica Denis, Kira Barbre, Keri Robinson, Ryan Wiegand, Katherine Gass, Christine Dubray. Project administration: Marisa A. Hast, Alain Javel, Eurica Denis, Jonas Rigodon, Tara A. Brant, Katherine Gass, Christine Dubray. Resources: Alain Javel, Eurica Denis, Kira Barbre, Jonas Rigodon, Keri Robinson, Katherine Gass, Marc Aurèle Telfort. Supervision: Alain Javel, Eurica Denis, Jonas Rigodon, Katherine Gass, Marc Aurèle Telfort, Christine Dubray. Visualization: Marisa A. Hast, Ryan Wiegand. Writing – original draft: Marisa A. Hast. Writing – review & editing: Alain Javel, Eurica Denis, Kira Barbre, Jonas Rigodon, Keri Rob- inson, Tara A. Brant, Ryan Wiegand, Katherine Gass, Marc Aurèle Telfort, Christine Dubray. References 1. Utzinger J, Becker SL, Knopp S, Blum J, Neumayr AL, Keiser J, et al. Neglected tropical diseases: diag- nosis, clinical management, treatment and control. Swiss Med Wkly. 2012; 142:w13727. Epub 2012/11/ 28. https://doi.org/10.4414/smw.2012.13727 PMID: 23180107. PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 16 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti 2. Mathew CG, Bettis AA, Chu BK, English M, Ottesen EA, Bradley MH, et al. The health and economic burden of lymphatic filariasis prior to mass drug administration programmes. Clin Infect Dis. 2019. Epub 2019/07/26. https://doi.org/10.1093/cid/ciz671 PMID: 31343064. 3. World Health Organization. Elimination of lymphatic filariasis as a public health problem—resolution of the executive board of the WHO. Geneva: World Health Organization, 1997. 4. World Health Organization. Monitoring and epidemiological assessment of mass drug administration in the global programme to eliminate lymphatic filariasis: a manual for national elimination programmes. Geneva: World Health Organization, 2011. 5. Chu BK, Deming M, Biritwum NK, Bougma WR, Dorkenoo AM, El-Setouhy M, et al. Transmission assessment surveys (TAS) to define endpoints for lymphatic filariasis mass drug administration: a multi- center evaluation. PLoS Negl Trop Dis. 2013; 7(12):e2584. Epub 2013/12/18. https://doi.org/10.1371/ journal.pntd.0002584 PMID: 24340120; PubMed Central PMCID: PMC3855047. 6. World Health Organization. Validation of elimination of lymphatic filariasis as a public health problem. Geneva: 2017. 7. World Health Organization. Global programme to eliminate lymphatic filariasis: progress report, 2018. Wkly Epidemiol Rec. 2019; 94(41):457–72. PMID: 28984121. 8. Lammie PJ, Gass KM, King J, Deming MS, Addiss DG, Biswas G, et al. Evolution of the monitoring and evaluation strategies to support the World Health Organization’s Global Programme to Eliminate Lym- phatic Filariasis. Int Health. 2020; 13(Supplement_1):S65–S70. Epub 2020/12/23. https://doi.org/10. 1093/inthealth/ihaa084 PMID: 33349883; PubMed Central PMCID: PMC7753166. 9. Michael E, Singh BK, Mayala BK, Smith ME, Hampton S, Nabrzyski J. Continental-scale, data-driven predictive assessment of eliminating the vector-borne disease, lymphatic filariasis, in sub-Saharan Africa by 2020. BMC Med. 2017; 15(1):176. Epub 2017/09/28. https://doi.org/10.1186/s12916-017- 0933-2 PMID: 28950862; PubMed Central PMCID: PMC5615442. 10. Prada JM, Davis EL, Touloupou P, Stolk WA, Kontoroupis P, Smith ME, et al. Elimination or Resur- gence: Modelling Lymphatic Filariasis After Reaching the 1% Microfilaremia Prevalence Threshold. J Infect Dis. 2019. Epub 2019/12/20. https://doi.org/10.1093/infdis/jiz647 PMID: 31853554. 11. Rao RU, Nagodavithana KC, Samarasekera SD, Wijegunawardana AD, Premakumara WD, Perera SN, et al. A comprehensive assessment of lymphatic filariasis in Sri Lanka six years after cessation of mass drug administration. PLoS Negl Trop Dis. 2014; 8(11):e3281. Epub 2014/11/14. https://doi.org/ 10.1371/journal.pntd.0003281 PMID: 25393404; PubMed Central PMCID: PMC4230885. 12. Sheel M, Sheridan S, Gass K, Won K, Fuimaono S, Kirk M, et al. Identifying residual transmission of lymphatic filariasis after mass drug administration: Comparing school-based versus community-based surveillance—American Samoa, 2016. PLoS Negl Trop Dis. 2018; 12(7):e0006583. Epub 2018/07/17. https://doi.org/10.1371/journal.pntd.0006583 PMID: 30011276; PubMed Central PMCID: PMC6062125. 13. Lau CL, Sheridan S, Ryan S, Roineau M, Andreosso A, Fuimaono S, et al. Detecting and confirming residual hotspots of lymphatic filariasis transmission in American Samoa 8 years after stopping mass drug administration. PLoS Negl Trop Dis. 2017; 11(9):e0005914. Epub 2017/09/19. https://doi.org/10. 1371/journal.pntd.0005914 PMID: 28922418; PubMed Central PMCID: PMC5619835. 14. Oscar R, Lemoine JF, Direny AN, Desir L, Beau de Rochars VE, Poirier MJ, et al. Haiti National Pro- gram for the elimination of lymphatic filariasis—a model of success in the face of adversity. PLoS Negl Trop Dis. 2014; 8(7):e2915. Epub 2014/07/18. https://doi.org/10.1371/journal.pntd.0002915 PMID: 25032697; PubMed Central PMCID: PMC4102456. 15. Institut Haïtien de Statistique et d’Informatique. Population Totale, Population de 18 Ans et Plus Menages et Densites Estimes en 2015 2015 [cited 2021 May 5]. Available from: http://www.ihsi.ht/. 16. Weil GJ, Curtis KC, Fakoli L, Fischer K, Gankpala L, Lammie PJ, et al. Laboratory and field evaluation of a new rapid test for detecting Wuchereria bancrofti antigen in human blood. Am J Trop Med Hyg. 2013; 89(1):11–5. Epub 2013/05/22. https://doi.org/10.4269/ajtmh.13-0089 PMID: 23690552; PubMed Central PMCID: PMC3748464. 17. Won KY, Sambou S, Barry A, Robinson K, Jaye M, Sanneh B, et al. Use of Antibody Tools to Provide Serologic Evidence of Elimination of Lymphatic Filariasis in The Gambia. Am J Trop Med Hyg. 2018; 98 (1):15–20. Epub 2017/11/23. https://doi.org/10.4269/ajtmh.17-0371 PMID: 29165213; PubMed Central PMCID: PMC5928708. 18. Parker RA, Erdman DD, Anderson LJ. Use of mixture models in determining laboratory criterion for identification of seropositive individuals: application to parvovirus B19 serology. J Virol Methods. 1990; 27(2):135–44. Epub 1990/02/01. https://doi.org/10.1016/0166-0934(90)90130-8 PMID: 2156877. 19. Gay NJ, Vyse AJ, Enquselassie F, Nigatu W, Nokes DJ. Improving sensitivity of oral fluid testing in IgG prevalence studies: application of mixture models to a rubella antibody survey. Epidemiol Infect. 2003; PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 17 / 18 PLOS NEGLECTED TROPICAL DISEASESLymphatic filariasis follow-up in Haiti 130(2):285–91. Epub 2003/05/06. https://doi.org/10.1017/s0950268802008051 PMID: 12729197; PubMed Central PMCID: PMC2869964. 20. Vyse AJ, Gay NJ, Hesketh LM, Pebody R, Morgan-Capner P, Miller E. Interpreting serological surveys using mixture models: the seroepidemiology of measles, mumps and rubella in England and Wales at the beginning of the 21st century. Epidemiol Infect. 2006; 134(6):1303–12. Epub 2006/05/03. https:// doi.org/10.1017/S0950268806006340 PMID: 16650326; PubMed Central PMCID: PMC2870519. 21. Pfeiffer RM, Gail MH, Brown LM. Probability of helicobacter pylori infection based on IgG levels and other covariates using a mixture model. J Epidemiol Biostat. 2000; 5(5):267–75. Epub 2001/01/06. PMID: 11142602. 22. Drexler N, Washington CH, Lovegrove M, Grady C, Milord MD, Streit T, et al. Secondary mapping of lymphatic filariasis in Haiti-definition of transmission foci in low-prevalence settings. PLoS Negl Trop Dis. 2012; 6(10):e1807. Epub 2012/10/17. https://doi.org/10.1371/journal.pntd.0001807 PMID: 23071849; PubMed Central PMCID: PMC3469481. 23. Krentel A, Fischer PU, Weil GJ. A review of factors that influence individual compliance with mass drug administration for elimination of lymphatic filariasis. PLoS Negl Trop Dis. 2013; 7(11):e2447. Epub 2013/11/28. https://doi.org/10.1371/journal.pntd.0002447 PMID: 24278486; PubMed Central PMCID: PMC3836848. 24. Adams AM, Vuckovic M, Birch E, Brant TA, Bialek S, Yoon D, et al. Eliminating Neglected Tropical Dis- eases in Urban Areas: A Review of Challenges, Strategies and Research Directions for Successful Mass Drug Administration. Trop Med Infect Dis. 2018; 3(4). Epub 2018/11/25. https://doi.org/10.3390/ tropicalmed3040122 PMID: 30469342; PubMed Central PMCID: PMC6306919. 25. Won KY, Robinson K, Hamlin KL, Tufa J, Seespesara M, Wiegand RE, et al. Comparison of antigen and antibody responses in repeat lymphatic filariasis transmission assessment surveys in American Samoa. PLoS Negl Trop Dis. 2018; 12(3):e0006347. Epub 2018/03/10. https://doi.org/10.1371/journal. pntd.0006347 PMID: 29522520; PubMed Central PMCID: PMC5862496. 26. Rao RU, Samarasekera SD, Nagodavithana KC, Dassanayaka TDM, Punchihewa MW, Ranasinghe USB, et al. Reassessment of areas with persistent Lymphatic Filariasis nine years after cessation of mass drug administration in Sri Lanka. PLoS Negl Trop Dis. 2017; 11(10):e0006066. Epub 2017/10/31. https://doi.org/10.1371/journal.pntd.0006066 PMID: 29084213; PubMed Central PMCID: PMC5679644. 27. Kubofcik J, Fink DL, Nutman TB. Identification of Wb123 as an early and specific marker of Wuchereria bancrofti infection. PLoS Negl Trop Dis. 2012; 6(12):e1930. Epub 2012/12/14. https://doi.org/10.1371/ journal.pntd.0001930 PMID: 23236529; PubMed Central PMCID: PMC3516582. 28. Riches N, Badia-Rius X, Mzilahowa T, Kelly-Hope LA. A systematic review of alternative surveillance approaches for lymphatic filariasis in low prevalence settings: Implications for post-validation settings. PLoS Negl Trop Dis. 2020; 14(5):e0008289. Epub 2020/05/13. https://doi.org/10.1371/journal.pntd. 0008289 PMID: 32396575; PubMed Central PMCID: PMC7217451. 29. Hamlin KL, Moss DM, Priest JW, Roberts J, Kubofcik J, Gass K, et al. Longitudinal monitoring of the development of antifilarial antibodies and acquisition of Wuchereria bancrofti in a highly endemic area of Haiti. PLoS Negl Trop Dis. 2012; 6(12):e1941. Epub 2012/12/14. https://doi.org/10.1371/journal. pntd.0001941 PMID: 23236534; PubMed Central PMCID: PMC3516578. 30. Wilson NO, Badara Ly A, Cama VA, Cantey PT, Cohn D, Diawara L, et al. Evaluation of Lymphatic Fila- riasis and Onchocerciasis in Three Senegalese Districts Treated for Onchocerciasis with Ivermectin. PLoS Negl Trop Dis. 2016; 10(12):e0005198. Epub 2016/12/08. https://doi.org/10.1371/journal.pntd. 0005198 PMID: 27926918; PubMed Central PMCID: PMC5142766. 31. de Souza DK, Owusu IO, Otchere J, Adimazoya M, Frempong K, Ahorlu CS, et al. An evaluation of Wb123 antibody elisa in individuals treated with ivermectin and albendazole, and implementation chal- lenges in Africa. Pan Afr Med J. 2017; 27:65. Epub 2017/08/19. https://doi.org/10.11604/pamj.2017.27. 65.11004 PMID: 28819487; PubMed Central PMCID: PMC5554654. 32. Dorkenoo AM, Koba A, Halatoko WA, Teko M, Kossi K, Yakpa K, et al. Assessment of the usefulness of anti-Wb123 antibody for post-elimination surveillance of lymphatic filariasis. Parasit Vectors. 2021; 14 (1):23. Epub 2021/01/08. https://doi.org/10.1186/s13071-020-04535-y PMID: 33407812; PubMed Cen- tral PMCID: PMC7789272. PLOS Neglected Tropical Diseases \| https://doi.org/10.1371/journal.pntd.0010231 February 25, 2022 18 / 18 PLOS NEGLECTED TROPICAL DISEASES	null
10.1371_journal.pmed.1002934.pdf	Data Availability Statement: A minimal dataset and STATA DO file are included as Supporting information files to enable readers to replicate our analysis.	A minimal dataset and STATA DO file are included as Supporting information files to enable readers to replicate our analysis.	RESEARCH ARTICLE Mortality and recovery following moderate and severe acute malnutrition in children aged 6–18 months in rural Jharkhand and Odisha, eastern India: A cohort study 1,2, Nirmala Nair2, Andrew Copas1, Hemanta Pradhan2, Naomi SavilleID 1, Audrey ProstID Prasanta Tripathy2, Rajkumar Gope2, Shibanand Rath2, Suchitra Rath2, Jolene SkordisID Sanghita BhattacharyyaID 4 2, Harshpal S. SachdevID 3, Anthony CostelloID 2, a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Prost A, Nair N, Copas A, Pradhan H, Saville N, Tripathy P, et al. (2019) Mortality and recovery following moderate and severe acute malnutrition in children aged 6–18 months in rural Jharkhand and Odisha, eastern India: A cohort study. PLoS Med 16(10): e1002934. https://doi. org/10.1371/journal.pmed.1002934 Academic Editor: Zulfiqar A. Bhutta, The Hospital for Sick Children, CANADA Received: January 28, 2019 Accepted: September 9, 2019 Published: October 15, 2019 Copyright: © 2019 Prost et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: A minimal dataset and STATA DO file are included as Supporting information files to enable readers to replicate our analysis. Funding: This study used data collected as part of a trial funded by a Global Health Trial grant funded by UK Medical Research Council (https://mrc.ukri. org; MR/K007270/1), Wellcome Trust (https:// wellcome.ac.uk; 099708/Z/12/Z), and UK Department for International Development (https:// 1 University College London, Institute for Global Health, London, United Kingdom, 2 Ekjut, Chakradharpur, Jharkhand, India, 3 Public Health Foundation of India, New Delhi, India, 4 Sitaram Bhartia Institute of Science and Research, New Delhi, India [email protected] Abstract Background Recent data suggest that case fatality from severe acute malnutrition (SAM) in India may be lower than the 10%–20% estimated by the World Health Organization (WHO). A contempo- rary quantification of mortality and recovery from acute malnutrition in Indian community set- tings is essential to inform policy regarding the benefits of scaling up prevention and treatment programmes. Methods and findings We conducted a cohort study using data collected during a recently completed cluster-ran- domised controlled trial in 120 geographical clusters with a total population of 121,531 in rural Jharkhand and Odisha, eastern India. Children born between October 1, 2013, and February 10, 2015, and alive at 6 months of age were followed up at 9, 12, and 18 months. We measured the children’s anthropometry and asked caregivers whether children had been referred to services for malnutrition in the past 3 months. We determined the incidence and prevalence of moderate acute malnutrition (MAM) and SAM, as well as mortality and recovery at each follow-up. We then used Cox-proportional models to estimate mortality hazard ratios (HRs) for MAM and SAM. In total, 2,869 children were eligible for follow-up at 6 months of age. We knew the vital status of 93% of children (2,669/2,869) at 18 months. There were 2,704 children-years of follow-up time. The incidence of MAM by weight-for- length z score (WLZ) and/or mid-upper arm circumference (MUAC) was 406 (1,098/2,704) per 1,000 children-years. The incidence of SAM by WLZ, MUAC, or oedema was 190 (513/ 2,704) per 1,000 children-years. There were 36 deaths: 12 among children with MAM and six among children with SAM. Case fatality rates were 1.1% (12/1,098) for MAM and 1.2% (6/513) for SAM. In total, 99% of all children with SAM at 6 months of age (227/230) were PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 1 / 16 Mortality and recovery following acute malnutrition in rural India www.gov.uk/government/organisations/ department-for-international-development; MR/ K007270/1), for which NN and AP were Co-PIs. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. Abbreviations: aOR, adjusted odds ratio; CMAM, community-based management of acute malnutrition; HR, hazard ratio; ICDS, Integrated Child Development Services; IQR, interquartile range; LiST, Lives Saved Tool; MAM, moderate acute malnutrition; MTC, malnutrition treatment centre; MUAC, mid-upper arm circumference; RUTF, ready-to-use therapeutic food; SAM, severe acute malnutrition; SD, standard deviation; STROBE, Strengthening the Reporting of Observational Studies in Epidemiology; TEM, technical error of measurement; UR, uncertainty range; WASH, water sanitation and hygiene; WHO, World Health Organization; WHZ, weight-for-height z score; WLZ, weight-for-length z score. alive 3 months later, 40% (92/230) were still SAM, and 18% (41/230) had recovered (WLZ � −2 standard deviation [SD]; MUAC � 12.5; no oedema). The adjusted HRs using all anthro- pometric indicators were 1.43 (95% CI 0.53–3.87, p = 0.480) for MAM and 2.56 (95% CI 0.99–6.70, p = 0.052) for SAM. Both WLZ < −3 and MUAC � 11.5 and < 12.5 were associ- ated with increased mortality risk (HR: 3.33, 95% CI 1.23–8.99, p = 0.018 and HR: 3.87, 95% CI 1.63–9.18, p = 0.002, respectively). A key limitation of our analysis was missing WLZ or MUAC data at all time points for 2.5% of children, including for two of the 36 children who died. Conclusions In rural eastern India, the incidence of acute malnutrition among children older than 6 months was high, but case fatality following SAM was 1.2%, much lower than the 10%–20% estimated by WHO. Case fatality rates below 6% have now been recorded in three other Indian studies. Community treatment using ready-to-use therapeutic food may not avert a substantial number of SAM-related deaths in children aged over 6 months, as mortality in this group is lower than expected. Our findings strengthen the case for prioritising prevention through known health, nutrition, and multisectoral interventions in the first 1,000 days of life, while ensuring access to treatment when prevention fails. Author summary Why was this study done? • Moderate acute malnutrition (MAM) and severe acute malnutrition (SAM) increase a child’s risk of dying. The World Health Organization (WHO) estimates that 10%–20% of children die following SAM, but this range comes from studies conducted over 15 years ago, mostly in African nations, and does not discriminate between children with and without medical complications. • Recent studies from India have suggested that acute malnutrition may carry a lower risk of mortality than estimated by WHO, with substantial implications for the number of lives that could be saved by treating children as outpatients with ready-to-use therapeu- tic food. Unfortunately, most Indian studies conducted to date have been cross-sec- tional, small, or did not examine mortality risk with mid-upper arm circumference (MUAC), a commonly used anthropometric indicator to identify children with MAM or SAM. What did the researchers do and find? • We conducted a cohort study as a secondary data analysis nested within a recently com- pleted cluster-randomised controlled trial in 120 geographical clusters of Jharkhand and Odisha, two states of eastern India with a high prevalence of child undernutrition. • We determined the incidence and prevalence of MAM and SAM, mortality, and recov- ery among 2,869 children followed up from 6 to 18 months of age. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 2 / 16 Mortality and recovery following acute malnutrition in rural India • We found that SAM carried a lower case fatality rate (1.2%) than expected from WHO estimates (10%–20%), echoing results from three other Indian studies, which found case fatality rates ranging from 2.7% to 5.2% among children older than 6 months. We also found that over 40% of children with MAM and SAM remained acutely malnourished for over 3 months. What do these findings mean? • Our data add to the accruing evidence that, in India, high rates of acute malnutrition in children over 6 months old are accompanied by lower case fatality rates than those reported in older, largely African studies. • Overall, our findings suggest a need to prioritise the prevention of undernutrition by supporting known health, nutrition, and multisectoral interventions in the first 1,000 days of life. Introduction In 2017, around 50 million children under 5 years globally were acutely malnourished [1]. Severe acute malnutrition (SAM) is defined as weight-for-height z score (WHZ)/weight-for- length z score (WLZ) < −3 standard deviation (SD) of the World Health Organization (WHO) 2006 Growth Standards, or mid-upper arm circumference (MUAC) < 11.5 cm [2]. Moderate acute malnutrition (MAM) is defined as WHZ/WLZ � −2 and < −3 or MUAC � 11.5 cm and < 12.5 cm [2]. Both MAM and SAM increase children’s risk of dying: a pooled analysis of 10 longitudinal studies from Asia, Africa, and South America found that children with moderate and severe wasting had 3- and 9-fold increased mortality rates, respec- tively, compared with children with WHZ > −1 [3]. Drawing on data collected mostly from hospitalised children before the advent of community-based management of acute malnutri- tion (CMAM), WHO estimated that between 10% and 20% of children with SAM generally die within 2–3 months without treatment, without discriminating between children with and without medical complications [4–6]. Nearly 18% of all children under 5 years globally live in India (121 out of 679 million), as do half of all children under 5 years affected by wasting (around 25.5 million out of 50.5 million) [1,7]. Indian policy makers, scholars, and activists relentlessly call for increased public funding to support the health, nutrition, sanitation, and social protection interventions that could reduce undernutrition [8–10]. Despite these efforts, scaling up preventive supplementary nutrition and detecting children with acute malnutrition has proved challenging. The Inte- grated Child Development Services’ (ICDS) Anganwadi (nutrition) workers should normally give supplementary food to pregnant and breastfeeding women, children under 3 years, and adolescent girls. They should also measure children’s WHZ monthly and refer severely wasted children to malnutrition treatment centres (MTCs) [10]. In 2014, however, a nationally repre- sentative survey found that only 21% of children aged 6–35 months received supplementary food at least 21 days per month [11]. When children with SAM were found and referred to MTC, families often feared the length and cost of hospital care [12]. There are currently no national-level data on the proportion of children with SAM who are identified and treated and few data on mortality following acute malnutrition in the community [9]. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 3 / 16 Mortality and recovery following acute malnutrition in rural India Using the Lives Saved Tool (LiST), the 2013 Lancet Commission on Maternal and Child Undernutrition estimated that scaling up outpatient feeding with ready-to-use therapeutic foods (RUTFs) for MAM and SAM at 90% coverage could save an estimated 435,210 children under 5 years every year [13,14]. Previous studies have suggested that mortality from acute malnutrition may be lower in India than in other settings [15–17]. Measuring actual mortality rates from MAM and SAM for children in India is critical to understand how many deaths could be prevented by scaling up treatment. In this analysis, we estimated the prevalence and incidence of MAM and SAM, referrals, mortality, and recovery in a cohort of children aged 6–18 months in Jharkhand and Odisha, two Indian states with high levels of undernutrition, in the context of ordinary detection and treatment protocols from the ICDS, and with no ongoing CMAM programme. Methods Design We conducted a cohort study as a secondary data analysis nested within a cluster-randomised controlled trial conducted between 2013 and 2016. The original trial tested a community inter- vention to improve linear growth among children under 2 years in rural eastern India through a participatory learning and action cycle of meetings with women’s groups and monthly visits for all mothers from the third trimester of pregnancy until a child’s second birthday. The trial was registered as ISCRTN 51505201 and Clinical Trials Registry of India number 2014/06/ 004664 [18]. This cohort study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 STROBE Checklist). Setting The trial from which our cohort data are drawn took place in two districts: West Singhbhum, in Jharkhand, and Kendujhar, in Odisha. Both districts had high levels of child undernutrition: in 2015–2016, 38% of children under 5 years in West Singhbhum were wasted (13.5% severely wasted), as were 19% of children in Kendujhar (5% severely wasted) [7]. Over 75% of the pop- ulation in the trial study areas belonged to Adivasi communities (Hindi: original inhabitants— i.e., indigenous), also known as Scheduled Tribes. Most families were engaged in agricultural labour, and 46% of women could read [18]. In total, 61% of households had electricity, and 1% had an improved toilet [18]. ICDS had unequal coverage in the study areas: we witnessed mul- tiple interruptions in the provision of iron and folic acid, deworming tablets, and supplemen- tary nutrition between 2013 and 2016. In the study clusters, village-based Anganwadi workers weighed children whose mothers attended monthly Village Health and Nutrition Days, offered supplementary nutrition to moderately underweight children, and referred those who were severely underweight to the MTC. Three treatment centres served the study communities: two in West Singhbhum and one in Kendujhar. They admitted children with WHZ < −3 SD or MUAC < 11.5 cm for inpatient treatment. Participants In the trial that provided data for this study, 120 clusters were purposively created to approxi- mate Anganwadi worker catchment areas of 1,000 people each [19]. Community-based infor- mants identified all women in the third trimester of pregnancy residing in these clusters and invited them to participate in the study. All women who agreed were visited again within 72 hours of the birth, as well as 3, 6, 9, 12, and 18 months after the birth. For this cohort study, eli- gible participants were all singleton children born to pregnant women recruited between PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 4 / 16 Mortality and recovery following acute malnutrition in rural India October 1, 2013, and February 10, 2015 (the trial recruitment period), and who were alive at 6 months of age. Data collectors made a minimum of three attempts to find children at each fol- low-up. We excluded data from infants younger than 6 months for three reasons. First, these infants are not currently targeted by community-based detection, referral, and treatment pro- grammes for acute malnutrition in India. Second, other studies on mortality and recovery fol- lowing MAM and SAM in India included children older than 6 months, and we wanted our analyses to be comparable with these. Finally, we had substantial missing data for WLZ or MUAC within 72 hours of birth (39.6%, or 1,188/3,001) because many mothers and infants were still in hospital or away from their homes; these missing data made it challenging to explore the association between birth anthropometry and mortality. Variables We defined an episode of SAM as being newly identified with WLZ < −3, MUAC < 11.5 cm, and/or bilateral pitting oedema at 6, 9, or 12 months after not being SAM or missing at the pre- vious follow-up. Similarly, we defined an episode of MAM as being newly identified with WLZ between −3 and −2 and/or MUAC between 11.5 and 12.5 cm at 6, 9, or 12 months after not being MAM or missing at the previous follow-up. Children could be acutely malnourished at two or more consecutive follow-ups; we reported such cases as a single episode of acute malnu- trition. For example, a child found with SAM at both 6 and 9 months was counted as having a single episode of SAM. A child with SAM at 6 months who recovered at 9 months but was SAM again at 12 months was counted as having two episodes. Recovery was defined as WLZ � 2 SD, MUAC � 12.5, and no oedema. We considered a child to have died from acute malnutrition if they were acutely malnourished at the last follow-up before death (e.g., if a child was acutely malnourished at 6 months and died before the next follow-up at 9 months). If a child was SAM at 6 months, recovered at 9 months, and died at 10 months, the child was not considered to have died from SAM. The case fatality rate was calculated from among inci- dent cases of MAM/SAM only. Measurement Data collectors received 5 days of training on anthropometry, followed by 7 days of field prac- tice. They measured children’s length using Shorr boards with a precision of 0.1 cm, their weight using Tanita BD-590 scales with 10-g graduations, and MUAC using UNICEF tapes with a precision of 0.1 cm. Data collectors also took part in three technical error of measure- ment (TEM) exercises: two before and one during the trial. In the first TEM, reliability coeffi- cient values (R) were consistently over 0.95 for weight but below 0.5 for length and MUAC, which led to retraining and two further TEM exercises for length and MUAC. In the second and third TEMs, the average R was 0.98 for length and 0.78 for MUAC [20]. The third TEM took place in November 2014—i.e., nearly halfway through data collection. At each follow-up, data collectors checked whether the child was alive and asked whether they had ever been referred to an Anganwadi worker or MTC since the last visit. In total, 11% of interviews and measurements were conducted with a supervisor present. We had information on whether the children were alive at the last visit before migration. Data collectors sought to find each mother at least three times at each follow-up. Data collectors referred all children with WLZ < −3, oedema, or MUAC < 11.5 cm to local Anganwadi workers. This method was chosen to ensure that we did not create a parallel system for referrals to onwards facility-based treatment in MTC or bypass supplementary nutrition programmes. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 5 / 16 Mortality and recovery following acute malnutrition in rural India Statistical methods We had no formal written analysis plan but took key decisions and drafted dummy tables prior to the analysis. All children recruited in the original trial were included in the analysis. We described the mothers’ and children’s socioeconomic characteristics and then examined differences between children retained in the cohort and those lost to follow-up using chi- squared tests for categorical variables and t tests for continuous variables. We used the zscore06 macro in Stata 14 to generate z scores of 2006 WHO Growth Standards and excluded WLZ values < −5 and > 5 as implausible, as recommended in WHO guidance [21]. We checked and flagged any cases in which height and weight did not increase over an interval of 3 months. We then calculated the prevalence and incidence of MAM and SAM at 6, 9, 12, and 18 months. Within each time period (6–9 months, 9–12 months, 12–18 months), we checked whether surviving children remained SAM, became MAM, or recovered from acute malnutri- tion at the next follow-up. We used Cox-proportional models to estimate minimally and fully adjusted hazard ratios (HRs) with 95% CIs for mortality following MAM and SAM by WLZ, MUAC, and oedema (as relevant) combined and then for individual indicators [22]. The variables indicating MAM and SAM were time varying and updated at each follow-up time to reflect their impact on the risk of death only during an episode. In this analysis, we included the effect of prevalent MAM and SAM. We used multiple imputation by chained equations to account for missing measure- ments of WLZ or MUAC for children who had measurements from at least one follow-up. Thirty imputations sets were created, with missing values imputed conditional on values for the same variable at other times and also on whether the child subsequently died and the time of death or censoring. In all Cox-proportional models, we used robust standard errors to account for correlation within clusters (villages). We included data from both the intervention and control arms of the trial to maximise power and present HRs both adjusted and unad- justed for trial allocation. The intervention reduced infant mortality (adjusted odds ratio [aOR]: 0.63; 95% CI 0.39–1.00, p = 0.05) and underweight at 18 months (aOR: 0.81; 95% CI 0.66–0.99, p = 0.04) but did not reduce wasting (aOR: 0.88; 95% CI 0.71–1.08, p = 0.22) or increase mean MUAC at 18 months (adjusted difference: 0.05; 95% CI −0.08 to 0.186, p = 0.44). Adjusted analyses included variables for child sex, trial allocation arm, and district. A minimal data set and STATA DO file are included to enable readers to replicate our analyses (S1 Data and S1 DO File). Ethical considerations We sought women’s individual informed consent in writing or by thumbprint during the enrolment interview in pregnancy and verbally before all subsequent interviews. The trial was reviewed and approved by the research ethics committee of the Public Health Foundation of India (June 2013, TRC-IEC-163/13), an independent ethics committee linked to Ekjut (May 2013, reference IEC/EKJUT/01), and University College London’s Research Ethics Committee (June 2013, reference 1881/002). The independent ethics committee linked to Ekjut approved this secondary data analysis after the completion of the trial. Results Retention and participants’ characteristics Fig 1 shows the retention of participants in the study. In total, 2,869 singleton infants were eli- gible at 6 months, of which 36 died between 6 and 18 months, and 56 migrated permanently. We found 93% (2,669/2,869) of all children eligible at 6 months. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 6 / 16 Mortality and recovery following acute malnutrition in rural India Fig 1. Flowchart describing the retention of children in the study. MUAC, mid-upper arm circumference; WLZ, weight-for-length z score. https://doi.org/10.1371/journal.pmed.1002934.g001 Table 1 describes mothers’ and children’s characteristics by follow-up status. Around half of infants were born in Kendujhar District in Odisha, and the other half were born in West Singhbhum, Jharkhand. Over two-thirds were born in Adivasi families. In total, 50.5% of infants were female, and 49.5% were male. Mothers’ mean height was 1.50 m (SD: 0.07), and children’s mean birth weight was 2.57 kg (SD: 0.4). We compared the characteristics of chil- dren with known vital status at 18 months with those of children lost to follow-up at 18 months. Children lost to follow-up were more likely to be from Jharkhand (72% of children lost to follow-up versus 48% of children with known status), female (59% of children lost to follow-up versus 50% of children with known status), and from poorer households (median multidimensional poverty score: −0.90 interquartile range [IQR]: −1.69, 0.63 for children lost to follow-up versus −0.56, IQR: −1.64, 1.39 for children with known status at 18 months). Case fatality and mortality incidence following MAM and SAM Children had 1,098 episodes of MAM beginning between 6 and 18 months and a MAM inci- dence rate of 406 per 1,000 children-years (1,098/2,704 children-years). In all, 12 children died during an incident MAM episode, five within 6 months of the start of the episode, and seven after 6 months. The case fatality from MAM was 0.4% (5/1,098) within 6 months and 1.1% (12/1,098) overall. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 7 / 16 Mortality and recovery following acute malnutrition in rural India Table 1. Characteristics of participants. Characteristics N District (state), n (%) West Singhbhum (Jharkhand) Kendujhar (Odisha) Class or caste status, n (%) Scheduled Tribe Scheduled Caste Other Backward Class Other Multidimensional poverty scorec, median (interquartile range) Mother’s age, mean (SD) Maternal height in m, mean (SD) Parity, mean (SD) Child sex Female Male Children eligible at 6 months 2,869 Children with known status at 18 monthsa 2,669 Children lost to follow-up at 18 monthsb 200 1,425 (49.7) 1,444 (50.3) 2,192 (76.4) 188 (6.5) 484 (16.9) 5 (0.2) 1,281 (48.0) 1,388 (52.0) 2,033 (76.2) 178 (6.7) 453 (17.0) 5 (0.2) 144 (72.0) 56 (28.0) 159 (79.5) 10 (5.0) 31 (15.5) 0 (0) −0.63 (−1.64,1.23) −0.56 (−1.64,1.39) −0.90 (−1.69, 0.63) 24.0 (4.6) 1.50 (0.07) 2.4 (1.6) 1,450 (50.5) 1,419 (49.5) 24.0 (4.6) 1.50 (0.07) 2.4 (1.6) 1,332 (49.9) 1,337 (50.1) 23.7 (4.5) 1.50 (0.07) 2.5 (1.5) 118 (59.0) 82 (41.0) aChildren with known status at 18 months include those found alive at 18 months and those who died during follow-up. bLost to follow-up is defined as not found at 18 months and not dead. cThe Multidimensional Poverty Index comprised data on household assets, roof and floor materials, any deaths to a child under 5 years in the household, and any school-age children out of school, following the methodology described by Alkire and colleagues, which involves principal component analysis [22]. Abbreviation: SD, standard deviation https://doi.org/10.1371/journal.pmed.1002934.t001 Children had 513 episodes of SAM beginning between 6 and 18 months and a SAM inci- dence rate of 190 per 1,000 children-years (513/2,704 children-years). In total, six children died during an incident SAM episode, four within 6 months of the start of the episode, and two after 6 months. The case fatality for SAM was 0.8% (4/513) within 6 months and 1.2% overall (6/513). Table 2 shows the mortality incidence rates for MAM and SAM, as well as unadjusted and adjusted associations between MAM and SAM incidence and short-term (i.e., within the epi- sode) mortality. A total of 68 children could not be included in analyses with multiple imputa- tion: 10 had no WLZ measurement at any follow-up, 55 had no MUAC measurement at any follow-up, and three were immediately censored at the 6 months’ visit. S1 Table shows the number of missing values for each WLZ, MUAC, and oedema variable. The mortality inci- dence rate for MAM was 19.1 per 1,000 children-years. The mortality incidence rate for SAM was 20.1 per 1,000 children-years. The adjusted HR for MAM using all anthropometric indica- tors was 1.43 (95% CI 0.53–3.87, p = 0.480) and 2.56 (95% CI 0.99–6.70, p = 0.052) for SAM (overall p = 0.073). Lower MUAC, lower WLZ, and the presence of oedema were associated with increased risk of mortality (p = 0.003, p = 0.060, and p = 0.059, respectively). Nutritional status among survivors S2 Table describes the prevalence of MAM and SAM at the start of each follow-up period (i.e., at 6, 9, and 12 months), along with the vital and nutritional status of children at the next fol- low-up. In total, 99% of all children with SAM at 6 months (227/230) were alive after 3 months, 40% (92/230) were still SAM, and 18% (41/230) had ‘recovered’ (WLZ � −2 SD; PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 8 / 16 Mortality and recovery following acute malnutrition in rural India Table 2. Associations between anthropometric indicators of MAM or SAM and short-terma mortality. Anthropometric indicators Mortality incidence per 1,000 children- years Minimally adjusted hazard ratios (95% CI), by categoryb p Minimally adjusted hazard ratios (95% CI), overallb p Fully adjusted hazard ratios (95% CI) by categoryc p Fully adjusted hazard ratios (95% CI), overallc p All indicators No acute MAM SAM WLZ �−2 �−3 and <−2 <−3 MUAC �12.5 �11.5 and <12.5 <11.5 Oedema No Yes 9.2 19.1 19.7 9.3 21.1 25.0 10.3 24.4 11.3 12.2 67.4 1 1.40 (0.52–3.77) 2.23 (0.80–6.19) 1 1.14 (0.46–2.83) 2.80 (0.94–8.37) 1 3.81 (1.57–9.24) 1.95 (0.24–15.5) 1 0.508 1.43 (0.53–3.87) 0.480 0.124 1.47 (0.89–2.43) 0.128 2.56 (0.99–6.70) 0.052 1.57 (0.96–2.55) 0.070 1 0.773 1.20 (0.47–3.04) 0.702 0.064 1.53 (0.88–2.65) 0.131 3.33 (1.23–8.99) 0.018 1.65 (0.98–2.78) 0.061 1 0.003 3.87 (1.63–9.18) 0.002 0.527 2.13 (1.25–3.62) 0.005 1.89 (0.23–15.2) 0.551 2.10 (1.25–3.53) 0.005 1 1 4.79 (0.61–37.6) 0.136 5.87 (0.94–36.5) 0.058 aHazard ratios reflect the impact of a malnutrition indicator until its value is updated at a subsequent follow-up visit. bAdjusted for clustering using robust standard errors. cAdjusted for child’s sex, trial allocation, district, and clustering using robust standard errors. Abbreviations: MAM, moderate acute malnutrition; MUAC, mid-upper arm circumference; SAM, severe acute malnutrition; WLZ, weight-for-length z score https://doi.org/10.1371/journal.pmed.1002934.t002 MUAC � 12.5; no oedema). Only 5% (32/584) of children with SAM at 6, 9, or 12 months were ever referred to an MTC. The coverage of preventive nutrition services was also unequal: for example, 66% (1,917/2,869) of children aged 6 months were weighed by an Anganwadi worker, but only 13% (382/2,869) had mothers who received nutrition counselling. Outcomes up to 12 months after MAM and SAM Fig 2 shows outcomes at 9, 12, and 18 months for children who were not undernourished or had MAM or SAM at 6 months. These were derived after multiple imputation for missing WLZ, MUAC, and oedema data. The proportions of children with MAM or SAM at 6 months who had died by 18 months were 1.3% and 1.1%, respectively. These were not very different, in absolute terms, to the proportion of non-acutely malnourished children who died (0.8%). Recovery from MAM and SAM was relatively low and stable over time. Recovery out of acute undernutrition was similar for MAM or SAM, particularly by 18 months. Discussion We found a high incidence of MAM and SAM among rural, largely tribal communities of Jharkhand and Odisha. A substantial proportion (>40%) of children with MAM and SAM remained acutely malnourished for over 3 months, and 99% of children with SAM at 6 months survived for 3 months or more, suggesting that, in this context, MAM and SAM are persistent. WLZ < −3 was associated with increased mortality risk, as expected, but MUAC � 11.5 and < 12.5 was associated with a greater mortality risk than MUAC < 11.5. The finding that children with MUAC � 11.5 and < 12.5 had a higher mortality risk is counterintuitive and PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 9 / 16 Mortality and recovery following acute malnutrition in rural India Fig 2. Long-term outcomes for children with and without acute malnutrition at 6 months, using a data set with multiple imputation. MAM, moderate acute malnutrition; SAM, severe acute malnutrition. https://doi.org/10.1371/journal.pmed.1002934.g002 deserves further investigation. Anganwadi workers referred only 5% of SAM cases to MTC, highlighting a considerable treatment gap. Our most salient finding is that SAM carried a much lower case fatality rate (1.2%) than expected from older, largely African studies and WHO estimates (10%–20%). If the two deceased children for whom we did not have any anthropometry data had died following SAM, the case fatality rate would have been 1.5% (8/513), still much below 10%–20%. Three other Indian studies with different designs but broadly similar inclusion criteria offer data for comparison. The first was a follow-up study of children initially screened as having an MUAC < 11.5 cm as part of a multicentric trial in poor urban and rural communities in New Delhi, Udaipur, and Vellore [23]. Case fatality among children aged 6–23 months with MUAC < 11.5 cm was 4.3% (6/140) after 12 months [24]. A second study conducted in rural Uttar Pradesh found a case fatality of 2.7% among 409 children aged 6 months to 5 years fol- lowed up between 0.6 and 17.8 months after severe wasting [25]. In a third study from Bihar, Burza and colleagues followed up defaulters from a WHO-standard RUTF programme who had MUAC < 11.5 cm [26]. Case fatality among these defaulters—who tended to be younger children and girls—was 5.2% within 18 months (36/692). All of these case fatality rates, with varying durations of follow-up, are below WHO estimates (10%–20%) and closer to the 4% (range: 2%–7%) case fatality rate that a 2013 expert panel considered normal in the context of a CMAM programme [27]. There are at least four possible explanations for the considerable difference between case fatality rates in these recent Indian studies and the 10%–20% estimates drawn from older, largely African studies. First, the studies upon which WHO estimates are based included mostly hospitalised children who were more likely to have SAM with medical complications and, therefore, higher mortality rates [6]. A second possible explanation is that the most vulnerable children in our rural study areas died before reaching 6 months. In the trial’s control clusters, the infant mortality rate was 64 PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 10 / 16 Mortality and recovery following acute malnutrition in rural India per 1,000 live births; 64% (116/181) of all infant deaths across both trial areas occurred in the first month of life, and 86% (157/181) occurred in the first 6 months [18]. In 2015, the preva- lence of low birth weight per 100 births was greater in southern Asia than in sub-Saharan Africa (26.4%; uncertainty range [UR]: 18.6–35.2 versus 14.0%, UR: 12.2–17.2) [28]. Around half of neonatal deaths in India (370,000 of 698,000 neonatal deaths in 2015) are caused by pre- maturity or low birth weight, and these factors also accounted for 46.1% of all under-5 deaths in 2017 (95% uncertainty interval: 44.5–47.6) [29,30]. In rural India, undersize is probably tak- ing its heaviest toll in the first 6 months of life, before treatment with RUTF becomes relevant. A third source of variation in case fatality rates may lie in differences in infection burdens between South Asian and African settings. In 2017, the disability-adjusted life years rate for communicable maternal, neonatal, child, and nutritional diseases was 26,492 per 100,000 in sub-Saharan Africa, roughly double that in South Asia (13,197 per 100,000) [31]. HIV preva- lence is 4.1% among 15–49-year-olds in sub-Saharan Africa versus 0.2% in South Asia, includ- ing India, and several studies have demonstrated that HIV infection interacts with undernutrition to increase the risk of death [31–33]. A final explanation may lie in the developmental origins of adult health and disease. Moth- ers in our study areas had an average height of 1.5 m [17]. During pregnancy, half of them had an MUAC < 23 cm, and one-third of them did not have diets with minimum diversity [17]. Around one-third (33%) of infants were low birth weight, and over 65% were stunted by 18 months [18]. These data reflect intergenerational, chronic undernutrition inflicted by poverty [34]. Indian infants often have a ‘thin-fat’ phenotype: they have small abdominal viscera and low muscle mass but preserve body fat in utero that can track into birth and young adulthood, depending on the nutritional environment [35]. Studies have found that body fat–related mea- surements such as skinfold thickness, abdominal circumference, and MUAC are often similar between Indian babies and Western infants, despite low birth weight [36].The hidden adipos- ity of Indian infants may give small infants a survival advantage by acting as a form of energy reserve available to maintain body temperature and brain development in times of nutritional deprivation, albeit with adverse consequences for noncommunicable disease risk in later life [37]. It is therefore possible that thinness in infancy and early childhood does not carry the same mortality risk in this setting as in others. The interplay between undernutrition, immune dysfunction, and mortality risk is complex, as summarised in recent reviews [38–40]. Integrat- ing longitudinal immune assessments into ongoing trials of nutrition and water sanitation and hygiene (WASH) interventions will help elucidate when and how immune dysfunction and undernutrition interact and help better target public health interventions to lower mortality risk. Our study had several strengths. To our knowledge, it is the largest Indian community- based longitudinal study of mortality following MAM and SAM in the absence of management of uncomplicated SAM with RUTF for children older than 6 months. The three TEM exercises undertaken by the data collection team enhance our trust in the accuracy of anthropometry data. We had information about vital status for 93% of infants eligible for follow-up at 6 months by the end of the study period. Our study had three main limitations. We had missing anthropometry data for some chil- dren, including fully missing data for two of the 36 children who died. We attempted to address possible bias caused by missing data by using multiple imputation to provide more accurate estimates of mortality risk but were not able to use imputation for 68 children who had no WHZ or MUAC measurements at any time point. A second limitation is that we did not follow children up until 59 months of age. The study from Uttar Pradesh by Kapil and col- leagues found a case fatality rate of 2.7% among children aged 6–60 months; this is higher than among the younger children in our study but still not within WHO’s estimated range [25]. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 11 / 16 Mortality and recovery following acute malnutrition in rural India Our results are therefore likely to be generalisable to young children aged 6–24 months in other rural, underserved Indian settings and may provide a modest underestimate of case fatal- ity for children aged 6–59 months. Finally, monitoring and referral systems for SAM in the original trial differed from those available in the general population, and this may have con- tributed to lowering mortality. We have little evidence to support this hypothesis, however: infant mortality in the trial’s control area (63 per 1,000 live births) was similar to that reported in the 2011–2012 Annual Health Surveys for our two study districts (56 per 1,000 live births in Kendujhar and 58 in West Singhbhum), suggesting that the trial monitoring systems did not lead to large reductions in mortality in and of themselves [41]. If data from poor urban and rural communities in Delhi, Bihar, Uttar Pradesh and our own data from Jharkhand and Odisha are representative of the contemporary Indian situation, then case fatality rates following SAM in Indian children older than 6 months are substantially lower than those estimated from earlier studies conducted in African countries. In other Indian studies assessing progress following inpatient and/or community care for SAM, rates of recovery were often lower than recommended in SPHERE guidance, and relapse was common [42–44]. For example, in a recent follow-up study of 150 children discharged from MTCs in Jharkhand, 52% of children had relapsed into severe wasting 2 months after discharge [45]. In light of these findings, scaling up outpatient RUTF treatment for children over 6 months with- out concurrently strengthening ICDS, WASH, and the social protection interventions that support the intergenerational prevention of undernutrition may not have the large effects on mortality predicted by the 2013 Lancet Maternal and Child Nutrition Series in India and other South Asian contexts [17]. Prevention can work: a recent intervention combining crèches, par- ticipatory meetings with women’s groups, and home visits reduced wasting, underweight, and stunting among children under 3 years old in Jharkhand and Odisha [46]. A recent UNICEF- led consultation aptly called for a new approach to wasting in South Asia, one that positions prevention as the ‘first priority’ and ensures access to treatment when prevention fails [47]. Our findings support such a strategy. In conclusion, we found a high incidence of acute malnutrition among children older than 6 months in rural eastern India but a lower than expected case fatality rate following SAM. This finding has been replicated in at least three other Indian studies. Given that the risk of mortality is lower than expected among children older than 6 months and that many deaths occur because of prematurity or low birth weight during the neonatal period, outpatient treat- ment for SAM using RUTF for children over 6 months may be too late to avert a substantial number of deaths from undernutrition in Indian children. This further strengthens the case for prioritising prevention through known health, nutrition, and multisectoral interventions in the first 1,000 days of life. Supporting information S1 STROBE Checklist. STROBE, Strengthening the Reporting of Observational Studies in Epidemiology. (DOCX) S1 Data. (CSV) S1 DO File. (DO) S1 Table. (DOCX) PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 12 / 16 Mortality and recovery following acute malnutrition in rural India S2 Table. (DOCX) Acknowledgments We thank the mothers and children who took part in the study, as well as the SPKs, ANMs, ASHAs, and Anganwadi workers in participating clusters. We thank Dr Hembrom, who sup- ported referrals for severely acutely malnourished children. We thank Professor Clive Osmond, from the University of Southampton, for comments on the statistical methods, and Kripa Harper, UCL MSc class of 2017, for sharing her literature review on SAM in India. Author Contributions Conceptualization: Audrey Prost, Andrew Copas, Harshpal S. Sachdev. Data curation: Audrey Prost, Hemanta Pradhan, Shibanand Rath. Formal analysis: Audrey Prost, Andrew Copas, Hemanta Pradhan, Harshpal S. Sachdev. Funding acquisition: Audrey Prost, Nirmala Nair, Andrew Copas, Naomi Saville, Jolene Skor- dis, Sanghita Bhattacharyya, Harshpal S. Sachdev. Investigation: Audrey Prost, Nirmala Nair, Andrew Copas, Prasanta Tripathy, Rajkumar Gope, Shibanand Rath, Suchitra Rath, Sanghita Bhattacharyya, Harshpal S. Sachdev. Methodology: Audrey Prost, Nirmala Nair, Andrew Copas, Prasanta Tripathy, Rajkumar Gope, Shibanand Rath, Suchitra Rath, Harshpal S. Sachdev. Project administration: Audrey Prost, Rajkumar Gope, Shibanand Rath. Resources: Audrey Prost, Nirmala Nair, Prasanta Tripathy. Supervision: Audrey Prost, Hemanta Pradhan, Rajkumar Gope, Shibanand Rath, Suchitra Rath, Anthony Costello, Harshpal S. Sachdev. Visualization: Audrey Prost. Writing – original draft: Audrey Prost. Writing – review & editing: Audrey Prost, Nirmala Nair, Andrew Copas, Naomi Saville, Pra- santa Tripathy, Jolene Skordis, Anthony Costello, Harshpal S. Sachdev. References 1. UNICEF, WHO, World Bank. Joint Child Malnutrition Estimates 2018. New York: UNICEF; 2019 [cited 2019 Jul 15]. Available from: https://data.unicef.org/resources/dataset/malnutrition-data/. 2. Black RE, Victora CG, Walker SP, Bhutta ZA, Christian P, de Onis M, et al. Maternal and child undernu- trition and overweight in low and middle-income countries. Lancet. 2013; 382: 427–451. https://doi.org/ 10.1016/S0140-6736(13)60937-X PMID: 23746772 3. Olofin I, McDonald CM, Ezzati M, Flaxman S, Black RE, Fawzi WW, et al. Associations of suboptimal growth with all-cause and cause-specific mortality in children under five years: a pooled analysis of ten prospective studies. PLoS ONE. 2013; 8(5): e64636. https://doi.org/10.1371/journal.pone.0064636 PMID: 23734210 4. World Health Organization. Community-based management of severe acute malnutrition: a joint state- ment by the World Health Organization, the World Food Programme, the United Nations System Stand- ing Committee on Nutrition and the United Nations Children’s Fund. Geneva: World Health Organization; 2007. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 13 / 16 Mortality and recovery following acute malnutrition in rural India 5. World Health Organization. Guideline: Updates on the management of severe acute malnutrition in infants and children. Geneva: World Health Organization; 2013. 6. Schofield C, Ashworth A. Why have mortality rates for severe malnutrition remained so high? Bull World Health Organ. 1996; 74: 223–229. PMID: 8706239 7. Ministry of Health and Family Welfare, Government of India. National Family Health Survey (NFHS-4), 2015–2016. Mumbai, India: International Institute of Population Sciences; 2017. 8. Menon P, McDonald CM, Chakrabarti S. Estimating the costs of delivering direct nutrition interventions at scale: national and subnational insights from India. Maternal and Child Nutrition. 2016; 12: 169–185. https://doi.org/10.1111/mcn.12257 PMID: 27187914 9. Rao S, Kaul V. India’s integrated child development services scheme: challenges for scaling up. Child: care, health and development. 2018; 34: 31–40. 10. Ministry of Women and Child Development. National Nutrition Mission: Administrative Guidelines. New Delhi, India: Government of India; 2018 Feb. 11. Ministry of Women and Child Development. Rapid Survey on Children: 2013–2014. New Delhi, India: Government of India; 2014. 12. Burtscher D, Burza S. Health-seeking behaviour and community perceptions of childhood undernutri- tion and community management of acute malnutrition (CMAM) programme in rural Bihar, India: a quali- tative study. Public Health Nutrition. 2015; 18: 3234–3243. https://doi.org/10.1017/ S1368980015000440 PMID: 25753193 13. Bhutta ZA, Das JK, Rizvi A, Gaffey MF, Walker N, Horton S, et al. Evidence-based interventions for improvement of maternal and child nutrition: what can be done and at what cost? Lancet. 2013; 382: 452–477. https://doi.org/10.1016/S0140-6736(13)60996-4 PMID: 23746776 14. Clermont A, Walker N. Nutrition Interventions in the Lives Saved Tool (LiST). The Journal of Nutrition. 2017; 11: 2132S–2140S. 15. Kapil U, Sareen N. Management of children with severe acute malnutrition. Indian Pediatrics. 2014; 51: 587–8. 16. Pelletier DL. The relationship between child anthropometry and mortality in developing countries: impli- cations for policy, programs and future research. J Nutr. 1994; 124: 2047S–81S. https://doi.org/10. 1093/jn/124.suppl_10.2047S PMID: 7931716 17. Sachdev HS, Sinha S, Sareen N, Pandey RM, Kapil U. Survival and Recovery in Severely Wasted Under-five Children Without Community Management of Acute Malnutrition Programme. Indian Pediat- rics. 2017; 10: 817–824. 18. Nair N, Tripathy P, Sachdev HS, Pradhan P, Bhattacharyya S, Gope R, et al. Effect of participatory women’s groups and counselling through home visits on children’s linear growth in rural eastern India (CARING trial): a cluster-randomised controlled trial. Lancet Global Health. 2017; 5: e1004–e1016. https://doi.org/10.1016/S2214-109X(17)30339-X PMID: 28911749 19. Nair N, Tripathy P, Sachdev HS, Bhattacharyya S, Gope R, Gagrai S, et al. Participatory women’s groups and counselling through home visits to improve child growth in rural eastern India: protocol for a cluster randomised controlled trial. BMC Public Health. 2015; 15: 384. https://doi.org/10.1186/s12889- 015-1655-z PMID: 25886587 20. Ulijaszek SJ, Kerr DA. Anthropometric measurement error and the assessment of nutritional status. British Journal of Nutrition. 1999; 82: 165–177. https://doi.org/10.1017/s0007114599001348 PMID: 10655963 21. WHO Multicentre Growth Reference Study Group. WHO Child Growth Standards: Length/height-for- age, weight-for-age, weight-for-length, weight-for-height and body mass index-for-age: Methods and development. Geneva: World Health Organization; 2006. 22. Alkire S, Foster JE, Seth S, Santos ME, Roche JM, Ballon P. Multidimensional Poverty Measurement and Analysis. Oxford: Oxford University Press; 2015. 23. Bhandari N, Mohan SB, Bose A, Iyengar SD, Taneja S, Mazumder S, et al. Efficacy of three feeding reg- imens for home-based management of children with uncomplicated severe acute malnutrition: a rando- mised trial in India. BMJ Global Health. 2016; 1: e000144. https://doi.org/10.1136/bmjgh-2016-000144 PMID: 28588982 24. Taneja S, Rongsen-Chandola T, Mohan SB, Mazumder S, Bhandari N, Kaur J, et al. Mid upper arm cir- cumference as a predictor of risk of mortality in children in a low resource setting in India. PLoS ONE. 2018; 13(6): e0197832. https://doi.org/10.1371/journal.pone.0197832 PMID: 29856757 25. Kapil U, Pandey RM, Bansal R, Pant B, Varshney AM, Yadav CP, et al. Mid-upper arm circumference in detection of weight-for-height z score below -3 in children aged 6–59 months. Public Health Nutr. 2018. https://doi.org/10.1017/S1368980017004165 PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 14 / 16 Mortality and recovery following acute malnutrition in rural India 26. Burza S, Mahajan R, Marino E, Sunyoto T, Shandilya C, Tabrez M, et al. Seasonal effect and long-term nutritional status following exit from a Community-Based Management of Severe Acute Malnutrition program in Bihar, India. Eur J Clin Nutr. 2016; 70: 437–44. https://doi.org/10.1038/ejcn.2015.140 PMID: 26330147 27. Lenters LM, Wazny K, Webb P, Ahmed T, Bhutta ZA. Treatment of severe and moderate acute malnu- trition in low- and middle-income settings: a systematic review, meta-analysis and Delphi process. BMC Public Health. 2013; 13: S23. https://doi.org/10.1186/1471-2458-13-S3-S23 PMID: 24564235 28. Blencowe H, Krasevec J, de Onis M, Black RE, An X, Stevens GA, Borghi E, et al. National, regional, and worldwide estimates of low birthweight in 2015, with trends from 2000: a systematic analysis. Lan- cet Glob Health. 2019; 7: e849–60. 29. Million Deaths Study Collaborators. Changes in cause-specific neonatal and 1-59-month child mortality in India from 2000 to 2015: a nationally representative survey. Lancet. 2017; 390: 1972–1980. https:// doi.org/10.1016/S0140-6736(17)32162-1 PMID: 28939096 30. India State-Level Disease Burden Initiative Malnutrition Collaborators. The burden of child and maternal malnutrition and trends in its indicators in the states of India: the Global Burden of Disease Study 1990– 2017. Lancet Child Adolesc Health. 2019; https://doi.org/10.1016/S2352-4642(19)30273-1 31. Global Burden of Disease Collaborative Network. Global Burden of Disease Study 2017 (GBD 2017) Results. Seattle: Institute for Health Metrics and Evaluation (IHME); 2018. 32. Fergusson P, Tomkins A. HIV prevalence and mortality among children undergoing treatment for severe acute malnutrition in sub-Saharan Africa: a systematic review and meta-analysis. Transactions of the Royal Society of Tropical Medicine and Hygiene. 2009; 103: 541–548. https://doi.org/10.1016/j. trstmh.2008.10.029 PMID: 19058824 33. Kerac M, Bunn J, Chagaluka G, Bahwere P, Tomkins A, Collins S, et al. Follow-Up of Post-Discharge Growth and Mortality after Treatment for Severe Acute Malnutrition (FuSAM Study): A Prospective Cohort Study. PLoS ONE. 2014; 9(6): e96030. https://doi.org/10.1371/journal.pone.0096030 PMID: 24892281 34. Shah A, Lerche J, Axelby R, Benbabaali D, Donegan B, Raj J, et al. Ground down by growth: tribe, caste, class and inequality in 21st Century India. London: Pluto Press; 2017. 35. Yajnik CS, Fall CHD, Coyaji KJ, Hirve SS, Rao S, Barker DJ, et al. Neonatal anthropometry: the thin-fat Indian baby. The Pune Maternal Nutrition Study. International Journal of Obesity. 2003; 27: 173–180. https://doi.org/10.1038/sj.ijo.802219 PMID: 12586996 36. Krishnaveni GV, Hill JC, Veena SR, Leary SD, Saperia J, Chachyamma KJ, et al. Truncal Adiposity is Present at Birth and in Early Childhood in South Indian Children. Indian Pediatrics. 2005; 42: 527–538. PMID: 15995269 37. Muthayya S, Dwarkanath P, Thomas T, Vaz M, Mhaskar A, Mhaskar R, et al. Anthropometry and body composition of south Indian babies at birth. Public Health Nutrition. 2006; 9: 896–903. PMID: 17010256 38. Bartz A, Mody A, Hornik C, Bain J, Muehlbauer M, Kiyimba T, et al. Severe Acute Malnutrition in Child- hood: Hormonal and Metabolic Status at Presentation, Response to Treatment, and Predictors of Mor- tality. J Clin Endocrinol Metab. 2014; 99: 2128–2137. https://doi.org/10.1210/jc.2013-4018 PMID: 24606092 39. Maurya R, Bhattacharya P, Dey R, Nakhasi HL. Leptin Functions in Infectious Diseases. Front. Immu- nol. 2018; 9:2741. https://doi.org/10.3389/fimmu.2018.02741 PMID: 30534129 40. Bourke CD, Berkley JA, Prendergast AJ. Immune dysfunction as a cause and consequence of malnutri- tion. Trends Immunol. 2016; 37: 386–398. https://doi.org/10.1016/j.it.2016.04.003 PMID: 27237815 41. Ministry of Home Affairs, Office of the Registrar General and Census Commissioner. Annual Health Survey (2011–2). New Delhi, India: Government of India; 2013. 42. Garg A, Badgaiyan N, Singh K. Integrated program achieves good survival but moderate recovery rates among children with severe acute malnutrition in India. Am J Clin Nutr. 2013; 98:1335–42. https://doi. org/10.3945/ajcn.112.054080 PMID: 24067666 43. Singh K, Badgaiyan N, Ranjan A, Dixit HO, Kaushik A, Kushwala KP, et al. Management of children with severe acute malnutrition: experience of Nutrition Rehabilitation Centers in Uttar Pradesh, India. Indian Pediatrics. 2014; 51:21–5. https://doi.org/10.1007/s13312-014-0328-9 PMID: 24277964 44. Aguayo VM, Jacob S, Badgaiyan N, Chandra P, Kumar A, Singh K. Providing care for children with severe acute malnutrition in India: new evidence from Jharkhand. Public Health Nutrition. 2012; 17: 206–211. https://doi.org/10.1017/S1368980012004788 PMID: 23137752 45. Chaturvedi A, Patwari AK, Soni D. Progress of children with severe acute malnutrition in the malnutrition treatment centre rehabilitation program: evidence from a prospective study in Jharkhand, India. Nutri- tion Journal. 2018; 17: 69. https://doi.org/10.1186/s12937-018-0378-2 PMID: 30021572 PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 15 / 16 Mortality and recovery following acute malnutrition in rural India 46. Gope R, Tripathy P, Prasad V, Pradhan H, Sinha RK, Panda R, et al. Effects of participatory learning and action with women’s groups, counselling through home visits and crèches on undernutrition among children under three years in eastern India: a quasi-experimental study. BMC Public Health. 2019; 19: 962. https://doi.org/10.1186/s12889-019-7274-3 PMID: 31319828 47. UNICEF. Wasting in South Asia: Consultation on building the evidence base for the policy and pro- gramme response. New York: UNICEF; 2018 Nov 16. PLOS Medicine \| https://doi.org/10.1371/journal.pmed.1002934 October 15, 2019 16 / 16	null
10.1088_1361-6552_ad001a.pdf	Data availability statement The data cannot be made publicly available upon publication due to legal restrictions preventing unrestricted public distribution. The data that sup- port the findings of this study are available upon reasonable request from the authors.	Data availability statement The data cannot be made publicly available upon publication due to legal restrictions preventing unrestricted public distribution. The data that support the findings of this study are available upon reasonable request from the authors.	P A P E R iopscience.org/ped Phys. Educ. 59 (2024) 015001 (12pp) Einsteinian gravitational concepts throughout secondary school ∗ Corey McInerney  and Phil Sutton University of Lincoln, School of Mathematics and Physics, Lincoln LN6 7TS, United Kingdom E-mail: [email protected] Abstract Einstein’s theory of relativity is largely thought of as one of the most important discoveries of the 20th century and continues to pass observational tests over 100 years later. Yet, it is Newtonian gravity, a 350 year old formalism proven to be less accurate than relativity, which is taught in schools. It has been shown that Einsteinian gravitational concepts can be well understood by students in both primary and secondary education. In this paper, a cross-section of students from Yr 7 to 13 enrolled in an English secondary school took part in an intervention designed to introduce the idea of gravity from spacetime curvature. The overall aim of this work is to assess the viability of including relativity in the secondary curriculum and to ascertain which year this material would be best placed in. We determine that all year groups where able to appreciate the effects of curvature to some extent. Visual demonstrations aided conceptual understanding at Yr 7–8 level, but this does not have a strong effect on their ideas around the source of the gravitational force. Participants in Yr 9–13 were able to understand concepts beyond those introduced in the demonstrations. However, a deeper understanding of curvature as the source of the gravitational force is not seen until years 12 and 13. We find that those in Yr 13 have the best overall understanding of the concepts introduced during our intervention. Keywords: Einsteinian gravity, relativity, secondary school, conceptual understanding 1. Introduction In 1687, Isaac Newton published his law of uni- versal gravitation [1]; it is commonly taught in schools because it combines the concept of mass ∗ Author to whom any correspondence should be addressed. and weight with everyday phenomenon like fall- ing objects [2]. Gravity is regarded as being one of the key threshold concepts in education used as an indicator of a students understanding of wider physics topics [3, 4]. However, research suggests that a large number of misconceptions about gravity are held by students. These include issues surrounding the direction of gravitational 1361-6552/24/015001+12$33.00 1 © 2023 IOP Publishing Ltd C McInerney and P Sutton force, gravity outside of the Earth and the role of the Sun in celestial dynamics [5–7]. These issues are tied to students not being able to fully grasp where the gravitational force comes from, with some students believing that objects only have a gravitational pull if they are heavy and spherical, or that air is needed in order for gravity to act between objects [8, 9]. Misconceptions can stay with students throughout their schooling, impact- ing confidence, ability and thus students overall enjoyment of science [10]. A new approach pioneered by the Einstein- First project could eliminate some of the con- ceptual issues faced by students. The idea is to introduce principals from modern physics, such as Einstein’s theory of general relativity (GR) into the Australian curriculum at primary and secondary level. A similar project, ReleQuant, in Norway also pioneers the teaching of relativ- ity and quantum mechanics at secondary school [11, 12]. Einstein’s theory of GR provides a mathem- atical description of gravity and offers a qualitat- ive description of how gravity works [13]. A key feature of GR is the non-Euclidean geometry of space and the interplay between time and space. The sophisticated mathematics involved in GR are seen as a barrier to students below degree level. The ReleQuant project bypasses this math- ematical rigor by placing conceptual language at the forefront of the learning process, explor- ing concepts via thought experiments and visual demonstrations. This study explores the impact of a one- off intervention on students’ understanding of Einsteinian gravitational concepts. Similar studies have been performed in Lebanon [14], Australia [15, 16], Indonesia [17] and Italy [18, 19]. Each of those works report that students have a better understanding of gravity from a GR stand point after being introduced to concepts such as curvature, time dilation and spacetime diagrams. The intervention used for this study was designed to fit into a standard 50 min lesson period at a single English state funded secondary school. Within that period, students were taught about how spacetime curvature produces the gravitational force with the help of two hands on activities. The host school is a fully government funded, selective, mixed grammar school located the town of Gainsborough, County Lincolnshire, England. This study took place during February and March 2023 and at that time a total of 1237 students were enrolled at the school. An identical intervention was delivered to all age groups (years 7–13) at this school. Students were issued with questionnaires before and after the intervention and by comparing the results across year groups, we hope to identify the optimum place for GR to be introduced into the secondary science curriculum. In the following section we describe the research methodology in comparison with other works, as well as the content and application method of our intervention. In section 3, we present the results of our research, which are also discussed and analysed. 2. Method 2.1. The host school The school where this research was performed is a fully government funded, selective, mixed grammar school located in England. At the time of this study, a total of 1237 students were enrolled at the school. The UK government web- site (www.find-school-performance-data.service. gov.uk) shows that in 2022, the school was ranked in the top 25% in the UK for academic perform- ance at GCSE level and that 82% of pupils at ∗ the school achieved A-level grades between A and C. 2.2. Comparison with other works In other works, interventions range from short, to full 20 session one off workshops [20], programmes [21, 22]. The offering a longer syl- labus means more content and concepts can be covered, students can learn at their own pace and there are opportunities for deeper learning experi- ences. Nonetheless, it has been shown that one-off interventions have a significant impact on students engagement with science topics and consequently, their future career choices [23, 24]. The intervention used in this paper was designed as a single, one-off session to introduce January 2024 2 Phys. Educ. 59 (2024) 015001 Einsteinian gravitational concepts throughout secondary school Table 1. Number of participants in this study by year group. Key stage Year group Number of students 7 8 9 10 11 12 13 KS3 KS4 A-level Total 59 41 24 20 13 12 14 183 Table 2. An outline of the different topics and concepts covered during our intervention. Topic Details Newton’s law of universal gravitation Gravitational field lines E = mc2 General relativity Every particle attracts every other particle in the Universe. Field lines on the surface of and around celestial bodies. Relation between mass and energy in special relativity, leading to relation between mass and gravity in GR. Space, time and spacetime curvature. Spacetime simulator Gravitational force in GR Force increase with curvature. Triangles on balloons Geometry in GR GPS corrections for Earth’s curvature. Positive and negative curvature. The curvature of the Universe. students to the subject of GR and to explore grav- itational phenomena in the realm of GR. While other works have used participants from a single class, a single year group or a particular key stage (KS), we offered this intervention to all physics students in the school. A total of 183 students from years 7 to 13, choose to take part in this study, the breakdown of which is shown in table 1. Surveying students from one single school allows us to see how students understanding of gravity changes throughout the educational years via this cross-sectional study and offers the oppor- tunity for longitudinal studies in the future follow- ing specific cohort(s). Gravity is taught in England at KS3 (11–14 years old), KS4 (14–16 years old) and A-level (16–18 years old), therefore, any mis- conceptions picked-up at KS3 have the potential to carry forward and impact a students under- standing at A-level. 2.3. The intervention The intervention implemented in this work consisted of 33 identical 50 min workshops in the presence of 8–30 students from a single year group. The same researcher conducted all ses- sions, covering the same material and activities. The intervention began with students being asked to describe gravity in their own words before GR was introduced. Table 2 shows the order that topics were introduced, the material spoken about and where any hands-on activities involved tie-in (highlighted in blue). A common 2.3.1. The spacetime simulator. analogy for the curvature of spacetime uses a sheet of stretchy, elastic material such as Lycra, and some different massed objects. This particular analogy is commonly referred to as ‘the spacetime January 2024 3 Phys. Educ. 59 (2024) 015001 C McInerney and P Sutton Figure 1. The stretched sheet of Lycra used as a space- time simulator. Three 1 kg masses placed in the centre of the sheet provide curvature. simulator’ [25] and has proven to be effective at dispelling misconceptions around gravity [26]. For our demonstration, a large Lycra sheet was stretched flat across a plastic frame as seen in figure 1. Three students were then invited forward to hold the sheet in the air and it was explained to the group that the sheet represents spacetime in the flat Euclidean plane (as per the Newtonian model). Students were then invited forward to roll tennis, squash, and golf balls across it, perceiving that they travel in straight lines and that mass does not affect motion in flat space. A mass heavy enough to deform the sheet was then placed on it and students observed that the balls now roll towards the central mass much like a ball rolling towards the bottom of a hill. By adding more mass to the centre of the sheet and increasing the curvature, students observe that the balls roll towards the central mass much quicker than before. This opens up a discussion around mass, curvature and grav- itational attraction. Additionally, while the sheet is deformed, squash balls were given some sideways velocity. Here, students observe the ball circling around the central mass with motion analogous to an orbit. It was important here to draw attention to the fact that the orbital motion decays due to the loss of energy from friction between the ball and the sheet, whereas orbits in the Universe conserve energy in general and will be continuous. To explore this further, students were shown a plot akin to figure 1(b) from Kaur et al [16] which Figure 2. An example of some of the balloons used by students to demonstrate the effects of curvature on geometry. illustrates that increasing curvature (adding mass to the sheet) changes the distance between two points in spacetime, thus influencing the geometry of the surrounding spacetime. Following the 2.3.2. The geometry of GR. introduction to curvature, the intervention moved to exploring the geometry of spacetime. Pairs of students were issued with differ- ent sized balloons. Each balloon had a triangle drawn on it and the students were tasked with carefully measuring the internal angles of the tri- angles by placing protractors on the balloons sur- face. An example is shown in figure 2. Students of a tri- then compared their results to the 180 angle in Euclidean space to that of one with increased curvature and how that changes the internal angles. These results were then related to applications of non-Euclidean geometry such as the spherical curvature of Earth, it is effect on lines of longitude (such as their convergence at the poles), GR corrections to GPS and the poten- tial geometries of the Universe (hyperbolic, flat, parabolic). ◦ 2.4. Pre/post intervention questionnaire Two sets of questionnaires were used in this research. One disseminated 2 weeks before the intervention (pre-), and one given to students January 2024 4 Phys. Educ. 59 (2024) 015001 Einsteinian gravitational concepts throughout secondary school Table 3. Questions from the pre/post intervention questionnaire. Questions Q1 Can parallel lines meet? Q2 Can the sum of the angles in a triangle be different from 180 Q3 What is gravity? Q4 How do objects move in gravitational fields? Q5 Q6 Does space have a shape? What about the space around heavy objects like stars and planets? Q7 Do you prefer to learn about physics by listening to your teacher, watching demonstrations or doing Isaac Newton is famous for his laws of motion and his law of gravity. What is Albert Einstein famous for? ? ◦ practical work? Q8 Did you enjoy learning about this topic? What did you like/not like? Q9 Are you interested in learning more about gravity and general relativity? Q10 Should Einsteinian physics, like general relativity be included in the curriculum? 2 weeks after the intervention (post-). Following the format of related studies [15, 20, 21], the open- ended questions in the questionnaires examined conceptual understandings and attitudes towards GR. The pre-questionnaire comprised seven ques- tions exploring students’ understanding of grav- ity, its origins, workings, their perspectives on space shape, and opinions on physics teaching methods. The post-questionnaire was nearly identical to the pre-questionnaire, allowing direct compar- ison. It included three additional questions to gain insight into students’ views on the intervention and Einsteinian gravity itself. The questions themselves are open-ended and can be found in table 3. 3. Results and discussion Figure 3 and table 4 show the responses to Q1. The majority of participants (93.8%) said that parallel lines cannot meet. Of the ten that said they can (three from Yr 7, four from Yr 10 and one each from Yr 11, 12 and 13), five of these responses (three Yr 10, one Yr 11 and one Yr 13) used the phrase ‘non-Euclidean’ or ‘not on flat space’, thus showing good knowledge before the intervention. It is likely that these participants have had previous exposure to curved geomet- ries. The total number of ‘yes’ responses moves from 6.2% to 44.2% after the intervention. This is not as much as the 11.3% to 83.6% rise in correct responses to Q2, shown in figure 4 and table 5. The intervention design contributed to this as Q2 relates to the balloon activity, whereas the information about parallel lines also being influ- enced by curved geometry was presented after this activity as part of the slideshow presentation. As such, participants had to pay more attention to grasp this information. Nonetheless, all year groups showed an increase in ‘yes’ responses for both Q1 and Q2 with Yr 9, 11, and 12 having the highest increase in percentage. This is per- haps expected as these cohorts have studied more geometry than younger students and so are better prepared for discussion’s around non-Euclidean geometry. These that participants indicate comprehend information better through visual and hands-on activities. Well-designed prac- tical activities are reported to increase students levels of understanding [27], with practical work seen by students as more interesting and enga- ging than listening to their teacher or watching demonstrations [28]. results Q3 is more open to predisposed misconcep- tions than Q1 or Q2. As seen in figure 5, there were many different responses to this question. We have categorised responses into those which described gravity as a force versus those which did not. It was expected that post-intervention, participants would shift their answers towards describing grav- ity as the ‘bending/warping of spacetime’. 82% of the pre-questionnaire responses described grav- ity as a force, with responses ranging from ‘the January 2024 5 Phys. Educ. 59 (2024) 015001 C McInerney and P Sutton Figure 3. Responses to Q1. Can parallel lines meet? Table 4. Difference in pre- and post-questionnaire responses to Q1 by year group. Pre Post No 93.62% 100.00% 100.00% 78.95% 91.67% 91.67% 90.91% Yes 8.51% 0.00% 0.00% 21.05% 8.33% 8.33% 9.09% No 79.66% 75.76% 33.33% 6.25% 20.00% 0.00% 11.11% Yes 20.34% 21.21% 61.11% 75.00% 80.00% 100.00% 88.89% Yr 7 Yr 8 Yr 9 Yr 10 Yr 11 Yr 12 Yr 13 force that keeps us on Earth’, ‘the force that stops us from floating/keeps us on the ground’ and ‘the force between objects/masses’. Only 1.8% of participants specified that gravity was the cause of spacetime curvature pre-intervention, increasing to only 8.9% post- intervention. Similar studies [26] also struggled to fully influence descriptions of gravity using the spacetime simulator. However looking at table 6 which shows the breakdown of Q3 by year group, reveals that responses related to spacetime curvature increased notably in years 10 (up 17.52%), 12 (up 20.24%), and 13 (up 35.35%). One positive result from table 6 is that par- ticipants associating gravity as an Earth-bound force decreased across all year groups (except Yr 10), particularly in Yr 8 and 11. The data indicates that the KS4 and A-level groups best understood the concept of gravity from space- time curvature. The majority of participants how- ever either did not grasp the concepts introduced (not indicated by the results to Q1 and Q2), or the intervention’s influence was limited by their pre-existing classroom-based learning and opinions. Q4 assess participants understanding of how gravity influences the motion of objects. Presented January 2024 6 Phys. Educ. 59 (2024) 015001 Einsteinian gravitational concepts throughout secondary school ◦ Figure 4. Responses to Q2. Can the sum of the angles in a triangle be different from 180 ? Table 5. Difference in pre- and post-questionnaire responses to Q2 by year group. Pre Post No 93.48% 88.89% 87.50% 84.21% 83.33% 83.33% 90.91% Yes 6.52% 11.11% 12.50% 15.79% 8.33% 16.67% 9.09% No 13.79% 33.33% 16.67% 6.25% 0.00% 0.00% 11.11% Yes 86.21% 63.64% 83.33% 75.00% 100.00% 100.00% 88.89% Yr 7 Yr 8 Yr 9 Yr 10 Yr 11 Yr 12 Yr 13 in figure 6, the responses to this question are extremely varied, ranging from high level answers like ‘conic sections’ and ‘along geodesics’ to less descriptive responses such as ‘they float’ or ‘fast’. Many responses show an understanding of the attractive powers of gravity, as well as know- ledge of other aspects of physics such as kinetic and gravitational potential energy. While many of these descriptions are not wholly incorrect, they fail to properly describe the motion of objects in gravitational fields. This is likely due to par- ticipants lack of formal education on gravitational orbits, a topic not covered in detail until Yr 13. It is possible that responses such as ‘around’ and ‘in circles’ are consequences of this as those in lower year groups attempt to describe orbits without knowledge of the correct scientific terminology. This is supported by the fact that we see a slight increase in these descriptors post-intervention as participants watched objects circle around masses on the spacetime simulator sheet. such as Positively, phrases ‘they fall’, ‘towards ground/source’ and ‘pushed/pulled’ decrease post-intervention, whereas the number of participants describing orbital motion doubles. Those describing motion in gravitational fields January 2024 7 Phys. Educ. 59 (2024) 015001 C McInerney and P Sutton Figure 5. Pre- (a) and post-questionnaire (b) responses to Q3. What is gravity? Responses in which gravity is described as a force are shown in the right-hand charts, with other opinions of gravity being shown in the left-hand charts. Table 6. Difference in pre- and post-questionnaire responses to Q3 by year group. Pre Force Post Force Earth Other Spacetime curvature Earth Earth Other Spacetime curvature Earth Yr 7 Yr 8 Yr 9 Yr 10 Yr 11 Yr 12 Yr 13 25.53% 48.94% 35.14% 54.05% 29.17% 87.50% 5.56% 72.22% 58.33% 41.67% 16.67% 75.00% 0.00% 81.82% 0.00% 0.00% 0.00% 5.56% 0.00% 8.33% 9.09% 6.38% 17.24% 48.28% 2.70% 25.81% 51.61% 0.00% 38.89% 38.89% 0.00% 15.38% 53.85% 0.00% 30.00% 40.00% 0.00% 0.00% 42.86% 0.00% 0.00% 44.44% 1.72% 3.23% 5.56% 23.08% 10.00% 28.57% 44.44% 6.90% 6.45% 0.00% 0.00% 0.00% 0.00% 0.00% January 2024 8 Phys. Educ. 59 (2024) 015001 Einsteinian gravitational concepts throughout secondary school Figure 6. Responses to Q4. How do objects move in gravitational fields? Figure 7. Responses to Q6. Does space have a shape? What about the space around heavy objects like stars and planets? as attractive also sees a significant increase post- intervention. This again is likely due to the use of the spacetime simulator demonstration where it was observed that objects on the sheet moved towards the source of curvature. This is also evid- enced by the added uses of terms such as ‘oval’ and ‘spirals’ which appear in post-questionnaire responses only. 58.2% of Pre-intervention, participants thought that space had no shape, often saying it is infinite or expanding (‘no because it is infinite/expanding’) to justify their response. This number decreases to 42.4% post-intervention, the intervention had limited indicating that success in conveying the geometry of space around celestial objects. Nonetheless, figure 7 reveals an increase in responses other than ‘yes’ or ‘no’ post-intervention with the addition of more conceptualised answers like ‘the shape changes with gravity’ or ‘it stretches’ demon- strating an understanding of the ideas presen- ted using the spacetime simulator. 9.2% of post- intervention participants shown deeper under- standing by correctly noting that space is curved around masses but flat otherwise. Coupling these responses with the other ‘yes’ responses gives a total percentage of conceptually correct answers of 58.9%. January 2024 9 Phys. Educ. 59 (2024) 015001 C McInerney and P Sutton Figure 8. Responses to Q9. Are you interested in learning more about gravity and general relativity? Figure 9. Responses to Q10. Should Einsteinian physics, like general relativity be included in the curriculum? Figure 8 shows responses to Q9, where 77.1% of participants said they would like to learn more about gravity and GR. The last question in our questionnaires addresses the main research question of this work and asked: should Einsteinian physics, like GR be January 2024 10 Phys. Educ. 59 (2024) 015001 Einsteinian gravitational concepts throughout secondary school included in the curriculum? results are shown in figure 9. A resounding 82.5% of participants said that it should be included. Of those, 10% indic- ated that, if added to the curriculum, it should be included at a specific level. 41% of those respond- ents said that this content should not be taught before A-level. The rest thought that it should not be covered before KS4. that 4. Conclusion In this work, 183 students from a state-funded English secondary school participated in a one- off intervention introducing them to GR and the concept the gravitational force is caused by the curvature of spacetime. This theory was introduced visually using the spacetime simulator demonstration. Once this idea was established, participants performed an experiment using bal- loons and triangles to investigate what happens to these shapes in curved spaces. The aim was to gauge students’ reception to GR and to investigate where these concepts would be best placed into a school curriculum. Results showed that 77.1% of our participants overall would like to delve further into the sub- ject. This is consistent with the opinions of teach- ers and the public who also encourage the addi- tion of Einsteinian physics into the curriculum [29, 30]. Noting that the largest percentage of our participants come from Yr 7 shows how eager young students are for opportunities to learn about cosmology. KS4 and A-level students demonstrated bet- ter understanding of the presented ideas through their responses to Q1–3 in the questionnaires. Years 11, 12 and 13 had the best improvement from pre- to post-questionnaire for Q1 and Q2, with Yr 10 also showing a good increase in under- standing for Q3. While many participants from the other year groups liked the material covered, their responses to Q3 showed little grasp of spacetime curvature. Years 7, 8 and 9 all showed good aware- ness of the deformation of triangles in curved spaces after the intervention, only Yr 9 parti- cipants showed a grasp of the curving of initially parallel lines. While Einsteinian gravitational concepts are taught completely distinct from the Newtonian gravity, it has been shown by other studies that students can understand the ideas of GR even at primary level [2, 15, 16, 18]. While the results to Q1 and Q2 show that participants were able to appreciate the effects of curvature, the results of Q3 demonstrate that it would take more than a one-off intervention to imbue a deeper under- standing of GR. This result aligns with that of other works [31]. Additionally, our intervention was not successful at dispelling misconceptions around the gravitational force. A multi-stage inter- vention may prove more beneficial for this [32]. Regarding GR and its associated geometry, our results show that this material is better suited to the A-level curriculum rather than lower year groups. Therefore, we conclude that introducing GR in the English secondary curriculum at Yr 13 aligns with students’ ability to recognise curvature as the source of gravitational force and matches the current placement of gravity study in the curriculum. Further insights into the effects of teaching GR could be obtained through a longitudinal study that follows a group or groups of students through late primary/early secondary school to KS4 and A-level. Data availability statement The data cannot be made publicly available upon publication due to legal restrictions preventing unrestricted public distribution. The data that sup- port the findings of this study are available upon reasonable request from the authors. Acknowledgments The authors would like to thank the staff and students of the science department at Queen Elizabeth’s High School for their support with and participation in this work. Ethical statement This study has been reviewed and given favour- able opinion by a University of Lincoln Research Ethics Committee. Reference Number: 12146. January 2024 11 Phys. Educ. 59 (2024) 015001 C McInerney and P Sutton ORCID iDs Corey McInerney  https://orcid.org/0000-0002- 2452-7259 Phil Sutton  https://orcid.org/0000-0003-3936- 4170 Received 5 June 2023, in final form 4 September 2023 Accepted for publication 4 October 2023 https://doi.org/10.1088/1361-6552/ad001a References [1] Newton I, Motte A and Chittenden N W 1848 Newton’s Principia. The Mathematical Principles of Natural Philosophy (D. Adee) [2] Elise B A 2007 Int. J. Sci. Educ. 29 1767–88 [3] Meyer J H F and Land R 2005 High. Educ. 49 373–88 [4] Ferreira A and Lemmer M 2021 J. Phys.: Conf. Ser. 1929 012003 [5] Syuhendri S 2019 J. Phys.: Conf. Ser. 1185 012047 [6] Smith C and Treagust D 1988 Aust. Sci. Teach. J. 33 21–24 [7] Piburn M D 1988 Misconceptions about gravity held by college students (ERIC Clearinghouse) (available at: https://nla.gov. au/nla.cat-vn5489406) [8] Galili I and Bar V 1997 Int. J. Sci. Educ. 19 317–40 [9] Ruggiero S, Cartelli A, Dupré F and Vicentini-Missoni M 1985 Int. J. Sci. Educ. 7 181–94 [10] Driver R 1981 Eur. J. Sci. Educ. 3 93–101 [11] Bungum B, Henriksen E K, Angell C, Tellefsen C W and Bøe M V 2015 Nordina 11 153–68 [12] Kersting M, Henriksen E K, Bøe M V and Angell C 2018 Phys. Rev. Phys. Educ. Res. 14 010130 [13] Moore T A 2012 A General Relativity Workbook (University Science Books) [14] Haddad W D and Pella M O 1972 J. Exp. Educ. 41 22–32 [15] Pitts M, Venville G, Blair D and Zadnik M 2014 Res. Sci. Educ. 44 363–88 [16] Kaur T, Blair D, Moschilla J, Stannard W and Zadnik M 2017 Phys. Educ. 52 065012 [17] Dua Y, Blair D, Kaur T and Choudhary R 2020 J. Pendidik. IPA Indones. 9 50–58 [18] Ruggiero M L, Mattiello S and Leone M 2021 Phys. Educ. 56 065011 [19] Ruggiero M L 2022 15th Marcel Grossmann Meeting pp 1329–34 [20] Choudhary R K, Foppoli A, Kaur T, Blair D G, Zadnik M and Meagher R 2018 Phys. Educ. 53 065020 [21] Kaur T, Blair D, Burman R, Stannard W, Treagust D, Venville G, Zadnik M, Mathews W and Perks D 2017 Evaluation of 14 to 15 year old students’ understanding and attitude towards learning Einsteinian physics (arXiv:1712.02063) [22] Kaur T, Blair D, Stannard W, Treagust D, Venville G, Zadnik M, Mathews W and Perks D 2020 Res. Sci. Educ. 50 2505–32 [23] Wynarczyk P and Hale S 2009 Improving take up of science and technology subjects in schools and colleges: a synthesis review Economic and Social Research Council (ESRC) and the Department for Children, Schools and Families [24] Laursen S, Liston C, Thiry H, Graf J and Schulz B 2007 CBE Life Sci. Educ. 6 49–64 [25] White G D, Mondragon T, Slaughter D and Coates D F 1993 Am. J. Phys. 61 367–71 [26] Postiglione A and Angelis I D 2021 Phys. Educ. 56 025019 [27] Abrahams I and Millar R 2008 Int. J. Sci. Educ. 30 1945–69 [28] Millar R 2010 Good Practice in Science Teaching: What Research has to say vol 2 pp 108–34 [29] Foppoli A, Choudhary R, Blair D, Kaur T, Moschilla J and Zadnik M 2018 Phys. Educ. 54 015001 [30] Balta N, Eryilmaz A and Oliveira A W 2022 Teach. Dev. 26 166–88 [31] Choudhary R, Kraus U, Kersting M, Blair D, Zahn C, Zadnik M and Meagher R 2019 Phys. Educ. 01 1950016 [32] Choudhary R, Foppoli A, Kaur T, Blair D, Burman R and Zadnik M A 2022 Res. Sci. Educ. 52 305–24 January 2024 12 Phys. Educ. 59 (2024) 015001	null
10.1371_journal.pone.0296688.pdf	Data Availability Statement: All data files are available from the OpenBU database (https://hdl. handle.net/2144/45321).	All data files are available from the OpenBU database ( https://hdl. handle.net/2144/45321 ).	RESEARCH ARTICLE Flanged males have higher reproductive success in a completely wild orangutan population Amy M. ScottID Wahyu Susanto6, Tatang Mitra Setia6, Cheryl D. Knott1,7 1,2, Graham L. Banes3,4, Wuryantari Setiadi5, Jessica R. Saragih5, Tri 1 Department of Anthropology, Boston University, Boston, Massachusetts, United States of America, 2 Department of Natural Resources and the Environment, University of New Hampshire, Durham, New Hampshire, United States of America, 3 Wisconsin National Primate Research Center, University of Wisconsin–Madison, Madison, Wisconsin, United States of America, 4 The Orang-Utan Conservation Genetics Project, Madison, Wisconsin, United States of America, 5 Eijkman Research Center for Molecular Biology, National Agency for Research and Innovation (BRIN), The Science and Technology Center of Soekarno, Cibinong, West Java, Indonesia, 6 Departemen of Biology, Faculty of Biology and Agricultural, Universitas Nasional, Kota Jakarta Selatan, Daerah Khusus Ibukota Jakarta, Indonesia, 7 Department of Biology, Boston University, Boston, Massachusetts, United States of America [email protected] Abstract Male orangutans (Pongo spp.) exhibit bimaturism, an alternative reproductive tactic, with flanged and unflanged males displaying two distinct morphological and behavioral pheno- types. Flanged males are larger than unflanged males and display secondary sexual char- acteristics which unflanged males lack. The evolutionary explanation for alternative reproductive tactics in orangutans remains unclear because orangutan paternity studies to date have been from sites with ex-captive orangutans, provisioning via feeding stations and veterinary care, or that lack data on the identity of mothers. Here we demonstrate, using the first long-term paternity data from a site free of these limitations, that alternative reproductive tactics in orangutans are condition-dependent, not frequency-dependent. We found higher reproductive success by flanged males than by unflanged males, a pattern consistent with other Bornean orangutan (Pongo pygmaeus) paternity studies. Previous paternity studies disagree on the degree of male reproductive skew, but we found low reproductive skew among flanged males. We compare our findings and previous paternity studies from both Bornean and Sumatran orangutans (Pongo abelii) to understand why these differences exist, examining the possible roles of species differences, ecology, and human intervention. Additionally, we use long-term behavioral data to demonstrate that while flanged males can displace unflanged males in association with females, flanged males are unable to keep other males from associating with a female, and thus they are unable to completely mate guard females. Our results demonstrate that alternative reproductive tactics in Bornean orangutans are condition-dependent, supporting the understanding that the flanged male morph is indicative of good condition. Despite intense male-male competition and direct sex- ual coercion by males, female mate choice is effective in determining reproductive out- comes in this population of wild orangutans. a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Scott AM, Banes GL, Setiadi W, Saragih JR, Susanto TW, Mitra Setia T, et al. (2024) Flanged males have higher reproductive success in a completely wild orangutan population. PLoS ONE 19(2): e0296688. https://doi.org/10.1371/journal. pone.0296688 Editor: Honnavalli Nagaraj Kumara, Salim Ali Centre for Ornithology and Natural History, INDIA Received: August 21, 2023 Accepted: December 17, 2023 Published: February 9, 2024 Copyright: © 2024 Scott et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All data files are available from the OpenBU database (https://hdl. handle.net/2144/45321). Funding: This research was supported by Adventure Travel Conservation Fund: adventuretravelconservationfund.org; Arcus Foundation (G-PGM-1708- 2235, G-PGM-1506- 1327, G-PGM- 1104-36, 1104-36/PID-01853, 0902-30): www.arcusfoundation.org; Association of Zoos and Aquariums Conservation Endowment Fund (13-1159, 11-1063) and Conservation Grants PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 1 / 20 PLOS ONEFund (15-1296): www.aza.org/conservation- funding; Balikpapan Orangutan Society-Canada: orangutan.ca; Conservation, Food and Health Foundation: cfhfoundation.grantsmanagement08. com; Disney Conservation Fund: impact.disney. com/environment/conservation; Focused on Nature: focusedonnature.org; Hollomon Price Foundation: hollomonpricefoundation.org; Houston Zoo: www.houstonzoo.org; Indonesia Climate Change Trust Fund: www.icctf.or.id; Keidanren Nature Conservation Fund: www.keidanren.net/ kncf/en; The Leakey Foundation: leakeyfoundation. org; Nacey Maggioncalda Foundation: formerly: www.naceymaggioncalda.org, currently: leakeyfoundation.org; National Geographic Society (ECO690-14, GEFNE68-13, 8564-08, C113-07): www.nationalgeographic.org/society; National Science Foundation (BCS-1638823, BCS- 0936199): nsf.gov; Ocean Park Conservation Fund: www.opcf.org.hk/en; Orangutan Conservancy: orangutan.com; Phoenix Zoo: www.phoenixzoo. org; Primate Conservation International; Sea World Busch Gardens Conservation Fund; swbg- conservationfund.org; Tides Foundation: tides.org; US Fish and Wildlife Service (F19AP00798, F18AP00898, F15AP00812, F13AP00920, F12AP00369, 96200-0-G249, 96200-9-G110, 98210-8-G661, 98210-7-G185): www.fws.gov; Wenner-Gren Foundation: wennergren.org; Whitley Fund for Nature: whitleyaward.org; Wildlife Conservation Network: wildnet.org; Woodland Park Zoo Partners for Wildlife: zoo.org; Zoo New England: www.zoonewengland.org; and Zoo Atlanta: zooatlanta.org grants to CDK. This research was supported by Boston University Graduate Research Abroad Fellowship: www.bu. edu/cas/admissions/phd-mfa/fellowship-aid/aid- for-phd-students; Boston University Graduate Student Organization Research Grant: www.bu. edu/gso/travelgrants; Boston University Women’s Guild: www.bu.edu/womensguild; Cora Du Bois Charitable Trust: library.harvard.edu/cora-du-bois- fellowship; The Leakey Foundation: leakeyfoundation.org/; and National Science Foundation Graduate Research Fellowship (Grant DGE-1247312): nsf.gov grants to AMS. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. Flanged male orangutans have higher reproductive success Introduction Alternative reproductive tactics (ARTs) are the existence of two distinct phenotypes within one sex in the context of reproduction [1]. ARTs occur throughout the animal kingdom and are expected to evolve when there is strong sexual selection [1,2], specifically intra-sexual com- petition [2]. There are two primary explanations for the existence of ARTs. The two pheno- types may be either frequency-dependent evolutionary stable strategies, where the relative fitness of each morph depends on its frequency in the population [3] or condition-dependent, due to difference in the quality (i.e. age, body condition, experience, nutritional state and/or genes) of individuals where one morph is ‘making the best of a bad lot’ [4]. Male orangutans (Pongo spp.) display two ARTs with males exhibiting distinct morphologi- cal (Fig 1) (expressed as bimaturism) and behavioral phenotypes [5–8]. Flanged males (50–90 kg) are up to twice the size of unflanged males, but some unflanged males reach flanged male body size (30–59 kg) [9,10]. Flanged males possess secondary sexual characteristics, including an enlarged throat sac and cheek flanges [6,7,11], and are the only morph capable of producing long calls [5,6]. Flanged males are intolerant of each other, either avoiding or fighting and wounding each other [5,12], but they are typically more tolerant of unflanged males [5]. Con- versely, unflanged males are generally tolerant of each other and tend to avoid flanged males [5,6]. Flanged males are dominant to unflanged males and displace unflanged males in con- sortships with females [5,6,12–14]. It has been suggested that flanged males use consortships to mate guard females, as a means to keep other males from mating with a female [12,15,16]. The male morphs also differ in activity patterns, with unflanged males traveling further per day than flanged males [14,17,18]. Due to these differences, the flanged male mating strategy has been described as “sit, call, and wait” and the unflanged male strategy described as “go, search, and find” [7,14]. For male orangutans, ARTs are plastic and sequential—an immature male first develops the unflanged male phenotype and may develop the flanged male phenotype later, but this transi- tion is irreversible [7,19]. There is tremendous variation in the age of flange development, with wild males reportedly developing flanges from ages 14 to 30, and some males never developing flanges [11,16,20]. Flanged males in poor condition exhibit shriveled flanges and are referred to as past-prime males [21]. Past-prime males are not regularly seen, suggesting that this phase is not reached by all males, and is likely short for the males who do become past-prime. Addi- tionally, the presence of past-prime males indicates that the flanged morph is so costly to maintain that some flanged males that cannot continue to maintain it enter the past-prime state [11]. Understanding how sexual selection acts on traits, such as ARTs, requires considering mul- tiple mechanisms of sexual selection simultaneously [22]. Both male and female reproductive strategies are expected to impact the relative reproductive success of each morph [22], and this is especially true for primates, where male and female strategies are closely tied [23]. Orangu- tans are semi-solitary with large home ranges and adults primarily range alone or adult females range with dependent offspring [11,16,24,25], so reproduction first requires finding a mate. It has been suggested that one function of flanged male long calls is to attract females [26,27], and it may also play a role in male-male competition [26,28,29]. Across study sites, female orangutans prefer flanged males [5,6,21,30,31]. Orangutans also have slow life histories, including the longest interbirth interval of any mammal (7.6 years) [32,33]. Slow life histories push the potential for sexual conflict to an extreme [34]. Both male morphs employ sexual coercion in the form of forced copulations [30,35]. Sexual coercion can override female mate choice, but it is unknown if it increases male reproductive success. Female orangutans do not display overt signals of ovulation, such as the sexual swellings typical of many cercopithecoids PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 2 / 20 PLOS ONEFlanged male orangutans have higher reproductive success Fig 1. Example of male orangutans displaying the two alternative reproductive tactics. An unflanged male (left) lacks cheek pads and a throat sac and has a smaller body size. A flanged male (right) has secondary sexual characteristics including large cheek pads (flanges), a large throat sac, and larger body size. Photos by Tim Laman. https://doi.org/10.1371/journal.pone.0296688.g001 [21,36] and ovulatory status appears to be effectively hidden from males [36,37]. Females pref- erentially mate with prime flanged males when they are ovulating and show increased willing- ness to mate with unflanged and non-prime males when the risk of conception is low [21]. Across primates this mating pattern—mating preferentially with preferred males when the likelihood of conception is highest and mating with non-preferred males when the likelihood of conception is lowest—is argued to be a paternity confusion strategy that reduces the likeli- hood of infanticide [21,38,39]. Quantifying the reproductive success of each morph is essential for testing hypotheses about the evolutionary pressures that resulted in orangutan ARTs. Previous studies of pater- nity in both Bornean (Pongo pygmaeus) and Sumatran (Pongo abelii) orangutans are limited by incomplete maternity data [40–43], the inclusion of ex-captive orangutans who may not display natural mating behaviors, or by provisioning from feeding stations and from veterinary care [13,20,44,45] (Table 1). The first orangutan (P. abelli) paternity study found that the two morphs had similar reproductive success and therefore concluded that the two morphs repre- sent alternative mating strategies that coexist as evolutionary stable strategies [20]. The subse- quent three orangutan (P. pygmaeus) paternity studies all concurred that flanged males had much higher reproductive success than unflanged males [13,44,45]. Each of these studies has unique limitations (Table 1). There are also important island or species differences to consider. P. abelii live in habitats with higher food availability, exist at higher densities, and are more social compared to P. pygmaeus [24,46]. We present paternity data from Cabang Panti Research Station in Gunung Palung National Park, Borneo, Indonesia (GPNP), the first from completely wild orangutans with known mothers. We compare our results against others to discern how study limitations and habitat differences explain contrasting results across sites. Orangutan paternity studies also differ in the degree of male reproductive skew—the degree to which reproduction is monopolized versus shared (Table 1). Characterizing male PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 3 / 20 PLOS ONEFlanged male orangutans have higher reproductive success Table 1. Paternity determination and paternity skew across study sites. Study Site No. offspring sired by flanged males No. offspring sired by unflanged males Ketambe Research Station, Gunung Leuser National Park [18] Kinabatangan Orang- utan Conservation Project, Lower Kinabatangan Wildlife Sanctuary [37] Camp Leakey, Tanjung Puting National Park [36] Sepilok Orangutan Rehabilitation Center [12] Cabang Panti Research Station, Gunung Palung National Park 4 9 10 4 5 6 1 3c 1 0 Total 32 11 Total no. offspring with father assigneda 10 10 14 6c 6c 46 Total no. offspring tested Total no. candidate sires testeda Study Periodb 11 16 25 8 11 16 17 4 1983– 1997 (11) 1985– 2000 (8) 1993– 2009 (13) 2010– 2014 (1) Most successful male’s share (mean) 48.18 Most successful male’s share (range) 33.33–100 33.32 18.18–50 56.57 14.29–100 57.14 NA 13 20 2008– 2014 (3) 33.33 20–40 Limitation Ex-captives in study population Limited population knowledge; Mothers genetically assigned Feeding station; Ex- captives in study population; veterinary care Feeding station; Ex- captives in study population; Only one flanged male sampled Gray background = P. abelii. White background = P. pygmaeus. a = Number of offspring and candidate sires tested are likely an underestimate of the total number of offspring born or candidate sires in the study site due to sampling difficulties. b = Number in parentheses is the number of 5-year periods during the study period. c = Number of offspring sired by flanged males and unflanged males do not add up to the total because there was a male of unknown morph who sired an offspring. https://doi.org/10.1371/journal.pone.0296688.t001 reproductive skew is important for understanding the evolution of ARTs in orangutans. Across primates, the degree of male reproductive skew in multi-male groups is best explained by the degree of female reproductive synchrony and the number of males in the group [47,48]. The orangutan social system, with a high fission-fusion dynamic (social associations vary in size, composition, and cohesion) [11,24,49], and a lack of group formation, makes defining the number of males in a "group" difficult. However, there is clearly a male biased operational sex ratio, with many males competing for a few conception opportunities, due in part to the long interbirth interval [16]. In terms of female reproductive synchrony, reproduction is asynchro- nous, although some sites do see increases in births following periods of high fruit availability [50]. Even without female reproductive synchrony, in a dispersed social system, low male reproductive skew is expected [47]. Additionally, male dominance can lead to higher repro- ductive success through priority-of-access [51], but this is not the case for all species [52]. Here we compare male reproductive skew across sites and use long-term behavioral data to test the ability of flanged males or a single dominant flanged male to mate guard females. We combine long-term behavioral observations and genetic paternity determination from a completely wild orangutan population at Cabang Panti Research Station in Gunung Palung National Park, Borneo, Indonesia, to investigate the evolution of male ARTs in Bornean orangutans. If male ARTs are frequency-dependent evolutionary stable strategies, we would expect the frequency of each morph to be stable and relative fitness of each morph to depend PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 4 / 20 PLOS ONEFlanged male orangutans have higher reproductive success on its frequency in the population [1,3], i.e. if 20% of males are flanged then 20% of offspring will be sired by flanged males. Conversely if the male morphs are condition-dependent strate- gies, then we would expect unequal fitness benefits for each morph, where the morph in ‘poor condition’ has lower reproductive success and takes advantage of alternative tactics [1,4]. First, we determine the relative reproductive success of the two morphs and measure male reproduc- tive skew. Second, we test the ability of flanged males to mate guard females. We then compare our results to those from prior studies in other populations to discern how study limitations and habitat differences might explain contrasting results across studies. Finally, we discuss the implications of these results for our understanding of the evolution of ARTs in orangutans and the interaction between male and female reproductive strategies. Materials and methods Study site and population Orangutans (Pongo pygmaeus wurmbii) were studied in Gunung Palung National Park (GPNP), West Kalimantan, Indonesia, based out of the Cabang Panti Research Station (CPRS) (1˚13´S, 1107´E) (3400 ha), as part of a study that began in 1994 [53]. Most orangutans encountered and followed were habituated and individually identifiable, but unknown and unhabituated individuals were also encountered, due to male dispersal and large home ranges [11,24,54,55]. Each month, phenology data were collected to characterize food availability of orangutan foods from 60 plots (totaling 9 ha) spread across 6 habitat types in the study site [56,57]. Fruit availability was calculated from the top 25 genera of plants that orangutans are known to consume most often at GPNP which represented 80% of fruit in their diet [57,58]. We then normalized that data by calculating modified Z scores from the percentage of stems that had mature or ripe fruits. Food availability was used as a control variable in our statistical models. Behavioral data collection We used long-term data (2008–2019) from orangutans in CPRS collected during focal follows [59] to assess the ability of the two male morphs to effectively mate guard females and to create a male dominance hierarchy. During orangutan follows, an association was recorded when- ever another orangutan came within 50 meters of the focal [60,61]. The identity and age-sex class of all orangutans was recorded. Males were classified by morph—flanged or unflanged. For this analysis, males who had small, developing flanges were classified as unflanged males. We used long-term follow data to tally the number of flanged and unflanged males that were seen in the study site one year prior to and following conception for each offspring, where we were able to identify a father and determine his morph. Females were classified as ‘sexually active’ or ‘non-sexually active’ based on the likelihood that they were fecund and actively mat- ing. The ‘sexually active’ category included nulliparous females, parous females without depen- dent offspring, mothers with offspring over age six, and pregnant females in the first trimester. Females in this population are most proceptive to mating during the first trimester of preg- nancy [21]. The non-sexually active category included parous females with dependent off- spring under age six and pregnant females in the second and third trimester. Since orangutans have a gestation period of approximately eight months [62] and an average interbirth interval of 7.6 years [32], females will on average conceive when a dependent offspring is 6.8 years old and will begin mating 6–12 months before she conceives. Therefore, we used six years as a cut- off because we expected females to begin mating again at approximately that time. Addition- ally, we have previously shown that male-female interactions change when the dependent off- spring reaches age six [63]. PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 5 / 20 PLOS ONEFlanged male orangutans have higher reproductive success We analyzed all adult male-female associations from 2008–2019 (N = 759), noting the occurrence and outcome of an encounter with a second or ‘extra-pair male’ (EPM). If the asso- ciation between the first male and female was terminated after the second male arrived, and the second male stayed with the female, we defined this as male displacement. Displacement did not necessarily involve agonism or aggression between the males, nor was it necessarily immediate. For each male-female association, the length of the association (in minutes) and all mating events were also recorded. We analyzed all adult male-male interactions from 2008–2014, the period with both behav- ioral data and with paternity determination data, to evaluate male dominance rank. Offspring with known paternities were conceived from January 2010 to August 2014. During this period, nine of the sampled flanged males (Bilbo was still unflanged), an additional three individually recognizable flanged males, and up to seven unknown flanged males were observed in the study site. We examined the outcome of all dyadic interactions between flanged males to eval- uate dominance rank. Dominance was defined by the outcomes of dyadic agonistic interac- tions [64]. We included avoidance, displacement, and chase interactions as dominance interactions with a clear dominant and subordinate individual. Sample collection Fecal samples were collected after observed defecation from known and unknown orangutans from 2008 through 2019. When possible, two samples were collected from one individual on separate occasions. Samples from mother and dependent offspring were collected in the same encounter. Samples were stored in either RNAlater, 70% ethanol, or dried using the two-step ethanol alcohol-silica desiccation method [65,66]. Dried samples were stored at ambient tem- perature (up to 40˚ C) until analysis. Samples stored in RNAlater or 70% ethanol were stored at -20˚ C or -80˚ C. Genotyping and paternity analysis We collected fecal samples from 42 orangutans for genotyping: 13 offspring, their 10 mothers, and 19 candidate fathers (8 unflanged, 10 flanged, and 1 observed as both unflanged and flanged males) in GPNP. Genomic DNA was extracted 2–3 times from each fecal sample using ChimerX stool DNA purification kits. Following Morin et al. [67], we quantified DNA content through qPCR12. We amplified a panel of 12 autosomal tetranucleotide microsatellites [20,44,68–71] (S1 Table). These were first co-amplified in an initial PCR reaction, with suffi- cient replicates to maintain error rates of less than 1% when scoring homozygotes, per Ara- ndjelovic et al. [72], before the products were re-amplified with labelled primers in panels of 3–5 loci. Fragment analysis was performed by the DNA sequencing unit at Eijkman Institute for Molecular Biology, using an Applied Biosystems 3130 Genetic Analyzer to size alleles against a GeneScan™ 500 LIZ™ internal size standard. Peaks were manually scored by two different peo- ple using GeneMapper (v3.7 and v4.0). Scores were concordant irrespective of software ver- sion. Heterozygotes were called when the same two alleles were observed in at least two independent amplifications, and homozygotes were called when only one allele was observed in up to five independent amplifications, per Arandjelovic et al. [72]. Prior to downstream analysis, CERVUS 3.0 [42] and MICRO-CHECKER 2.2.3 [73] were used to assess genotypes for null alleles, allelic dropout, and scoring errors due to stuttering, and to confirm that all 12 microsatellites were in Hardy-Weinberg equilibrium (S2 Table). Individual identity analysis was performed in CERVUS 3.0 to ensure that purported replicates derived from the same individual. Individuals genotyped at a minimum of nine loci were PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 6 / 20 PLOS ONEFlanged male orangutans have higher reproductive success subsequently used in parentage analysis, having met the minimum number of loci needed to tell full siblings apart (PID-sibs <0.001) for the mean observed heterozygosity in our panel of microsatellites (sensu Waits et al. [74]). Paternity analyses were performed in CERVUS 3.0 [42] and in COLONY 2.0.6.7 [75], using both an exclusionary approach and a likelihood approach. In the exclusionary approach, off- spring are required to share one allele at each locus with the known mother and the other allele must be shared with the father. On the other hand, the likelihood approach in CERVUS 3.0 allows for genotyping errors, null alleles, and potential mutations. The advantage of COLONY 2.0.6.7 is that it uses a full-pedigree likelihood approach, rather than dyadic relationships, when inferring both parentage and sibship. Field observation of mother-offspring pairs was confirmed using exclusionary maternity analysis CERVUS 3.0. Mothers were then used as known parents in CERVUS 3.0, increasing the statistical power of paternity assignment. All sampled males were considered candidate fathers for each offspring. Paternity was simulated using 100,000 offspring to obtain critical values of Delta at confidence levels of 80% (relaxed) and 95% (strict), sensu Marshall et al. [43]. For simulation in CERVUS 3.0, the proportion of candidate fathers sampled was inferred at three different values: 0.2, 0.5, 0.65 to simulate the possibility that an unsampled sire fathered offspring. The values 0.65 and 0.2 represent the upper and lower limits of ‘unknown’ males being entirely ‘known’ males or entirely ‘unknown, unique’ males, respec- tively. Each value produced the same results, so we report values using 0.5 as the proportion of candidate fathers. In COLONY 2.0.6.7, analysis was run with the following parameters: female polygamy and male polygamy without inbreeding or clones, ‘long’ length of run, ‘high’ likelihood precision, no updating of allele frequency, and no sibship prior. Reported paternity results take known maternal genotype into account. Again, all sampled males were considered candidate fathers for each offspring. Cross-site comparisons We compared our paternity data from CBRS in GPNP to published paternity results from four other orangutan study sites: Kinabatangan Orangutan Conservation Project, Lower Kinaba- tangan Wildlife Sanctuary [45]; Ketambe Research Station, Gunung Leuser National Park [20]; Camp Leakey, Tanjung Puting National Park [44]; and Sepilok Orangutan Rehabilitation Cen- ter [13]. Reproductive skew We calculated male reproductive skew (2008–2014) using two measures: Nonacs B index [76,77] and the most successful sire’s share as a percentage. We calculated both measures for our study population and the most successful sire’s share for all published orangutan paternity data. Due to the male dispersal and long lives, we were unable to accurately estimate adult male ages required for the multinomial skew index [78], and Nonacs B index was calculated with the Skew calculator 2013 (https://www.dropbox.com/home/2013%20Version, accessed December 2021) [79]. The B index takes residency and number of offspring into account (see S1 File for details of interpretation). We included only sampled males and the two unsampled fathers (of the two offspring for whom we could not identify a father) in our calculation. For the unsampled fathers, we used the average male residency time across this study period, 3 years. PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 7 / 20 PLOS ONEFlanged male orangutans have higher reproductive success Statistical analysis To test the ability of flanged and unflanged males to mate guard females, we used two-sided Fisher’s exact tests to compare the rate at which the two male morphs are displaced by an EPM in association with a female (N = 63). We additionally tested this hypothesis using two-sided Fisher’s exact tests to compare the rate at which the two male morphs are displaced by an EPM in association with only sexually active females (N = 36). Fisher’s exact tests are appropriate for comparisons when values in some categories are less than five [80]. Further, we tested the possibility that flanged male presence alone acts as a deterrent pre- venting EPM from encountering the flanged male-female pair using Chi-square tests of equal proportions and a binomial generalized linear mixed model (GLMM). We used Chi-square tests of equal proportions to compare the rate at which sexually active female (N = 486 associa- tions), non-sexually active female (N = 201 associations), and total female associations (N = 706 associations) with flanged versus unflanged males encounter an additional or EPM. We also tested whether male-female association grouping (sexually active female-flanged male, sexually active female-unflanged male, non-sexually active female-flanged male, and non-sexu- ally active female-unflanged male) impacted the chance of encountering an EPM using a bino- mial GLMM. Data exploration and model residuals revealed no violations of the assumptions of the binomial GLMM [81]. The response variable was the occurrence of an encounter with an EPM (yes/no). We used the length of a male-female association as an offset variable and included the identity of the male and female as random effects. Fruit availability (see Study Site and Population) was included as a fixed effect (control variable) because some study sites show that orangutans are more social during periods of high fruit [82–84] (but see [85,86]). We compared AIC values between models that excluded fixed effects to determine the best model and how to code male morph and female reproductive class (S3 Table). We performed all statistical procedures in R [87]. For the nonparametric post-hoc tests, we used the package PMCMR [88]. For the binomial GLMMs, we used the packages lme4 [89] and arm [90] to calculate confidence intervals. Graphs were made in the packages ggplot2 [91] and cowplot [92]. This study followed the American Society of Primatologists’ ‘Ethical Treatment of Non- Human Primates’ principles. It was non-invasive and observational. All protocols were approved by The Eijkman Institute Research Ethics Commission, Boston University IACUC (protocol no. 11–045 and 14–043) or deemed exempt by Boston University IACUC. All proto- cols were approved by the Indonesian State Ministry for Research and Technology (RISTEK), the Ministry of Home Affairs and the Indonesian Institute of Sciences (LIPI), the Center for Research and Development in Biology (PPPB), and Balai Taman Nasional Gunung Palung (BTNGP). Sample collection was approved by Balai Taman Nasional Gunung Palung (BTNGP), permit numbers: 86/YPPN/SK/XII/2009-2019. Results Male reproductive success Each of the three methods of paternity determination (exclusionary and likelihood approaches in CERVUS and full-pedigree likelihood approach in COLONY) were concordant in paternity assignment (Table 2). Paternity could be assigned for five out of seven offspring conceived during the sampling period (2008–2019) and one individual conceived prior to the sampling period. The flange-status of this sire at the time of conception (ca. 2005) is unknown, but he was flanged at first observation in 2009. Over a six-year period (2009–2014), four flanged males sired five offspring, indicating that male morph plays an important role in male PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 8 / 20 PLOS ONEFlanged male orangutans have higher reproductive success Table 2. Paternity assignment at cabang panti research station in GPNP. Offspring Est. Birth Year Mother Trio mis-match Pe Next best mis-match Delta Assigned Father CERVUS Exclusion Likelihood COLONY Prob. Dagul Rossa Berani Ijal Telur Uok Januari Benny Dolia Hannah Vanna Tawni Bayasa 2002 Delly 2004 Veli 2005 Bibi 2005 Irmaa 2007 Tari 2007 Umi 2009 JT 2010 Beth 2011 Dewi 2012 Heraa 2012 Veli 2014 Tari 2015 Bibi — — 0 — — — — 0 0 — 0 1 0 — — — — — — — 0.999 0.999 0.999 0.999 0.999 0.998 — — 4 — — — — 3 2 — 1 4 2 — — — — — — — 13.7 ‡ 7.76 4.30 6.50 4.08 5.29 ‡ ‡ ‡ ‡ ‡ — — 0.999 — — — — 0.961 0.023 — 0.880 0.023 0.072 — — Codet — — — — Prabu Senjaa — Prabu Mandab Moris Gray = conceptions within the sampling (fecal and behavioral data collected) period. a = genotyped at 11 loci, not all 12 loci. b = genotyped at 10 loci, not all 12 loci. Bold = male is known to have been flanged at the time of conception. Non-bolded father means that his phenotype was unknown at the time of conception. Trio mismatch = the number of loci that are a mismatch in the trio of offspring, mother and assigned father. Pe = exclusion probability, calculated in CERVUS 3.0 using allele frequencies from all 48 individuals genotyped. Next best mismatch = refers to the trio of offspring, mother, and the male with the closest match after the assigned father. ‡ = the trio delta value meets the strict (95%) confidence level. https://doi.org/10.1371/journal.pone.0296688.t002 reproductive success (Table 2). During each of these conception periods, there were never more flanged males than unflanged males observed in the study site (S4 Table). The mothers of these five offspring were parous at the time of conception. For the two offspring for whom fathers could not be determined, the COLONY pedigree results inferred different fathers. Male reproductive skew We found low reproductive skew in our study population. From 2008 to 2014, the period with both behavioral data and offspring genetic sampling, the most successful sire’s share was 28.57% and Nonacs B index was 0.0004 (Npotential sires = 17, Noffspring = 6, P = 0.502, 95% CI = -0.121–0.188) (Table 1). Our Nonacs B values indicated either a random or equal distribution of male reproductive skew (equalB = 0.121, monopolyB = 0.922, see S1 File for details of inter- pretation). Unfortunately, we were not able to construct a male dominance hierarchy because only six interactions between flanged males were observed during this same period, with 3434 flanged male observation hours (S5 Table). Even with few observations, we did not find a strict relationship between male dominance and reproductive success. For example, Senja, who sired one known offspring during this period, was subordinate to Codet, who did not sire any known offspring during this period (Tables 1 and S5). Cross-site comparisons Combining paternity assignment data across the five study sites showed that, overall, flanged males sired a greater proportion of offspring (69.57%) than did unflanged males (23.91%) PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 9 / 20 PLOS ONEFlanged male orangutans have higher reproductive success (Table 1). This was especially true for Bornean sites, where flanged males sired 77.78% of off- spring (Table 1). There was also variation in the degree of male reproductive skew across study sites (Table 1). Across study sites, the mean most successful sire’s share (for 5-year periods) ranged from 33.33%-57.57% (Table 1). Mate guarding Our paternity results demonstrated that a single male was unable to monopolize paternity within our study site during any time period. To examine this from a behavioral perspective, we examined the ability of males to mate-guard females. We tested whether the presence of a male in association with a female served to deter a second or ‘extra-pair male’ (EPM) from interacting with that female. Further, we examined the outcome of those interactions to deter- mine if flanged males were able to displace unflanged males. We observed no significant difference in the rate at which female associations with flanged and unflanged males encountered an EPM (χ2 = 0.120, df = 1, P = 0.730, N = 706). On average, an EPM was encountered every 56.65 hours of unflanged male-female associations, and every 54.08 hours of flanged male-female associations (S6 Table). Likewise, there was not a signifi- cant difference in the rate at which sexually active female associations (χ2 = 2.412, df = 1, P = 0.120, N = 486) or non-sexually active female associations (χ2 = 0.746, df = 1, P = 0.388, N = 201) with flanged and unflanged males encountered an EPM. On average, sexually active females in association with flanged males encountered an EPM every 95.85 hours and in asso- ciation with unflanged males encountered an EPM every 70.78 hours (S6 Table). In contrast, non-sexually active females in association with flanged males encountered an EPM every 15.28 hours and in association with unflanged males encountered an EPM every 43.08 hours, on average (S6 Table). Our best binomial Generalized Linear Mixed Model (GLMM) found that food availability and both male and female age-sex classes significantly impacted the likelihood that a male-female association would encounter an EPM (S7 Table). Flanged males with non- sexually active females were significantly more likely to encounter an EPM than were either flanged or unflanged males with sexually active females (S7 Table and Fig 2). However, after an EPM was encountered, there was a statistically significant difference between male morphs in the proportion of encounters in which the first male associating with a female was displaced (Fisher’s Exact Test, N = 50, P = 0.0004) (Fig 3A). Unflanged males were displaced in 60% of encounters with an EPM. Of the 18 times that unflanged males were displaced, 61% of the EPM were flanged. Flanged males were only displaced in 10% of encoun- ters with an EPM and they were never displaced by unflanged males. This 10% represents one instance in which one flanged male chased off another flanged male in the presence of two non-sexually active females. When considering only sexually active females, flanged males were statistically significantly less likely to be displaced than unflanged males (Fisher’s Exact Test, N = 36, P = 0.003) (Fig 3B). Conversely, when considering only non-sexually active females, there is no difference in the rate of displacement between male morphs (Fisher’s Exact Test, N = 20, P = 0.379) (Fig 3C). Thus, both female reproductive state and male morph are important determinants of orangutan mating behavior [21]. Discussion Male reproductive success and skew Our paternity results (6 assigned paternities over 10 years) most closely align with those of Kinabatangan [45], finding low reproductive skew among flanged males. While Kinabatangan inferred maternity from genetic data, our results confirm the same overall pattern. Only these two studies are from wholly wild and unprovisioned orangutans in primary rainforest habitat PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 10 / 20 PLOS ONEFlanged male orangutans have higher reproductive success Fig 2. The proportion of male-female associations in which the dyad encounters an ‘Extra-Pair Male’ (EPM) by male-female association group type. FL-SA = flanged male/sexually active female association. FL-NSA = flanged male/non-sexually active female association. UF-SA = unflanged male/sexually active female association. UF-NSA = unflanged male/non-sexually active female association. N values at the top of each column show the number of male-female associations in each group type. Dark gray represents encounters with an EPM and light gray represents no encounter with an EPM. Significance values from the binomial GLMM (* = P < 0.05). https://doi.org/10.1371/journal.pone.0296688.g002 without feeding stations, ex-captive orangutans, or veterinary care. This suggests that flanged males have higher reproductive success than unflanged males in completely wild Bornean pop- ulations, and that a single flanged male cannot monopolize paternity. In contrast, provisioning from feeding stations at Tanjung Putting [44] and Sepilok [13], in conjunction with veterinary interventions, may explain why a single flanged male was able to monopolize paternity at these two sites. Feeding stations may create an unnaturally high concentration of female orangutans in one area, increasing the ability of a single male to monopolize females. One unexpected out- come of feeding stations may be a reduction in genetic diversity in subsequent generations due to high male reproductive skew. It is likely that without feeding stations, either (1) dominant males are unable to monopolize females across large areas or (2) male dominance hierarchies are less strict when males are not competing over access to a feeding station. Due to the rarity of interactions between flanged males (6 interactions in 7 years), we could not construct a dominance hierarchy, but all observed interactions suggest a linear hierarchy with no PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 11 / 20 PLOS ONEFlanged male orangutans have higher reproductive success Fig 3. Male displacement in male-female associations. The proportion of (a) all male-female associations, (b) a subset of male-female associations where the female is a sexually active, and (c) a subset of male-female associations where the female is a non-sexually active in which the first male is displaced by an ‘extra-pair male’ (EPM). Male displacement is represented by darker shading. N values at the top of each column show the number of male-female associations by male morph. Dark gray represents encounters with displacement and light gray represents encounter with no displacement. Significance values from Fisher’s exact tests (* = P < 0.05, ns = P > 0.05). https://doi.org/10.1371/journal.pone.0296688.g003 observations of rank reversals or rank instability. Continued study of paternity and male inter- actions at GPNP and more orangutan study sites could differentiate between these two possibilities. Paternity results from Ketambe, Sumatra [20], starkly contrast with those from Borneo— unflanged males at Ketambe had higher reproductive success than flanged males [20]. It is unclear if this represents a species difference or unusual population parameters. The combina- tion of male rank instability, first-time mothers, and ex-captive females in that study [20,93] may have resulted in an inflated reproductive advantage for unflanged males. For instance, Sepilok and Tanjung Puting also found that the offspring of first-time mothers were sired by unflanged males [13,44], although in GPNP nulliparous females formed preferential mating relationships with flanged males [94]. If there truly is a species difference between the relative reproductive success of flanged and unflanged males, it is likely due to differences in the dura- tion of the unflanged stage and variation in the relative proportions of each morph between the islands [19,86]. However, it is important to note that orangutan paternity studies are limited by small sam- ple sizes (Table 1) due to their long interbirth intervals and semi-solitary social structure. Smaller samples are more subject to random stochasticity, which may also play a factor in explaining the differences between sites, but comparison of data across five different sites adds robustness to these comparisons. Small sample sizes may contribute to the finding of lower reproductive skew. In this comparative perspective, the two Bornean sites with completely wild orangutans (GPNP and Kinabatangan) agree that flanged males have higher reproductive success than unflanged males and reproductive success is spread broadly across many flanged males. But with only one Sumatran site in the sample [20], where there are also ex-captives, it is unclear if that pattern holds for Sumatran orangutans. Half of the offspring (5 out of 10) in Ketambe were born to matrilines with ex-captive mothers, and 4 of these 5 offspring were PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 12 / 20 PLOS ONEFlanged male orangutans have higher reproductive success sired by unflanged males [93]. If mating strategies are learned through the observation of the mother, then ex-captive matrilines might not display the same mate choice preferences as wild orangutans. Orangutan reproductive strategies Orangutans exhibit male-male competition, sexual coercion by males, and female mate choice. Reproductive success is impacted by the interaction between each of these male and female reproductive strategies. Due to the highly dispersed spatial distribution of female orangutans [11,24,25], it is expected that a single male cannot monopolize females or conceptions, result- ing in low male reproductive skew [47]. Our paternity data concluded that a single male can- not monopolize paternity, and the behavioral data on male-female association encounter rates with EPMs further addresses the inability of a single flanged male to monopolize females. Con- sistent with other studies [6,16,30], we found that flanged males at GPNP were able to displace unflanged males associating with females, specifically with sexually active females. However, the mere presence of a flanged male with a female does not keep other males away. But once a sexually active female is an association with a male, regardless of male morph, the pair is less likely to encounter an EPM. Female orangutans at GPNP display a mixed mating strategy pref- erentially associating with prime, flanged males when they are most likely to conceive and with non-prime, unflanged males when they are less likely to conceive [21,38]. Thus, females may be choosing who to associate with based on their probability of conception [21,38]. The limited ability of flanged males to mate guard further highlights the importance of female choice in facultative associations and mating. Therefore, female preference for flanged males, coupled with the flanged male ability to displace unflanged males, operate in parallel leading to higher reproductive success for flanged males, and flanged male inability to completely mate guard females, leads to low reproductive skew among these flanged males. Because both flanged and unflanged males perform forced copulations [11,35], our results cannot speak to the efficacy of that form of sexual coercion in leading to reproductive success. Since unflanged males have lower reproductive success, it appears that harassment by unflanged males is not a successful reproductive strategy. Instead, female preference for flanged males and flanged male competi- tive ability are operating in the same direction, leading to higher reproductive success for flanged males. Alternative reproductive strategies Our results also have important implications for understanding ARTs in male orangutans. ARTs have been hypothesized to be either frequency-dependent evolutionary stable strategies [3] or due to difference in the quality of individuals [1,4]. The first published study of orangu- tan paternity, and still the only study in Sumatran orangutans, found that the two morphs had comparable reproductive success at Ketambe, and thus argued that the two morphs were evo- lutionary stable strategies [7,20]. Now, 20 years later with data from an additional four Bor- nean sites, it is clear, that at least in Borneo, flanged males have higher reproductive success than unflanged males. The relative numbers of flanged and unflanged males are pivotal to our interpretation of paternity data, but accurate counts of the numbers of males in a study site are difficult to obtain due to large home ranges and the difficulty of visually identifying orangutans who tran- sition from unflanged to flanged males. Cross-site comparisons agree that in Sumatra there are approximately twice as many unflanged males as flanged males, whereas in Borneo there is more inter-site variation, but the morphs exist in approximately equal proportions [12,19]. Long-term demographic data at GPNP agrees with these approximations [35]. These PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 13 / 20 PLOS ONEFlanged male orangutans have higher reproductive success proportions indicate that males remain in the unflanged morph longer in Sumatra than in Bor- neo [19]. Because the first paternity study found that unflanged males sired 60% of offspring in a Sumatran population, it was argued that flanged and unflanged male morphs represent alter- native mating strategies that coexist, representing evolutionary stable strategies [7,20]. If this result is representative of all of Sumatra, it could explain why Sumatran males remain unflanged for longer than Bornean males. In this case, unflanged males avoid the energetic and competitive costs of becoming a flanged male [11], while achieving reproductive success. However, the paternity data from Borneo does not support the understanding that male alter- native reproductive strategies are frequency-dependent evolutionary stable strategies. In Bor- neo, 77% of offspring are sired by flanged males while flanged males only represent approximately 50% of all males, demonstrating that the flanged morph is absolutely and rela- tively more successful. Because the reproductive success of the morphs is not related to their proportion in the population, Bornean orangutan ARTs are not frequency-dependent. For Bornean orangutans, ARTs are due to individual differences in quality, with the flanged morph indicating higher quality and unflanged morph indicating lower quality. Our results support the view that the unflanged morph is a transitional stage, where unflanged males are in a ‘waiting room’, avoiding the costs associated with the flanged morph, and ‘making the best of a bad situation’ until they are able to flange [7,93]. The variation in results at different study sites highlights the dynamic nature of ARTs; the ability of each morph to attain repro- ductive success is likely highly dynamic, depending on the relative proportions of each male morph and density of orangutans, which is influenced by food availability [24,95]. More data on the relative proportion of each male morph, male dominance hierarchies, and paternity data from additional study sites of P. abelii in Sumatra will clarify if island differences are due to ecological factors or if there are true species differences. Additionally, studies of the recently described Tapanuli orangutan (Pongo tapanuliensis)—who live on Sumatra, south of Lake Toba, but inhabit less productive forests, live at lower densities, and are less social and thus more similar to Bornean orangutans than Sumatran orangutans [84]—will further help to clar- ify the relative roles of ecology and species differences. These combined paternity results across sites align with the model of developmental arrest in male orangutans developed by Pradhan et al. [95] which explains differences in the ratio of flanged to unflanged males across sites through ecology. According to this model, longer delays in the development of flanges are expected when females are monopolizable by the dominant male because, in this situation, non-dominant flanged males will have lower repro- ductive success; thus, males should remain unflanged to avoid the costs of the flanged male morph [24,46,95]. Dominant male monopolization of females is expected when orangutans live at higher densities and are more gregarious, which is related to increased food availability [24,46]. Thus, where orangutans live at higher densities (i.e., Ketambe, Sumatra), a dominant flanged male is expected to be able to monopolize females, and a smaller proportion of flanged males are expected. It is worth noting that at Ketambe, there are periods of both high repro- ductive skew, where a single flanged male sires many offspring, and periods of low reproduc- tive skew which correspond to times of rank instability. Conversely, where orangutan habitat is less productive and orangutans live at lower densities (i.e., Borneo), a dominant flanged male is not expected to be able to monopolize females, and a greater proportion of flanged males are expected. Paternity results from the two Bornean sites with completely wild orangu- tans agree with this model, showing that in the absence of feeding stations, a single male is not able to monopolize females. And in the Bornean case, we also see short developmental arrest, resulting in relatively more flanged males. In Borneo, with low reproductive skew among flanged males, there is a reproductive benefit to flanging. PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 14 / 20 PLOS ONEFlanged male orangutans have higher reproductive success Despite intense male-male competition [5,12,96] and sexual coercion [11,35], female choice remains an important factor in determining orangutan reproductive outcomes. The impor- tance of female choice may explain why it is that in all primate species with male ARTs, the morphs are attributed to individual differences in quality [23]. In taxa where the costs of repro- duction are disproportionately borne by one sex, we expect strong sexual selection, including mate choice to evolve [97]. In the case of mammals where obligate female gestation and lacta- tion mean that females must invest heavily in reproduction and parental care, we expect female choice to evolve [98], and thus is not surprising that male ARTs in orangutans are signals of male quality that are subject to female choice. We predict that in species with both strong mate choice, driven by differential costs of reproduction, and ARTs, the ARTs will be condition- dependent, rather than frequency-dependent. This study of orangutan paternity determination in GPNP is the first study of orangutan paternity from a completely wild population in a primary rainforest site, without feeding sta- tions, rehabilitant orangutans, or veterinary care, and with known maternal-offspring relation- ships. This enables us to better understand why previous orangutan paternity studies disagree on which morph has higher reproductive success and the degree of male reproductive skew. We show that the ARTs are condition-dependent. Flanged males have higher reproductive success, and unflanged males are ‘making the best of a bad situation’. Supporting information S1 Table. Panel of microsatellite primers used for genotyping. (DOCX) S2 Table. Summary statistics for the 12 microsatellite loci used. (DOCX) S3 Table. GLMM comparisons. (DOCX) S4 Table. Ratio of flanged to unflanged males during each of the five conception periods where paternity and male morph were determined. (DOCX) S5 Table. Flanged male-flanged male dyadic dominance interactions from 2008 to 2014. (DOCX) S6 Table. Rates at which male-female associations encounter an additional or Extra-Pair Male (EPM), expressed as the average number of hours of association per encounter with an EPM. (DOCX) S7 Table. GLMM testing the probability that a male-female pair encountered an EPM. (DOCX) S1 File. Reproductive skew calculation and nonac’s b interpretation. (DOCX) Acknowledgments We thank the Universitas Nasional (UNAS), the Eijkman Research Center for Molecular Biol- ogy, National Research and Innovation Agency–BRIN (Eijkman Institute for Molecular Biol- ogy–EIMB), the Universitas Tanjungpura (UNTAN), the Directorate of Natural Resource PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 15 / 20 PLOS ONEFlanged male orangutans have higher reproductive success Conservation and Ecosystems (KSDAE)–Ministry of Environment and Forestry, the Gunung Palung National Park office (BTNGP), National Research and Innovation Agency (BRIN, for- merly LIPI), the Center for Research and Development in Biology (PPPB), and the Ministry of Research and Technology/National Research and Innovation Agency (RISTEK/BRIN) for their sponsorship, collaboration, and permissions to conduct research in Gunung Palung National Park. We are grateful to all the dedicated field assistants, research assistants, field managers, and students who assisted with project maintenance and data collection. Thank you to the Genome Diversity and Disease Laboratory and DNA Sequencing Unit at EIMB and Hannah Gorman, Morgana Haub, and Jay Coogan for help with genotyping. We would like to thank Carolyn Hodges-Simeon, Christopher Schmitt, and Larissa Swedell for constructive sug- gestions on the manuscript. Author Contributions Conceptualization: Amy M. Scott, Cheryl D. Knott. Data curation: Amy M. Scott, Cheryl D. Knott. Formal analysis: Amy M. Scott, Cheryl D. Knott. Funding acquisition: Amy M. Scott, Cheryl D. Knott. Investigation: Amy M. Scott, Jessica R. Saragih. Methodology: Amy M. Scott, Graham L. Banes, Cheryl D. Knott. Project administration: Amy M. Scott, Tri Wahyu Susanto, Cheryl D. Knott. Resources: Wuryantari Setiadi, Tri Wahyu Susanto, Tatang Mitra Setia, Cheryl D. Knott. Supervision: Wuryantari Setiadi, Tatang Mitra Setia, Cheryl D. Knott. Validation: Amy M. Scott, Graham L. Banes. Visualization: Amy M. Scott. Writing – original draft: Amy M. Scott. Writing – review & editing: Amy M. Scott, Graham L. Banes, Cheryl D. Knott. References 1. Taborsky M, Oliveira RF, Brockmann HJ. The evolution of alternative reproductive tactics: Concepts and questions. In: Oliveira RF, Taborsky M, Brockmann HJ, editors. Alternative Reproductive Tactics: An Integrated Approach. New York City, New York: Cambridge University Press; 2008. p. 1–21. 2. Schuster SM, Wade MJ. Mating Systems and Strategies. Princeton University Press; 2003. 3. Maynard Smith J. Evolution and the Theory of Games. Cambridge University Press; 1982. 4. Eberhard WG. Beetle horn dimorphism: Making the best of a bad lot. Am Nat. 1982; 119(3):420–6. 5. Galdikas BMF. Adult male sociality and reproductive tactics among orangutans at Tanjung Puting. Folia Primatologica. 1985; 45(1):9–24. 6. Mitani JC. Mating behaviour of male orangutans in the Kutai Game Reserve, Indonesia. Anim Behav. 1985; 33:392–402. 7. Utami Atmoko SS, van Hooff JARAM. Alternative male reproductive strategies: Male bimaturism in orangutans. In: Kappeler PM, van Schaik CP, editors. Sexual Selection in Primates: New and Compara- tive Perspectives. Cambridge: Cambridge University Press; 2004. p. 196–207. 8. MacKinnon J. The behaviour and ecology of wild orang-utans (Pongo pygmaeus). Anim Behav. 1974; (22):3–74. 9. Markham R, Groves CP. Weights of wild orang utans. Am J Phys Anthropol. 1990; 81:1–3. https://doi. org/10.1002/ajpa.1330810102 PMID: 2405688 PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 16 / 20 PLOS ONEFlanged male orangutans have higher reproductive success 10. Kralick AE, O’Connell CA, Bastian ML, Hoke MK, Zemel BS, Schurr TG, et al. Beyond dimorphism: Body size variation among adult orangutans is not dichotomous by sex. Integr Comp Biol. 2023 Apr 14; icad015. 11. Knott CD, Kahlenberg SM. Orangutans: Understanding forced copulations. In: Campbell CJ, Fuentes A, MacKinnion KC, Bearder SK, Stumpf RM, editors. Primates in Perspective. 2nd ed. New York: Oxford University Press; 2011. p. 313–26. 12. Utami Atmoko SS, Singleton I, van Noordwijk MA, van Schaik CP, Mitra Setia T. Male-male relation- ships in orangutans. In: Wich SA, Utami Atmoko SS, Mitra Setia T, van Schaik CP, editors. Orangutans: Geographic variation in behavioral ecology and conservation. Oxford University Press; 2009. p. 225– 34. 13. Tajima T, Malim TP, Inoue E. Reproductive success of two male morphs in a free-ranging population of Bornean orangutans. Primates. 2018; 59(2):127–33. https://doi.org/10.1007/s10329-017-0648-1 PMID: 29387973 14. Utami SS. Bimaturism in orang-utan males: Reproductive and ecological strategies [Doctoral Disserta- tion]. [Utrecht]: Universiteit Utrecht; 2000. 15. Rijksen HD. A field study of Sumatran orang-utans (Pongo pygmaeus abelii Lesson 1827): Ecology, behavior, and conservation. Wageningen University and Research; 1978. 16. Rodman PS, Mitani JC. Orangutans: Sexual dimorphism in a solitary species. In: Smuts BB, Cheney DL, Seyfarth RM, Wrangham RW, Struhsaker TT, editors. Primate Societies. Chicago: University of Chicago Press; 1987. p. 146–54. 17. van Schaik CP, van Noordwijk MA, Vogel ER. Ecological sex differences in wild orangutans. In: Wich SA, Utami Atmoko SS, Mitra Setia T, van Schaik CP, editors. Orangutans: Geographic Variation in Behavioral Ecology and Conservation. Oxford University Press; 2009. p. 255–68. 18. Mitani JC. Orangutan activity budgets: Monthly variations and the effect of body size, parturition, and sociality. Am J Primatol. 1989; 18:87–100. 19. Dunkel LP, Arora N, van Noordwijk MA, Utami Atmoko SS, Putra AP, Kru¨ tzen M, et al. Variation in developmental arrest among male orangutans: A comparison between a Sumatran and a Bornean pop- ulation. Front Zool. 2013; 10:1–12. 20. Utami SS, Bruford MW, de Ruiter JR, Hooff JARAM. Male bimaturism and reproductive success in Sumatran orang-utans. Behavioral Ecology. 2002; 13(5):643–52. 21. Knott CD, Emery Thompson M, Stumpf RM, McIntyre MH. Female reproductive strategies in orangu- tans, evidence for female choice and counterstrategies to infanticide in a species with frequent sexual coercion. Proceedings of the Royal Society B. 2010 Jan 7; 277(1678):105–13. https://doi.org/10.1098/ rspb.2009.1552 PMID: 19812079 22. Hunt J, Breuker CJ, Sadowski JA, Moore AJ. Male-male competition, female mate choice and their interaction: Determining total sexual selection. J Evol Biol. 2009; 22(1):13–26. https://doi.org/10.1111/j. 1420-9101.2008.01633.x PMID: 19120810 23. Setchell JM. Alternative reproductive tactics in primates. In: Oliveira RF, Taborsky M, Brockmann HJ, editors. Alternative Reproductive Tactics: An Integrative Approach. Cambridge University Press; 2008. p. 373–98. 24. Delgado RA, van Schaik CP. The behavioral ecology and conservation of the orangutan (Pongo pyg- maeus): A tale of two islands. Evol Anthropol. 2000; 9(5):201–18. 25. Mitra Setia T, Delgado RA, Utami Atmoko SS, Singleton I, van Schaik CP. Social organization and male-female relationships. In: Orangutans: Geographic Variation in Behavioral Ecology and Conserva- tion. 2009. 26. Galdikas BMF. The orangutan long call and snag crashing at Tanjung Puting Reserve. Primates. 1983; 24(3):371–84. 27. Delgado RA. The function of adult male long calls in wild orangutans (Pongo pygmaeus) [Doctoral Dis- sertation]. [Durham, North Carolina]: Duke University; 2003. 28. Spillmann B, Willems EP, van Noordwijk MA, Mitra Setia T, van Schaik CP. Confrontational assessment in the roving male promiscuity mating system of the Bornean orangutan. Behav Ecol Sociobiol. 2017; 71:20. 29. Mitani JC. Sexual selection and adult male orangutan long calls. Anim Behav. 1985; 33:272–83. 30. Fox EA. The function of female choice in the Sumatran orangutan (Pongo pygmaeus abelii) [Doctoral Dissertation]. Duke University; 1998. 31. Schurmann CL. Mating behaviour of wild orang utans. In: de Boer LEM, editor. The orangutan Its biol- ogy and conservation. The Hague: Dr W. Junk Publishers; 1982. p. 269–84. PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 17 / 20 PLOS ONEFlanged male orangutans have higher reproductive success 32. van Noordwijk MA, Utami Atmoko SS, Knott CD, Kuze N, Morrogh-Bernard HC, Oram F, et al. The slow ape: High infant survival and long interbirth intervals in wild orangutans. Vol. 125, Journal of Human Evolution. Elsevier Ltd; 2018. p. 38–49. 33. Galdikas BMF, Wood JW. Birth spacing patterns in humans and apes. Am J Phys Anthropol. 1990; 83:185–91. https://doi.org/10.1002/ajpa.1330830207 PMID: 2248378 34. Aloise King ED, Banks PB, Brooks R. Sexual conflict in mammals: Consequences for mating systems and life history. Mamm Rev. 2013; 43:47–58. 35. Knott CD. Orangutans: Sexual coercion without sexual violence. In: Muller MN, Wrangham RW, editors. Sexual Coercion in Primates and Humans. New York: Harvard University Press; 2009. p. 81–111. 36. Kunz JA, Duvot GJ, van Noordwijk MA, Willems EP, Townsend M, Mardianah N, et al. The cost of asso- ciating with males for Bornean and Sumatran female Orangutans: A hidden form of sexual conflict? Behav Ecol Sociobiol. 2021; 75(6). https://doi.org/10.1007/s00265-020-02948-4 PMID: 33408436 37. Nadler RD. Sexual and reproductive behavior. In: Schwartz JH, editor. Orang-utan Biology. New York: Oxford University Press; 1988. p. 105–16. 38. Stumpf RM, Emery Thompson M, Knott CD. A comparison of female mating strategies in Pan troglo- dytes and Pongo spp. Int J Primatol. 2008 Aug 13; 29(4):865–84. 39. 40. 41. van Schaik CP, Hodges JK, Nunn CL. Paternity confusion and the ovarian cycle of female primates. In: van Schaik CP, Janson CH, editors. Infanticide by Males and its Implications. Cambridge University Press; 2000. p. 361–87. Jones AG, Ardren WR. Methods of parentage analysis in natural populations. Vol. 12, Molecular Ecol- ogy. 2003. p. 2511–23. https://doi.org/10.1046/j.1365-294x.2003.01928.x PMID: 12969458 Jones AG, Small CM, Paczolt KA, Ratterman NL. A practical guide to methods of parentage analysis. Vol. 10, Molecular Ecology Resources. 2010. p. 6–30. https://doi.org/10.1111/j.1755-0998.2009. 02778.x PMID: 21564987 42. Kalinowski ST, Taper ML, Marshall TC. Revising how the computer program CERVUS accommodates genotyping error increases success in paternity assignment. Mol Ecol. 2007; 16(5):1099–106. https:// doi.org/10.1111/j.1365-294X.2007.03089.x PMID: 17305863 43. Marshall TC, Slate J, Kruuk LEB, Pemberton JM. Statistical confidence for likelihood-based paternity inference in natural populations. Mol Ecol. 1998; 7(5):639–55. https://doi.org/10.1046/j.1365-294x. 1998.00374.x PMID: 9633105 44. Banes GL, Galdikas BMF, Vigilant L. Male orang-utan bimaturism and reproductive success at Camp Leakey in Tanjung Puting National Park, Indonesia. Behav Ecol Sociobiol. 2015; 69(11):1785–94. 45. Goossens B, Setchell JM, James SS, Funk SM, Chikhi L, Abulani A, et al. Philopatry and reproductive success in Bornean orang-utans (Pongo pygmaeus). Mol Ecol. 2006 Aug; 15(9):2577–88. 46. van Schaik CP, Marshall AJ, Wich SA. Geographic variation in orangutan behavior and biology. In: Wich SA, Utami Atmoko SS, Mitra Setia T, van Schaik CP, editors. Orangutans: Geographic Variation in Behavioral Ecology and Conservation. Oxford University Press; 2009. p. 352–61. 47. Ostner J, Nunn CL, Schu¨ lke O. Female reproductive synchrony predicts skewed paternity across pri- mates. Behavioral Ecology. 2008; 19(6):1150–8. https://doi.org/10.1093/beheco/arn093 PMID: 19018288 48. Gogarten JF, Koenig A. Reproductive seasonality is a poor predictor of receptive synchrony and male reproductive skew among nonhuman primates. Behav Ecol Sociobiol. 2013; 67(1):123–34. 49. Aureli F, Schaffner CM, Boesch C, Bearder SK, Call J, Chapman CA, et al. Fission-fusion dynamics. Curr Anthropol. 2008; 49(4):627–54. 50. Knott CD, Thompson ME, Wich SA. The ecology of female reproduction in wild orangutans. In: Wich SA, Atmoko SSU, Setia TM, Van Schaik CP, editors. Orangutans: Geographic Variation in Behavioral Ecology and Conservation. New York: Oxford University Press; 2009. p. 171–88. 51. Altmann SA. A field study of the sociobiology of rhesus monkeys, Macaca mulatta. Ann N Y Acad Sci. 1962; 102(2):338–435. 52. Port M, Kappeler PM. The utility of reproductive skew models in the study of male primates, a critical evaluation. Evol Anthropol. 2010; 19(2):46–56. 53. Knott CD, Kane EE, Achmad M, Barrow EJ, Bastian ML, Beck J, et al. The Gunung Palung Orangutan Project: Twenty-five years at the intersection of research and conservation in a critical landscape in Indonesia. Biol Conserv. 2021; 255(October 2020):108856. 54. Singleton I, Knott CD, Morrogh-Bernard HC, Wich SA, van Schaik CP. Ranging behavior of orangutan females and social organization. In: Wich SA, Utami Atmoko SS, Mitra Setia T, van Schaik CP, editors. Orangutans: Geographic Variation in Behavioral Ecology and Conservation. Oxford University Press; 2009. p. 205–13. PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 18 / 20 PLOS ONEFlanged male orangutans have higher reproductive success 55. Morrogh-Bernard HC, Morf N, Chivers DJ, Kru¨ tzen M. Dispersal patterns of orang-utans (Pongo spp.) in a Bornean peat-swamp forest. Int J Primatol. 2011; 32(2):362–76. 56. Marshall AJ, Beaudrot LH, Wittmer HU. Responses of primates and other frugivorous vertebrates to plant resource variability over space and time at Gunung Palung National Park. Int J Primatol. 2014; 35:1178–201. 57. Knott CD, Scott AM, O’Connell CA, Scott KS, Laman TG, Riyandi, et al. Possible male infanticide in wild orangutans and a re-evaluation of infanticide risk. Sci Rep. 2019; 9:7806. https://doi.org/10.1038/ s41598-019-42856-w PMID: 31127126 58. Knott CD. Energetic responses to food availability in the great apes: Implications for hominin evolution. In: Brockman DK, van Schaik CP, editors. Seasonality in Primates: Studies of Living and Extinct Human and Non-Human Primates. New York: Cambridge University Press; 2005. p. 351–78. 59. Scott A, Banes G, Setiadi W, Saragih J, Susanto W, Mitra Setia T, et al. OpenBU. 2022. Dataset for: Mate guarding by male orangutans in Gunung Palung National Park, Knott Lab. Available from: https:// hdl.handle.net/2144/45321 60. Knott CD, Beaudrot LH, Snaith T V., White S, Tschauner H, Planansky G. Female-female competition in Bornean orangutans. Int J Primatol. 2008 Aug 2; 29(4):975–97. 61. Mitani JC, Grether GF, Rodman PS, Priatna D. Associations among wild orang-utans: Sociality, passive aggregations or chance? Anim Behav. 1991; 42(1):33–46. 62. Graham CE. Reproductive physiology. In: Schwartz JH, editor. Orang-utan Biology. Oxford University Press; 1988. p. 91–103. 63. Scott AM, Knott CD, Susanto TW. Are male orangutans a threat to infants? Evidence of mother–off- spring counterstrategies to infanticide in Bornean orangutans (Pongo pygmaeus wurmbii). Int J Prima- tol. 2019; 40(3):435–55. 64. Drews C. The concept and definition of dominance in animal behaviour. Behaviour. 1993; 125(3– 4):283–313. 65. Roeder AD, Archer FI, Poinar HN, Morin PA. A novel method for collection and preservation of faeces for genetic studies. Mol Ecol Notes. 2004 Dec; 4(4):761–4. 66. Nsubuga AM, Robbins MM, Roeder AD, Morin PA, Boesch C, Vigilant L. Factors affecting the amount of genomic DNA extracted from ape faeces and the identification of an improved sample storage method. Mol Ecol. 2004 Jul; 13(7):2089–94. https://doi.org/10.1111/j.1365-294X.2004.02207.x PMID: 15189228 67. Morin PA, Chambers KE, Boesch C, Vigilant L. Quantitative polymerase chain reaction analysis of DNA from noninvasive samples for accurate microsatellite genotyping of wild chimpanzees (Pan troglodytes verus). Mol Ecol. 2001 Dec 21; 10(7):1835–44. 68. Kanthaswamy S, Smith DG. Population subdivision and gene flow among wild orangutans. Primates. 2002; 43(4):315–27. https://doi.org/10.1007/BF02629605 PMID: 12426465 69. Nietlisbach P, Nater A, Greminger MP, Arora N, Kru¨tzen M. A multiplex-system to target 16 male-spe- cific and 15 autosomal genetic markers for orang-utans (genus: Pongo). Conserv Genet Resour. 2010; 2:153–8. 70. Zhang Y wu Morin PA, Ryder OA Zhang YP. A set of human tri- and tetra-nucleotide microsatellite loci useful for population analyses in gorillas (Gorilla gorilla gorilla) and orangutans (Pongo pygmaeus). Conservation Genetics. 2001; 2(4):391–5. 71. Di Fiore A. A rapid genetic method for sex assignment in non-human primates. Conservation Genetics. 2005; 6(6):1053–8. 72. Arandjelovic M, Guschanski K, Schubert G, Harris TR, Thalmann O, Siedel H, et al. Two-step multiplex polymerase chain reaction improves the speed and accuracy of genotyping using DNA from noninva- sive and museum samples. Mol Ecol Resour. 2009 Jan; 9(1):28–36. https://doi.org/10.1111/j.1755- 0998.2008.02387.x PMID: 21564562 73. Van Oosterhout C, Hutchinson WF, Wills DPM, Shipley P. MICRO-CHECKER: Software for identifying and correcting genotyping errors in microsatellite data. Mol Ecol Notes. 2004; 4(3):535–8. 74. Waits LP, Luikart G, Taberlet P. Estimating the probability of identity among genotypes in natural popu- lations: Cautions and guidelines. Mol Ecol. 2001; 10(1):249–56. https://doi.org/10.1046/j.1365-294x. 2001.01185.x PMID: 11251803 75. Jones OR, Wang J. COLONY: A program for parentage and sibship inference from multilocus genotype data. Mol Ecol Resour. 2010; 10(3):551–5. https://doi.org/10.1111/j.1755-0998.2009.02787.x PMID: 21565056 76. Nonacs P. Measuring and using skew in the study of social behavior and evolution. Am Nat. 2000; 156 (6):577–89. https://doi.org/10.1086/316995 PMID: 29592547 PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 19 / 20 PLOS ONEFlanged male orangutans have higher reproductive success 77. Nonacs P. Measuring the reliability of skew indices is there one best index. Vol. 65, Animal Behaviour. Academic Press; 2003. p. 615–27. 78. Ross CT, Jaeggi A V., Borgerhoff Mulder M, Smith JE, Smith EA, Gavrilets S, et al. The multinomial index: A robust measure of reproductive skew: The Multinomial Index. Proceedings of the Royal Society B: Biological Sciences. 2020 Oct 14; 287(1936). 79. Nonacs P. Skew Calculator 2013 [Internet]. 2013 [cited 2021 Nov 30]. Available from: 19/04/ 2023https://www.dropbox.com/home/2013%20Version 80. McDonald JH. Handbook of Biological Statistics. 3rd ed. Sparky House Publishing; 2014. 81. Zuur AF, Ieno EN, Walker NJ, Saveliev AA, Smith GM. Mixed Effects Models and Extensions in Ecology with R. Springer; 2009. 82. Knott CD. Reproductive, physiological and behavioral responses of orangutans in Borneo to fluctua- tions in food availability [Doctoral Dissertation]. Harvard University; 1999. 83. Sugardjito J, te Boekhorst IJA, Vanhooff J. Ecological constraints on the grouping of wild orangutans (Pongo pygmaeus) in the Gunung-Leuser-National-Park, Sumatra, Indonesia. Int J Primatol. 1987; 8 (1):17–41. 84. Roth TS, Rianti P, Fredriksson GM, Wich SA, Nowak MG. Grouping behavior of Sumatran orangutans and Tapanuli orangutans living in forests with low fruit abundance. Am J Primatol. 2020;e23123. 85. Wich SA, Geurts ML, Mitra Setia T, Utami Atmoko SS. Influences of food availability on Sumatran orangutan sociality and reproduction. In: Hohmann G, Robbins MM, Boesch C, editors. Feeding Ecol- ogy in Apes and Other Primates. Cambridge: Cambridge University Press; 2006. p. 337–58. 86. van Schaik CP. The socioecology of fission-fusion sociality in orangutans. Primates. 1999; 40(1):69– 86. https://doi.org/10.1007/BF02557703 PMID: 23179533 87. R Core Team. R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing; 2021. 88. Pohlert T. The Pairwise Multiple Comparison of Mean Ranks Package (PMCMR). [Internet]. R pack- age. 2014. Available from: http://CRAN.R-project.org/package=PMCMR 89. Bates D, Machler M, Bolker BM, Walker SC. Fitting linear mixed-effects models using lme4. J Stat Softw. 2014;arXiv preprint arXiv:1406.5823. 90. Gelman A, Su Y, Yajima M, Hill J, Grazia Pittau M, Kerman J, et al. Arm: Data analysis using regression and multilevel/hierarchical models [Internet]. R Package; 2020. Available from: https://cran.r-project. org/web/packages/arm/index.html 91. Wickham H. ggplot2: Elegant graphics for data analysis [Internet]. R Package; 2016. Available from: https://ggplot2.tidyverse.org 92. Claus O. Wilke. cowplot: Streamlined plot theme and plot annotations for “ggplot2” [Internet]. 2017. Available from: https://CRAN.R-project.org/package=cowplot 93. Utami Atmoko SS, Mitra Setia T, Goossens B, James SS, Knott CD, Morrogh-Bernard HC, et al. Orang- utan mating behavior and strategies. In: Orangutans: Geographic Variation in Behavioral Ecology and Conservation. 2009. 94. O’Connell CA, Susanto TW, Knott CD. Sociosexual behavioral patterns involving nulliparous female orangutans (Pongo sp.) reflect unique challenges during the adolescent period. Am J Primatol. 2019; (September):1–12. 95. Pradhan GR, van Noordwijk MA, van Schaik CP. A model for the evolution of developmental arrest in male orangutans. Am J Phys Anthropol. 2012; 25:18–25. 96. van Schaik CP, van Hooff JARAM. Toward an understanding of the orangutan’s social system. In: Marchant LF, editor. Great Ape Societies. Cambridge University Press; 1996. p. 3–15. 97. Kokko H, Monaghan P. Predicting the direction of sexual selection. Ecol Lett. 2001; 4(2):159–65. 98. Kokko H, Jennions MD. Parental investment, sexual selection and sex ratios. J Evol Biol. 2008; 21 (4):919–48. https://doi.org/10.1111/j.1420-9101.2008.01540.x PMID: 18462318 PLOS ONE \| https://doi.org/10.1371/journal.pone.0296688 February 9, 2024 20 / 20 PLOS ONE	null
10.1103_physrevd.107.023530.pdf	null	null	Editors' Suggestion Featured in Physics PHYSICAL REVIEW D 107, 023530 (2023) Joint analysis of Dark Energy Survey Year 3 data and CMB lensing from SPT and Planck. II. Cross-correlation measurements and cosmological constraints C. Chang ,1,2 Y. Omori,1,2,3,4 E. J. Baxter,5 C. Doux,6 A. Choi,7 S. Pandey,6 A. Alarcon,8 O. Alves,9,10 A. Amon,4 F. Andrade-Oliveira,9 K. Bechtol,11 M. R. Becker,8 G. M. Bernstein,6 F. Bianchini,3,4,12 J. Blazek,13,14 L. E. Bleem,8,2 H. Camacho,15,10 A. Campos,16 A. Carnero Rosell,10,17,18 M. Carrasco Kind,19,20 R. Cawthon,21 R. Chen,22 J. Cordero,23 T. M. Crawford,1,2 M. Crocce,24,25 C. Davis,4 J. DeRose,26 S. Dodelson,16,27 A. Drlica-Wagner,1,2,28 K. Eckert,6 T. F. Eifler,29,30 F. Elsner,31 J. Elvin-Poole,32,33 S. Everett,34 X. Fang,29,35 A. Fert´e,30 P. Fosalba,24,25 O. Friedrich,36 M. Gatti,6 G. Giannini,37 D. Gruen,38 R. A. Gruendl,19,20 I. Harrison,23,39,40 K. Herner,28 H. Huang,29,41 E. M. Huff,30 D. Huterer,9 M. Jarvis,6 A. Kovacs,17,18 E. Krause,29 N. Kuropatkin,28 P.-F. Leget,4 P. Lemos,31,42 A. R. Liddle,43 N. MacCrann,44 J. McCullough,4 J. Muir,45 J. Myles,3,4,46 A. Navarro-Alsina,47 Y. Park,48 A. Porredon,32,33 J. Prat,1,2 M. Raveri,6 R. P. Rollins,23 A. Roodman,4,46 R. Rosenfeld,10,49 A. J. Ross,32 E. S. Rykoff,4,46 C. Sánchez,6 J. Sanchez,28 L. F. Secco,2 I. Sevilla-Noarbe,50 E. Sheldon,51 T. Shin,6 M. A. Troxel,22 I. Tutusaus,24,25,52 T. N. Varga,53,54 N. Weaverdyck,9,26 R. H. Wechsler,3,4,46 W. L. K. Wu,4,46 B. Yanny,28 B. Yin,16 Y. Zhang,28 J. Zuntz,55 T. M. C. Abbott,56 M. Aguena,10 S. Allam,28 J. Annis,28 D. Bacon,57 B. A. Benson,1,2,28 E. Bertin,58,59 S. Bocquet,60 D. Brooks,31 D. L. Burke,4,46 J. E. Carlstrom,1,2,8,61,62 J. Carretero,37 C. L. Chang,1,2,8 R. Chown,63,64 M. Costanzi,65,66,67 L. N. da Costa,10,68 A. T. Crites,1,2,69 M. E. S. Pereira,70 T. de Haan,71,72 J. De Vicente,50 S. Desai,73 H. T. Diehl,28 M. A. Dobbs,74,75 P. Doel,31 W. Everett,76 I. Ferrero,77 B. Flaugher,28 D. Friedel,19 J. Frieman,2,28 J. García-Bellido,78 E. Gaztanaga,24,25 E. M. George,79,72 T. Giannantonio,36,80 N. W. Halverson,76,81 S. R. Hinton,82 G. P. Holder,20,75,83 D. L. Hollowood,34 W. L. Holzapfel,72 K. Honscheid,32,33 J. D. Hrubes,84 D. J. James,85 L. Knox,86 K. Kuehn,87,88 O. Lahav,31 A. T. Lee,26,89 M. Lima,10,90 D. Luong-Van,84 M. March,6 J. J. McMahon,1,2,61,62 P. Melchior,91 F. Menanteau,19,20 S. S. Meyer,1,2,61,92 R. Miquel,37,93 L. Mocanu,1,2 J. J. Mohr,94,95,96 R. Morgan,11 T. Natoli,1,2 S. Padin,1,2,97 A. Palmese,35 F. Paz-Chinchón,19,80 A. Pieres,10,68 A. A. Plazas Malagón,91 C. Pryke,98 C. L. Reichardt,12 M. Rodríguez-Monroy,50 A. K. Romer,42 J. E. Ruhl,99 E. Sanchez,50 K. K. Schaffer,2,61,100 M. Schubnell,9 S. Serrano,24,25 E. Shirokoff,1,2 M. Smith,101 Z. Staniszewski,99,30 A. A. Stark,102 E. Suchyta,103 G. Tarle,9 D. Thomas,57 C. To,32 J. D. Vieira,20,83 J. Weller,53,54 and R. Williamson104,1,2 (DES & SPT Collaborations) 1Department of Astronomy and Astrophysics, University of Chicago, Chicago, Illinois 60637, USA 2Kavli Institute for Cosmological Physics, University of Chicago, Chicago, Illinois 60637, USA 3Department of Physics, Stanford University, 382 Via Pueblo Mall, Stanford, California 94305, USA 4Kavli Institute for Particle Astrophysics & Cosmology, P. O. Box 2450, Stanford University, Stanford, California 94305, USA 5Institute for Astronomy, University of Hawai‘i, 2680 Woodlawn Drive, Honolulu, Hawaii 96822, USA 6Department of Physics and Astronomy, University of Pennsylvania, Philadelphia, Pennsylvania 19104, USA 7California Institute of Technology, 1200 East California Blvd, MC 249-17, Pasadena, California 91125, USA 8Argonne National Laboratory, 9700 South Cass Avenue, Lemont, Illinois 60439, USA 9Department of Physics, University of Michigan, Ann Arbor, Michigan 48109, USA 10Laboratório Interinstitucional de e-Astronomia - LIneA, Rua Gal. Jos´e Cristino 77, Rio de Janeiro, RJ - 20921-400, Brazil 11Physics Department, 2320 Chamberlin Hall, University of Wisconsin-Madison, 1150 University Avenue Madison, Wisconsin 53706-1390 12School of Physics, University of Melbourne, Parkville, Victoria 3010, Australia 13Department of Physics, Northeastern University, Boston, Massachusetts 02115, USA 14Laboratory of Astrophysics, École Polytechnique F´ed´erale de Lausanne (EPFL), Observatoire de Sauverny, 1290 Versoix, Switzerland 15Instituto de Física Teórica, Universidade Estadual Paulista, São Paulo, Brazil 16Department of Physics, Carnegie Mellon University, Pittsburgh, Pennsylvania 15312, USA 17Instituto de Astrofisica de Canarias, E-38205 La Laguna, Tenerife, Spain 18Universidad de La Laguna, Dpto. Astrofísica, E-38206 La Laguna, Tenerife, Spain 2470-0010=2023=107(2)=023530(25) 023530-1 © 2023 American Physical Society C. CHANG et al. PHYS. REV. D 107, 023530 (2023) 19Center for Astrophysical Surveys, National Center for Supercomputing Applications, 1205 West Clark St., Urbana, Illinois 61801, USA 20Department of Astronomy, University of Illinois at Urbana-Champaign, 1002 W. Green Street, Urbana, Illinois 61801, USA 21Physics Department, William Jewell College, Liberty, Missouri, 64068 22Department of Physics, Duke University Durham, North Carolina 27708, USA 23Jodrell Bank Center for Astrophysics, School of Physics and Astronomy, University of Manchester, Oxford Road, Manchester, M13 9PL, United Kingdom 24Institut d’Estudis Espacials de Catalunya (IEEC), 08034 Barcelona, Spain 25Institute of Space Sciences (ICE, CSIC), Campus UAB, Carrer de Can Magrans, s/n, 08193 Barcelona, Spain 26Lawrence Berkeley National Laboratory, 1 Cyclotron Road, Berkeley, California 94720, USA 27NSF AI Planning Institute for Physics of the Future, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213, USA 28Fermi National Accelerator Laboratory, P. O. Box 500, Batavia, Illinois 60510, USA 29Department of Astronomy/Steward Observatory, University of Arizona, 933 North Cherry Avenue, Tucson, Arizona 85721-0065, USA 30Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Drive, Pasadena, California 91109, USA 31Department of Physics & Astronomy, University College London, Gower Street, London, WC1E 6BT, United Kingdom 32Center for Cosmology and Astro-Particle Physics, The Ohio State University, Columbus, Ohio 43210, USA 33Department of Physics, The Ohio State University, Columbus, Ohio 43210, USA 34Santa Cruz Institute for Particle Physics, Santa Cruz, California 95064, USA 35Department of Astronomy, University of California, Berkeley, 501 Campbell Hall, Berkeley, California 94720, USA 36Kavli Institute for Cosmology, University of Cambridge, Madingley Road, Cambridge CB3 0HA, United Kingdom 37Institut de Física d’Altes Energies (IFAE), The Barcelona Institute of Science and Technology, Campus UAB, 08193 Bellaterra (Barcelona) Spain 38University Observatory, Faculty of Physics, Ludwig-Maximilians-Universitat, Scheinerstrasse 1, 81679 Munich, Germany 39Department of Physics, University of Oxford, Denys Wilkinson Building, Keble Road, Oxford OX1 3RH, United Kingdom 40School of Physics and Astronomy, Cardiff University, CF24 3AA, United Kingdom 41Department of Physics, University of Arizona, Tucson, Arizona 85721, USA 42Department of Physics and Astronomy, Pevensey Building, University of Sussex, Brighton, BN1 9QH, United Kingdom 43Instituto de Astrofísica e Ciências do Espaço, Faculdade de Ciências, Universidade de Lisboa, 1769-016 Lisboa, Portugal 44Department of Applied Mathematics and Theoretical Physics, University of Cambridge, Cambridge CB3 0WA, United Kingdom 45Perimeter Institute for Theoretical Physics, 31 Caroline St. North, Waterloo, Ontario N2L 2Y5, Canada 46SLAC National Accelerator Laboratory, Menlo Park, California 94025, USA 47Instituto de Física Gleb Wataghin, Universidade Estadual de Campinas, 13083-859, Campinas, SP, Brazil 48Kavli Institute for the Physics and Mathematics of the Universe (WPI), UTIAS, The University of Tokyo, Kashiwa, Chiba 277-8583, Japan 49ICTP South American Institute for Fundamental Research Instituto de Física Teórica, Universidade Estadual Paulista, São Paulo, Brazil 50Centro de Investigaciones Energ´eticas, Medioambientales y Tecnológicas (CIEMAT), Madrid, Spain 51Brookhaven National Laboratory, Bldg 510, Upton, New York 11973, USA 52D´epartement de Physique Th´eorique and Center for Astroparticle Physics, Universit´e de Gen`eve, 24 quai Ernest Ansermet, CH-1211 Geneva, Switzerland 53Max Planck Institute for Extraterrestrial Physics, Giessenbachstrasse, 85748 Garching, Germany 54Universitäts-Sternwarte, Fakultät für Physik, Ludwig-Maximilians Universität München, Scheinerstrasse 1, 81679 München, Germany 55Institute for Astronomy, University of Edinburgh, Edinburgh EH9 3HJ, United Kingdom 023530-2 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) 56Cerro Tololo Inter-American Observatory, NSF’s National Optical-Infrared Astronomy Research Laboratory, Casilla 603, La Serena, Chile 57Institute of Cosmology and Gravitation, University of Portsmouth, Portsmouth, PO1 3FX, United Kingdom 58CNRS, UMR 7095, Institut d’Astrophysique de Paris, F-75014, Paris, France 59Sorbonne Universit´es, UPMC Univ Paris 06, UMR 7095, Institut d’Astrophysique de Paris, F-75014, Paris, France 60Ludwig-Maximilians-Universität, Scheiner- strasse 1, 81679 Munich, Germany 61Enrico Fermi Institute, University of Chicago, 5640 South Ellis Avenue, Chicago, Illinois 60637, USA 62Department of Physics, University of Chicago, 5640 South Ellis Avenue, Chicago, Illinois 60637, USA 63Department of Physics & Astronomy, The University of Western Ontario, London Ontario N6A 3K7, Canada 64Institute for Earth and Space Exploration, The University of Western Ontario, London Ontario N6A 3K7, Canada 65Astronomy Unit, Department of Physics, University of Trieste, via Tiepolo 11, I-34131 Trieste, Italy 66INAF-Osservatorio Astronomico di Trieste, via G. B. Tiepolo 11, I-34143 Trieste, Italy 67Institute for Fundamental Physics of the Universe, Via Beirut 2, 34014 Trieste, Italy 68Observatório Nacional, Rua Gal. Jos´e Cristino 77, Rio de Janeiro, RJ - 20921-400, Brazil 69Department of Astronomy & Astrophysics, University of Toronto, 50 St George St, Toronto, Ontario, M5S 3H4, Canada 70Hamburger Sternwarte, Universität Hamburg, Gojenbergsweg 112, 21029 Hamburg, Germany 71High Energy Accelerator Research Organization (KEK), Tsukuba, Ibaraki 305-0801, Japan 72Department of Physics, University of California, Berkeley, California 94720, USA 73Department of Physics, IIT Hyderabad, Kandi, Telangana 502285, India 74Department of Physics and McGill Space Institute, McGill University, 3600 Rue University, Montreal, Quebec H3A 2T8, Canada 75Canadian Institute for Advanced Research, CIFAR Program in Gravity and the Extreme Universe, Toronto, Ontario M5G 1Z8, Canada 76Department of Astrophysical and Planetary Sciences, University of Colorado, Boulder, Colorado 80309, USA 77Institute of Theoretical Astrophysics, University of Oslo, P.O. Box 1029 Blindern, NO-0315 Oslo, Norway 78Instituto de Fisica Teorica UAM/CSIC, Universidad Autonoma de Madrid, 28049 Madrid, Spain 79European Southern Observatory, Karl-Schwarzschild-Straße 2, 85748 Garching, Germany 80Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, United Kingdom 81Department of Physics, University of Colorado, Boulder, Colorado, 80309, USA 82School of Mathematics and Physics, University of Queensland, Brisbane, QLD 4072, Australia 83Department of Physics, University of Illinois Urbana-Champaign, 1110 West Green Street, Urbana, Illinois 61801, USA 84University of Chicago, 5640 South Ellis Avenue, Chicago, Illinois 60637, USA 85Center for Astrophysics \| Harvard & Smithsonian, 60 Garden Street, Cambridge, Massachusetts 02138, USA 86Department of Physics, University of California, One Shields Avenue, Davis, California 95616, USA 87Australian Astronomical Optics, Macquarie University, North Ryde, New South Wales 2113, Australia 88Lowell Observatory, 1400 Mars Hill Rd, Flagstaff, Arizona 86001, USA 89Department of Physics, University of California, Berkeley, California, 94720, USA 90Departamento de Física Matemática, Instituto de Física, Universidade de São Paulo, CP 66318, São Paulo, SP, 05314-970, Brazil 91Department of Astrophysical Sciences, Princeton University, Peyton Hall, Princeton, New Jersey 08544, USA 92Department of Physics, University of Chicago, 5640 South Ellis Avenue, Chicago, Illinois, 60637, USA 93Institució Catalana de Recerca i Estudis Avançats, E-08010 Barcelona, Spain 94Ludwig-Maximilians-Universität, Scheiner- str. 1, 81679 Munich, Germany 95Excellence Cluster Universe, Boltzmannstr. 2, 85748 Garching, Germany 96Max-Planck-Institut fur extraterrestrische Physik, Giessenbachstrasse 85748 Garching, Germany 97California Institute of Technology, 1200 East California Boulevard, Pasadena, California 91125, USA 98School of Physics and Astronomy, University of Minnesota, 116 Church Street SE Minneapolis, Minnesota 55455, USA 99Department of Physics, Case Western Reserve University, Cleveland, Ohio 44106, USA 023530-3 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) 100Liberal Arts Department, School of the Art Institute of Chicago, Chicago, Illinois USA 60603 101School of Physics and Astronomy, University of Southampton, Southampton, SO17 1BJ, United Kingdom 102Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, Massachusetts 02138, USA 103Computer Science and Mathematics Division, Oak Ridge National Laboratory, Oak Ridge, Tennessee 37831 104Jet Propulsion Laboratory, California Institute of Technology, Pasadena, California 91109, USA (Received 4 April 2022; accepted 30 September 2022; published 31 January 2023) Cross-correlations of galaxy positions and galaxy shears with maps of gravitational lensing of the cosmic microwave background (CMB) are sensitive to the distribution of large-scale structure in the Universe. Such cross-correlations are also expected to be immune to some of the systematic effects that complicate correlation measurements internal to galaxy surveys. We present measurements and modeling of the cross-correlations between galaxy positions and galaxy lensing measured in the first three years of data from the Dark Energy Survey with CMB lensing maps derived from a combination of data from the 2500 deg2 SPT-SZ survey conducted with the South Pole Telescope and full-sky data from the Planck satellite. The CMB lensing maps used in this analysis have been constructed in a way that minimizes biases from the thermal Sunyaev Zel’dovich effect, making them well suited for cross-correlation studies. The total signal-to-noise of the cross-correlation measurements is 23.9 (25.7) when using a choice of angular scales optimized for a linear (nonlinear) galaxy bias model. We use the cross-correlation measurements to obtain constraints on cosmological parameters. For our fiducial galaxy sample, m ¼ 0.272þ0.032 which consist of four bins of magnitude-selected galaxies, we find constraints of Ω −0.052 and S8 ≡ σ8 −0.028 ) when assuming linear (nonlinear) galaxy bias in our modeling. Considering only the cross-correlation of galaxy shear with −0.061 and S8 ¼ 0.740þ0.034 CMB lensing, we find Ω −0.029 . Our constraints on S8 are consistent with recent cosmic shear measurements, but lower than the values preferred by primary CMB measurements from Planck. −0.044 and S8 ¼ 0.734þ0.035 m ¼ 0.245þ0.026 m ¼ 0.270þ0.043 ¼ 0.736þ0.032 ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ m=0.3 Ω −0.028 (Ω p DOI: 10.1103/PhysRevD.107.023530 I. INTRODUCTION Significant progress has been made recently in using cross-correlations between galaxy imaging and cosmic microwave background (CMB) surveys to constrain cos- mological parameters. These developments have come naturally as ongoing galaxy and CMB surveys collect increasingly sensitive data across larger and larger over- lapping areas of the sky. The Dark Energy Survey [DES, 1] is the largest galaxy weak lensing survey today, covering ∼5000 deg2 of sky that is mostly in the southern hemi- sphere. By design, the DES footprint overlaps with high- resolution CMB observations from the South Pole Telescope [SPT, 2], enabling a large number of cross- correlation analyses [3–12]. Although CMB photons originate from the high-redshift Universe, their trajectories are deflected by low-redshift structures as a result of gravitational lensing—these are the same structures traced by the distributions of galaxies and the galaxy weak lensing signal measured in optical galaxy surveys. Cross-correlating CMB lensing with galaxy sur- veys therefore allows us to extract information stored in the large-scale structure. t κ CMBi, CMB, and hγ In this work, we analyze both hδgκ the cross correlation of the galaxy density field δg and the CMB CMBi,1 the weak lensing convergence field κ cross correlation of the galaxy weak lensing shear field γ and κ CMB. Notably, these two two-point functions correlate measurements from very different types of surveys (galaxy surveys in the optical and CMB surveys in the millimeter), and are therefore expected to be very robust to systematic biases impacting only one type of survey. Furthermore, CMB lensing is sensitive to a broad range of redshift, with peak sensitivity at redshift z ∼ 2; galaxy lensing, on the other hand, is sensitive to structure at z ≲ 1 for current surveys. As a result, the CMB lensing cross-correlation functions, hδgκ CMBi, are expected to increase in signal-to-noise relative to galaxy lensing correlations as one considers galaxy samples that extend to higher redshift. Our analysis relies on the first three years (Y3) of galaxy observations from DES and a CMB lensing map con- structed using data from the 2500 deg2 SPT-SZ survey [13] CMBi þ hγ κ t 1The “t” subscript denotes the tangential component of shear, which will be discussed in Sec. IV. 023530-4 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) t κ CMBi þ hγ and Planck [14]. The combined signal-to-noise of the hδgκ CMBi measurements used in the present cosmological analysis is roughly a factor two larger than in the earlier DES þ SPT results presented in [11], which used first year (Y1) DES data. This large improvement in signal-to-noise derives from two main advancements: (1) We have adopted a different methodology in con- structing the CMB lensing map, which results in much lower contamination from the thermal Su- nyaev Zel’dovich (tSZ) effect, allowing small-scale information to be used in the cosmological analysis. This methodology is described in [15]. (2) Data from DES Y3 covers an area approximately three times larger than DES Y1 and is slightly deeper. Along with the significant increase in signal-to-noise, we have also updated our models for the correlation functions to include a number of improvements following [16]. These include an improved treatment of galaxy intrinsic align- ments, inclusion of magnification effects on the lens galaxy density, and application of the so-called lensing ratio likelihood described in [17]. The analysis presented here is the second of a series of three papers: In [15] (Paper I) we describe the construction of the combined, tSZ-cleaned SPT þ Planck CMB lensing map and the methodology for the cosmological analysis. In this paper (Paper II), we present the data measurements of the cross-correlation probes hδgκ CMBi, a series of diagnostic tests, and cosmological constraints from this cross-correlation combination. In [18] (Paper III), we will present the joint cosmological constraints from hδgκ CMBi and the DES-only 3 × 2 pt probes,2 and tests of consistency between the two, as well as constraints from a joint analysis with the CMB lensing auto-spectrum. CMBi þ hγ CMBi þ hγ κ κ t t Suprime-Cam Subaru Similar analyses have recently been carried out using different galaxy imaging surveys and CMB data. [19] studied the cross-correlation of the galaxy weak lensing from the Hyper Strategic Program Survey [HSC-SSP, 20] and the Planck lensing map [21]; [22] used the same HSC galaxy weak lensing measurement to cross-correlate with CMB lensing from the POLARBEAR experiment [23]; [24] cross-correlated gal- axy weak lensing from the Kilo-Degree Survey [KiDS, 25] and the CMB lensing map from the Atacama Cosmology Telescope [ACT, 26]; and [27] cross-correlated the galaxy density measured in unWISE data [28] with Planck CMB lensing. Compared to these previous studies, in addition to the new datasets, this paper is unique in that we combine hδgκ CMBi. Moreover, our analysis uses the CMBi and hγ κ t 2The 3 × 2 pt probes refer to a combination of three two-point functions of the galaxy density field δg and the weak lensing shear field γ: galaxy clustering hδgδgi, galaxy-galaxy lensing hδgγti and cosmic shear hγγi. same modeling choices and analysis framework as in [16], making it easy to compare and combine our results later (i.e. Paper III). The structure of the paper is as follows. In Sec. II we briefly review the formalism of our model for the two cross- correlation functions and the parameter inference pipeline (more details can be found in Paper I). In Sec. III we review the data products used in this analysis. In Sec. IV we introduce the estimators we use for the correlation func- tions. In Sec. V we describe out blinding procedure and unblinding criteria. In Sec. VI we present constraints on cosmological parameters as well as relevant nuisance parameters when fitting to the cross-correlation functions. Finally we conclude in Sec. VII. II. MODELING AND INFERENCE We follow the theoretical formalism laid out in Paper I and [29] for this work. Here, we summarize only the main equations relevant to this paper. Following standard con- vention, we refer to the galaxies used to measure δg as lens galaxies, and the galaxies used to measure γ as source galaxies. power spectra: Using Angular Limber approximation3 the cross-spectra between CMB lensing convergence and galaxy density/shear can be related to the matter power spectrum via: [31], the Z CκCMBXiðlÞ ¼ dχ qκCMBðχÞqi XðχÞ χ2 (cid:3) l þ 1=2 χ (cid:4) ; zðχÞ ; PNL ð1Þ where X ∈ fδg; γg, i labels the redshift bin, PNLðk; zÞ is the nonlinear matter power spectrum, which we compute using CAMB and HALOFIT [32,33], and χ is the comoving distance to redshift z. The weighting functions, qXðχÞ, describe how the different probes respond to large-scale structure at different distances, and are given by qκCMBðχÞ ¼ 3Ω mH2 0 2c2 χ aðχÞ χ(cid:2) − χ χ(cid:2) ; qi δg ðχÞ ¼ biðzðχÞÞni δg ðzðχÞÞ dz dχ γðχÞ ¼ qi 3H2 0Ω m 2c2 χ aðχÞ Z χh χ dχ0ni γðzðχ0ÞÞ dz dχ0 χ0 − χ χ0 ; ð2Þ ð3Þ ð4Þ 3In [30], the authors showed that at DES Y3 accuracy, the Limber approximation is sufficient for galaxy-galaxy lensing and cosmic shear but insufficient for galaxy clustering. Given the primary probe in this work, hδgκ CMBi, are at much lower signal-to-noise than galaxy-galaxy lensing and cosmic shear, we expect that Limber approximation is still a valid choice. CMBi þ hγ κ t 023530-5 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) where H0 and Ω m are the Hubble constant and matter density parameters, respectively, aðχÞ is the scale factor corresponding to comoving distance χ, χ(cid:2) denotes the comoving distance to the surface of last scattering, bðzÞ is the galaxy bias as a function of redshift, and ni δg=γðzÞ are the normalized redshift distributions of the lens/source galaxies in bin i. We note that the above equations assumes linear galaxy bias, which is our fiducial model. Modeling the nonlinear galaxy bias involves changes to both Eq. (1) and Eq. (3) (see below). Correlation functions: The angular-space correlation functions are then computed via X wδi gκCMBðθÞ ¼ wγi t κCMBðθÞ ¼ l X l 2l þ 1 4π FðlÞPlðcosðθÞÞCδi gκCMBðlÞ; 2l þ 1 4πlðl þ 1Þ FðlÞP2 lðcosθÞCκi γκCMBðlÞ; ð5Þ ð6Þ where Pl and P2 l are the lth order Legendre polynomial and associated Legendre polynomial, respectively, and FðlÞ describes filtering applied to the κ CMB maps. For correlations with the κ CMB maps, we set FðlÞ ¼ BðlÞHðl − l − lÞ, where HðlÞ is a step function and BðlÞ ¼ expð−0.5lðl þ 1Þσ2Þ with σ ≡ θ FWHM= FWHM describes the beam applied to the CMB lensing maps (see discussion of l max, and θ FWHM choices in Sec. III, and further discussion in Paper I). minÞHðl ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 8 ln 2 , and θ min, l max p Galaxy bias: We consider two models for the galaxy bias. Our fiducial choice is a linear bias model where bðzÞ ¼ bi is not a function of scale and is assumed to be a free parameter for each tomographic bin i. The second bias model is described in [34] and is an effective 1-loop model with renormalized nonlinear galaxy bias parameters: bi 1 (linear bias), bi 2 (local quadratic bias), bi s2 (tidal quadratic bias) and bi 3nl (third-order non-local bias). The latter two parameters can be derived from bi 1, making the total number of free parameters for this bias model two per tomographic bin i. To use this model, we replace the combination of biPNL in Equation (1) with Pgm described in [34]. Intrinsic alignment (IA): Galaxy shapes can be intrinsi- cally aligned as a result of nearby galaxies evolving in a common tidal field. IA modifies the observed lensing signal. We adopt the five-parameter (a1,η1,a2,η2,bta) tidal alignment tidal torquing model (TATT) of [35] to describe galaxy IA. a1 and η1 characterize the amplitude and redshift dependence of the tidal alignment; a2 and η2 characterize the amplitude and redshift dependence of the tidal torquing effect; bta accounts for the fact that our measurement is weighted by the observed galaxy counts. In Sec. VI B, we will also compare our results using the nonlinear alignment model a simpler IA model, [NLA, 36]. The TATT model is equivalent to the NLA model in the limit that a2 ¼ η2 ¼ bta ¼ 0. Impact of lensing magnification on lens galaxy density: Foreground structure modulates the observed galaxy den- sity as a result of gravitational magnification. The effect of magnification can be modeled by modifying Eq. (3) to include the change in selection and geometric dilution quantified by the lensing bias coefficients Ci g: qi δg ðχÞ → qi δg ðχÞð1 þ Ci gκi gÞ; ð7Þ where κi in [29] and the values of Ci to the values listed in Table I. g is the tomographic convergence field, as described g are estimated in [37] and fixed Uncertainty in redshift distributions: We model uncer- tainty in the redshift distributions of the source galaxies with shift parameters, Δzi, defined such that for each redshift bin i, niðzÞ → niðz − Δi zÞ: ð8Þ For the lens sample, we additionally introduce a stretch parameter (σz) when modeling the redshift distribution, as motivated by [38]: niðzÞ → σi zniðσi z½z − hzi(cid:3) þ hzi − Δi zÞ; ð9Þ where hzi is the mean redshift. Uncertainty in shear calibration: We model uncertainty in the shear calibration with multiplicative factors defined such that the observed CκCMBγ is modified by CκCMBγiðlÞ → ð1 þ miÞCκCMBγiðlÞ; ð10Þ where mi is the shear calibration bias for source bin i. they can, however, Lensing ratio (or shear ratio, SR): The DES Y3 3 × 2 pt analysis used a ratio of small-scale galaxy lensing measure- ments to provide additional information, particularly on source galaxy redshift biases and on IA parameters. These ratios are not expected to directly inform the cosmological constraints; improve constraints via degeneracy breaking with nuisance parameters. The lensing ratios can therefore be considered as another form of systematic calibration, in a similar vein to, e.g., spectro- scopic data used to calibrate redshifts, and image simulations used to calibrate shear biases. In [17], it was demonstrated that the lensing ratio measurements are approximately independent of the 3 × 2 pt measurements, making it trivial to combine constraints from 3 × 2 pt and lensing ratios at the likelihood level. Unless otherwise mentioned, all our analy- ses will include the information from these lensing ratios. We investigate their impact in Sec. VI B. Angular scale cuts: The theoretical model described above is uncertain on small scales due to uncertainty in our understanding of baryonic feedback and the galaxy-halo 023530-6 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) TABLE I. Prior values for cosmological and nuisance param- eters included in our model. For the priors, U½a; b(cid:3) indicates a uniform prior between a and b, while N ½a; b(cid:3) indicates a Gaussian prior with mean a and standard deviation b. δðaÞ is a Dirac Delta function at value a, which effectively means that the parameter is fixed at a. Note that the fiducial lens sample is the first 4 bins of the MAGLIM sample. The two high-redshift MAGLIM bins and the REDMAGIC sample are shown in gray to indicate they are not part of the fiducial analysis. Parameter inference: We assume a Gaussian likelihood4 for the data vector of measured correlation functions, ⃗d, given a model, ⃗m, generated using the set of parameters ⃗p: ln Lð⃗dj ⃗mð ⃗pÞÞ ¼ − 1 2 XN ij ðdi − mið ⃗pÞÞTC−1 ij ðdj − mjð ⃗pÞÞ; ð11Þ Parameter Ωm As × 109 ns Ωb h Ωνh2 × 104 a1 a2 η1 η2 bta MAGLIM b1(cid:4)(cid:4)(cid:4)6 b1(cid:4)(cid:4)(cid:4)6 1 b1(cid:4)(cid:4)(cid:4)6 2 C1(cid:4)(cid:4)(cid:4)6 l Δ1…6 z × 102 σ1…6 z REDMAGIC b1(cid:4)(cid:4)(cid:4)5 b1(cid:4)(cid:4)(cid:4)5 1 b1(cid:4)(cid:4)(cid:4)5 2 C1(cid:4)(cid:4)(cid:4)5 l z × 102 Δ1…5 σ1…4 z Prior U½0.1; 0.9(cid:3) U½0.5; 5.0(cid:3) U½0.87; 1.07(cid:3) U½0.03; 0.07(cid:3) U½0.55; 0.91(cid:3) U½6.0; 64.4(cid:3) U½−5.0; 5.0(cid:3) U½−5.0; 5.0(cid:3) U½−5.0; 5.0(cid:3) U½−5.0; 5.0(cid:3) U½0.0; 2.0(cid:3) U½0.8; 3.0(cid:3) U½0.67; 3.0(cid:3) U½−4.2; 4.2(cid:3) δð1.21Þ, δð1.15Þ, δð1.88Þ, δð1.97Þ, δð1.78Þ, δð2.48Þ N ½−0.9; 0.7(cid:3), N ½−3.5; 1.1(cid:3), N ½−0.5; 0.6(cid:3), N ½−0.7; 0.6(cid:3), N ½0.2; 0.7(cid:3), N ½0.2; 0.8(cid:3) N ½0.98; 0.062(cid:3), N ½1.31; 0.093(cid:3), N ½0.87; 0.054(cid:3), N ½0.92; 0.05(cid:3), N ½1.08; 0.067(cid:3), N ½0.845; 0.073(cid:3) U½0.8; 3.0(cid:3) U½0.67; 2.52(cid:3) U½−3.5; 3.5(cid:3) δð1.31Þ, δð−0.52Þ, δð0.34Þ, δð2.25Þ, δð1.97Þ N ½0.6; 0.4(cid:3), N ½0.1; 0.3(cid:3), N ½0.4; 0.3(cid:3), N ½−0.2; 0.5(cid:3), N ½−0.7; 1.0(cid:3) δð1Þ, δð1Þ, δð1Þ, δð1Þ, N ½1.23; 0.054(cid:3) METACALIBRATION m1…4 × 103 Δ1…4 z × 10−2 N ½−6.0; 9.1(cid:3), N ½−20.0; 7.8(cid:3), N ½−24.0; 7.6(cid:3), N ½−37.0; 7.6(cid:3) N ½0.0; 1.8(cid:3), N ½0.0; 1.5(cid:3), N ½0.0; 1.1(cid:3), N ½0.0; 1.7(cid:3) connection (or, nonlinear galaxy bias). We take the approach of only fitting the correlation functions on angular scales we can reliably model. In Paper I we determined the corresponding angular scale cuts by requiring the cosmo- logical constraints to not be significantly biased when prescriptions for unmodeled effects are introduced. In Figs. 2 and 19 the scale cuts are marked by the gray bands. where the sums run over all of the N elements in the data and model vectors. The posterior on the model parameters is then given by: Pð ⃗mð ⃗pÞj⃗dÞ ∝ Lð⃗dj ⃗mð ⃗pÞÞPpriorð ⃗pÞ; ð12Þ where Ppriorð ⃗pÞ is a prior on the model parameters. Our choice of priors is summarized in Table I. The covariance matrix used here consists of an analytical lognormal covariance combined with empirical noise estimation from simulations. The covariance has been extensively validated in Paper I. In Appendix A Fig. 11 we show that the diagonal elements of our final analytic covariance are in excellent agreement with a covariance estimated from data using jackknife resampling. Our modeling and inference framework is built within the COSMOSIS package [40] and is designed to be consistent with those developed as part of [16]. We generate parameter samples using the nested sampler POLYCHORD [41]. III. DATA A. CMB lensing maps There are two major advances in the galaxy and CMB data used here relative to the DES Y1 and SPT analysis presented in [9,10]. First, for the CMB map in the SPT footprint, we used the method developed in [42] and described in Paper I to remove contamination from the tSZ effect by combining data from SPT and Planck. Such contamination was one of the limiting factors in our Y1 analysis. Second, the DES Y3 data cover a significantly larger area on the sky than the DES Y1 data. Consequently, the DES Y3 footprint extends beyond the SPT footprint, necessitating the use of the Planck-only lensing map [14] over part of the DES Y3 patch. As discussed in Paper I, the different noise properties and filtering of the two lensing maps necessitates separate treatment throughout. The “SPT þ Planck” lensing map, which overlaps with the DES footprint at < −40 degrees in declination, is filtered by l max ¼ 5000 and a Gaussian smooth- ing of θ FWHM ¼ 6 arcmin. This map is produced using the combination of 150 GHz data from the 2500 deg2 SPT-SZ survey [e.g., 13], Planck 143 GHz data, and the min ¼ 8, l 4See e.g., [39] for tests of the validity of this assumption in the context of cosmic shear, which would also apply here. 023530-7 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) tSZ-cleaned CMB Planck temperature map generated using the Spectral Matching Independent Component Analysis the SMICA-noSZ map). The (SMICA) algorithm (i.e. “Planck” lensing map, which overlaps with the DES footprint at > −39.5 degrees in declination, is filtered by max ¼ 3800 and a Gaussian smoothing of min ¼ 8, l l θ FWHM ¼ 8 arcmin is applied. This map is reconstructed using the Planck SMICA-noSZ temperature map alone. We leave a small 0.5 deg gap between the two lensing maps to reduced the correlation between structures on the boundaries. The resulting effective overlapping areas with DES are 1764 deg2 and 2156 deg2 respectively for the SPT þ Planck and Planck patches respectively. B. The DES Y3 data products is mostly in the southern hemisphere, DES [43] is a photometric survey in five broadband filters (grizY), with a footprint of nearly 5000 deg2 of sky that imaging hundreds of millions of galaxies. It employs the 570- megapixel Dark Energy Camera [DECam, 1] on the Cerro Tololo Inter-American Observatory (CTIO) 4 m Blanco telescope in Chile. We use data from the first three years (Y3) of DES observations. The foundation of the various DES Y3 data products is the Y3 Gold catalog described in [44], which achieves S=N ∼ 10 for extended objects up to i ∼ 23.0 over an unmasked area of 4143 deg2. In this work we use three galaxy samples: two lens samples for the galaxy density-CMB lensing correlation, hδgκ CMBi, and one source sample for the galaxy shear-CMB lensing correlation, hγ CMBi. We briefly describe each sample below. These samples are the same as those used in [16] and we direct the readers to a more detailed description of the samples therein. κ t 1. Lens samples: MAGLIM and REDMAGIC We will show results from two lens galaxy samples named MAGLIM and REDMAGIC. Following [16], the first four bins of the MAGLIM sample will constitute our fiducial sample, though we show results from the other bins and samples to help understand potential systematic effects in the DES galaxy selection. The MAGLIM sample consists of 10.7 million galaxies selected with a magnitude cut that evolves linearly with the photometric redshift estimate: i < 4zphot þ 18. zphot is deter- mined using the Directional Neighborhood Fitting algorithm [DNF, 45]. [46] optimized the magnitude cut to balance the statistical power of the sample size and the accuracy of the photometric redshifts for cosmological constraints from galaxy clustering and galaxy-galaxy lensing. MAGLIM is divided into six tomographic bins. The top panel of Fig. 1 shows the per-bin redshift distributions, which have been validated using cross-correlations with spectroscopic gal- axies in [38]. Weights are derived to account for survey systematics, as described in [47]. FIG. 1. Redshift distribution for the tomographic bins for the galaxy samples used in this work: the MAGLIM lens sample (top), the REDMAGIC lens sample (middle) and the METACAL source sample (bottom). The fiducial lens sample only uses the first four bins of the MAGLIM sample, or the solid lines. We perform tests with the nonfiducial samples (dashed lines) for diagnostic purposes. The REDMAGIC sample consists of 2.6 million luminous red galaxies (LRGs) with small photometric redshift errors [48]. REDMAGIC is constructed using a red sequence template calibrated via the REDMAPPER algorithm [49,50]. The lens galaxies are divided into five tomographic bins. The redshift distributions are shown in the middle panel of Fig. 1. These distributions are estimated using draws from the redshift probability distribution functions of the individual REDMAGIC galaxies. As with MAGLIM, [38] validates the redshift distributions, and [47] derives sys- tematics weights. We note that in [16] the two high-redshift bins were excluded in MAGLIM due to poor fits in the 3 × 2 pt analysis, while the REDMAGIC sample was excluded due to an internal tension between galaxy-galaxy lensing and galaxy clustering. With the addition of CMB lensing cross- correlations, one of the aims of this work will be to shed light on potential systematic effects in the lens samples. We briefly discuss this issue in Sec. VI D but there will be a more in-depth discussion in Paper III when we combine with the 3 × 2 pt probes. 2. Source sample: METACALIBRATION For the source sample, we use the DES Y3 shear catalog presented in [51], which contains over 100 million galaxies. 023530-8 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) The galaxy shapes are estimated using the METACALIBRATION algorithm [52,53]. The shear catalog has been thoroughly tested in [51,54]. In [54], the authors used realistic image simulations to constrain the multiplicative bias of the shear estimate to be at most 2%–3%, primarily attributed to a shear- dependent detection bias coupled with object blending effects. The residual shear calibration biases are folded into the modeling pipeline and are listed in Table I. The source galaxies are divided into four tomographic bins based on the SOMPZ algorithm described in [55], utilizing deep field data described in [56] and image simulations described in [57]. The bottom panel of Fig. 1 shows the redshift distributions, which have been validated in [58,17]. IV. CORRELATION FUNCTION ESTIMATORS Our estimator for the galaxy-CMB lensing correlation [Eq. (5)] is hδgκ CMBðθαÞi ¼ hδgκ CMBðθαÞi0 − hδRκ CMBðθαÞi; ð13Þ where hδgκ CMBðθαÞi0 ¼ 1 δgκCMB θα N XNg XNpix i¼1 j¼1 ηδg i ηκCMB j κ CMB;jΘαðj ˆθi − ˆθjjÞ ð14Þ and Our estimator for the galaxy shear-CMB lensing corre- lation [Eq. (6)] is PNgal i¼1 hγ κ CMBðθαÞi ¼ t PNpix j¼1 ηe κ P ηκCMB i j sðθαÞ CMB;jeij t ηκCMB ηe i j Θαðj ˆθi − ˆθjjÞ ; ð16Þ where eij is the component of the corrected ellipticity t oriented orthogonally to the line connecting pixel j and the CMB value in the pixel is κj source galaxy. The κ CMB and ηe i and ηκCMB are the weights associated with the source galaxy j and the κ CMB pixel, respectively. The weights for the source galaxies are derived in Gatti et al. [51] and combines the signal-to-noise and size of each galaxy. sðθαÞ is the METACALIBRATION response. We find that sðθÞ is approx- imately constant over the angular scales of interest, but different for each redshift bin. We carry out these mea- surements using the TREECORR package5 [59] in the angular range 2.50 < θ < 250.00. Note that Eq. (16) does not require subtracting a random component as in Eq. (13) since unlike a density field, the mask geometry cannot generate an artificial signal in a shear field. κ CMBi and hγ CMBi correla- tion functions are shown in Fig. 2. The hδgκ CMBi mea- surements using the REDMAGIC sample are shown in Appendix C. The signal-to-noise (S/N) of the different measurements are listed in Table II. Here, signal-to-noise is calculated via The measured MAGLIM hδgκ t v u u t ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ XN C−1 ij dj dT i ij ; ð17Þ hδRκ CMBðθαÞi ¼ 1 NRκCMB θα XNrand XNpix i¼1 j¼1 ηδR i ηκCMB j κ CMB;jΘαðj ˆθi − ˆθjjÞ; S=N ≡ ð15Þ where the sum in i is over all galaxies and the sum in j is δgκCMB over all pixels in the CMB convergence map; N θα ) is the number of galaxy-κ δRκCMB (N CMB pixel (random- θα CMB pixel) pairs that fall within the angular bin θα; ηδg, κ ηδR and ηκCMB are the weights associated with the galaxies, the randoms and the κ CMB pixels. The weights for the galaxies/randoms are derived in Rodríguez-Monroy [47] using a combination of maps of survey properties (e.g., seeing, depth, airmass) to correct for any spurious signals in the large-scale structure, while the κ CMB weights account for differences in the noise levels of pixels in the κ CMB map. The random catalog is used to sample the selection function of the lens galaxies, and has a number density much higher than the galaxies. ˆθi (ˆθj) is the angular position of galaxy i (pixel j), and Θα is an ˆθi − ˆθjj falls in the angular indicator function that is 1 if j bin θα and 0 otherwise. t κ where d is the data vector of interest and C is the covariance matrix. The final signal-to-noise of the fiducial hδgκ CMBi þ hγ CMBi data vector after the linear bias scale cuts is 23.9, about two times larger than in the Y1 study [11]—the main improvement, in addition to the increased sky area, comes from extending our analysis to smaller scales, enabled by the tSZ-cleaned CMB lensing map. The tSZ signal is correlated with large-scale structure, and can propagate into a bias in the estimated κ CMB if not mitigated. In the DES Y1 analysis presented in [11], tSZ cleaning was not imple- mented at the κ CMB map level, necessitating removal of small-scale CMB lensing correlation measurements from the model fits. This problem was particularly severe for hγ CMBi. Comparing results for the SPT þ Planck and Planck patches in Table II, the SPT þ Planck area domi- nates the signal-to-noise before scale cuts in all the probes, even with a smaller sky area. This is due to the lower noise level of the SPT maps. However, since the higher signal-to- κ t 5https://github.com/rmjarvis/TreeCorr. 023530-9 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) FIG. 2. Measurement of the MAGLIM galaxy density-CMB lensing correlation (top) and galaxy shear-CMB lensing correlation (bottom). For each set of measurements, the upper row shows measurement with the SPT þ Planck CMB lensing map and the lower row shows measurement with the Planck CMB lensing map. The shapes and amplitudes are different due to the difference in the L cut and smoothing of the CMB lensing map. The light (dark) shaded regions in the hδgκ CMBi panels indicate the data points removed when assuming linear (nonlinear) galaxy bias, while the shaded regions in the hγ CMBi panels show the data points removed in all cases (only two bins require scale cuts). The dashed dark gray line shows the best-fit fiducial model for the fiducial lens sample, while the χ2 per degree of freedom (ν) evaluated at the best-fit model with scale cuts for linear galaxy bias model is shown in the upper left corner of each panel. κ t t κ CMBi, even though hδgκ noise necessitates a more stringent scale cut, the resulting signal-to-noise after scale cuts is only slightly higher for the SPT þ Planck patch. Finally, comparing hδgκ CMBi and hγ CMBi starts with ∼75% more signal-to-noise before scale cuts compared to hγ CMBi, the scale cuts remove significantly more signal in hδgκ CMBi compared to hγ CMBi. This is due to limits in our ability to model nonlinear galaxy bias on small scales—indeed we see that the signal-to-noise in hδgκ CMBi increases by 13% when switching from linear to nonlinear galaxy bias model. Overall, these signal-to-noise levels are consistent with the forecasts in Paper I. κ κ t t V. BLINDING AND UNBLINDING Following [16], we adopt a strict, multilevel blinding procedure in our analysis designed to minimize the impact of experimenter bias. The first level of blinding occurs at the shear catalog level, where all shears are multiplied by a 023530-10 secret factor [51]. The second level of blinding occurs at the two-point function level, where we follow the procedure outlined in [60] and shift the data vectors by an unknown amount while maintaining the degeneracy between the different parts of the data vector under the same cosmology. The main analyses in this paper were conducted after the unblinding of the shear catalog, so the most relevant blinding step is the data vector blinding. Below we outline the list of tests that were used to determine whether our measurement is sufficiently robust to unblind: (i) Pass all tests described in Appendix B, which indicate no outstanding systematic contamination in the data vectors. These tests include: (1) check for spurious correlation of our signal with survey property maps, (2) check the cross-shear component of hγ CMBi, (3) check the impact of weights used for the lens galaxies, (4) check the effect of the point-source mask in the CMB lensing map on our measurements, and (5) check that cross-correlating an external large-scale κ t JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) TABLE II. Signal-to-noise for the hδgκCMBi þ hγtκCMBi data vector when different scale cuts are applied. Rows involving the two high-redshift MAGLIM bins and the REDMAGIC sample are shown in gray to indicate that they are not part of the fiducial analysis. the different parts of Scale cuts None Linear bias Nonlinear bias SPT þ Planck hδgκCMBi MAGLIM hδgκCMBi MAGLIM 6 bin hδgκCMBi REDMAGIC hγ CMBi MAGLIM κ t 26.8 30.2 23.7 15.0 17.9 20.5 17.0 10.4 CMBi MAGLIM CMBi MAGLIM 6 bin CMBi REDMAGIC Planck hδgκ hδgκ hδgκ hγtκCMBi MAGLIM Combined hδgκCMBi MAGLIM 32.2 hγtκCMBi MAGLIM 18.2 hδgκCMBi þ hγtκCMBi MAGLIM 34.8 14.5 17.4 14.2 13.4 13.1 15.9 12.5 10.4 19.6 16.9 23.9 17.3 20.0 15.7 13.4 13.8 16.8 12.8 10.4 22.2 16.9 25.7 structure tracer (the cosmic infrared background in this case) with different versions of our CMB lensing maps yields consistent results. (ii) With unblinded chains, use the posterior predictive distribution (PPD) method developed in [61] to evaluate the consistency between the two subsets of the data vectors that use different CMB lensing maps (i.e. the SPT þ Planck patch and the Planck patch). The p-value should be larger than 0.01. (iii) With unblinded chains, verify that the goodness-of- fit of the data with respect to the fiducial model has a p-value larger than 0.01 according to the same PPD framework. CMBi þ hγ Except for the first step, all the above are applied to the hδgκ CMBi data vectors with the fiducial analysis choices (ΛCDM cosmology and linear galaxy bias scale cuts), for the first four bins of the MAGLIM lens sample. κ t VI. PARAMETER CONSTRAINTS FROM CROSS-CORRELATIONS OF DES WITH CMB LENSING CMBi þ hγ Following the steps outlined in the previous section, we found (1) no evidence for significant systematic biases in our measurements, as shown in Appendix B, (2) we obtain a p-value greater than 0.01 when comparing the hδgκ CMBi constraints from the Planck region to constraints from the SPT þ Planck region, and (3) the goodness-of-fit test of the fiducial hδgκ CMBi unblinded chain has a p-value greater than 0.01. In the following, we will quote the precise p-values obtained from these tests using the updated covariance matrix. CMBi þ hγ κ κ t t With all the unblinding tests passed, we froze all analysis choices and unblinded our cosmological constraints. We then updated the covariance matrix to match the best-fit parameters from the cosmological analysis.6 The results we present below use the updated covariance matrix. The main constraints on cosmological parameters are summarized in Table III. A. Cosmological constraints from cross-correlations In Fig. 3 we show constraints from hδgκ CMBi using the first 4 bins of the MAGLIM sample. For compari- son, we also show constraints from hγ CMBi-only, cosmic shear (from [62,63]), and 3 × 2 pt (from [16]). κ t We find that our analysis of hδgκ CMBi þ hγ CMBi þ hγ CMBi gives κ κ t t the following constraints: m ¼ 0.272þ0.032 Ω −0.052 ; σ8 ¼ 0.781þ0.073 −0.073 ; S8 ¼ 0.736þ0.032 −0.028 : κ t m constraints. While hδgκ As can be seen from Fig. 3 and expected from Paper I, the constraints are dominated by hγ CMBi slightly improving the Ω CMBi by itself does not tightly constrain cosmology because of the degeneracy with galaxy bias, the shape information in hδgκ CMBi provides additional information on Ωm when combined with hγ CMBi, with hδgκ κ CMBi. t t t κ κ CMBi þ hγ Figure 3 also shows constraints from DES-only probes, including cosmic shear and 3 × 2 pt. We find that the constraints on S8 from hδgκ CMBi are compa- rable to those from cosmic shear and 3 × 2 pt, and in reasonable agreement. The uncertainties of the hδgκ CMBi þ hγ CMBi constraints on S8 are roughly 30% (70%) larger than that of cosmic shear (3 × 2 pt). We will perform a complete assessment of consistency between these probes in Paper III. We can also see that the degeneracy direction of the hδgκ CMBi constraints are slightly differ- ent from 3 × 2 pt, which will help in breaking degeneracies when combined. CMBi þ hγ κ t We consider constraints from the SPT þ Planck and Planck patches separately in Fig. 4. As discussed earlier in Sec. V, the consistency of these two patches was part of the unblinding criteria, thus these two constraints are consistent 6This procedure is the same as in [16]. Since we cannot know the cosmological and nuisance parameters exactly before running the full inference, a set of fiducial parameters were used to generate the first-pass of the covariance that was used for all blinded chains. After unblinding, we update the parameters to values closer to the best-fit parameters from the data. After confirming that the 5 × 2 pt best-fit constraints Paper III are consistent with the 3 × 2 pt best-fit constraints, we chose to use the 3 × 2 pt best-fit parameters for evaluating the covariance matrix, as this makes our modeling choices more consistent with that of [16]. 023530-11 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) TABLE III. ΛCDM constraints on Ω CMBi þ hγ CMBi and different lens samples. We show the constraints using both linear and nonlinear galaxy bias. The last column shows the p-value corresponding to the goodness of fit for the chain. The parts shown in gray indicate that they are not part of the fiducial samples. m, σ8 and S8 using hδgκ κ t Dataset hγtκCMBi MAGLIM hδgκCMBi þ hγtκCMBi MAGLIM 4 bin linear galaxy bias hδgκCMBi þ hγtκCMBi MAGLIM 4 bin nonlinear galaxy bias hδgκCMBi þ hγtκCMBi MAGLIM 6 bin linear galaxy bias hδgκCMBi þ hγtκCMBi MAGLIM 6 bin nonlinear galaxy bias hδgκCMBi þ hγtκCMBi REDMAGIC linear galaxy bias hδgκ CMBi REDMAGIC nonlinear galaxy bias CMBi þ hγ κ t under the PPD metric. We find a p-value of 0.37 (0.33) when comparing the Planck (SPT þ Planck) results to constraints from SPT þ Planck (Planck). We also observe that the constraints are somewhat tighter in the SPT þ Planck patch in S8, consistent with the slightly larger signal-to-noise (see Table II). We note however, that the signal-to-noise before scale cuts of the SPT þ Planck patch is significantly larger than the Planck patch due to the lower noise and smaller beam size of the SPT lensing map (for hδgκ CMBi: 26.8 vs. 17.9; for hγ CMBi: 15.0 vs. 10.4), though most of the signal-to-noise is on the small scales which we had to remove due to uncertainties in the theoretical modeling. This highlights the importance of improving the small-scale modeling in future work. κ t FIG. 3. Constraints on cosmological parameters Ω from hδgκ κ show the corresponding constraints from hγ shear and 3 × 2 pt for comparison. m, σ8, and S8 CMBi using the MAGLIM sample. We also CMBi-only, cosmic t CMBi þ hγ κ t σ8 0.790þ0.080 −0.092 0.781þ0.073 −0.073 0.820þ0.079 −0.067 0.755þ0.071 −0.071 0.769þ0.071 −0.071 0.793þ0.072 −0.083 0.794þ0.069 −0.069 Ωm 0.270þ0.043 −0.061 0.272þ0.032 −0.052 0.245þ0.026 −0.044 0.288þ0.037 −0.053 0.273þ0.034 −0.047 0.266þ0.036 −0.050 0.253þ0.030 −0.046 S8 0.740þ0.034 −0.029 0.736þ0.032 −0.028 0.734þ0.035 −0.028 0.732þ0.032 −0.029 0.727þ0.035 −0.028 0.738þ0.034 −0.030 0.723þ0.033 −0.030 PPD p-value 0.72 0.50 0.51 0.45 0.45 0.39 0.41 B. Lensing ratio and IA modeling As discussed in Sec. II, we have included the lensing ratio likelihood in all our constraints. As was investigated in detail in [17], the inclusion of the lensing ratio informa- tion mainly constrains the IA parameters and source galaxy redshift biases. The TATT IA model adopted here is a general and flexible model that allows for a large range of possible IA contributions. As such, it is expected that including the lensing ratio could have a fairly large impact for data vectors that are not already constraining the IA parameters well. We now examine the effect of the lensing ratio on our fiducial hδgκ CMBi constraints by first removing the lensing ratio prior in our fiducial result, and then doing the same comparison with a different, CMBi þ hγ κ t m, σ8, and S8 FIG. 4. Constraints on cosmological parameters Ω using the hδgκ κ CMBi probes. We also show the con- t straints only using the SPT þ Planck area and only using the Planck area. CMBi þ hγ 023530-12 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) t κ CMBi þ hγ other probes in the plot. We note that this degeneracy is likely sourced by the lensing ratio likelihood, which on its own is degenerate in the η1 − η2 plane. This is consistent with what we have seen in the simulations in Paper I. The fact that it appears more prominent in hδgκ CMBi than in the other probes is partly related to the fact that a1 and a2 are constrained to be further away from zero in the case of hδgκ CMBi, allowing η1 and η2 (the redshift evolution of the terms associated with a1 and a2) to be constrained better. Another relevant factor is that hδgκ CMBi probes slightly larger redshift ranges than cosmic shear and 3 × 2 pt due to the CMB lensing kernal, which allows for a longer redshift lever arm to constrain η1 and η2, resulting in qualitatively different behaviors in the η1 − η2 parameter space. CMBi þ hγ CMBi þ hγ κ κ t t C. Nonlinear galaxy bias As discussed in Sec. II, we test a nonlinear galaxy bias model in addition to our baseline linear galaxy bias analysis. With a nonlinear galaxy bias model we are able to use somewhat smaller scales and utilize more signal in the data (see Table II). In Fig. 7 we show the cosmological constraints of our fiducial hδgκ CMBi data vector with the nonlinear galaxy bias model. We find that the constraints between the two different galaxy bias models are consistent. There is a small improvement in the Ω m direction, which is not surprising given that nonlinear bias impacts hδgκ CMBi improves the Ω CMBi, and hδgκ m con- straints relative to hγ κ CMBi alone. The overall improvement t is nevertheless not very significant, as hγ CMBi is domi- nating the constraints. CMBi þ hγ κ κ t t D. Comparison with alternative lens choices t κ CMBi þ hγ We have defined our fiducial lens sample to be the first four bins of the MAGLIM sample. This choice is informed by the 3 × 2 pt analysis in [16], where alternative lens samples were also tested but were deemed to be potentially contaminated by systematic effects and therefore not used in the final cosmology analysis. Here, we examine the hδgκ CMBi constraints using the two alternative choices for lenses: (1) including the two high-redshift bins in MAGLIM to form a 6-bin MAGLIM sample, and (2) the REDMAGIC lens sample. As we have emphasized through- out the paper, since the galaxy-CMB lensing cross- correlation is in principle less sensitive to some of the systematic effects, these tests could potentially shed light on the issues seen in [16]. We only examine the hδgκ CMBi constraints here, but will carry out a more extensive investigation in combination with the 3 × 2 pt probes in Paper III. CMBi þ hγ κ t In Fig. 8 we show constraints from hδgκ CMBi lens samples: 4-bin MAGLIM CMBi þ hγ κ t using the three different FIG. 5. Constraints on cosmological parameters Ωm, σ8, and S8 using the hδgκCMBi þ hγtκCMBi probes with and without includ- ing the lensing ratio (SR) likelihood, and when assuming the NLA IA model instead of our fiducial IA model TATT. more restrictive IA model, the NLA model (see Sec. II). These results are shown in Fig. 5. We make several observations from Fig. 5. First, the lensing ratio significantly tightens the constraints in the S8 direction (roughly a factor of 2), as expected from Paper I. Second, without the lensing ratio, different IA models result in different S8 constraints, with TATT resulting in ∼40% larger uncertainties than NLA. This is expected given that TATT is a more general model with three more free parameters to marginalize over compared to NLA. That being said, the constraints are still fully consistent when using the different IA models. Third, when lensing ratio is included, there is very little difference in the constraints between the two different IA models. This suggests that the IA constraints coming from the lensing ratio are sufficient to make the final constraints insensitive to the particular IA model of choice. Finally, it is interesting to look at the constraints on the IA parameters for our fiducial hδgκ CMBi analy- sis with and without the lensing ratio. We show this in Fig. 6, and compare them with constraints from cosmic shear [62,63] and 3 × 2 pt [16]. We find two noticeable degeneracies in these parameters: CMBi þ hγ κ t (i) The lensing ratio restricts the a1 − a2 parameter space to a narrow band. This is seen in the cosmic shear and 3 × 2 pt results, as well as the hδgκ CMBi þ hγ CMBi results, although hδgκ CMBi pre- fers somewhat higher a2 values. CMBi þ hγ κ κ (ii) There is a noticeable η1 − η2 degeneracy that shows CMBi þ hγ CMBi and not in the up uniquely in hδgκ κ t t t 023530-13 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) FIG. 6. Constraints on S8 and the IA parameters from our fiducial hδgκ include the hδgκ CMBi þ hγ CMBi constraints without the lensing ratio (SR) likelihood for comparison. CMBi þ hγ t κ CMBi results, cosmic shear and 3 × 2 pt. We also κ t t κ (fiducial), 6-bin MAGLIM and REDMAGIC. The best-fit parameters as well as the goodness-of-fit are listed in Table III. Broadly, all three constraints appear to be very consistent with each other. This is not surprising given that the constraining power is dominated by hγ CMBi as we discussed earlier. In [16] it was shown that for the 3 × 2 pt analysis, both the 6-bin MAGLIM and the REDMAGIC samples give goodness-of-fits that fail our criteria, while for hδgκ CMBi all three samples give acceptable goodness-of-fits values as seen in Table III. This could imply that the systematic effects that contaminated the other correlation functions in 3 × 2 pt are not affecting the hδgκ CMBi results strongly. Compared to the fiducial constraints, the constraining power of the 6-bin MAGLIM sample is slightly higher in the Ω m direction due to the added signal-to-noise from the high-redshift bins, while CMBi þ hγ CMBi þ hγ κ κ t t m and S8. the constraining power of the REDMAGIC sample is slightly lower in both Ω The DES Y3 3 × 2 pt analyses found that the poor fits for the alternative lens samples can be explained by inconsistent galaxy bias between galaxy-galaxy lensing hδgγ ti and galaxy clustering hδgδgi. That is, when allowing the galaxy bias to be different in galaxy-galaxy lensing and galaxy clustering, improves signifi- the goodness-of-fit cantly. Operationally, this is achieved in [16] by adding a free parameter, Xlens, defined such that Xi lens ¼ bi hδgγti=bi hδgδgi; ð18Þ hδgγti (bi where bi hδgγ hδgδgi) is the linear galaxy bias parameter for ti (hδgδgi) in lens galaxy redshift bin i. Xlens is expected 023530-14 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) FIG. 7. Fiducial hδgκ logical parameters Ω galaxy bias models. CMBi þ hγ CMBi constraints on cosmo- m, σ8, and S8 using linear and nonlinear κ t in general the constraints from hδgκ Our CMB lensing cross-correlation analysis provides an interesting way to explore this systematic effect. In essence, with fixed cosmology, we can fit for galaxy bias using hδgκ CMBi and compare with the galaxy bias derived from hδgγti and hδgδgi. Our results are shown in Fig. 9. We find that CMBi on galaxy bias are weaker than both galaxy-galaxy lensing and galaxy clustering, this is expected due to the lower signal-to-noise. As such, the hδgκ CMBi-inferred galaxy bias values are largely consistent with both galaxy-galaxy lensing and galaxy clustering. There are a few bins, though, where hδgκ CMBi does show a preference for the galaxy bias values to agree more with one of the two probes. Noticeably, for the last two MAGLIM bins, hδgκ CMBi prefers a galaxy bias value that is closer to that inferred by galaxy clustering. On the other hand, for the highest two REDMAGIC bins, hδgκ CMBi prefers galaxy bias values that are closer to galaxy-galaxy lensing. These findings are investigations on Xlens consistent with the various described in [34,64] and suggest potential issues in the measurements or modeling of galaxy-galaxy lensing in the two high-redshift MAGLIM bins and galaxy clustering in the REDMAGIC sample.7 However, we caution that these results can be cosmology-dependent, and change slightly if a different cosmology is assumed. t t κ E. Implications for S8 tension In Fig. 10, we compare our constraints on S8 from hγ κ CMBi to those from recent measurements of cosmic shear from galaxy surveys (light blue circles) as well as other recent hγ CMBi constraints (dark blue squares). We κ show only the constraint CMBi (rather than t hδgκ κ to compare only CMBi) since we want measurements of gravitational lensing. These lensing measurements are not sensitive to the details of galaxy bias, unlike hδgκ CMBi. We see that the constraints on S8 obtained from hγ κ CMBi in this work (gray band) are for the t first time comparable to the state-of-the-art cosmic shear measurements. CMBi þ hγ from hγ t FIG. 8. Fiducial constraints on cosmological parameters Ω m, σ8, and S8 using the hδgκ CMBi þ hγ CMBi probes compared with using the REDMAGIC lens sample instead of the MAGLIM lens sample. κ t Figure 10 also shows the inferred value of S8 from the primary CMB (black triangles), as measured by Planck [21], ACT DR4 [65], combining ACT DR4 and the Wilkinson Microwave Anisotropy Probe [WMAP, 65], and SPT-3G [66]. As discussed in several previous works [e.g., 62,63,67] and can be seen in the figure, there is a ∼2.7σ tension8 between the S8 value inferred from cosmic to equal 1 in the case of no significant systematic effects. ≠ 1 for the two high-redshift In [16] it was found that Xlens bins in the MAGLIM sample and for all bins in the REDMAGIC though there was not enough information to sample, determine whether the systematic effect was in hδgγ ti or hδgδgi. 7In particular, [34] tested an alternative REDMAGIC sample and suggested potential remedies to the systematic effect in REDMAGIC that will be explored in future work. 8Here we are quoting the 1D parameter difference in S8, or 8 − S2 , where the superscript 1 and 2 refer ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 8Þ þ σ2ðS2 σ2ðS1 8Þ ðS1 to the two datasets we are comparing. 8Þ= p 023530-15 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) FIG. 9. With fixed cosmological parameters, the inferred galaxy bias from hδgκ the MAGLIM sample (top) and the REDMAGIC sample (bottom). CMBi, galaxy-galaxy lensing and galaxy clustering, for FIG. 10. Comparison of late-time measurements of S8 from lensing-only data (cosmic shear hγγi and galaxy shear-CMB lensing cross- correlation hγ CMBi) to the inferred value of S8 from the primary CMB. κ t shear and the Planck primary CMB constraint—cosmic shear results prefer a lower S8 value. This is intriguing given that it could indicate an inconsistency in the ΛCDM model. We also see that the other CMB datasets are currently much less constraining, but show some variation, with the lowest S8 value from SPT-3G fairly consistent with all the cosmic shear results. With this work, we can now meaningfully add hγ CMBi into this comparison, and as we see in Fig. 10, the hγ CMBi constraints on S8 are also largely below that coming from the primary CMB. This is potentially exciting, since the hγ CMBi measurements come from a cross-correlation between two very different surveys, and are therefore κ κ κ t t t expected to be highly robust to systematic errors. Our results therefore lend support to the existence of the S8 tension. In Paper III we will perform a more rigorous and complete analysis of the consistency of our constraints here with other datasets. VII. SUMMARY We have presented measurements of two cross- correlations between galaxy surveys and CMB lensing: the galaxy position-CMB lensing correlation (hδgκ CMBi), and the galaxy shear-CMB lensing correlation (hγ κ CMBi). t These measurements are sensitive to the statistics of 023530-16 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) large-scale structure, and are additionally expected to be very robust to many observational systematics. Our measurements make use of the latest data from the first three years of observations of DES, and a new CMB lensing map constructed explicitly for cross-correlations using SPT and Planck data. In particular, our fiducial results are from four tomographic bins of the MAGLIM lens galaxy sample. The signal-to-noise of the full data vector without angular scale cuts is ∼30; the part of the data vector used for cosmological inference has a signal-to- noise of ∼20. The main reduction of the signal-to-noise comes from uncertainty in the modeling of nonlinear galaxy bias, which necessitates removal of the small-angle hδgκ CMBi correlation measurements. Compared to the DES Y1 analysis, the signal-to-noise increased by a factor of ∼2 and we are no longer limited by contamination of tSZ in the CMB lensing map. −0.044 ; S8 ¼ 0.734þ0.035 The joint analysis of these two cross-correlations results −0.052 ; S8 ¼ 0.736þ0.032 m ¼ 0.272þ0.032 in the constraints Ω −0.028 m ¼ 0.245þ0.026 (Ω −0.028 ) when assuming lin- ear (nonlinear) galaxy bias in our modeling. For S8, these constraints are more than a factor of 2 tighter than our DES Y1 results, ∼30% looser than constraints from DES Y3 cosmic shear and ∼70% looser than constraints from DES Y3 3 × 2 pt. We highlight here several interesting findings from this work: t t t κ κ κ (i) We find that hγ κ CMBi þ hγ CMBi þ hγ CMBi dominates the constraints in the hδgκ CMBi combination, confirming t our findings from the simulated analysis in Paper I. (ii) We find that the lensing ratio has a large impact on the hδgκ CMBi constraints, improving the S8 constraints by ∼40%. In addition, the hδgκ CMBi þ hγ CMBi data vector constrains the η1 − η2 degen- eracy direction, something not seen in the DES Y3 3 × 2 pt data vectors. (iii) We investigate the use of two alternative lens samples for the analysis: the 6-bin MAGLIM sample and the REDMAGIC sample. In contrast to the fiducial DES Y3 3 × 2 pt analysis, we find that the hδgκ CMBi analysis using the two alter- native lens samples pass our unblinding criteria and show no signs of systematic contamination. (iv) With fixed cosmology, we use the hδgκ CMBi þ hγ κ t t κ CMBi þ hγ CMBi data vector to constrain the galaxy bias values using the 6-bin MAGLIM sample and the REDMAGIC sample. For the two high-redshift MAGLIM bins, we find bias values that agree more with galaxy clustering. On the other hand, for the REDMAGIC sample, we find bias values more con- sistent with galaxy-galaxy lensing. These provide additional information for understanding the sys- tematic effect seen in [16] from these two alternative lens samples. κ (v) Comparing with previous cosmic shear and hγ CMBi constraints, we find that in line with previous findings, our hγ CMBi constraint on S8 is lower than the primary CMB constraint from Planck. In addition, for the first time, hγ CMBi has achieved comparable precision to state-of-the-art cosmic shear constraints. κ κ t t t The constraints derived in this paper from hδgκ CMBi þ hγ κ CMBi can now be compared and combined with the t DES Y3 3 × 2 pt probes [16], which we will do in Paper III. We will present therein our final combined results along with tests for consistency with external datasets. It is however intriguing that with the galaxy-CMB lensing cross-correlation probes alone, our datasets provide very competitive constraints on the late-time large-scale struc- ture compared to galaxy-only probes. Due to the relative insensitivity to certain systematic effects, this additional constraint is especially important for cross-checking and significantly improving the robustness of the galaxy-only results. Another unique aspect of this work compared to other cross-correlation analyses is that we have carried out our work in an analysis framework that is fully coherent with the galaxy-only probes, making it easy to compare and combine. Looking forward to the final datasets from DES, SPT, and ACT, as well as datasets from the Vera C. Rubin Observatory’s Legacy Survey of Space and Time9 (LSST), the ESA’s Euclid mission,10 the Roman Space Telescope,11 (SO), and CMB Stage-413 the Simons Observatory12 (CMB-S4), our results show that there are significant opportunities for combining the galaxy and CMB lensing datasets to both improve the constraints on cosmological parameters and to make the constraints themselves more robust to systematic effects. ACKNOWLEDGMENTS C. C. and Y. O. are supported by DOE grant No. DE- SC0021949. The South Pole Telescope program is sup- ported by the National Science Foundation (NSF) through the grant No. OPP-1852617. Partial support is also pro- vided by the Kavli Institute of Cosmological Physics at the University of Chicago. Argonne National Laboratory’s work was supported by the U.S. Department of Energy, Office of Science, Office of High Energy Physics, under contract No. DE-AC02- 06CH11357. Work at Fermi National Accelerator Laboratory, a DOE-OS, HEP User 9https://www.lsst.org. 10https://www.euclid-ec.org. 11https://roman.gsfc.nasa.gov. 12https://simonsobservatory.org/. 13https://cmb-s4.org/. 023530-17 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) support the National Center the University of Facility managed by the Fermi Research Alliance, LLC, was supported under Contract No. DE-AC02- 07CH11359. The Melbourne authors acknowledge support from the Australian Research Council’s Discovery Projects scheme (No. DP200101068). The McGill authors acknowledge funding from the Natural Sciences and Engineering Research Council of Canada, Canadian Institute for Advanced research, and the Fonds de recherche du Qu´ebec Nature et technologies. The CU Boulder group acknowledges support from NSF Grant No. AST-0956135. The Munich group acknowledges the support by the ORIGINS Cluster (funded by the Deutsche Forschungsge- meinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy—EXC-2094–390783311), the MaxPlanck-Gesellschaft Faculty Fellowship Program, and the Ludwig-Maximilians-Universität München. J. V. acknowledges from the Sloan Foundation. Funding for the DES Projects has been provided by the U.S. Department of Energy, the U.S. National Science the Ministry of Science and Education of Foundation, Spain, the Science and Technology Facilities Council of the United Kingdom, the Higher Education Funding Council for England, for Supercomputing Illinois at Urbana- Applications at Champaign, the Kavli Institute of Cosmological Physics at the University of Chicago, the Center for Cosmology and Astro-Particle Physics at the Ohio State University, the Mitchell Institute for Fundamental Physics and Astronomy at Texas A&M University, Financiadora de Estudos e Projetos, Fundação Carlos Chagas Filho de Amparo `a Pesquisa do Estado do Rio de Janeiro, Conselho Nacional de Desenvolvimento Científico e Tecnológico and the Minist´erio da Ciência, Tecnologia the Deutsche Forschungsgemeinschaft and the Collaborating Institutions in the Dark Energy Survey. The Collaborating Institutions are Argonne National Laboratory, the University of California at Santa Cruz, the University of Cambridge, Centro de Investigaciones Energ´eticas, Medioambientales y Tecnológicas-Madrid, the University of Chicago, University College London, the DES-Brazil Consortium, the University of Edinburgh, the Eidgenössische Technische Hochschule (ETH) Zürich, Fermi National Accelerator Laboratory, the University of Illinois at Urbana-Champaign, the Institut de Ci`encies de l’Espai (IEEC/CSIC), the Institut de Física d’Altes Energies, Lawrence Berkeley National Laboratory, the Ludwig-Maximilians Universität München and the associated Excellence Cluster Universe, the University of Michigan, NFS’s NOIRLab, the University of Nottingham, The Ohio State University, the University of Pennsylvania, the University of Portsmouth, SLAC National Accelerator Laboratory, Stanford University, the University of Sussex, Inovação, e system is Texas A&M University, and the OzDES Membership Consortium. Based in part on observations at Cerro Tololo Inter-American Observatory at NSF’s NOIRLab (NOIRLab Prop. ID 2012B-0001; PI: J. Frieman), which is managed by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation. The DES data management supported by the National Science Foundation under Grants No. AST-1138766 and No. AST-1536171. The DES participants from Spanish institutions are partially supported by MICINN under grants No. ESP2017-89838, No. PGC2018-094773, No. PGC2018-102021, No. SEV-2016-0588, No. SEV- 2016-0597, and No. MDM-2015-0509, some of which include ERDF funds from the European Union. I. F. A. E. is partially funded by the CERCA program of the Generalitat de Catalunya. Research leading to these results has received funding from the European Research Council under the European Union’s Seventh Framework Program (No. FP7/2007-2013) including ERC grant agreements No. 240672, No. 291329, and No. 306478. We acknowl- edge support from the Brazilian Instituto Nacional de Ciência e Tecnologia (INCT) do e-Universo (CNPq grant No. 465376/2014-2). This manuscript has been authored by Fermi Research Alliance, LLC under Contract No. DE- AC02-07CH11359 with the U.S. Department of Energy, Office of Science, Office of High Energy Physics. APPENDIX A: JACKKNIFE COVARIANCE MATRIX We have performed extensive validation tests on our methodology of modeling in the covariance matrix in Paper I. The ultimate check, however, is to compare the covariance matrix with a data-driven jackknife covariance matrix. The jackknife covariance incorporates naturally the noise in the data as well as any non-cosmological spatial variation in the data that might be important. This comparison was done after unblinding and the update of the covariance described in footnote 6, and is only used as a confirmation—that is, we cannot change any analysis choices based on this check. into 80 patches) In Fig. 11 we show the diagonal elements of the covariance matrix (calculated using the jackknife delete-one block jackknife method by dividing the foot- print lens sample, compared with our fiducial covariance matrix. We find excellent agreement between them on all scales, both hδgκ κ CMBi and hγ CMBi, and on both the SPT þ Planck t and Planck patch. the fiducial for 023530-18 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) FIG. 11. Comparison between the diagonal elements of the jackknife covariance and our fiducial covariance matrix (analytical covariance with noise-noise correction applied). APPENDIX B: DIAGNOSTIC TESTS We perform a number of diagnostic tests to make sure that our measurements are not significantly contaminated by potential systematic effects. As we have discussed in Sec. I, cross-survey correlations like those presented here to possible are expected to be inherently more robust systematic effects. In addition, extensive tests have been done on both the galaxy and CMB data products in [13,47,51,62,63,68]. We perform a series of diagnostic tests specific to the cross-correlation probes. 1. Cross-correlation with survey property maps If a given contaminant associated with some survey property simultaneously affects the galaxy and the CMB fields that we are cross-correlating, the cross-correlation signal will contain a spurious component is not cosmological. A possible example is dust, which could simultaneously contaminate the CMB lensing map (through thermal emission in CMB bands) and the galaxy density field (through extinction). In addition to dust, we consider several other possible survey properties. This test is designed to detect such effects. We calculate the correlation statistic, Xf S, between the observables of interest and various survey property maps: that Xf SðθÞ ¼ hκ CMB SðθÞihfSðθÞi hSSðθÞi ; ðB1Þ CMBi and Xf where S is the survey property map of interest, and f is either δg or γ t. This expression captures correlation of the systematic with both κ CMB and f, and is normalized to have the same units as hfκ the CMBi. Henceforth, we omit θ-dependence in the notation for simplicity, but note that all the factors in Eq. (B1) are functions of θ. Unless the systematic map is correlated with both f and κ CMB, it will not bias hfκ S will be consistent with zero. Note that Xf S should also be compared with the statistical uncertainty of hfκ CMBi, as a certain systematic could be significantly detected but have little impact on the final results if it is much smaller than the statistical uncertainty. CMBi, we consider two S fields: stellar density and extinction. For hγ CMBi, we look in addition at two fields associated with PSF modeling errors. The quantities q and w measure the point-spread function (PSF) modeling residuals as introduced in [51], q ¼ e(cid:2) is the difference of the true ellipticity of the PSF as measured by stars and that inferred by the PSF model, and − TmodelÞ=T(cid:2), where T is a measure of size w ¼ e(cid:2)ðT(cid:2) of the PSF, is the impact on the PSF model ellipticity when For hδgκ − emodel κ t 023530-19 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) FIG. 12. The measured systematic contamination of hδgκCMBi for the MAGLIM lens sample, as assessed by Eq. (B1), for the SPT þ Planck field (top) and the Planck field (bottom) and for different redshift bins. For reference, the gray band shows 10% of the statistical uncertainties for the corresponding data vectors. In all cases, the measured bias is significantly below the statistical uncertainties on the hδgκCMBi measurements. FIG. 13. Same as Fig. 12 but for the REDMAGIC lens sample. − Tmodel. As both q and w are the PSF size is wrong by T(cid:2) spin-2 quantities like the ellipticity, we first decompose them into E and B modes using the same method used for generating weak lensing convergence maps in [69]. We then use the E-mode maps as the S maps to perform the cross-correlation test. The rationale here is that if there is a nontrivial E-mode component, it could signify con- tamination in the shear signal and will correlate with the shear field. Figures 12–14 show the result of our measured Xf S for the different parts of the data vector. For comparison, we also plot the statistical uncertainty on the data vector; given that the statistical uncertainties are much larger than the measured biases in all cases, we scale the statistical uncertainties by 0.1 (hδgκ CMBi) and 0.01 CMBi). The χ2 values per degree of freedom for (hγ κ t the Xf S measurements with respect to the null model are shown in Tables IV–VI together with the probability-to- exceed (PTE) values. The χ2 as well as the error bars on the plots are derived from jackknife resampling where we use 65 equal-area jackknife patches for the SPT þ Planck footprint and 85 patches for the Planck area. To obtain a more reliable jackknife covariance, we measure Xf S using 10 angular bins instead of the 20 bins used for the data vectors. In general, most of the systematic effects are very consistent with zero. For hδgκ CMBi, we find that the absolute level of the potential systematic effects as quantified by Xf S is 1-2 orders of magnitudes smaller than the statistical errors. There appears to be more cross-correlation for the SPT þ Planck area, especially with extinction. All of the PTE values of these cross-correlations are above our threshold 023530-20 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) FIG. 14. The measured systematic contamination of hγtκCMBi, as assessed by Eq. (B1), for the SPT þ Planck field (top) and the Planck field (bottom) and for different redshift bins. The gray band shows 1% of the statistical uncertainties for the corresponding data vectors. t κ for concern of 0.01, so we deem these results acceptable. CMBi, we find that the absolute levels of the Xf For hγ S measurements is much lower (> 2 orders of magnitude)— this is expected as it is much less obvious how the survey property maps will leave an imprint on the shear field. Interestingly, we also find that overall the error bars are larger in the SPT þ Planck patch compared to the Planck patch. This can be due to the survey property maps containing higher fluctuation in the SPT þ Planck area as part of the footprint is close to the galactic plane or the Large Magellanic Cloud (LMC). spatial 2. Cross-shear component t × κ κ κ During the measurement of hγ CMBi, we additionally measure its cross-shear counterpart hγ CMBi. We replace et in Eq. (16) with e×, the corrected ellipticity oriented 45° to the line connecting map pixel and the source galaxy. The correlation hγ CMBi should be consistent with zero. Any × significant detection of hγ CMBi could signal systematic effects in the hγ κ × CMBi measurements. Our results are shown in Fig. 15 with the χ2 per degree of freedom and PTE values listed in Table VI. We find no significant detection of hγ κ CMBi in all parts of the data vector. κ × t TABLE IV. The χ2 per degree of freedom for the systematics diagnostics quantity [Eq. (B1)] for the MAGLIM hδgκ CMBi mea- surements. The different columns represent the different survey properties S, whereas the different rows are for the tomographic bins in both the SPT þ Planck patch and the Planck patch. The corresponding PTE values are listed in the parentheses. 3. hδgκCMBi measurements with and without weights As discussed in [47], weights are applied to the lens galaxies in order to remove correlations with various survey properties. When performing the hδgκ CMBi measurement in SPT þ Planck Planck S Bin 1 2 3 4 5 6 1 2 3 4 5 6 Stellar density Extinction TABLE V. Same as Table IV but for the REDMAGIC lens sample. χ2=d:o:f: (PTE) 0.42 (0.85) 0.10 (0.99) 0.21 (0.98) 0.13 (0.99) 0.22 (0.98) 0.36 (0.93) 0.02 (0.99) 0.12 (0.99) 0.15 (0.99) 0.08 (0.99) 0.06 (0.99) 0.05 (0.99) 0.90 (0.49) 0.65 (0.71) 0.64 (0.72) 1.12 (0.34) 1.34 (0.21) 1.66 (0.10) 0.40 (0.87) 0.26 (0.96) 0.28 (0.96) 0.33 (0.93) 0.21 (0.98) 0.18 (0.98) SPT þ Planck Planck S Bin 1 2 3 4 5 1 2 3 4 5 Stellar density Extinction χ2=d:o:f: (PTE) 0.09 (0.99) 0.50 (0.83) 0.42 (0.88) 0.28 (0.96) 0.73 (0.64) 0.09 (0.99) 0.09 (0.99) 0.05 (0.99) 0.04 (0.99) 0.04 (0.99) 0.20 (0.97) 0.56 (0.78) 0.38 (0.91) 0.76 (0.62) 1.13 (0.33) 0.38 (0.89) 0.16 (0.99) 0.19 (0.98) 0.16 (0.99) 0.16 (0.99) 023530-21 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) TABLE VI. The χ2 per degree of freedom for the systematics diagnostics quantity [Eq. (B1)] for the hγ CMBi measurements. The different columns represent the different survey properties S, whereas the different rows are for the tomographic bins in both the SPT þ Planck patch and the Planck patch. The corresponding PTE values are listed in the parentheses. The last column lists the corresponding numbers for the cross-shear measurement described in Sec. B 2. κ t SPT þ Planck Planck S Bin 1 2 3 4 1 2 3 4 Stellar density Extinction PSF model error q χ2=d:o:f: (PTE) PSF model error w γ× 0.12 (0.99) 0.17 (0.99) 0.32 (0.94) 0.20 (0.97) 0.09 (0.99) 0.09 (0.99) 0.12 (0.99) 0.16 (0.99) 0.12 (0.99) 0.38 (0.95) 0.39 (0.90) 0.41 (0.86) 0.06 (0.99) 0.04 (0.99) 0.07 (0.99) 0.18 (0.99) 0.34 (0.96) 0.20 (0.99) 0.40 (0.89) 0.19 (0.97) 0.11 (0.99) 0.25 (0.98) 0.19 (0.99) 0.27 (0.98) 0.15 (0.99) 0.18 (0.99) 0.30 (0.95) 0.15 (0.98) 0.08 (0.99) 0.17 (0.99) 0.14 (0.99) 0.18 (0.99) 1.11 (0.34) 1.18 (0.29) 0.60 (0.75) 1.91 (0.07) 1.15 (0.31) 1.28 (0.23) 1.16 (0.31) 1.12 (0.33) the effect of Eq. (14), these weights are applied (i.e. the ηδg). In a cross- these weights will be non- correlation, negligible if the systematic effect that is being corrected by the weights also correlates with the CMB lensing map. We note that this test is not always a null-test, as we consider it more correct to use the weights. Rather, it shows qualitatively the level of the correction from these weights—naively, the smaller the correction to start with, the less likely the residual contamination will be. In Fig. 16 we show the difference between the hδgκ CMBi measurements with and without using the lens weights, for the two lens samples. To understand the significance of these results, we calculate the Δχ2 between the data vectors with and without weights for the fiducial MAGLIM sample, using the analytic covariance for the data vector and find a Δχ2 of 1.23 after scale cuts. Propagating this into cosmo- logical constraints by running two chains using hδgκ CMBi with and without weights (fixing galaxy bias) gives a negligible 0.02σ shift in the Ω − S8 plane. It is also worth m pointing out that we see that the weights most significantly affect the two high-redshift bins in the MAGLIM sample, this is likely due to the fact that the high-redshift bins are fainter and more affected by the spatially varying observing conditions. 4. Biases from source masking In constructing the CMB lensing maps for this analysis, we apply a special procedure at the locations of bright point sources to reduce their impact on the output lensing maps. As described in more detail in Paper I, the CMB lensing estimator that we use involves two CMB maps, or “legs.” One of these is high-resolution map (i.e. the SPT þ Planck temperature map), and the other is a low-resolution tSZ-cleaned map (i.e. the Planck SMICAnosz temperature map). To reduce the impact of point sources, we inpaint the point sources with fluxes 6.4 < F < 200 mJy using the method described in [70]. The total inpainted area is roughly 3.6% of the map. The corresponding location in the tSZ-cleaned map are left untouched. We expect this procedure to result in a reasonable estimate of κ CMB at the FIG. 15. Cross-correlation between than cross-component of shear with CMB lensing for the SPT þ Planck field (top) and the Planck field (bottom) and for different redshift bins. The gray band shows the statistical uncertainties for hγ κ CMBi. t 023530-22 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) FIG. 18. Cross-correlation between CIB and the Planck lensing map in the North patch (solid gray), the Planck lensing map in the South patch (open red), and the SPT þ Planck lensing map (black). FIG. 16. The difference in the hδgκCMBi cross-correlation between the two lens galaxy samples MAGLIM and REDMAGIC with CMB lensing when using weights and without weights, over the statistical uncertainty of the measurement σ. locations of the point sources, given that only one leg is inpainted, and the area being inpainted is small (such that Gaussian constrained inpainting predicts the pixels values of the inpainted region well) although it is possible that the noise properties of these regions differ somewhat from the map as a whole. To test whether the inpainting procedure results in any bias, we also measure the cross-correlation with the lensing map after masking (i.e. completely removing) all the point sources down to 6.4 mJy. We show in Fig. 17 the difference in the data vectors using the alternative mask and the fiducial one. We find that there is no coherent difference in the correlation measurements across the range of angular scales considered. There is, however, some scatter about our nominal measurements. The level of this scatter is small, roughly 0.25 and 0.50σ across the full range of angular scales for hδgκ CMBi and hγ CMBi respectively.14 Given that such scatter is expected to have negligible impact on our results, and since some scatter between the data points is expected simply due to the different selection of pixels in the masked and unmasked CMB lensing maps, we do not find this to be a cause for worry. Our baseline results will use the unmasked version of the CMB lensing map. κ t 5. Variations in the CMB lensing map Our fiducial analysis uses the SPT þ Planck map in the Dec < −40° region and the Planck lensing map in the region Dec > −39.5°. We left a 0.5° gap between the two maps to avoid correlation between the large-scale structure on the boundary. Here we like to verify that the cross- correlation of our CMB lensing maps with another large- scale structure tracer is consistent between the two patches and the two CMB lensing data sets. We choose to use the cosmic infrared background (CIB) map from [71]15 as this large-scale structure tracer. We carry out the following two 14This scatter results from the slightly higher-noise region caused by the half-leg lensing reconstruction, with the point sources left in the non-inpainted map effectively behaving as noise. 15Here we use the nH ¼ 2.5e20 cm−1 maps as defined in [71]. FIG. 17. Difference in the data vectors using the alternative mask and the fiducial one. This test is only done for the SPT þ Planck patch, as it is specific to the SPT lensing reconstruction. 023530-23 C. CHANG et al. PHYS. REV. D 107, 023530 (2023) FIG. 19. Same as Fig. 2 but for the REDMAGIC sample. tests: (1) we compare the cross-correlations between the CIB map and the Planck lensing map split into two sub- regions (the “North” region with DEC > −39.5° and the “South” region with DEC < −40°), and verify that they are consistent; (2) we compare in the South patch the cross- correlations between the CIB map with either the Planck CMB lensing map or our SPT þ Planck lensing map, and verify that they are consistent. The resulting correlation measurements are shown in Fig. 18—the high signal-to-noise is expected due to the significant overlap in the kernels of the two tracers. We make two comparisons: (1) CIB × Planck North vs. CIB × Planck South: We find a two-sample χ2=ν of 24.28=20 with a PTE of 0.23. This demonstrates that the two patches are consistent with each other. (2) SPT þ Planck vs. Planck South: We compute the two-sample χ2, and find χ2=ν ¼ 23.9=20., with a PTE of 0.25. This demonstrates that the two mea- surements are consistent with each other. We note that there are two caveats associated with these that, at cross-correlation measurements. The first is there may be residuals. Second, 545 GHz, galactic emission is non-negligible, and while the CIB maps from [71] are intended to be free of galactic dust, CMB correlation is most sensitive to redshifts higher than those probed by DES galaxies, thus we are extrapolating the results above to lower redshift. the CIB-κ APPENDIX C: REDMAGIC RESULTS In this appendix we show the results for the second lens sample—the REDMAGIC sample. The data vector is shown in Fig. 19 with signal-to-noise values listed in Table II. We find that (1) no significant systematic effects were found as described in Appendix B, (2) we get a p-value greater than 0.01 when comparing the hδgκ CMBi con- straints from Planck to constraints from SPT þ Planck, and (3) the goodness-of-fit of the fiducial hδgκ CMBi þ hγ CMBi unblinded chain corresponds to a p-value greater than 0.01. These results allowed us to unblind our results, and the final constraints are listed in Table III and the fiducial constraints are shown in Fig. 8. CMBi þ hγ κ κ t t [1] B. Flaugher, H. T. Diehl, K. Honscheid et al. (DES [6] E. Baxter, J. Clampitt, T. Giannantonio et al., Mon. Not. R. Collaboration), Astron. J. 150, 150 (2015). Astron. Soc. 461, 4099 (2016). [2] J. E. Carlstrom, P. A. R. Ade, K. A. Aird et al., Publ. Astron. [7] E. J. Baxter, S. Raghunathan, T. M. Crawford et al., Mon. Soc. Pac. 123, 568 (2011). Not. R. Astron. Soc. 476, 2674 (2018). [3] B. Soergel, S. Flender, K. T. Story et al., Mon. Not. R. [8] J. Prat, E. Baxter, T. Shin et al., Mon. Not. R. Astron. Soc. Astron. Soc. 461, 3172 (2016). 487, 1363 (2019). [4] D. Kirk, Y. Omori, A. Benoit-L´evy et al., Mon. Not. R. [9] Y. Omori, T. Giannantonio, A. Porredon et al., Phys. Rev. D Astron. Soc. 459, 21 (2016). 100, 043501 (2019). [5] T. Giannantonio, P. Fosalba, R. Cawthon et al., Mon. Not. [10] Y. Omori, E. J. Baxter, C. Chang et al., Phys. Rev. D 100, R. Astron. Soc. 456, 3213 (2016). 043517 (2019). 023530-24 JOINT ANALYSIS OF …. II. CROSS-CORRELATION … PHYS. REV. D 107, 023530 (2023) [11] T. M. C. Abbott et al. (DES and SPT Collaborations), Phys. [42] M. S. Madhavacheril and J. C. Hill, Phys. Rev. D 98, Rev. D 100, 023541 (2019). 023534 (2018). [12] M. Costanzi, A. Saro, S. Bocquet et al., Phys. Rev. D 103, 043522 (2021). [43] B. Flaugher, Int. J. Mod. Phys. A 20, 3121 (2005). [44] I. Sevilla-Noarbe, K. Bechtol, M. Carrasco Kind et al., [13] Y. Omori, R. Chown, G. Simard et al., Astrophys. J. 849, Astrophys. J. Suppl. Ser. 254, 24 (2021). 124 (2017). [45] J. De Vicente, E. Sánchez, and I. Sevilla-Noarbe, Mon. Not. [14] N. Aghanim, Y. Akrami et al. (Planck Collaboration), R. Astron. Soc. 459, 3078 (2016). Astron. Astrophys. 641, A8 (2020). [46] A. Porredon et al. (DES Collaboration), Phys. Rev. D 103, [15] Y. Omori et al., preceding paper, Phys. Rev. D 107, 023529 043503 (2021). (2023). [16] T. M. C. Abbott et al. (DES Collaboration), Phys. Rev. D 105, 023520 (2022). [47] M. Rodríguez-Monroy, N. Weaverdyck, J. Elvin-Poole et al., Mon. Not. R. Astron. Soc. 511, 2665 (2022). [48] E. Rozo, E. S. Rykoff, A. Abate et al., Mon. Not. R. Astron. [17] C. Sánchez, J. Prat, G. Zacharegkas et al., Phys. Rev. D 105, Soc. 461, 1431 (2016). 083529 (2022). [49] E. S. Rykoff, E. Rozo, M. T. Busha et al., Astrophys. J. 785, [18] T. M. C. Abbott et al. (DES and SPT Collaborations), following paper, Phys. Rev. D 107, 023531 (2022). [19] G. A. Marques, J. Liu, K. M. Huffenberger, and J. C. Hill, 104 (2014). [50] E. S. Rykoff, E. Rozo, D. Hollowood et al., Astrophys. J. Suppl. Ser. 224, 1 (2016). Astrophys. J. 904, 182 (2020). [51] M. Gatti, E. Sheldon, A. Amon et al., Mon. Not. R. Astron. [20] H. Aihara, N. Arimoto, R. Armstrong et al., Publ. Astron. Soc. 504, 4312 (2021). Soc. Jpn. 70, S4 (2018). [21] P. A. R. Ade, N. Aghanim et al. (Planck Collaboration), [52] E. Huff and R. Mandelbaum, arXiv:1702.02600. [53] E. S. Sheldon and E. M. Huff, Astrophys. J. 841, 24 Astron. Astrophys. 594, A15 (2016). (2017). [22] T. Namikawa, Y. Chinone, H. Miyatake et al., Astrophys. J. [54] N. MacCrann, M. R. Becker, J. McCullough et al., Mon. 882, 62 (2019). Not. R. Astron. Soc. 509, 3371 (2022). [23] Z. D. Kermish, P. Ade, A. Anthony et al., Proc. SPIE Int. [55] J. Myles, A. Alarcon, A. Amon et al., Mon. Not. R. Astron. Soc. Opt. Eng. 8452, 1C (2012). Soc. 505, 4249 (2021). [24] N. C. Robertson, D. Alonso, J. Harnois-D´eraps et al., [56] W. G. Hartley, A. Choi, A. Amon et al., Mon. Not. R. Astron. Astrophys. 649, A146 (2021). Astron. Soc. 509, 3547 (2022). [25] J. T. A. de Jong, G. A. Verdoes Kleijn, K. H. Kuijken et al., [57] S. Everett, B. Yanny, N. Kuropatkin et al., Astrophys. J. Exp. Astron. 35, 25 (2013). Suppl. Ser. 258, 15 (2022). [26] D. S. Swetz, P. A. R. Ade, M. Amiri et al., Astrophys. J. [58] M. Gatti, G. Giannini, G. M. Bernstein et al., Mon. Not. R. Suppl. Ser. 194, 41 (2011). Astron. Soc. 510, 1223 (2022). [27] A. Krolewski, S. Ferraro, and M. White, J. Cosmol. [59] M. Jarvis, G. Bernstein, and B. Jain, Mon. Not. R. Astron. Astropart. Phys. 12 (2021) 028. Soc. 352, 338 (2004). [28] E. F. Schlafly, A. M. Meisner, and G. M. Green, Astrophys. [60] J. Muir, G. M. Bernstein, D. Huterer et al., Mon. Not. R. J. Suppl. Ser. 240, 30 (2019). Astron. Soc. 494, 4454 (2020). [29] E. Krause, X. Fang, S. Pandey et al., arXiv:2105.13548. [30] X. Fang, E. Krause, T. Eifler, and N. MacCrann, J. Cosmol. [61] C. Doux, E. Baxter, P. Lemos et al., Mon. Not. R. Astron. Soc. 503, 2688 (2021). Astropart. Phys. 05 (2020) 010. [62] A. Amon, D. Gruen, M. A. Troxel et al., Phys. Rev. D 105, [31] D. N. Limber, Astrophys. J. 117, 134 (1953). [32] A. Lewis, A. Challinor, and A. Lasenby, Astrophys. J. 538, 023514 (2022). [63] L. F. Secco, S. Samuroff, E. Krause et al., Phys. Rev. D 105, 473 (2000). 023515 (2022). [33] R. Takahashi, M. Sato, T. Nishimichi, A. Taruya, and M. [64] A. Porredon, M. Crocce, J. Elvin-Poole et al., arXiv: Oguri, Astrophys. J. 761, 152 (2012). 2105.13546. [34] S. Pandey, E. Krause, J. DeRose et al., Phys. Rev. D 106, [65] S. Aiola, E. Calabrese, L. Maurin et al., J. Cosmol. 043520 (2022). Astropart. Phys. 12 (2020) 047. [35] J. A. Blazek, N. MacCrann, M. A. Troxel, and X. Fang, [66] D. Dutcher, L. Balkenhol, P. A. R. Ade et al., Phys. Rev. D Phys. Rev. D 100, 103506 (2019). 104, 022003 (2021). [36] S. Bridle and L. King, New J. Phys. 9, 444 (2007). [37] J. Elvin-Poole, N. MacCrann et al., arXiv:2209.09782. [38] R. Cawthon et al. (DES Collaboration), Mon. Not. R. [67] M. Asgari, C. Heymans, H. Hildebrandt et al., Astron. Astrophys. 624, A134 (2019). [68] J. Prat, J. Blazek, C. Sánchez et al., Phys. Rev. D 105, Astron. Soc. 513, 5517 (2022), arXiv:2012.12826. 083528 (2022). [39] C.-H. Lin, J. Harnois-D´eraps, T. Eifler, T. Pospisil, R. Mandelbaum, A. B. Lee, and S. Singh, Mon. Not. R. Astron. Soc. 499, 2977 (2020). [40] J. Zuntz, M. Paterno, E. Jennings et al., Astron. Comput. 12, 45 (2015). [69] N. Jeffrey, M. Gatti, C. Chang et al., Mon. Not. R. Astron. Soc. 505, 4626 (2021). [70] A. Benoit-L´evy, T. D´echelette, K. Benabed, J.-F. Cardoso, D. Hanson, and S. Prunet, Astron. Astrophys. 555, A37 (2013). [41] W. J. Handley, M. P. Hobson, and A. N. Lasenby, Mon. Not. [71] D. Lenz, O. Dor´e, and G. Lagache, Astrophys. J. 883, 75 R. Astron. Soc. 453, 4384 (2015). (2019). 023530-25	null
10.1371_journal.pcbi.1007117.pdf	Data Availability Statement: All relevant data are within the manuscript and its Supporting Information files.	All relevant data are within the manuscript and its Supporting Information files.	RESEARCH ARTICLE Development of a physiologically-based pharmacokinetic pediatric brain model for prediction of cerebrospinal fluid drug concentrations and the influence of meningitis Laurens F. M. VerscheijdenID M. RusselID 1* 1, Jan B. Koenderink1, Saskia N. de WildtID 1,2, Frans G. 1 Department of Pharmacology and Toxicology, Radboud University Medical Center, Radboud Institute for Molecular Life Sciences, Nijmegen, The Netherlands, 2 Intensive Care and Department of Pediatric Surgery, Erasmus MC-Sophia Children’s Hospital, Rotterdam, The Netherlands * [email protected] Abstract Different pediatric physiologically-based pharmacokinetic (PBPK) models have been described incorporating developmental changes that influence plasma drug concentrations. Drug disposition into cerebrospinal fluid (CSF) is also subject to age-related variation and can be further influenced by brain diseases affecting blood-brain barrier integrity, like menin- gitis. Here, we developed a generic pediatric brain PBPK model to predict CSF concentra- tions of drugs that undergo passive transfer, including age-appropriate parameters. The model was validated for the analgesics paracetamol, ibuprofen, flurbiprofen and naproxen, and for a pediatric meningitis population by empirical optimization of the blood-brain barrier penetration of the antibiotic meropenem. Plasma and CSF drug concentrations derived from the literature were used to perform visual predictive checks and to calculate ratios between simulated and observed area under the concentration curves (AUCs) in order to evaluate model performance. Model-simulated concentrations were comparable to observed data over a broad age range (3 months–15 years postnatal age) for all drugs investigated. The ratios between observed and simulated AUCs (AUCo/AUCp) were within 2-fold difference both in plasma (range 0.92–1.09) and in CSF (range 0.64–1.23) indicating acceptable model performance. The model was also able to describe disease-mediated changes in neonates and young children (<3m postnatal age) related to meningitis and sep- sis (range AUCo/AUCp plasma: 1.64–1.66, range AUCo/AUCp CSF: 1.43–1.73). Our model provides a new computational tool to predict CSF drug concentrations in children with and without meningitis and can be used as a template model for other compounds that pas- sively enter the CNS. a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Verscheijden LFM, Koenderink JB, de Wildt SN, Russel FGM (2019) Development of a physiologically-based pharmacokinetic pediatric brain model for prediction of cerebrospinal fluid drug concentrations and the influence of meningitis. PLoS Comput Biol 15(6): e1007117. https://doi.org/10.1371/journal.pcbi.1007117 Editor: James Gallo, University at Buffalo - The State University of New York, UNITED STATES Received: February 1, 2019 Accepted: May 21, 2019 Published: June 13, 2019 Copyright: © 2019 Verscheijden et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the manuscript and its Supporting Information files. Funding: The authors received no specific funding for this work. Competing interests: The authors have declared that no competing interests exist. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 1 / 19 Development of a pediatric brain PBPK model Author summary Developmental processes in children affect pharmacokinetics and should ideally be taken into account when establishing drug dosing regimens. One way to incorporate develop- mental differences is by making use of physiologically-based pharmacokinetic (PBPK) models in which kinetic equations are used to describe drug disposition processes and developmental biology. With these equations the absorption of drugs into the model, the flow of drugs between different compartments (representing major organs/tissues), and excretion from the model are predicted. PBPK models can also be used to describe drug concentrations in different target tissues, which often correlate better with the clinical effects. Here, we developed a generic pediatric PBPK model of drug disposition in the cerebrospinal fluid (CSF), that was able to describe clinically measured drug concentra- tions of several drugs in neonates and children. The model could be useful in predicting CSF concentrations of other drugs in pediatric populations where clinical data is often sparse or absent and by this means guide first-in-child dose recommendations. Introduction Growth and development significantly impact handling of drugs in children across the pediat- ric age range. Simple linear bodyweight-based extrapolations from adult to pediatric doses have resulted in toxicity or therapy failure [1]. Taking developmental changes of the processes involved in drug disposition into account in dosing guidelines will lead to improved therapeu- tic efficacy and safe exposure in children of different ages [2]. Physiologically-based pharmacokinetic (PBPK) modeling is an important tool to simulate drug exposure and design dosing guidelines. PBPK models are compartmental kinetic models in which physiological and drug-specific parameters are as much as possible separated [3]. Physiological parameters describe biological values and processes, and if sufficient data describing developmental biology is available, they can be used to predict plasma drug concen- trations in pediatric populations. By this means, PBPK models can guide first-in-child dosing regimens for drugs of which pediatric clinical drug concentrations are scarcely available, resulting in more focused, data-rich clinical trials. There are multiple examples of the success- ful application of pediatric PBPK models in the drug development process [4]. PBPK models also allow predictions of drug concentrations in target tissues, which often correlate better with the clinical effect. This is especially the case for organs like the brain that are characterized by permeability-limited disposition of various drugs, leading to a significant lag time between their peak plasma and tissue concentrations [5–7]. For drugs acting in the brain, differences in blood and cerebrospinal fluid (CSF) dynamics, blood-brain barrier (BBB) and blood-CSF barrier (BCSFB) permeability, brain and CSF compartment volumes, as well as disease-mediated changes could influence the amount of drug entering the different parts of the central nervous system and should be included to facilitate robust predictions. Recently, brain PBPK models for adult populations were developed to allow predictions of drug concentrations in brain parenchyma and CSF. For example, Gaohua et al. developed an adult brain PBPK model consisting of 4 compartments, which was validated using measured paracetamol CSF concentrations [8]. Yamamoto et al. used a different approach by first build- ing a rat model that incorporated multiple brain and CSF compartments [9, 10]. Physiological parameters were in turn adjusted in order to develop a human version [11]. Finally, they extended this adult model into a morphine pediatric brain PBPK model. This model allowed simulations of morphine extracellular fluid concentrations after traumatic brain injury in six PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 2 / 19 Development of a pediatric brain PBPK model children older than 2 years of age [12]. To become more widely accepted, pediatric brain PBPK models should be validated with observed data of different drugs and more individuals across the pediatric age range to increase confidence in the physiological parameters included. In addition to the validation of a brain PBPK model for a relatively healthy pediatric popu- lation, including pathophysiological changes associated with conditions known to affect brain drug concentrations is important to demonstrate model performance in diseased children. Meningitis is a severe condition that leads to impaired BBB and enhanced penetration of drugs into brain and CSF, particularly in newborns and young children [13, 14]. The aim of our study was to describe drug CSF concentrations and describe the effects of meningitis in children by developing and validating a generic pediatric brain PBPK model based on different drugs that enter the brain by passive transfer. Our model provides a new computational tool to predict CSF concentrations in children of drugs that undergo passive transfer and it could serve as a good template for further extension to carrier-mediated trans- port and disposition in different regions of the central nervous system. Results The full code of the pediatric PBPK model for paracetamol is available in Rstudio format (S1 File). Summary tables of physiological and drug-related parameters are reported in S1 Table and S2 Table, respectively, and can be used to adapt the model to the adult situation and for other drugs. The healthy pediatric brain model was built for paracetamol and subsequently val- idated using the nonsteroidal anti-inflammatory drugs (NSAIDs) ibuprofen, flurbiprofen and naproxen. BBB permeability of the antibiotic meropenem was optimized in an adult popula- tion suffering from meningitis and subsequently simulations were performed for a pediatric population with meningitis and sepsis (methods). Building a healthy adult brain PBPK model using paracetamol First, the adult brain PBPK model was built and validated using published paracetamol data [7, 15]. Model simulations of plasma and CSF concentrations largely overlaid the observed data and the ratios of the respective AUCs were within twofold difference (Fig 1, Table 1). Building a healthy pediatric brain PBPK model using paracetamol After inclusion of the age-related data, simulations of paracetamol concentrations were per- formed for children aged between 3 months and 13 years. The pediatric model simulations also largely overlaid with observed data (Fig 1). Ratios of simulated over observed plasma and CSF AUCs were also within twofold difference (Table 1). An equivalent dose of approximately 15 mg/kg resulted in a CSF AUC0-6h that was 34% higher in children between 3 months and 13 years of age compared to adults. Because CSF production rate is possibly influenced by co- medication, a sensitivity analysis was performed to investigate the effect of CSF production rate on paracetamol concentration-time profiles [16]. Twofold differences in CSF production had a significant impact on the paracetamol CSF profiles (Fig 2A and 2B), while leaving the shape of the plasma concentration-time curves virtually unaffected (difference in AUC, Cmax and Tmax <1% from original concentration-time profile). Validation of the pediatric brain PBPK model using ibuprofen, flurbiprofen and naproxen To validate the physiological parameters included in the paracetamol model for children aged between 3 months and 15 years, simulations were performed with another set of drugs that PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 3 / 19 Development of a pediatric brain PBPK model Fig 1. Simulations of paracetamol concentration-time profiles. Plasma and CSF concentration-time profiles of paracetamol after a single intravenous dose of 1000 mg in adults (A, B) and (15 mg/kg) (C,D) in children. Solid black lines indicate simulation of median profiles, the grey area represents 90% CI and dotted lines indicate the minimum and maximum simulation. Dots indicate measured data derived from clinical studies together with the reported S. E.M (adult) or individual observations (pediatric). Log-transformed results are depicted in the right upper corners. https://doi.org/10.1371/journal.pcbi.1007117.g001 were not used to build the pediatric model, i.e. ibuprofen, flurbiprofen and naproxen. The model predicted clinically observed data reasonably well, except for flurbiprofen for which plasma volume of distribution seemed to be overestimated. For this drug a Kp scalar of 0.33 Table 1. Ratio observed/predicted AUC for plasma and CSF drug concentrations. Drug Paracetamol Paracetamol Ibuprofen (oral) Ibuprofen (IV) Flurbiprofen Naproxen Meropenem Meropenem Meropenem Population Plasma Adult Paediatric Paediatric Paediatric Paediatric Paediatric Adult Pediatric(sepsis) Pediatric(meningitis) 1.01 1.09 0.94 0.92 0.99 1.07 1.07 1.66 1.64 https://doi.org/10.1371/journal.pcbi.1007117.t001 CSF 1.18 1.23 1.05 0.64 0.76 0.99 1.23 1.73 1.41 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 4 / 19 Development of a pediatric brain PBPK model Fig 2. The effect of CSF production rate on paracetamol CSF concentration-time profiles. The red lines indicate the simulations using the default CSF production rate. The black lines indicate a twofold increase or reduction in CSF production rate, respectively. The grey lines indicate a fivefold increase or reduction in CSF production rate, respectively. https://doi.org/10.1371/journal.pcbi.1007117.g002 (i.e. Kp of every compartment multiplied by 0.33) was introduced to adjust volume of distribu- tion such that simulations of plasma concentrations better correlated with observations (Fig 3). The difference between observed and simulated AUCs for both plasma and CSF were within 2-fold (Table 1). Meropenem permeability in the adult brain meningitis PBPK model Drug-related parameters of meropenem were included in the adult model. No individual doses were reported in the study of Lu et al., which was used for validation [17]. Therefore, an intravenous dose of 1500 mg/8h was chosen for our virtual population, which was in the range of the doses used by Lu et al. (1000 mg/8h, 1000 mg/6h, 2000 mg/8h). Simulations of plasma concentration-time profiles matched well with observed data points. After optimizing perme- ability in this model, a BBB permeability surface area product of 0.003 L/h (PS 0.0015 L/h for BCSFB) together with a CV of 150% was found (Fig 4, Table 1). Meropenem permeability in the pediatric brain meningitis PBPK model The BBB/BCSFB permeability values estimated in the adult model were also used in the pediat- ric model for children aged between 1 day and 3 months, after correction for brain weight. This resulted in a <2 fold overlay between pediatric plasma and CSF concentrations both for patients with meningitis and sepsis (Fig 5, Table 1). Discussion The pediatric brain PBPK model described here could predict CSF concentrations of the anal- gesics paracetamol, ibuprofen and naproxen, and the antibiotic meropenem over a wide age range. The CSF AUCs were simulated within 2-fold error of clinically observed values without the need of changing system parameters of the compartments describing the healthy brain. This shows that parameterization of these compartments was sufficiently robust to allow for simulations of other drugs as well, and that the model could be useful in deriving mechanisti- cally-informed dosing regimens for children. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 5 / 19 Development of a pediatric brain PBPK model PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 6 / 19 Development of a pediatric brain PBPK model Fig 3. Simulations of pediatric concentration-time profiles for ibuprofen, flurbiprofen and naproxen. Simulations of oral ibuprofen (10 mg/kg in suspected sepsis patients) (A,B), IV ibuprofen (10 mg/kg in surgery patients) (C,D), IV flurbiprofen (0.9 mg/kg in surgery patients) (E,F), and oral naproxen (10 mg/kg in surgery patients) (G,H) concentration-time profiles. Solid black lines indicate simulation of median profiles, the grey area represents 90% CI and dotted lines indicate the minimum and maximum simulation. Dots indicate measured data derived from clinical studies. Log- transformed concentration-time data are depicted in the right upper corners (0 values were discarded). https://doi.org/10.1371/journal.pcbi.1007117.g003 Pharmacokinetic simulations were validated in relatively healthy children between 3 months and 15 years of age, which resulted in accurate AUC estimates (Table 1). In addition, an attempt was made to perform simulations in children younger than 3 months, including premature neonates, suffering from meningitis/sepsis. Although also in this case the simulated AUCs were within 2-fold of observations, a trend towards an overestimation of plasma and CSF levels could be observed (Table 1). It remains to be elucidated whether this is the result of an influence of age and/or disease, or of the pharmacokinetic simulations in the validation study, which were limited by the availability of only peak and through concentrations. PBPK models are inherently complex, due to the many different drug-specific and (physio- logical) system-specific parameters. Their robustness and reliability remain a challenge and there is a clear need to validate model performance with sound experimental data. In this study, most of the physiological processes could be incorporated in an age-appropriate man- ner, however, due to absence of data, the relative flows between brain compartments expressed as percentage of CSF production rate, were assumed to be the same as in adults. Our pediatric brain model is structurally similar to that described by Gaohua et al. for an adult population [8]. Recently, another pediatric brain PBPK model was developed to simulate concentrations of morphine in brain extracellular fluid of traumatic brain injury patients, but it was only validated with experimental data derived from six children [12]. Although this model provides proof of principle for prediction of brain drug concentrations in children, the pathophysiological changes associated with severe traumatic brain injury may have impacted morphine disposition and hence extrapolation to other patient populations. Other models are based on animal studies and, although they have been validated more extensively, translation to the human situation remains difficult [18, 19]. Depending on the pathophysiological data available, the pediatric model described in this study can be extended with different disease conditions. We simulated drug concentrations in Fig 4. Simulations of adult concentration-time profiles for intravenous meropenem in meningitis patients. Simulations of adult plasma and CSF concentration-time profiles (i.v. 1500 mg/8h) (A,B). Solid black lines indicate simulated median profiles, the grey area represents 90% CI and dotted lines indicate the minimum and maximum simulation. Dots indicate measured data derived from clinical studies. Log-transformed results are depicted in the left upper corners. https://doi.org/10.1371/journal.pcbi.1007117.g004 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 7 / 19 Development of a pediatric brain PBPK model Fig 5. Simulations of pediatric concentration-time profiles for intravenous meropenem. Simulations of plasma and CSF for patients suffering from sepsis (i.v. 20 mg/kg/8h) (A,B), or meningitis (i.v. 40 mg/kg/8h) (C,D). Solid black lines indicate simulated median profiles, the grey area represents 90% CI and dotted lines indicate the minimum and maximum simulation. Dots indicate measured data derived from clinical studies. Log-transformed results are depicted in the left upper corners. https://doi.org/10.1371/journal.pcbi.1007117.g005 a population of children with meningitis and sepsis. Blood-brain barrier permeability has been described to initially increase during meningitis, although at later times it will return to normal values due to antibiotic-mediated recovery of patients [13]. Since quantitative data on mem- brane permeability of meropenem were not available from literature, the brain permeability parameter was obtained for the adult population by empirical optimization. The value we derived this way (0.003 L/h) is in the same order of magnitude as intercompartmental clear- ances that have been described in population PK models for adults and children (0.0017 and 0.0007 L/h, respectively) [17, 20]. This estimated permeability was subsequently applied in simulations with the pediatric model, which is also in accordance with studies in which a cor- relation between TNFα and blood-brain barrier damage was found, but not between TNFα and age [21]. Simulated and measured maximum concentrations and AUCcsf/AUCserum ratios were higher in patients suffering from sepsis/meningitis compared to relatively healthy individuals. The highest measured concentration in the healthy population was 1.6 mg/L and the AUCcsf/ AUCserum ratio was 0.047[22]. In our simulations the upper bound of the 90% confidence interval was around 5mg/L and AUCcsf/AUCserum ratios ranged from 0.09–0.12, indicating that inflammation increased BBB permeability. In the studies of Lu et al. and Germovsek et al. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 8 / 19 Development of a pediatric brain PBPK model that we used for validation, samples were taken at different moments after the start of dosing and patients likely differed in disease severity, which might explain the large variability in BBB and BCSFB permeability observed [17, 20]. Time-dependent effects on permeability could, however, not be estimated from the available data. In septic children (without meningitis) the same estimate on meropenem blood-brain barrier permeability resulted in an acceptable over- lay between simulations and observed values, which can possibly be explained by a sepsis- induced increase in BBB permeability [23]. In the current model, drugs were included for which carrier-mediated transport does not play a major role in BBB and BCSFB. A next step will be to incorporate membrane transporters in the model for relevant drug substrates. Data on quantitative proteomics of transporter abundance in adults could form the bases for in vitro-in vivo extrapolation (IVIVE) of trans- porter-mediated transport in an adult brain PBPK model [24, 25]. However, for pediatric pop- ulations absolute protein expression of BBB/BCSFB transporters has not yet been quantified. Immunohistochemistry studies indicate that expression may not be fully matured in young children as has been described for the ABC transporter P-glycoprotein [26, 27]. Only for very few other transporters information on human ontogeny is available [27]. A limitation of the current study is that data used for validation were obtained in children over a broad age range, which could not be further stratified. Children were suffering from a clinical condition and/or receiving co-medication that could have influenced the pharmacoki- netic profiles. Next, CSF drug concentrations were used to validate the simulations and although this is a relevant compartment for antibitiotics during meningitis, the parenchymal extracellular or intracellular fluid is probably more important for other drugs, like analgesics. Future research should be aimed at refining the model by dividing the brain into an intracellu- lar and extracellular space and by expanding the CSF compartments, to better describe the continuum between cranial and spinal CSF. This requires more clinical data on drug disposi- tion in brain tissue and age-appropriate physiological parameters. It has become more widely accepted that drug research should have an increased focus on pediatric populations to improve safety and reduce off-label dosing [28]. Simple body weight- based scaling ignores developmental processes as illustrated by the simulations in this study with normalized paracetamol doses that did not result in equivalent CSF concentration-time profiles. Modeling and simulation have been recognized as a way to make optimal use of exist- ing data, which could result in more focused clinical trials in children. Moreover, in PBPK modeling, parameters can be partly or completely estimated from in vitro or in silico studies, further reducing the need for in vivo studies [29]. The ultimate goal of pediatric PBPK model- ing would be to build models without making use of clinical data and although this might be difficult due to the uncertainty in underlying physiological processes, it could provide guid- ance for dose selection in first-in-child drug studies. Thereafter, clinical data can be used to improve model performance in a “learn and confirm” cycle [30]. Several published models have shown that it is possible to use a bottom-up in vitro-in vivo extrapolation (IVIVE) approach for the estimation of rates of plasma absorption and elimination, which was not yet incorporated in the current model. Linking the brain pediatric PBPK model developed in this study to the existing IVIVE-based models could eventually provide the possibility to mechanis- tically predict brain concentrations, which would facilitate dosing based on higher quality data as compared to simply scaling from adult dosing regimens. In conclusion, a mechanistic pediatric PBPK model was developed incorporating 4 differ- ent brain compartments, which was used to simulate the plasma and CSF pharmacokinetics of different drugs. The model can be valuable to predict CSF concentrations in cases where clini- cal data in this compartment is restricted. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 9 / 19 Development of a pediatric brain PBPK model Materials and methods Approach A five-step approach was used to build and validate the PBPK brain model for a pediatric pop- ulation (Fig 6). These steps are briefly summarized below. Step 1. Building of an adult PBPK brain model (paracetamol) The adult PBPK brain model developed by Gaohua et al. was used as a template and para- cetamol as model compound because it is not a substrate for drug transporter proteins expressed in the blood-brain barrier (BBB) or blood CSF barrier (BCSFB) [8]. Step 2. Building of a pediatric PBPK brain model (paracetamol) Physiological parameters in the adult model were changed to age-appropriate pediatric parameters. Paracetamol was used to allow simulations in relatively healthy children. Step 3. Validation of the pediatric PBPK brain model (ibuprofen, flurbiprofen, naproxen) Physiological parameters included in the model were validated further by simulation of plasma and CSF levels of ibuprofen, flurbiprofen and naproxen in relatively healthy children. These drugs were chosen because they are no known substrates for BBB drug transporters, and pediatric CSF concentrations were available in the literature. Step 4. Building of an adult meningitis PBPK brain model (meropenem) To the best of our knowledge, no mechanistic data is currently available quantifying the influence of meningitis on brain drug permeability. The impact of meningitis on the passage of meropenem across the BBB and BCSFB was therefore estimated in the adult brain PBPK model (described below step 1) using empirical optimization. Step 5. Building of a pediatric meningitis PBPK brain model (meropenem) Meropenem BBB passage estimated in the adult meningitis PBPK model was incorporated in a pediatric model to allow simulations for children with meningitis (combining steps 2, 3 and 4). Because children suffering from sepsis (without meningitis) were also included in the clinical study used for validation, simulations were also done for this population. Fig 6. Workflow used for the building and validation of the pediatric brain PBPK model. https://doi.org/10.1371/journal.pcbi.1007117.g006 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 10 / 19 Development of a pediatric brain PBPK model At each step, the model was validated using published plasma and CSF concentration data, as described below. Step 1 Adult PBPK brain model Plasma model. A PBPK model was coded in R software Version 1.1.442 and consisted of 14 compartments representing major organs and tissues (Fig 7). Average physiological param- eter values and inter-individual variability were derived from literature [31], or values and equations reported in the Simcyp simulator (Version 17 Release 1) [32–35]. As these are based on weight, height, body surface area, sex and/or age-related equations, correlation between parameters was partly accounted for, because variability in the original parameters is propa- gated to the estimated/predicted parameter (e.g. a high body weight will on average result in higher organ volumes). Residual parameter variability was assumed to be log-normally distrib- uted. Organ partitioning coefficients were based on previous publications taking into account both logP and ionization of compounds [36–38]. Plasma elimination from the model was included using in vivo-measured clearance values reported in literature. Clearance were not extrapolated from in vitro experiments, because in this way less robust plasma concentration- time profiles would be generated, which would impede proper assessment of predicted CSF concentrations. Because total body clearance in the adult model was not attributed to specific organs or patient characteristics, only uncorrelated variability in clearance was included. This was described in the adult model as: Pi ¼ Ppop � eZZ ð1Þ Fig 7. Schematic outline of the PBPK model including four brain compartments (modified from Gaohua et al.) [8]. Qsin and Qsout represent CSF shuttle flow between cranial CSF and spinal CSF compartments. Qssink and Qcsink are the flows from CSF compartments to blood. Qbulk represents bulk flow from brain mass to cranial CSF. PSB, PSC and PSE represent permeability surface area products between brain blood and brain mass, brain blood and cranial CSF, and brain mass and cranial CSF, respectively. Subscripts lu, br, ad, bo, he, ki, mu, sk, li, re, gu, sp, ha denote lung, brain, adipose tissue, bone, heart, kidney, muscle, skin, liver, rest tissue, gut, spleen and hepatic artery, respectively. CL is the total clearance from the model. IV is an intravenous dose and oral is an oral dose route of administration. https://doi.org/10.1371/journal.pcbi.1007117.g007 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 11 / 19 Development of a pediatric brain PBPK model where Pi is the parameter for an individual, Ppop is the population average, Z is the standard normal variable, and η is the variance. Brain model. The brain part of the model was connected to the plasma PBPK model as described before and subdivided into 4 compartments, consisting of brain blood, brain mass, cranial CSF and spinal CSF (Fig 7) [8]. The time-based differential equations used to describe concentration changes in these brain compartments were as follows: Vbb � dCbb dt ¼ Qbrain � Cbla (cid:0) Cbb ð � ð Þ þ PSb � fubm � Cbm (cid:0) � fubb � Cbb Þ þ PSc fuccsf � Cccsf (cid:0) fubb � Cbb þ Qssink � Cscsf þ Qcsink � Cccsf ð2Þ Vbm dCbm dt Vccsf � dCccsf dt ð ¼ PSb � fubb � Cbb (cid:0) fubm � Cbm � Þ þ PSe � fuccsf � Cccsf (cid:0) � fubm � Cbm (cid:0) Qbulk � Cbm � ¼ PSe � fubm � Cbm (cid:0) fuccsf � Cccsf ð3Þ � � þ PSc � fubb � Cbb (cid:0) � fuccsf � Cccsf þ Qbulk � Cbm þ Qsout � Cscsf (cid:0) Qsin � Cccsf (cid:0) Qcsink � Cccsf Vscsf � dCscsf dt ¼ Qsin � Cccsf (cid:0) Qsout � Cscsf (cid:0) Qssink � Cscsf ð4Þ ð5Þ Where Vbb, Vbm, Vccsf, Vscsf, fubm, fubb, fuccsf, fuscsf, Cbb, Cbm, Cccsf and Cscsf represent volumes, unbound fractions and concentrations in brain blood, brain mass, cranial CSF and spinal CSF, respectively. Qbrain denotes brain blood flow and Cbla denotes concentration in arterial blood. Qsin and Qsout represent CSF shuttle flow between cranial CSF and spinal CSF compartments. Qssink and Qcsink are the flows from CSF compartments to blood. Qbulk represents bulk flow from brain mass to cranial CSF. PSb, PSc and PSe represent permeability surface area products between brain blood and brain mass, brain blood and cranial CSF, and brain mass and cranial CSF, respectively. The following assumptions were made: (1) The BBB is a barrier between brain mass and brain blood, and the BCSFB between cranial CSF and brain blood. The barrier separating brain mass and cranial CSF is of high permeability. No barrier exists between cranial CSF and spinal CSF. (2) Compartments are of constant volume and well-stirred. (3) Permeability sur- face area products of the BCSFB are two-fold smaller than corresponding BBB values, as was previously described [8, 11, 39]. (4) As drug metabolism was included as a total-body clear- ance, a specific contribution of brain metabolism is assumed absent and drug entering the brain mass compartment is returned into brain blood to preserve mass balance. (5) Trans- porter-mediated transfer across barriers is considered negligible for the studied drugs, as they are no known substrates for drug transporters expressed in BBB or BCSFB. Therefore, barrier penetration is considered to occur by passive diffusion and described by permeability surface area products [24]. (6) CSF production rate was multiplied by 2 in populations receiving spinal ketamine anes- thesia, based on a two-fold increase that was observed in cats [16]. (7) A brain tissue binding parameter was derived from literature or estimated using the prediction option incorporated in Simcyp [32–35]. BBB permeability was estimated for each drug separately from rat carotid artery perfusion data or cell-based passive permeability assays, and in the latter case scaled to in vivo values by PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 12 / 19 Development of a pediatric brain PBPK model using the equation: PSbbb ¼ in vitro permeability � BBB surface ð6Þ where in vitro permeability is expressed as dm/h, and BBB surface in dm2. A sensitivity analysis was performed to investigate the effect of CSF production rate change on the drug concentra- tion-time curve in the spinal CSF compartment. Step 2 Pediatric PBPK brain model Plasma model. To translate the adult model to a pediatric model estimates of height, weight, body surface area, organ volumes, tissue blood flows, hematocrit and albumin concen- trations were adjusted using previously reported equations from literature [31] or the pediatric Simcyp simulator [32–34]. This resulted in an age-appropriate set of parameters for each simu- lated patient. Child-specific tissue composition was also incorporated, which affects the predic- tion of organ partitioning coefficients. Pediatric plasma clearances were not estimated by the model, but derived from literature (S2 Table, partially also from the same studies used for vali- dation of the brain PBPK model) and consisted of body weight-based relations. The age range for the patients in these studies overlapped with that of the simulated patients almost completely, to reduce unjustified extrapolation of clearance values across ages. For oral dosing studies a rate constant (ka) and variability was derived from literature, and if this value was unavailable an estimate was made by using the MechPeff model incorporated in Simcyp [32– 34]. Unexplained variability for clearance and absorption was incorporated using Eq 1 (previ- ous section). Brain model. The time-based differential equations and assumptions made in the adult brain PBPK model were also applied to the pediatric model. In addition, physiological values for brain parameters were derived from literature leading to the following considerations: (1) Brain volume, brain blood flow, spinal CSF volume and CSF production rate were adjusted as a function of age [31, 35, 40–43]. (2) Cranial CSF volume was not expected to further increase after birth [35, 44]. (3) Relative CSF flows between compartments, expressed as percentage of CSF production, were assumed to be similar to adult values. (4) Since BBB surface area per gram of brain is similar for adults and children, total surface area was smaller in children because of a lower brain mass compared to adults [31, 43, 45] (Table 2). Step 3 Validation of pediatric brain PBPK model To validate the physiological parameters included in the pediatric PBPK model, plasma and CSF concentrations were simulated with different drugs not used to build the model, namely ibuprofen, flurbiprofen, and naproxen using both adult and pediatric models. Only drug-spe- cific parameters were adjusted to perform simulations for these drugs, of which brain penetra- tion is like paracetamol known not to be affected by drug transporters. Step 4 Incorporation of meningitis in the adult model In case of meningitis, blood-brain barrier function is known to be impaired [47]. To have an estimate on permeability of meropenem through the BBB the permeability surface area prod- uct was empirically optimized in the adult meropenem model, both for the population average and coefficient of variation. Effects of meningitis on plasma clearance was already incorpo- rated in the plasma parameter derived from the study that was also used for validation of the model [17]. In addition, effects of meningitis on free fraction and erythrocyte-plasma parti- tioning coefficients were not incorporated, because differences in albumin concentrations and PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 13 / 19 Table 2. Pediatric brain physiological parameters. Equation ageþ0:315 9þ6:92�age 1:04 Vbrain total ¼ 10 � Vbrain blood = 0.05 � Vbrain Vccsf = 0.143 Vscsf ¼ 1:94�body weightþ0:13 Limit Vscsf � 0:143 0:8 � 0:2 Vendothelial = Vbrain total Vbrain mass = Vbrain total−Vendothelial−Vbrain blood−Vccsf−Vscsf � 0.005 1000 Equation Qcsfproductionrate(3m−18y) = 0.024 Qcsfproductionrateð<3mÞ ¼ 4:007�logðageÞþ7:088 Qcsfproductionrate (CV%) = 10 Qbulk = 0.25�Qcsfproductionrate 1000 Qbulk (CV%) = 8 Qsin = Qssink+Qsout Qsout = 0.9�Qssink Qsout (CV%) = 100 Qcsink = 0.75 � Qcsfproductionrate+Qbulk−Qsin+Qsout Qssink = 0.38�(0.75�Qcsfproductionrate+Qbulk) Qssink (CV%) = 30 Qbrain ¼ Qcarout � 10þ2290�ð10(cid:0) 0:608�age(cid:0) 10(cid:0) 0:639�ageÞ 100 BBB ¼ BBBadult � Vbrain total child Vbrain total adult Development of a pediatric brain PBPK model Volumes(L)a Notes Description Brain volume Brain blood volume Cranial CSF volume Spinal CSF volume Endothelial cell volume Brain mass volume Spinal CSF volume capped at 20% of total CSF volume (same as in adult) Fluid flow rates(L/h) Description Notes CSF production rate If ketamine used in clinical study. Production rate multiplied by 2. Bulk flow brain mass to cranial CSF Relative CSF flows (as part of Qcsfproductionrate) assumed to be the same for adults and children. Flow from cranial CSF to spinal CSF Flow from spinal CSF to cranial CSF Flow from cranial CSF to brain blood Flow from spinal CSF to brain blood Brain blood flow Cardiac output � fractional tissue flow BBB surface area (m2) BBB surface area Ref. [31] [8, 35] [35, 44] [35, 40] [35] [35] Ref. [8, 35, 41, 42] [8, 35] [8, 35] [8, 35] [8, 35] [8, 35] [8, 35] [8, 35] [8, 35] [35] [31, 43, 45] a Tissue volumes were converted to liters. Adult organ densities reported in Abduljalil et al. [46] were used to convert equations predicting organ weight to organ volumes if needed. https://doi.org/10.1371/journal.pcbi.1007117.t002 hematocrit levels were assumed to be of minor importance due to low protein binding and low cell penetration of meropenem [48–51]. Step 5 Incorporation of meningitis in the pediatric model The value for blood-brain barrier permeability estimated in the adult model was also used in the pediatric model after correction for the difference in blood-brain barrier surface area. The effect of meningitis on plasma clearance was already incorporated in the clearance parameter derived from the NeoMero study, in which only patients were included suffering from sepsis or meningitis [20]. Effects of meningitis on free fraction and erythrocyte-plasma partitioning coefficients were not incorporated, as also for the pediatric population this was assumed to be of minor importance (see previous section). PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 14 / 19 Development of a pediatric brain PBPK model Table 3. Characteristics of studies included for validation. Study Drug Number of patients Co-medication Age(y) Indication CSF collection Singla et al., 2012 [15] Bannwarth et al., 1992 [7] Kumpulainen et al., 2007 [52] Va¨litalo et al., 2012 [53] Kumpulainen et al., 2010 [54] Kokki et al., 2007 [55] Har-Even et al., 2014 [56] Lu et al., 2016 [17] Germovsek et al., 2018 [20] Paracetamol Propacetamol (similar to 50% dose of paracetamol) [58] Paracetamol Naproxen Flurbiprofen Ibuprofen Ibuprofen Meropenem 7 43 32 53 27 36 28 82 Meropenem 167 https://doi.org/10.1371/journal.pcbi.1007117.t003 - - Midazolam, ketamine, propofol, thiopental Midazolam, ketamine, propofol, thiopental Midazolam, ketamine, propofol, thiopental, paracetamol, ketoprofen, fentanyl 19–44 Healthy Spinal catheter 31–73 Nerve-root compression pain Diagnostic lumbar puncture 0.25–13 Elective surgery 0.25–12 0.25–13 Surgery lower part body Surgery lower part body Lumbar puncture for spinal anesthesia Lumbar puncture for spinal anesthesia Lumbar puncture for spinal anesthesia Midazolam, ketamine 0.25–12 Surgery - 0.42–15 Suspected sepsis Lumbar puncture for spinal anesthesia Lumbar puncture for sepsis assessment - not reported - not reported 17–77 Meningitis Lumbar drainage 0.0027– 0.25 Sepsis/ Meningitis Opportunistic lumbar puncture for sepsis/ meningitis assessment Percentage male (%) 100 56 59 74 78 69 64 61 53 In vivo observations and model evaluation Clinical studies describing concentration-time profiles of paracetamol in adults (step 1) and paracetamol (step 2), ibuprofen, flurbiprofen and naproxen (step 3), in children, were used for validation and extracted from original publications with WebPlotDigitizer version 4.1 [7, 15, 17, 20, 52–56]. As plasma and CSF samples were taken in the setting of clinical care, the major- ity of data was derived from individuals having a clinical condition, however, except for patients suffering from sepsis and/or meningitis, this was not expected to affect brain perme- ability (Table 3). Studies included healthy adult volunteers (1 study [15]), adult nerve root compression pain/arthritis patients (1 study [7]), pediatric surgery patients receiving spinal anesthesia (4 studies [52–55]) and pediatric (suspected) sepsis patients (1 study [56]). For step 4 (adults) and step 5 (children) meropenem PK studies, all patients suffered from (suspected) sepsis and/or meningitis (2 studies [17, 20]). In the study of Germovsek et al. concentrations were described as ‘time after dose’ at steady state, which was expected to be reached after 24h [57]. Because in this study meropenem CSF values for children suffering from sepsis (without meningitis) were available, simulations were performed for this population as well. Simulations were run for 1000 virtual individuals who were matched with the individuals in the original studies for dosing regimen, age range, and percentage male/female. Simulations of CSF concentration-time profiles were visualized for the spinal compartment as clinical mea- surements were performed by lumbar puncture. Results were compared with observations using visual predictive checks in which median, 5th percentile, 95th percentile, minimum, and maximum values were overlaid with clinical observations derived from literature. Also, plasma and CSF AUC0-last were calculated for median simulated data and for observed data using a non-compartmental (linear trapezoidal) approach. A naïve pooling approach was used for the population data because variability between subjects was expected to cancel out in the analysis. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 15 / 19 Development of a pediatric brain PBPK model In case of two observations at the same time, the average was used. AUCs were compared between simulated data and observed data according to: fold error ¼ observed AUC predicted AUC ð7Þ and model simulations were considered acceptable if the ratio was within two-fold difference [59]. Only plasma peak and through levels were available for the pediatric population receiving meropenem and few high CSF values largely influenced AUC. This means that requirements for the use of the naïve pooling approach are not met [60]. In contrast, AUC was for this study calculated below the simulated concentration time profile reported. Supporting information S1 File. Rstudio code. (R) S1 Table. Physiological parameters. (PDF) S2 Table. Drug-related parameters. (PDF) Author Contributions Conceptualization: Laurens F. M. Verscheijden, Jan B. Koenderink, Saskia N. de Wildt, Frans G. M. Russel. Data curation: Laurens F. M. Verscheijden. Formal analysis: Laurens F. M. Verscheijden. Supervision: Jan B. Koenderink, Saskia N. de Wildt, Frans G. M. Russel. Validation: Laurens F. M. Verscheijden. Visualization: Laurens F. M. Verscheijden. Writing – original draft: Laurens F. M. Verscheijden, Saskia N. de Wildt, Frans G. M. Russel. Writing – review & editing: Jan B. Koenderink, Saskia N. de Wildt, Frans G. M. Russel. References 1. Johnson TN. The problems in scaling adult drug doses to children. Arch Dis Child. 2008; 93(3):207–11. Epub 2007/03/17. https://doi.org/10.1136/adc.2006.114835 PMID: 17363397. 2. 3. 4. Templeton IE, Jones NS, Musib L. Pediatric Dose Selection and Utility of PBPK in Determining Dose. AAPS J. 2018; 20(2):31. Epub 2018/02/15. https://doi.org/10.1208/s12248-018-0187-8 PMID: 29441439. Lu J, Goldsmith MR, Grulke CM, Chang DT, Brooks RD, Leonard JA, et al. Developing a Physiologi- cally-Based Pharmacokinetic Model Knowledgebase in Support of Provisional Model Construction. PLoS Comput Biol. 2016; 12(2):e1004495. Epub 2016/02/13. https://doi.org/10.1371/journal.pcbi. 1004495 PMID: 26871706; PubMed Central PMCID: PMCPMC4752336. Jamei M. Recent Advances in Development and Application of Physiologically-Based Pharmacokinetic (PBPK) Models: a Transition from Academic Curiosity to Regulatory Acceptance. Curr Pharmacol Rep. 2016; 2:161–9. https://doi.org/10.1007/s40495-016-0059-9 PMID: 27226953; PubMed Central PMCID: PMCPMC4856711. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 16 / 19 Development of a pediatric brain PBPK model 5. Kim A, McCully C, Cruz R, Cole DE, Fox E, Balis FM, et al. The plasma and cerebrospinal fluid pharma- cokinetics of sorafenib after intravenous administration in non-human primates. Invest New Drugs. 2012; 30(2):524–8. https://doi.org/10.1007/s10637-010-9585-1 PMID: 21072558. 6. Hain RD, Hardcastle A, Pinkerton CR, Aherne GW. Morphine and morphine-6-glucuronide in the plasma and cerebrospinal fluid of children. Br J Clin Pharmacol. 1999; 48(1):37–42. https://doi.org/10. 1046/j.1365-2125.1999.00948.x PMID: 10383558; PubMed Central PMCID: PMCPMC2014881. 7. Bannwarth B, Netter P, Lapicque F, Gillet P, Pere P, Boccard E, et al. Plasma and Cerebrospinal-Fluid Concentrations of Paracetamol after a Single Intravenous Dose of Propacetamol. Brit J Clin Pharmaco. 1992; 34(1):79–81. https://doi.org/10.1111/j.1365-2125.1992.tb04112.x WOS:A1992JC22000013. 8. Gaohua L, Neuhoff S, Johnson TN, Rostami-Hodjegan A, Jamei M. Development of a permeability-lim- ited model of the human brain and cerebrospinal fluid (CSF) to integrate known physiological and bio- logical knowledge: Estimating time varying CSF drug concentrations and their variability using in vitro data. Drug Metab Pharmacokinet. 2016; 31(3):224–33. https://doi.org/10.1016/j.dmpk.2016.03.005 PMID: 27236639. 9. Yamamoto Y, Valitalo PA, van den Berg DJ, Hartman R, van den Brink W, Wong YC, et al. A Generic Multi-Compartmental CNS Distribution Model Structure for 9 Drugs Allows Prediction of Human Brain Target Site Concentrations. Pharm Res. 2017; 34(2):333–51. https://doi.org/10.1007/s11095-016- 2065-3 PMID: 27864744; PubMed Central PMCID: PMCPMC5236087. 10. Yamamoto Y, Valitalo PA, Huntjens DR, Proost JH, Vermeulen A, Krauwinkel W, et al. Predicting Drug Concentration-Time Profiles in Multiple CNS Compartments Using a Comprehensive Physiologically- Based Pharmacokinetic Model. CPT Pharmacometrics Syst Pharmacol. 2017; 6(11):765–77. https:// doi.org/10.1002/psp4.12250 PMID: 28891201; PubMed Central PMCID: PMCPMC5702903. 11. Yamamoto Y, Valitalo PA, Wong YC, Huntjens DR, Proost JH, Vermeulen A, et al. Prediction of human CNS pharmacokinetics using a physiologically-based pharmacokinetic modeling approach. Eur J Pharm Sci. 2018; 112:168–79. https://doi.org/10.1016/j.ejps.2017.11.011 PMID: 29133240. 12. Ketharanathan N, Yamamoto Y, Rohlwink UK, Wildschut E, Mathot RAA, de Lange ECM, et al. Combin- ing brain microdialysis and translational pharmacokinetic modeling to predict drug concentrations in pediatric severe traumatic brain injury: the next step towards evidence-based pharmacotherapy? J Neu- rotrauma. 2018. https://doi.org/10.1089/neu.2017.5588 PMID: 30019622. 13. Nau R, Sorgel F, Eiffert H. Penetration of drugs through the blood-cerebrospinal fluid/blood-brain barrier for treatment of central nervous system infections. Clin Microbiol Rev. 2010; 23(4):858–83. https://doi. org/10.1128/CMR.00007-10 PMID: 20930076; PubMed Central PMCID: PMCPMC2952976. 14. Kim KS. Acute bacterial meningitis in infants and children. Lancet Infect Dis. 2010; 10(1):32–42. https:// doi.org/10.1016/S1473-3099(09)70306-8 PMID: 20129147. 15. Singla NK, Parulan C, Samson R, Hutchinson J, Bushnell R, Beja EG, et al. Plasma and Cerebrospinal Fluid Pharmacokinetic Parameters After Single-Dose Administration of Intravenous, Oral, or Rectal Acetaminophen. Pain Pract. 2012; 12(7):523–32. https://doi.org/10.1111/j.1533-2500.2012.00556.x WOS:000308540000005. PMID: 22524979 16. Wood JH. Neurobiology of Cerebrospinal Fluid 1 1st ed. Philadelphia: University of Pennsylvania Press1980. 17. Lu C, Zhang Y, Chen M, Zhong P, Chen Y, Yu J, et al. Population Pharmacokinetics and Dosing Regi- men Optimization of Meropenem in Cerebrospinal Fluid and Plasma in Patients with Meningitis after Neurosurgery. Antimicrob Agents Chemother. 2016; 60(11):6619–25. https://doi.org/10.1128/AAC. 00997-16 PMID: 27572392; PubMed Central PMCID: PMCPMC5075067. 18. Ball K, Bouzom F, Scherrmann JM, Walther B, Decleves X. A physiologically based modeling strategy during preclinical CNS drug development. Mol Pharm. 2014; 11(3):836–48. https://doi.org/10.1021/ mp400533q PMID: 24446829. 19. Ball K, Bouzom F, Scherrmann JM, Walther B, Decleves X. Physiologically based pharmacokinetic modelling of drug penetration across the blood-brain barrier—towards a mechanistic IVIVE-based approach. AAPS J. 2013; 15(4):913–32. https://doi.org/10.1208/s12248-013-9496-0 PMID: 23784110; PubMed Central PMCID: PMCPMC3787211. 20. Germovsek E, Lutsar I, Kipper K, Karlsson MO, Planche T, Chazallon C, et al. Plasma and CSF phar- macokinetics of meropenem in neonates and young infants: results from the NeoMero studies. J Antimi- crob Chemother. 2018. https://doi.org/10.1093/jac/dky128 PMID: 29684147. 21. Sharief MK, Ciardi M, Thompson EJ. Blood-brain barrier damage in patients with bacterial meningitis: association with tumor necrosis factor-alpha but not interleukin-1 beta. J Infect Dis. 1992; 166(2):350–8. https://doi.org/10.1093/infdis/166.2.350 PMID: 1634806. 22. Nau R, Lassek C, Kinzig-Schippers M, Thiel A, Prange HW, Sorgel F. Disposition and elimination of meropenem in cerebrospinal fluid of hydrocephalic patients with external ventriculostomy. Antimicrob PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 17 / 19 Development of a pediatric brain PBPK model Agents Chemother. 1998; 42(8):2012–6. Epub 1998/08/04. PMID: 9687399; PubMed Central PMCID: PMCPMC105725. 23. Kuperberg SJ, Wadgaonkar R. Sepsis-Associated Encephalopathy: The Blood-Brain Barrier and the Sphingolipid Rheostat. Front Immunol. 2017; 8:597. https://doi.org/10.3389/fimmu.2017.00597 PMID: 28670310; PubMed Central PMCID: PMCPMC5472697. 24. Uchida Y, Ohtsuki S, Katsukura Y, Ikeda C, Suzuki T, Kamiie J, et al. Quantitative targeted absolute proteomics of human blood-brain barrier transporters and receptors. J Neurochem. 2011; 117(2):333– 45. https://doi.org/10.1111/j.1471-4159.2011.07208.x PMID: 21291474. 25. 26. Li J, Wu J, Bao X, Honea N, Xie Y, Kim S, et al. Quantitative and Mechanistic Understanding of AZD1775 Penetration across Human Blood-Brain Barrier in Glioblastoma Patients Using an IVIVE- PBPK Modeling Approach. Clin Cancer Res. 2017; 23(24):7454–66. Epub 2017/09/21. https://doi.org/ 10.1158/1078-0432.CCR-17-0983 PMID: 28928160; PubMed Central PMCID: PMCPMC6097184. Lam J, Baello S, Iqbal M, Kelly LE, Shannon PT, Chitayat D, et al. The ontogeny of P-glycoprotein in the developing human blood-brain barrier: implication for opioid toxicity in neonates. Pediatr Res. 2015; 78 (4):417–21. https://doi.org/10.1038/pr.2015.119 PMID: 26086643. 27. Daood M, Tsai C, Ahdab-Barmada M, Watchko JF. ABC transporter (P-gp/ABCB1, MRP1/ABCC1, BCRP/ABCG2) expression in the developing human CNS. Neuropediatrics. 2008; 39(4):211–8. https:// doi.org/10.1055/s-0028-1103272 PMID: 19165709; PubMed Central PMCID: PMCPMC2821654. 28. Frattarelli DA, Galinkin JL, Green TP, Johnson TD, Neville KA, Paul IM, et al. Off-label use of drugs in children. Pediatrics. 2014; 133(3):563–7. https://doi.org/10.1542/peds.2013-4060 PMID: 24567009. 29. Barrett JS, Della Casa Alberighi O, Laer S, Meibohm B. Physiologically based pharmacokinetic (PBPK) modeling in children. Clin Pharmacol Ther. 2012; 92(1):40–9. https://doi.org/10.1038/clpt.2012.64 PMID: 22669290. 30. Rostami-Hodjegan A. Physiologically based pharmacokinetics joined with in vitro-in vivo extrapolation of ADME: a marriage under the arch of systems pharmacology. Clin Pharmacol Ther. 2012; 92(1):50– 61. Epub 2012/05/31. https://doi.org/10.1038/clpt.2012.65 PMID: 22644330. 31. Price PS, Conolly RB, Chaisson CF, Gross EA, Young JS, Mathis ET, et al. Modeling interindividual var- iation in physiological factors used in PBPK models of humans. Crit Rev Toxicol. 2003; 33(5):469–503. PMID: 14594104. 32. 33. 34. Jamei M, Dickinson GL, Rostami-Hodjegan A. A framework for assessing inter-individual variability in pharmacokinetics using virtual human populations and integrating general knowledge of physical chem- istry, biology, anatomy, physiology and genetics: A tale of ’bottom-up’ vs ’top-down’ recognition of covariates. Drug Metab Pharmacokinet. 2009; 24(1):53–75. Epub 2009/03/03. PMID: 19252336. Jamei M, Marciniak S, Edwards D, Wragg K, Feng K, Barnett A, et al. The simcyp population based sim- ulator: architecture, implementation, and quality assurance. In Silico Pharmacol. 2013; 1:9. Epub 2013/ 01/01. https://doi.org/10.1186/2193-9616-1-9 PMID: 25505654; PubMed Central PMCID: PMCPMC4230310. Jamei M, Marciniak S, Feng K, Barnett A, Tucker G, Rostami-Hodjegan A. The Simcyp population- based ADME simulator. Expert Opin Drug Metab Toxicol. 2009; 5(2):211–23. Epub 2009/02/10. https:// doi.org/10.1517/17425250802691074 PMID: 19199378. 35. Simcyp. 17.0.90.0 ed. available from: https://www.certara.com/. 36. Rodgers T, Leahy D, Rowland M. Physiologically based pharmacokinetic modeling 1: predicting the tis- sue distribution of moderate-to-strong bases. J Pharm Sci. 2005; 94(6):1259–76. https://doi.org/10. 1002/jps.20322 PMID: 15858854. 37. Rodgers T, Rowland M. Physiologically based pharmacokinetic modelling 2: predicting the tissue distri- bution of acids, very weak bases, neutrals and zwitterions. J Pharm Sci. 2006; 95(6):1238–57. https:// doi.org/10.1002/jps.20502 PMID: 16639716. 38. Rodgers T, Rowland M. Mechanistic approaches to volume of distribution predictions: understanding the processes. Pharm Res. 2007; 24(5):918–33. https://doi.org/10.1007/s11095-006-9210-3 PMID: 17372687. 39. Skipor J, Thiery JC. The choroid plexus—cerebrospinal fluid system: undervaluated pathway of neuro- endocrine signaling into the brain. Acta Neurobiol Exp (Wars). 2008; 68(3):414–28. PMID: 18668165. 40. Rochette A, Malenfant Rancourt MP, Sola C, Prodhomme O, Saguintaah M, Schaub R, et al. Cerebro- spinal fluid volume in neonates undergoing spinal anaesthesia: a descriptive magnetic resonance imag- ing study. Br J Anaesth. 2016; 117(2):214–9. https://doi.org/10.1093/bja/aew185 PMID: 27440633. 41. Blomquist HK, Sundin S, Ekstedt J. Cerebrospinal fluid hydrodynamic studies in children. J Neurol Neu- rosurg Psychiatry. 1986; 49(5):536–48. https://doi.org/10.1136/jnnp.49.5.536 PubMed Central PMCID: PMCPMC1028808. PMID: 3711916 PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 18 / 19 Development of a pediatric brain PBPK model 42. Yasuda T, Tomita T, McLone DG, Donovan M. Measurement of cerebrospinal fluid output through external ventricular drainage in one hundred infants and children: correlation with cerebrospinal fluid production. Pediatr Neurosurg. 2002; 36(1):22–8. https://doi.org/10.1159/000048344 PMID: 11818742. 43. Dekaban AS. Changes in brain weights during the span of human life: relation of brain weights to body heights and body weights. Ann Neurol. 1978; 4(4):345–56. https://doi.org/10.1002/ana.410040410 PMID: 727739. 44. Andescavage NN, DuPlessis A, McCarter R, Vezina G, Robertson R, Limperopoulos C. Cerebrospinal Fluid and Parenchymal Brain Development and Growth in the Healthy Fetus. Dev Neurosci. 2016; 38 (6):420–9. https://doi.org/10.1159/000456711 PMID: 28315866. 45. Ballabh P, Braun A, Nedergaard M. Anatomic analysis of blood vessels in germinal matrix, cerebral cor- tex, and white matter in developing infants. Pediatr Res. 2004; 56(1):117–24. https://doi.org/10.1203/ 01.PDR.0000130472.30874.FF PMID: 15128918. 46. Abduljalil K, Jamei M, Johnson TN. Fetal Physiologically Based Pharmacokinetic Models: Systems Information on the Growth and Composition of Fetal Organs. Clin Pharmacokinet. 2018. https://doi.org/ 10.1007/s40262-018-0685-y PMID: 29987449. 47. Weiss N, Miller F, Cazaubon S, Couraud PO. The blood-brain barrier in brain homeostasis and neuro- logical diseases. Biochim Biophys Acta. 2009; 1788(4):842–57. https://doi.org/10.1016/j.bbamem. 2008.10.022 PMID: 19061857. 48. Craig WA. The pharmacology of meropenem, a new carbapenem antibiotic. Clin Infect Dis. 1997; 24 Suppl 2:S266–75. https://doi.org/10.1093/clinids/24.supplement_2.s266 PMID: 9126702. 49. Nicolau DP. Pharmacokinetic and pharmacodynamic properties of meropenem. Clin Infect Dis. 2008; 47 Suppl 1:S32–40. https://doi.org/10.1086/590064 PMID: 18713048. 50. Nicholson JP, Wolmarans MR, Park GR. The role of albumin in critical illness. Br J Anaesth. 2000; 85 (4):599–610. https://doi.org/10.1093/bja/85.4.599 PMID: 11064620. 51. Radke C, Horn D, Lanckohr C, Ellger B, Meyer M, Eissing T, et al. Development of a Physiologically Based Pharmacokinetic Modelling Approach to Predict the Pharmacokinetics of Vancomycin in Criti- cally Ill Septic Patients. Clin Pharmacokinet. 2017; 56(7):759–79. https://doi.org/10.1007/s40262-016- 0475-3 PMID: 28039606. 52. Kumpulainen E, Kokki H, Halonen T, Heikkinen M, Savolainen J, Laisalmi M. Paracetamol (acetamino- phen) penetrates readily into the cerebrospinal fluid of children after intravenous administration. Pediat- rics. 2007; 119(4):766–71. https://doi.org/10.1542/peds.2006-3378 WOS:000245406200039. PMID: 17403848 53. Valitalo P, Kumpulainen E, Manner M, Kokki M, Lehtonen M, Hooker AC, et al. Plasma and Cerebrospi- nal Fluid Pharmacokinetics of Naproxen in Children. Journal of Clinical Pharmacology. 2012; 52 (10):1516–26. https://doi.org/10.1177/0091270011418658 WOS:000308763200006. PMID: 22067196 54. Kumpulainen E, Valitalo P, Kokki M, Lehtonen M, Hooker A, Ranta VP, et al. Plasma and cerebrospinal fluid pharmacokinetics of flurbiprofen in children. Br J Clin Pharmacol. 2010; 70(4):557–66. https://doi. org/10.1111/j.1365-2125.2010.03720.x PMID: 20840447; PubMed Central PMCID: PMCPMC2950990. 55. Kokki H, Kumpulainen E, Lehtonen M, Laisalmi M, Heikkinen M, Savolainen J, et al. Cerebrospinal fluid distribution of ibuprofen after intravenous administration in children. Pediatrics. 2007; 120(4):E1002– E8. https://doi.org/10.1542/peds.2007-0064 WOS:000249870000079. PMID: 17908721 56. Har-Even R, Stepensky D, Britzi M, Soback S, Bar Chaim A, Brandriss N, et al. Plasma and cerebrospi- nal fluid concentrations of ibuprofen in pediatric patients and antipyretic effect: Pharmacokinetic-phar- macodynamic modeling analysis. Journal of Clinical Pharmacology. 2014; 54(9):1023–30. https://doi. org/10.1002/jcph.307 WOS:000340447400008. PMID: 24733245 57. Moon YS, Chung KC, Gill MA. Pharmacokinetics of meropenem in animals, healthy volunteers, and patients. Clin Infect Dis. 1997; 24 Suppl 2:S249–55. Epub 1997/02/01. https://doi.org/10.1093/clinids/ 24.supplement_2.s249 PMID: 9126700. 58. Flouvat B, Leneveu A, Fitoussi S, Delhotal-Landes B, Gendron A. Bioequivalence study comparing a new paracetamol solution for injection and propacetamol after single intravenous infusion in healthy subjects. Int J Clin Pharmacol Ther. 2004; 42(1):50–7. Epub 2004/02/06. PMID: 14756388. 59. Sager JE, Yu J, Ragueneau-Majlessi I, Isoherranen N. Physiologically Based Pharmacokinetic (PBPK) Modeling and Simulation Approaches: A Systematic Review of Published Models, Applications, and Model Verification. Drug Metab Dispos. 2015; 43(11):1823–37. Epub 2015/08/25. https://doi.org/10. 1124/dmd.115.065920 PMID: 26296709; PubMed Central PMCID: PMCPMC4613950. 60. Mahmood I. Naive pooled-data approach for pharmacokinetic studies in pediatrics with a very small sample size. Am J Ther. 2014; 21(4):269–74. Epub 2012/06/21. https://doi.org/10.1097/MJT. 0b013e31824ddee3 PMID: 22713529. PLOS Computational Biology \| https://doi.org/10.1371/journal.pcbi.1007117 June 13, 2019 19 / 19	null
10.1088_1402-4896_ad16b7.pdf	Data availability statement The data that support the ﬁndings of this study are available upon reasonable request from the authors.	Data availability statement The data that support the findings of this study are available upon reasonable request from the authors.	RECEIVED 15 July 2023 REVISED 11 December 2023 ACCEPTED FOR PUBLICATION 18 December 2023 PUBLISHED 29 December 2023 Phys. Scr. 99 (2024) 015246 https://doi.org/10.1088/1402-4896/ad16b7 PAPER Global and multistable dynamics in calcium oscillations model Rajes Kannan Subramanian1, Zeric Tabekoueng Njitacke2,∗ Karthikeyan Rajagopal4 1 Center for Artiﬁcial Intelligence, Chennai Institute of Technology, Chennai 600069, Tamil Nadu, India 2 Department of Electrical and Electronic Engineering, College of Technology (COT), University of Buea, PO Box 63, Buea, Cameroon 3 Department of Automation, Biomechanics and Mechatronics, Lodz University of Technology, ul. Stefanowskiego 1/15, 90-537 Lodz, , Jan Awrejcewicz3 and Poland 4 Centre for Nonlinear Systems, Chennai Institute of Technology, Chennai 600069, Tamil Nadu, India ∗ Author to whom any correspondence should be addressed. E-mail: [email protected] Keywords: calcium oscillation, two-parameter charts, coexistence of bifurcations, multistability Abstract In most animal and plant cells, the information’s processing is insured by calcium ions. This contribution studies the global dynamics of a model of calcium oscillation. From the stability analysis, it is found that the oscillations of that model are self-excited since they are generated from unstable equilibria. Using two-parameter charts, the general behavior of the model is explored. From the hysteresis analysis using bifurcation diagrams with their related Largest Lyapunov Exponent (LLE) graphs, the coexisting oscillation modes are recorded. This phenomenon is characterized by the simultaneous existence of periodic and chaotic oscillations in the considered model by just varying the initial conditions. Using a set of parameters for which the model exhibits multistability, the basins of attraction related to each coexisting solution are computed and enable the capture of any coexisting pattern. 1. Introduction ) - Ca ( IP 3 +Ca2 +Ca2 +Ca 2 [ released on cytosolic ] which constitutes most cell types [7–9]. In [10], , cross-coupling model (ICC) of [11], the study of the +Ca2 concentration. From the built model, under the variation of the +Ca2 oscillations was introduced. That model was derived from the study of the Calcium oscillations have a key role in the function of cells since they act in them from fertilization to death. The oscillation processes are found in a variety of cell types, such as neuron cells [1], astrocyte cells [2], T-lymphocyte cells [3], oocyte cells [4]and hepatocyte cells [5, 6]. From experimental work, it has been shown that oscillations are related to the variation in free cytosolic an improved model of the + 2 mechanism in inositol trisphosphate calcium-induced calcium release (CICR) mechanism in [12] and the consideration of the bell-shaped dependence of agonist, complex phenomena such as period adding, period doubling, mixed mode states, and chaos were found, and they match well with experimental results. Reference [13] introduced three different models of calcium oscillations, all of which are derived from one- or two-pool models centered on ICR. The ﬁrst model was created by taking into account the suppression of the release channel on a single intracellular store when cytosolic 1,4,5-trisphosphate (IP3) that is activated by periodic and chaotic behaviors. Another model of conducted on hepatocytes. This new model considered the feedback inhibition of calcium and activated phospholipase C on the initial agonist receptor complex, along with the receptor type-dependent self-enhanced behavior of the activated aG subunit. From the simulation, the authors showed the model was able to exhibit rich dynamical behavior ranging from simple periodic to chaotic dynamics. The fractional version of the model proposed in [14] has been introduced in [5]. Exploiting the modiﬁed trapezoidal rule for fractional integrals, the bifurcation analyses of the model were addressed. Using the numerical simulation, different types of bursting ] levels are high. In the second model, the authors incorporated the degradation of inositol +Ca2 oscillations was proposed in [14], based on experiments +Ca .2 Through numerical simulations, these models revealed +Ca2 [ +Ca2 © 2023 IOP Publishing Ltd Phys. Scr. 99 (2024) 015246 R K Subramanian et al calcium oscillations in non-excitable cells have been highlighted in [15]. After that, it is also found in the work of [16] that focused on the fractional investigation of the calcium oscillations based on the Atangana-Baleanu operator. As can be seen in the quoted literature, many works have been done in the area of designing and +Ca2 oscillations, taking into consideration several additional hypotheses [5, 10–15]. Using time investigating delay with additive Gaussian colored noised, intracellular calcium oscillation (ICO) was investigated [17]. From the work of the authors, the found that time delay was able to induce saw-tooth calcium wave and transition in intracellular calcium oscillation. In the intracellular calcium oscillation system, which is deﬁned by the processes of active and passive intracellular coherence resonance were examined [18]. The authors discovered that (i) a speciﬁc time delay or correlation time value caused stochastic or reverse synchronization; (ii) a phenomenon known as reverse resonance was obtained in the function of reciprocal coefﬁcient of variance versus time delay or versus strength of noises as time delay increased; and (iii) both stochastic and reverse resonance were observed in the function of reciprocal coefﬁcient of variance versus correlation time with varying strength of noises. Also, under the effect of the time delay, the phenomenon of oscillatory coherence in the intracellular calcium oscillation system was captured in [19]. transport driven by colored sounds, the effects of time delay on the +Ca2 Therefore, a plethora of dynamical behaviors such as period adding, period doubling, mixed mode states, bursting, and chaos were recorded. Some of those behaviors have already been found in some other biological and mechanical systems, such as pancreatic beta-cells [20], electromechanical transducers [21], modiﬁed Rayleigh-Dufﬁng oscillators [22], neuron models [23–26]. Since the aforementioned works are all related to dynamical systems, one striking and widespread phenomenon found in those systems is the coexistence of patterns (multistability). Multistability in a model is characterized by the presence of multiple stable states, which can be achieved by modifying the initial conditions while keeping the parameters constant [27, 28]. Despite the large number of works already addressed on the calcium oscillation model [23–26], there is currently no study that speciﬁcally focuses on analyzing the coexisting bifurcations with more than two oscillating states in such model. Since multistability analysis is important in calcium oscillation models because it helps to identify the different stable states that the system can exhibit. This is important because many biological systems are able to exhibit multiple stable states, which can have different functional consequences. For example, in pancreatic cells, calcium oscillations are involved in regulating insulin secretion. Multistability analysis has shown that these cells can exhibit different stable states of calcium oscillations, which may correspond to different levels of insulin secretion [29]. Understanding the multistability of calcium oscillation models can also help to identify potential therapeutic targets for diseases that involve abnormal calcium signaling, such as epilepsy or cardiac arrhythmias [30, 31]. Therefore, in order to contribute to the existing literature and uncover previously unknown behaviors of the calcium oscillation model proposed in [14] based on experiments in hepatocytes, this study will reexamine the model using adapted computation techniques. The remainder of this study is organized as follows: section 2 provides a review of the mathematical model for calcium oscillations and examines its stability near equilibria. Section 3 employs numerical simulation techniques to explore the overall dynamics of the model, including one- and two-parameter charts and basins of attraction. Finally, section 4 offers a conclusion that summarizes the ﬁndings and presents potential directions for future research. 2. Model and its features In the cells of plants and animals, calcium ions are the main element involved in information processing since cytoplasmic free calcium ions enable the second messenger under the effect of a variety of stimuli. From the experimental investigations of hepatocytes [32], the mathematical model of the calcium oscillation of equation (1) has been introduced. d x dt d u dt d y dt d v dt ⎧ ⎪ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎪ ⎩ = + k 1 k x 2 - k 3 xu + k 4 x - k 5 xy + x k 6 = k x 7 - k 8 u + u k 9 k u 12 + k x 13 - k 14 y + y k 15 - k 16 y + y k 17 = k 10 v = - k 10 + uyv + k 11 uyv + k 11 + k 16 y + k 17 In equation (1), the ﬁrst variable x represents the concentration of the aG ( which is one of three guanine nucleotide-binding protein components, all of which are heterotrimeric membrane-associated G proteins) v y ( ) 1 2 Phys. Scr. 99 (2024) 015246 R K Subramanian et al Table 1. Equilibrium points for varying k10 and their stability. Control parameter Equilibrium points Eigenvalues and their stability =k 10 3.63 =k 10 5 - ) ( 7.814; 29.575; 0.775; 0.155 - ) ( 10.054; 53.709; 0.051 1.318; - ) ( 19.124; 153.502; 32.724; 0.001 - - - ) ( 2.664 0.00; 0.056; 0.029; - - - ( ) 2.613 0.005; 0.646; 0.352; - - - ) ( 2.535 0.017; 2.155; 1.240; - - - - ) ( 7.814; 29.575; 0.775; 0.11 - ) ( 0.037 10.054; 3.709; 1.318; - ) ( 153.502; 32.724; 0 19.124; - - ( ) 0.056; 0; 0.029; 2.662 - - - - ( ) 2.593 0.646; 0.352; - - - - ) ( 2.488 2.155; 1.240; - 0.005; 0.017; 2 l = 0.189 1,2 l = 96.749, 2 1 l = 6833.018, 1 l = 3.103, 1 l = 5.390, 1 l = 29.063, 1 unstable node l =  0.191 1,2 l = 132.911, 1 l = 9410.44, 1 l = 3.104, 1 l = 5.390, 1 l = 29.066, 1 unstable node 2 2 l = - 3,4 1.184j  l = -15.589 l = -0.164 2 2 l = -1.172, l = -0.718, 2 l = -1.237, unstable  0.335 1.224 24.155 l = - 3,4 l = 3,4 l = -8.225, l = -10.252, 3 l = -18.566, 3 3 12.283j   unstable unstable 1.404j 7.902j l = -962.771 4 l = -1035.832 4 l = -1276.513 4 unstable node unstable node 1.186j l = - 3,4 2 l = -15.429 l = -0.1642, 2 unstable  0.336 30.025 l = - 3,4 l = 3,4 l = -5.981, l = -7.562, 4 3 l = -13.996, l = -1276.211 11.752 j unstable + j 1.405 7.902 j unstable l = -962.769 l = -1035.797 1.224  3 4 4 3 l = -1.171, l = -0.718, 2 l = -1.237, unstable node unstable node subunits. u stands for the concentration of active PLC (phospholipase C, which is a membrane-associated enzyme family that cleaves phospholipids immediately before the phosphate group). The third variable y represents the free calcium concentration in the cytosol. The fourth variable v is considered as the concentration of calcium in the endoplasmic reticulum (ER). In addition, the constant k1 stands for the spontaneous activation of the aG sub-unit, k a2 mimicking the accelerated development of the active aG after the binding if agonist to the membrane receptor. In other word, k2 can stands for the concentration of the agonist. In general, =k i i ( description of those coefﬁcient is found in [29]. ) are constant values use to setup the calcium model under consideration more detailed 17 1  v 0.  = = = = y The equilibrium points of the considered model are obtained by solving the equation given as  u  x computed and recorded in the second column of table 1 for the following values of the coefﬁcient: =k =k =k 6 2 =k =k 16 12 With the help of MAPLE, the equilibria on the calcium oscillation model have been =k 3 =k 13 2.7738, 0.7, 0.64, 13.58, =k 7 =k 17 2.08, 0.05. =k 4 =k 14 =k 5 =k 15 1.18, 4.85, 0.19, 153, 4.88, 0.16, 1 =k 11 32.24, 29.09, =k 9 =k 8 2.67, 0.09, The Jacobian matrix around the any of the equilibrium of the considered calcium model is given by equation (2). J = a k 7 - ¯ u k 13 0 ⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣ With = - a k 2 ¯ x + ¯ k y 14 + k 15 - 2 - ¯ k x 3 + k 4 ¯ x k 8 + + + ( ¯ u k 9 ¯ ¯ k yv 10 + ¯ k v 11 ¯ ¯ k yv 10 + ¯ k v 11 - ¯ k u 8 + k 2 ) 9 k 12 4 ¯ k u 3 + k k 16 + k 17 + ( ¯ x + 2 ¯ ¯ k ux 3 + ) k 4 ¯ k y 16 + k 17 - ¯ k x 5 + k 6 ¯ x 0 b - ¯ ¯ k uv 10 + ¯ k v 11 + k 16 + k 17 ¯ y - ( ¯ y ¯ k y 16 + k 17 2 ) 0 0 ¯ ¯ k yu 10 + ¯ k v 11 ¯ ¯ k uy 10 + ¯ k v 11 - - + 2 ¯ ¯ ¯ k yuv 10 + ) ( ¯ k v 11 ¯ ¯ ¯ k uyv 10 + ) ( ¯ v k 11 2 ⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦ - ¯ k y 5 + k 6 ¯ x + ( ¯ x ¯ ¯ k yx 5 + k 6 2 ) , b = ¯ ¯ k uv 10 + ¯ k v 11 - k 14 + k 15 ¯ y ( ¯ y Afterward, the eigenvalues with their related stability have been obtained and discussed in the third line of ( ¯ y ¯ y ) ) 2 table 1 by solving equation (2). det ( J l- I ) = 0 With I representing a fourth order, identify matrix. From the computation, it is found that for each discrete value of the control parameter, the calcium oscillation model has six equilibria. The determination of the eigenvalues related to each of those equilibria revealed that the model displays self-excited oscillation from the nature of the eigenvalues, which can be either unstable saddle-focus or unstable nodes. The properties of having a large number of ﬁxed points can be the origin of some coexisting states that may be found during the investigation of the considered model. 3 ( ) 2 ( ) 3 Phys. Scr. 99 (2024) 015246 R K Subramanian et al Figure 1. Two-dimensional charts showing the global behavior the calcium oscillation model under the variation of the paramter in =k =k ( various planes. The plane k The 16 8 data in red are associated with chaotic motion while those in blue are associated with periodic motion. Initial conditions are ( 0.01; 0.01; 0.01; 20 by decreasing the control parameters. ) in addition, the left diagrams are obtained by increasing the control parameter while the right one are obtained ) obtained for =k ) obtained for =k ( the plane k 32.24, 32.24, 4.88, 4.85. k,10 k,10 16 5 8 5 3. Numerical results Since most of the previous works addressed on that calcium model were just some routes that gave birth to its chaotic dynamics. Our goal in this work is the investigation of the global behavior dynamics of that calcium oscillation model, using some two-dimensional charts that will exploit the sign of the Largest Lyapunov Exponent (LLE). The largest Lyapunov exponent of our model have been computed using the formula of equation (4) l max = lim t¥ 1 t ⎡ ⎣ ln ( 2 d ¯ x + 2 d ¯ u + 2 d ¯ y + d 2 ¯ ) v ⎤ ⎦ ( ) 4 and computed from the variational equations obtained by perturbing the solutions of the considered calcium d d  + oscillation model as follows: v, with the help of the u u, algorithm developed in [33]. 2 2 + d d ¯ is the distance between neighboring trajectories; ¯ v y 2 l ( exp asymptotically Using some bifurcation diagrams that will be computed using a hysteresis approach based on the  + d x, x 2 2 + d d ¯ ¯ u x d d + ¯ ¯ v y u + =  + v  + y d y, and d ¯ u d ¯ x max + + ) . t x v y 2 2 2 continuation method or diagrams using a ﬁxed initial condition based parallel bifurcation branch approach, the possible existence of the coexisting dynamics will be addressed. Almost all these characterization tools will be integrated using the Runge–Kutta algorithm with a ﬁxed time step of ´ - 10 3 with the format of the variable extended precision mode. 1 3.1. Two-dimensional charts The two-parameter charts from which the global behavior of the calcium oscillation model is examined are captured by varying simultaneously in the backward and upward directions two control parameters of the model. The two-parameter charts and the attraction basin, are obtained by numerically computing the LE spectrum on a grid of 400 values of the chosen space parameters, in order to have diagrams with uniform ´400 4 Phys. Scr. 99 (2024) 015246 R K Subramanian et al Figure 2. Bifurcation diagrams (a(i) and b(i)) showing the behavior of the calcium model under the variation of one parameters with their corresponding Largest Lyapunov Exponents (a(ii) and b(ii)). There, the local maxima of the concentration of the aG subunits ) are recorded under the variation of the coefﬁcient k .10 This particular diagram enables to see the rate of change of the ( xmax concentration of the aG subunits under the variation of the coefﬁcient k10 The enlargement of ﬁgures 2(a(i) and (b(i)) and found in ﬁgures 2(a(ii) and (b(ii)). Figure 3. Coexisting patterns in the calcium model for the value of the parameter initial states 5.25; 2.25; 0; 0 with coexisting phase portraits, while ﬁgure 3 (b) is associated with the related time series showing the temporal evolution of the coexisting concentration of the aG subunits. =k 10 ) while the periodic one sis obtained using the initial states The chaotic pattern is obtained using the - ) Figure 3(a) is associated 3.63. ( 5.25; 8.25; 0; 0 . ( shapes. The integration was realized in Turbo pascal with variables in extended precision mode and ran on an octa-core workstation with an Intel core i7 processor. For each variation of the control parameter, the LLE of the model is computed and used to differentiate the two main behaviors of the model. On the one hand, the model exhibits chaotic behavior supported by a positive LLE, and on the other, periodic behavior supported by a null LLE. More importantly, this upward and backward computational technique represents an effective way to track windows in which the model can experience coexisting patterns. Upon analysis of the two-parameter charts presented in ﬁgure 1, it is evident that the diagrams in the left panel are generated by simultaneously increasing 5 Phys. Scr. 99 (2024) 015246 R K Subramanian et al Figure 4. Demarcation diagram showing the set of initial conditions related to each of the coexisting patterns of ﬁgure 3. The red domain is associated with chaotic patterns, blue with periodic patterns and yellow with unbounded motion. Figure 4 (a) is obtained ( ) for 0 while ﬁgure 4(b) is obtained for ( ) 0 ( ) 0 ( ) 0 = = = = u 0 0 x v y . both control parameters, whereas the diagrams in the right panel are derived by decreasing both control parameters. Moreover, it can be observed that the predominant behaviors depicted on these charts are periodic, as their occurrence area surpasses that of the chaotic behaviors. y v 0. = ( ) 0 ( ) 0 3.63 ] and ] and 15, 0 =k 10 and computing the ( )) when 0 15, 15 u0 , Î - [ Î - [ ] only chaotic and periodic patterns can captured while for 3.2. Coexisting oscillation modes Coexisting behaviors occur when, for the same set of system parameters, the model is able to display two or more completely different patterns by just changing the initial conditions. This phenomenon can be observed in the calcium oscillation model, as shown in ﬁgures 2 (a(i) and b(i)), along with the corresponding LLE graph (a(ii) and b(ii)). These ﬁgures. present two sets of data, with the blue set obtained by increasing the control parameter K10 and the red set obtained by decreasing the control parameters. Figure 2 clearly demonstrates that the diagrams in b (i, ii) are magniﬁcations of those in a (i, ii). Depending on the sweeping direction, the calcium oscillation model has two different routes to chaotic patterns, which are related to the hysteresis of the model. In that region of coexisting bifurcation, taking a discrete value of the parameters model starting from different initial conditions, as can be seen in the caption of ﬁgure 3, the model experiences the coexistence of the chaotic (red) and the periodic (blue) patterns (see ﬁgure 3). To obtain probability that any of the coexisting attractors can be found, the basins of attraction of ﬁgure 4 have been computed. Figure 4(a) in = ( ( ) obtained in the plane x As it can be seen on that demarcation diagram, for Î [ ( ) ( ) x 0 0, 15 u 0 Î - ( ) [ ( ) u 0 x 0 same line as it can be seen in ﬁgure 4(b) when varying y 0( ) between -5, 30 each of the three coexisting behavior occupy a speciﬁc domain of initial condition where it can ( ) x 0 be found. Therefore, the high sensitivity of the calcium oscillation model to the variation of the initial condition is demonstrated and validated by the previous investigation. In the work of the authors, such as Pankratova et al [34], the author investigated the bistability of the calcium oscillation. That bistability involves the coexistence of a steady state (point attractor) and an oscillatory mode. In contrast, our coexisting patterns found multistability since more than two patterns coexist in the model depending on the choice of the initial conditions. In our model, we have the coexistence of a steady state, two oscillatory modes (periodic and chaotic), and an unbounded region, which are well supported using the basin of attraction of ﬁgure 4. To the best of the author’s knowledge, there is no work related to the experimental observation of the multistability in the calcium oscillation, but concerning some neural circuits [35] and some nonlinear oscillators such as Chua [36] or Jerk [37], multistability has been revealed experimentally, and it has been found that the coexisting patterns were able to move randomly from one to another without any modiﬁcation of a parameter of the circuit only by switching ON and OFF the power supply. In addition, the sudden changing of patterns was related to environmental conditions such as noise temperatures or magnetic ﬁelds. Therefore, such multistability can also be studied experimentally in the calcium oscillation based on environmental conditions since it has already been proved numerically. ] only chaotic patterns and unbounded motion can be recorded. In the ] with ] and v 0( ) between -5, 20 15, 15 ( ) 0 = = u 0 [ [ 6 Phys. Scr. 99 (2024) 015246 4. Conclusion R K Subramanian et al This contribution was devoted to the further analysis of the calcium oscillation model obtained from hepatocyte cells. From the computation of the equilibrium points, it has been found that the model exhibits self-excited dynamics. Based on the numerical simulations exploiting the fourth-order Runge–Kutta algorithm, the global dynamics of the model under the variation of parameters have been revealed through two-parameter charts. From the hysteresis analysis, it has been found that model-supported multistability, characterized by the coexistence of up to three distinct dynamical states for the same set of parameters. So far, the result on the multistable dynamics involving more than two coexisting oscillatory modes in such a model has not yet been reported and represents an enriching contribution regarding the dynamics of such a model. Therefore, the effect of ﬁelds on such models as well as their network investigations will represent the scope of our future work. Acknowledgments This work was supported by the Centre for Nonlinear Systems, Chennai Institute of Technology (CIT), India via grant number CIT/CAI/2023/RP/012. Jan Awrejcewicz has been supported by the Polish National Science Centre under the Grant OPUS 18No.2019/ 35/B/ST8/00980. Data availability statement The data that support the ﬁndings of this study are available upon reasonable request from the authors. Declarations Conﬂict of interest The authors declare that they have no conﬂict of interest. ORCID iDs Zeric Tabekoueng Njitacke Jan Awrejcewicz https://orcid.org/0000-0003-0387-921X https://orcid.org/0000-0001-7797-8929 References [1] Pawar A and Pardasani K R 2023 Computational model of calcium dynamics-dependent dopamine regulation and dysregulation in a dopaminergic neuron cell The European Physical Journal Plus 138 30 [2] Jha B K, Adlakha N and Mehta M 2014 Two-dimensional ﬁnite element model to study calcium distribution in astrocytes in presence of excess buffer International Journal of Biomathematics 7 1450031 [3] Naik P A and Zu J 2020 Modeling and simulation of spatial-temporal calcium distribution in T lymphocyte cell by using a reaction- diffusion equation Journal of Bioinformatics and Computational Biology 18 2050013 [4] Panday S and Pardasani K R 2014 Finite element model to study the mechanics of calcium regulation in oocyte J. Mech. Med. Biol. 14 1450022 [5] Suansook Y and Paithoonwattanakij K 2015 Fractional dynamics in calcium oscillation model Mathematical Problems in Engineering 2015 276059 [6] Evans N H, McAinsh M R and Hetherington A M 2001 Calcium oscillations in higher plants Curr. Opin. Plant Biol. 4 415–20 [7] Dupont G et al 2011 Calcium oscillations Cold Spring Harb. Perspect. Biol. 3 a004226 [8] Lewis R 2003 Calcium oscillations in T-cells: mechanisms and consequences for gene expression Biochem. Soc. Trans. 31 925–9 [9] Sun S et al 2007 Physical manipulation of calcium oscillations facilitates osteodifferentiation of human mesenchymal stem cells FASEB J. 21 1472–80 [10] Shen P and Larter R 1995 Chaos in intracellular Ca2+ [11] Meyer T and Stryer L 1991 Calcium spiking Annual Review of Biophysics and Biophysical Chemistry 20 153–74 [12] Goldbeter A, Dupont G and Berridge M J 1990 Minimal model for signal-induced Ca2+ oscillations and for their frequency encoding oscillations in a new model for non-excitable cells Cell calcium 17 225–32 through protein phosphorylation Proc. Natl Acad. Sci. 87 1461–5 [13] Borghans J M, Dupont G and Goldbeter A 1997 Complex intracellular calcium oscillations A theoretical exploration of possible mechanisms Biophys. Chem. 66 25–41 [14] Kummer U et al 2000 Switching from simple to complex oscillations in calcium signaling Biophys. J. 79 1188–95 [15] Perc M and Marhl M 2003 Different types of bursting calcium oscillations in non-excitable cells Chaos, Solitons Fractals 18 759–73 7 Phys. Scr. 99 (2024) 015246 R K Subramanian et al [16] Gómez-Aguilar J et al 2019 Chaos in a calcium oscillation model via atangana-baleanu operator with strong memory Euro. Phys. J. Plus 134 140 [17] Duan P-F et al 2020 Time delay induces saw-tooth calcium wave and transition in intracellular calcium oscillation Chin. J. Phys. 66 150–6 [18] Duan W-L, Long F and Li C 2014 Reverse resonance and stochastic resonance in intracellular calcium oscillations Physica A 401 52–7 [19] Duan W-L 2014 Time delay induces oscillatory coherence in intracellular calcium oscillation system Physica A 405 10–6 [20] Fallah H 2016 Symmetric fold/super-Hopf bursting, chaos and mixed-mode oscillations in Pernarowski model of pancreatic beta-cells Int. J. Bifurcation Chaos 26 1630022 [21] Jiang W-A et al 2021 Exploiting bursting oscillations to improve energy capture from slowly changing excitation Journal of Vibration Engineering & Technologies 9 1923–39 [22] Zhang C, Ma X and Bi Q 2022 Complex mixed-mode oscillations based on a modiﬁed Rayleigh-Dufﬁng oscillator driven by low- frequency excitations Chaos, Solitons Fractals 160 112184 [23] Bao B et al 2018 AC-induced coexisting asymmetric bursters in the improved Hindmarsh–Rose model Nonlinear Dyn. 92 1695–706 [24] Bao H et al 2019 Hidden bursting ﬁrings and bifurcation mechanisms in memristive neuron model with threshold electromagnetic induction IEEE Trans Neural Netw. Learn. Syst. 31 502–11 [25] Njitacke Z T et al 2020 Hidden electrical activity of two neurons connected with an asymmetric electric coupling subject to electromagnetic induction: Coexistence of patterns and its analog implementation. Chaos, Solitons Fractals 137 109785 [26] Njitacke Z T et al 2021 Hamiltonian energy and coexistence of hidden ﬁring patterns from bidirectional coupling between two different neurons Cognitive Neurodynamics 1–18 [27] Fozin T F et al 2022 Coexistence of hyperchaos with chaos and its control in a diode-bridge memristor based MLC circuit with experimental validation Phys. Scr. 97 075204 [28] Njitacke Z T et al 2022 Complex dynamics and autapse-modulated information patterns in memristive Wilson neurons Nonlinear Dyn. 110 2793–804 [29] Falcke M 2004 Reading the patterns in living cells—the physics of Ca2+ [30] Morciano G et al 2022 Calcium dysregulation in heart diseases: targeting calcium channels to achieve a correct calcium homeostasis signaling Adv. Phys. 53 255–440 Pharmacol. Res. 177 106119 [31] Romito O et al 2022 Calcium signaling: a therapeutic target to overcome resistance to therapies in cancer Cell Calcium 108 102673 [32] Abbott L F 1999 Lapicque’s introduction of the integrate-and-ﬁre model neuron (1907)Brain Res. Bull. 50 303–4 [33] Wolf A et al 1985 Determining lyapunov exponents from a time series Physica. D 16 285–317 [34] Pankratova E V et al 2022 Bistability and chaos emergence in spontaneous dynamics of astrocytic calcium concentration Mathematics 10 1337 [35] Bao B et al 2023 Memristor-induced mode transitions and extreme multistability in a map-based neuron model Nonlinear Dyn. 111 3765–79 [36] Bao B et al 2016 Multistability in Chua’s circuit with two stable node-foci. Chaos 26 043111 [37] Kengne J, Njitacke Z and Fotsin H 2016 Dynamical analysis of a simple autonomous jerk system with multiple attractors Nonlinear Dyn. 83 751–65 8	null
10.1016_j.isci.2021.102204.pdf	Data and code availability The data that support the ﬁndings of this study and sequences were submitted to the NCBI sequence read archive (SRA) under BioProject ID: PRJNA633714. Code for analysis is available on github repository: https://github.com/galud27.	Data and code availability The data that support the findings of this study and sequences were submitted to the NCBI sequence read archive (SRA) under BioProject ID: PRJNA633714. Code for analysis is available on github repository: https://github.com/galud27 .	UC San Diego UC San Diego Previously Published Works Title Sensitivity of the mangrove-estuarine microbial community to aquaculture effluent Permalink https://escholarship.org/uc/item/3229890c Journal iScience, 24(3) ISSN 2589-0042 Authors Erazo, Natalia G Bowman, Jeff S Publication Date 2021-03-01 DOI 10.1016/j.isci.2021.102204 Copyright Information This work is made available under the terms of a Creative Commons Attribution- NonCommercial-NoDerivatives License, available at https://creativecommons.org/licenses/by-nc-nd/4.0/ Peer reviewed eScholarship.org Powered by the California Digital Library University of California iScience ll OPEN ACCESS Article Sensitivity of the mangrove-estuarine microbial community to aquaculture efﬂuent Natalia G. Erazo, Jeff S. Bowman [email protected] HIGHLIGHTS In near-intact mangrove forests, we observed the presence of nitrogen ﬁxers Calothrix could play a role in increasing nitrogen inventories via nitrogen ﬁxation Disturbed sites were correlated with increased nitrogen and reduction in diversity Disturbed sites were dominated by nitriﬁers, denitriﬁes, and sulfur- oxidizing bacteria Erazo & Bowman, iScience 24, 102204 March 19, 2021 ª 2021 The Authors. https://doi.org/10.1016/ j.isci.2021.102204 iScience ll OPEN ACCESS Article Sensitivity of the mangrove-estuarine microbial community to aquaculture efﬂuent Natalia G. Erazo1,3,4,* and Jeff S. Bowman1,2,3 SUMMARY Mangrove-dominated estuaries host a diverse microbial assemblage that facili- tates nutrient and carbon conversions and could play a vital role in maintaining ecosystem health. In this study, we used 16S rRNA gene analysis, metabolic infer- ence, nutrient concentrations, and d13C and d15N isotopes to evaluate the impact of land use change on near-shore biogeochemical cycles and microbial community structures within mangrove-dominated estuaries. Samples in close proximity to 3(cid:1); lower in mi- active shrimp aquaculture were high in NH4 crobial community and metabolic diversity; and dominated by putative nitriﬁers, denitriﬁes, and sulfur-oxidizing bacteria. Near intact mangrove forests we observed the presence of potential nitrogen ﬁxers of the genus Calothrix and or- der Rhizobiales. We identiﬁed possible indicators of aquaculture efﬂuents such as Pseudomonas balearica, Ponitmonas salivibrio, family Chromatiaceae, and genus Arcobacter. These results highlight the sensitivity of the estuarine-mangrove mi- crobial community, and their ecosystem functions, to land use changes. (cid:1), and PO4 +, NO3 (cid:1) NO2 INTRODUCTION Mangrove forests are among the most productive ecosystems in the world, harbor signiﬁcant biodiversity, and provide numerous ecosystem services (Ewel et al., 1998). These forests aid in the exchange of carbon and nutrients with the coastal marine environment (Robertson et al., 2011), with an estimated export of 10% of the marine dissolved organic matter to adjacent ecosystems (Dittmar and Lara, 2001). These forests act as carbon sinks by sequestering CO2, help stabilize coastlines, and support coastal ﬁsheries by acting as nursery grounds for a range of marine species (Kathiresan and Bingham, 2001). Despite their ecological and economic importance they have suffered severe losses in the past years (Duke et al., 2007). Although deforestation rates have declined (Friess et al., 2020), mangrove forests are still threatened by pollution, overextraction, conversion to aquaculture, agriculture, and the overall degradation of the environment (Lovelock et al., 2004; Reef et al., 2010; Friess et al., 2019). A key driver of the reduction in mangrove forest area is the expansion of shrimp aquaculture. Within Ecuador, the expansion of aquaculture exceeds the global trend with deforestation rates higher than 80% (Hamilton and Lovette, 2015). Here, shrimp aquaculture has grown to a $1.3 billion industry by 2012 and represents the second largest component of the Ecuadorian economy after fossil fuels (Hamilton and Lovette, 2015). Shrimp aquaculture efﬂuent is associated with the input of excess nutrients to adjacent coastal ecosystems; consequently, it can lead to changes in microbial community structure, biogeochemical cycles, and eutrophication (Maher et al., 2016; Rosentreter et al., 2018). Changes in nutrient ﬂuxes can indi- rectly alter the redox state of the water column and sediment. This can shift mangrove forests from acting as sinks to sources of greenhouse gases such as CO2, nitrous oxide, and methane (Maher et al., 2016). Microorganisms (here meaning single-celled members of the domains bacteria, archaea, and eukarya) are a key component of the mangrove forest and are present in the sediment, the water column, and as bioﬁlms on mangrove roots (Vazquez et al., 2000; Holguin et al., 2001). These microbes interact with mangroves as co-dependent ecosystem engineers and are responsible for many of the biogeochemical processes attrib- uted to mangrove forests (Holguin et al., 2006; Reis et al., 2017; Shiau and Chiu, 2020). Mangrove forest productivity, for example, is dependent on the microbial recycling mechanisms that keep nitrogen and 1Scripps Institution of Oceanography, UC San Diego, 8622 Kennel Way, La Jolla, CA 92037, USA 2Center for Microbiome Innovation, UC San Diego, La Jolla, CA, USA 3Center for Marine Biodiversity and Conservation, UC San Diego, La Jolla, CA, USA 4Lead contact Correspondence: [email protected] https://doi.org/10.1016/j.isci. 2021.102204 iScience 24, 102204, March 19, 2021 ª 2021 The Authors. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 1 ll OPEN ACCESS iScience Article Table 1. Environmental properties for high, intermediate, and low disturbed mangrove forests Disturbance Low Phosphate (mM)a 0.23 G 0.23 Nitrate+nitrite (mM)a 0.46 G 0.54 Ammonia (mM)a 0.39 G 0.36 Chlorophyll (mg L-1)a 11.52 G 5.86 Intermediate 0.33 G 0.33 0.87 G 0.46 1.77 G 0.60 8.80 G 2.58 High 2.41 G 1.01 9.91 G 8.75 12.79 G 7.50 p Valuef 9.10 3 1011 8.2 3 10 (cid:1)12 2.20 3 10 (cid:1)16 30.75 G 23.52 1.70 3 10 (cid:1)6 d13C (range)b (cid:1)18.45, (cid:1)27.76 (cid:1)18.49, (cid:1)29.00 (cid:1)27.01, (cid:1)32.08 – d15N (range)b 0.36, 11.08 0.54, 8.84 0.73, 5.86 – Samples (n)c, d, e 89 34 29 – aMean value. bLow and high values provided. cLow disturbance (Cayapas-Mataje = 88, Muisne = 1). dIntermediate disturbance (Cayapas-Mataje = 33, Muisne = 1). eHigh disturbance (Muisne = 29). fp Value (Kruskal-Wallis test). other nutrients within the system (Alongi, 1994). Because of the dependence of ecosystem functions on mi- crobes, microbes can be used as sensitive indicators of environmental change and stress. The planktonic microbial community in mangrove forests has been understudied when compared with the sediment community (Gomes et al., 2011; Imchen et al., 2017; Zhang et al., 2017; Gong et al., 2019). In this study, we evaluated the impact of land use change (mangrove forest converted to aquaculture) on micro- bial community structure and key biogeochemical parameters in the water column. We tested the hypoth- esis that shrimp aquaculture facilities are correlated with increased nitrogen inputs, altered microbial struc- ture, and alpha diversity. We identiﬁed speciﬁc microbial taxa that were differentially present between more and less perturbed sites associated with different levels of nutrient enrichment due to land use change. These taxa can be further developed as indicators of perturbation and mangrove forest health. The observed changes in the microbial community structure of the more and less disturbed sites high- lighted the sensitivity of the mangrove forest to aquaculture efﬂuent, with implications for coastal biogeo- chemical cycling and carbon and nitrogen subsidies to adjacent ecosystems. RESULTS Physicochemical properties (cid:1)1 for phosphate, 9.91 G 8.75 mmol L (cid:1)1 for nitrate + nitrite, 12.79 G 7.50 mmol L The disturbed sites (Muisne) were associated with higher levels of ammonia, nitrate + nitrite, phosphate, and chlorophyll a near aquaculture efﬂuent sites (Figure 1). The mean concentrations were 2.41 G (cid:1)1 for ammonia, 1.01 mmol L (cid:1)1 for chlorophyll a (Table 1 and Figure 2). These biogeochemical parameters were and 30.75 G 23.52 mg L (cid:1)6, respectively) in signiﬁcantly lower (Kruskal-Wallis test, p = 9.1 3 10 (cid:1)1 for phosphate, 0.46 G the low disturbance forest (Cayapas-Mataje) with values of 0.23 G 0.23 mmol L (cid:1)1 chlorophyll 0.54 mmol L a. Areas of intermediate disturbance were found around limited aquaculture facilities where the mean con- (cid:1)1 for nitrate + nitrite, 1.77 G centrations were 0.33 G 0.33 mmol L 0.60 mmol L (cid:1)1 for nitrate + nitrite, 0.39 G 0.36 mmol L (cid:1)1 for phosphate, 0.87 G 0.46 mmol L (cid:1)1 for ammonia, and 8.80 G 2.58 mg L (cid:1)1 for chlorophyll (Table 1, Figure 2). (cid:1)1 ammonia, and 11.52 G 5.86 mg L (cid:1)16, 1.7 3 10 (cid:1)12, 2.2 3 10 (cid:1)11, 8.2 3 10 C and N isotope values ranged from (cid:1)18.45 to (cid:1)27.76&d13C in the low disturbed sites, (cid:1)18.94 to (cid:1)29.00&d13C in the intermediate disturbed sites, and (cid:1)27.01 to (cid:1)32.08&d13C in the high disturbed sites (Table 1, Figure 2). The d15N values ranged from 0.36 to 11.08& in the low disturbed sites, 0.54 to 8.84& in the intermediate disturbed sites, and 0.73 to 5.86& in the high disturbed sites (Table 1, Figure 2). The N (cid:1)1; for low and intermediate distur- value for the high disturbed sites ranged from (cid:1)43.68 to (cid:1)4.44 mmol L (cid:1)1 (Figure 2). We identiﬁed higher N:P ratios associated bance sites it ranged from (cid:1)28.10 to 0.21 mmol L with high disturbance and lower ratios with low disturbance sites, and we observed a negative correlation (cid:1)8) and 16S rRNA gene copy number (Spearman’s with genome size (Spearman’s rho = (cid:1)0.46, p = 9.3 3 10 (cid:1)6) (Figure 2). The taxa most associated with smaller predicted genomes were Can- rho = (cid:1)0.5, p = 1.7 3 10 didatus Dependentiae (1.14 Mb), Candidatus Nasuia deltocephalinicola (1.12 Mb), and Candidatus Pelagi- bacter sp. IMCC9063 (1.28 Mb). The taxa most associated with larger predicted genomes were genera 2 iScience 24, 102204, March 19, 2021 ll OPEN ACCESS iScience Article A B C D Figure 1. Map of study site in coastal Ecuador (A) Study site in Esmeraldas, Ecuador, South America. (B) Location of the two ecological reserves: Cayapas-Mataje (CM) and Muisne (M). (C and D) (C) Map of land use changes in CM and (D) map of land use changes in M; green shows mangrove forest cover, pink shows shrimp aquaculture cover, and yellow circles show sampling locations. The base maps were generated from data obtained in Hamilton (2020). Calothrix (12.05 Mb), Oscillatoria acuminata (7.80 Mb), Moorea producens PAL-8-15-08-1 (9.71 Mb), San- daracinus amylolyticus (10.33 Mb), and Singulisphaera acidiphila (9.76 Mb). Alpha diversity For the bacterial community, the inverse Simpson’s indicator of diversity was signiﬁcantly lower in the highly disturbed sites when compared with the intermediate and low sites with mean G SD values of 36.08 G 26.41, (cid:1)9) (Figure 3). The mean diversity 30.05 G 17.56, and 56.73 G 19.82 respectively, (Kruskal-Wallis, p = 3.5 3 10 for the archaeal community was 5.00 G 1.34 for high, 6.38 G 2.29 for intermediate, and 6.85 G 2.87 for low distur- bance sites, and low and intermediate disturbance sites had signiﬁcant higher diversity than high disturbance (cid:1)6) (Figure 3). Alpha diversity for the archaeal community was lower than for sites (Kruskal-Wallis, p = 1.4 3 10 the bacterial community. Low disturbance sites had higher diversity than intermediate disturbance sites for the bacterial community, but no difference was observed between low and intermediate sites for the archaeal community (Figure 3). We also evaluated the predicted metabolic diversity for the bacterial community; the mean metabolic diversity for low disturbance was 244.01 G 8.72 (mean G SD); for intermediate disturbance, was 235.22 G 8.02; and for high disturbance, was 237.87 G 9.29. The low disturbance sites had higher metabolic diversity (Kruskal-Wallis, p = 2.2 3 10 (cid:1)7) when compared with intermediate and high disturbance sites. Differentiated abundance of bacterial and archaeal communities and metabolic pathways Unique reads are represented at the strain (closest completed genome or [CCG]) or clade level (closest estimated genome [CEG]) depending on the point of placement by paprica. The bacterial community iScience 24, 102204, March 19, 2021 3 ll OPEN ACCESS iScience Article A D B E C F Figure 2. Biogeochemical and bacterial signatures (A–C) (A) Nitrogen (ammonia and nitrate + nitrite) and phosphate species concentrations, (B) mean of genome size versus N:P ratio, (C) mean of number of 16S copies versus N:P ratio and Spearman’s correlation. (D and E) (D) N* value and (E) chlorophyll values of three levels of disturbance. Kruskal-Wallis test and p values with Dunn post-test. p < 0.01, p < 0.001. (F) d13C and d15N isotopic signatures. composition was dominated by the class Actinobacteria: Rhodoluna lacicola (CEG), Actinobacteria bacte- rium IMCC26256 (CCG), Acidimicrobium ferrooxidans DSM 10331 (CCG); family Pelagibacteraceae: Can- didatus Pelagibacter sp. IMCC9063 (CCG), Candidatus Puniceispirillum marinum IMCC1322 (CCG), Candi- family Flavobacteriaceae: Kordia sp. SMS9 (CCG), datus Pelagibacter ubique HTCC1062 (CCG); Owenweeksia hongkongensis DSM 17368 (CCG); cyanobacteria: Synechococcus sp. WH 7803 (CCG); and family Rhodobacteraceae: Thalassococcus sp. S3 (CCG) and Sulﬁtobacter sp. AM1-D1(CCG). The archaeal community was dominated by the most abundant class Thermoplasmata: Candidatus Methano- massiliicoccus intestinalis Issoire-Mx1 (CCG), class Methanococci: Methanococcales (CEG), and phylum Thaumarchaeota (Figure S1). Our DESeq2 results identiﬁed 333 amplicon sequence variants or ASVs that were signiﬁcantly different between sites separated by level of disturbance. Here we focus on the top 60 most abundant differentially present ASVs that were signiﬁcantly differentially present across our entire dataset (Figure 4). Members of Chromatiaceae bac- (cid:1)9), genus Delf- terium 2141T.STBD.0c.01a (CCG) (p = 2.02 3 10 (cid:1)7), Steroidobacter denitri- tia (CEG) (p = 1.03 3 10 (cid:1)8) were the most ﬁcans (CEG) (p = 6.17 3 10 signiﬁcantly most abundant taxa in the high disturbed site than in the low disturbed site. Cyanobacteria such as (cid:1)29), and M. producens PAL-8-15-08-1 (CCG) (p = 2.45 3 10 (cid:1)13), and Pseudomonas balearica DSM 6083 (CCG) (p = 2.52 3 10 (cid:1)14), Arcobacter nitroﬁgilis DSM 7299 (CCG) (p = 1.31 3 10 (cid:1)7), order Nostococales (CEG) (p = 7.84 3 10 (cid:1)19), family Planctomycetes (CEG) (p = 1.62 3 10 4 iScience 24, 102204, March 19, 2021 iScience Article ll OPEN ACCESS A B C Figure 3. Alpha diversity (A–C) (A) Bacterial community diversity, (B) archaeal diversity, and (C) metabolic diversity for the three levels of disturbance using InvSimpson metric. Kruskal- Wallis test and p values with Dunn post-test; *p < 0.001. Cyanobium gracile PCC 6307 (CCG) (p = 2.41 3 10 low disturbance sites were characterized by a higher abundance of SAR11 (CEG) (p = 2.17 3 10 dobacteraceae (CEG) (p = 2.30 3 10 yloceanibacter (CEG) (p = 1.42 3 10 (p = 9.12 3 10 othrix sp. NIES-4071 (CCG) (p = 1.79 3 10 (cid:1)38) were also more abundant in the high disturbed sites. The (cid:1)14), family Rho- (cid:1)40), family Flavobacteriaceae (CEG) (p = 8.26 3 10 (cid:1)28), and genus Meth- (cid:1)8). Oscillatoria species such as Oscillatoria nigroviridis PCC 7112 (CCG) (cid:1)7) were more abundant in the low and intermediate disturbed sites as well as cyanobacteria Cal- (cid:1)20) (Figure 4). For domain archaea we identiﬁed a total of seven (CEG) taxa that were the most abundant and differentially (cid:1)13) was associated with high disturbed samples. Candi- present. Candidatus Korarchaeota (p = 1.09 3 10 (cid:1)68), genus Meth- datus Mancarchaeum acidiphilum (p = 4.28 3 10 (cid:1)56) were more abundant anomassiliicoccus (p = 5.08 3 10 in low disturbed sites (Figure 4). (cid:1)28), and genus Methanococcales (p = 3.79 3 10 (cid:1)40), genus Nitrosopumilus (p = 2.92 3 10 A correspondence analysis (CA) of bacterial and archaeal community structures depicted the dissimilar relationship of samples for bacteria and archaea in terms of level of disturbance associated with aquacul- ture (Figure 5). For bacteria, the ﬁrst axis explained 30.6%, and the second axis, 18.3%. The top contributing taxa to the difference were Betaproteobacteria (cos2 = 0.86), Acidothermus cellulolyticus 11B (cos2 = 0.83), and S. denitriﬁcans (cos2 = 0.91) (Figure 5). For the archaeal community, the ﬁrst dimension accounted for 19.9% and the second dimension accounted for 11.8% of variability. Among the top contributors to the two dimensions were class Thermoplasmata (cos2 = 0.61), Candidatus Methanomassiliicoccus intestinalis (cos2 = 0.73), and Candidatus Mancarchaeum acidiphilum (cos2 = 0.63) (Figure 5). The results of our ANO- SIM test showed that the bacterial and archaeal communities were signiﬁcantly different for low and high disturbance mangrove forests (R = 0.52 and p value = 0.001, R = 0.45 and p value = 0.001). We also observed clear association of location of samples with ammonia concentration in dimension 1 (Spearman’s rho = (cid:1)9) for bacteria, and for archaea 0.56, p = 1.4 3 10 only dimension 2 showed a signiﬁcant correlation (Spearman’s rho = 0.49, p = 7.2 3 10 (cid:1)10) and dimension 2 (Spearman’s rho = 0.54, p = 1.2 3 10 (cid:1)5) (Figure S2). A canonical correspondence analysis (CCA) was further performed to examine the relationships between metabolic pathways and environmental factors. This showed that the biogeochemical parameters associ- ated with nitrogen species, phosphate, N:P, chlorophyll a, and d13C and d15N together accounted for 20% of the variability in the metabolic pathways. Nitrogen, phosphorus, and chlorophyll were factors that inﬂu- enced the metabolic pathways in the high disturbance sites. The ﬁrst dimension accounted for 26.1%, and the second dimension accounted for 9.1% of the variability. Here, the top contributors’ predicted meta- bolic pathways of dimethylsulfoniopropionate (DMSP) degradation III methylation (cos2 = 0.61) and glycine betaine (GBT) degradation I (cos2 = 0.62) were associated with low and intermediate disturbance. Taxa associated with DMSP degradation III methylation were Candidatus Puniceispirillum marinum IMCC1322 and Thalassococcus sp. S3, and for GBT degradation, taxa were Alphaproteobacterium HIMB59, cyano- bacteria, and Pelagibacteraceae. Other metabolic pathways with high contributions were arsenate iScience 24, 102204, March 19, 2021 5 ll OPEN ACCESS A B iScience Article Figure 4. Microbial and archaeal signatures of disturbance (A and B) (A) Differentially abundant bacterial taxa (top 60) in high, intermediate, and low disturbance result from DESeq2 analysis; (B) differentially abundant archaeal taxa result from DESeq2. Samples and taxa were clustered using Bray-Curtis dissimilarity distance. detoxiﬁcation (cos2 = 0.75) and methylphosphonate degradation (cos2 = 0.83), both associated with high disturbance sites (Figure 6). Taxa associated with these pathways were Erythrobacter atlanticus and Can- didatus Puniceispirillum marinum IMCC1322 for arsenate detoxiﬁcation and Starkeya novella DSM 506 (or- der Rizobiales) and Oceanicola sp. 3 for methylphosphonate degradation. Weighted gene correlation network analysis (cid:1)17, pink: r = 0.86, p = 2 3 10 (cid:1)50) Moorea producens PAL-8-15-08-01 (CCG) (r = 0.89, p = 5.18 3 10 Weighted gene correlation network analysis (WGCNA) found clusters of highly correlated taxa across samples. We related these clusters to ammonia and nitrate + nitrite to better understand the impact of aquaculture efﬂuent on microbial community structure. We identiﬁed eight major modules or subnetworks. Each module was as- signed a particular color (Figure S3). The blue and pink modules were positively correlated with ammonia and (cid:1)43). The yellow module was negatively nitrate + nitrite (blue: r = 0.64, p = 6.00 3 10 (cid:1)6) (Figure S3). Taxa associated with the pink correlated with ammonia, nitrate, and nitrite (r = (cid:1)0.42, p = 6.00 3 10 (cid:1)53), Actinobacteria bacterium module (Figure S4) included Sulfurivermis fontis (CEG) (r = 0.90, p = 1.45 3 10 (cid:1)53), Candidatus Methylopumilus planktonicus (CEG) (r = 0.89, p = IMCC26256 (CCG) (r = 0.90, p = 9.62 3 10 (cid:1)50), Phycisphaera mikurensis 1.98 3 10 (cid:1)30), and NBRC 102666 (r = 0.85, p = 1.19 3 10 (cid:1)30). All these taxa were signiﬁcantly correlated with Steroidobacter denitriﬁcans (CEG) (r = 0.79, p = 2.31 3 10 ammonia, and with nitrate + nitrite (Table 2). Taxa most strongly associated with the blue module consisted of (cid:1)8), A. cellulolyticus 11B (CCG) (r = 0.69, p = C. bacterium 2141T.STBD.0c.01a (CCG) (r = 0.49, p = 8.88 3 10 (cid:1)14), Pontimonas salivibrio (CEG) (r = 0.59, p = 3.37 3 10 (cid:1)20) (Table 2, Figure S4). The taxa 1.52 3 10 that were most negatively correlated with ammonia in the yellow module were: O. acuminata PCC 6304 (CCG) (cid:1)8), Candidatus pelagi- (cid:1)3, Synechococcus sp. WH 7803 (CCG) (r = (cid:1)0.49, p = 4.25 3 10 (r = (cid:1)0.34, p = 9.57 3 10 (cid:1)7), and Coraliomargarita akajimensis DSM 45221 (CCG) (r = bacter sp. IMCC9063 (CCG) (r = (cid:1)0.37, p = 1.17 3 10 (cid:1)0.36, p = 7.51 3 10 (cid:1)20), S. denitriﬁcans (CEG) (r = 0.61, p = 3.62 3 10 (cid:1)15), M. producens PAL-8-15-08-01 (CCG) (r = 0.69, p = 5.07 3 10 (cid:1)40), Cyanobium gracile PCC 6307 (CCG) (r = 0.78, p = 7.76 3 10 (cid:1)6) (Table 2, Figure S4). We further explored taxa that correlated with salinity to better understand the impact of tide on the micro- (cid:1)8) bial community structure. The red module was positively correlated with salinity (r = 0.47, p = 7.00 3 10 6 iScience 24, 102204, March 19, 2021 iScience Article A C B D ll OPEN ACCESS E Figure 5. Bacterial and archaeal community structure (A–E) Correspondence analysis (CA) ordination of (A) the bacterial community for samples that cluster in ordination space have similar community compositions, whereas those that are dispersed are less similar. (B) Square cosine components for samples; large value of cos2 shows a relatively large contribution to the total distance for bacterial community. (C) CA ordination for archaeal community. (D) Square cosine components for samples for archaeal community. (E) Contribution of top 10 taxa with highest cos2 values for archaeal community (see Figure S2). (Figure S3). Taxa associated with the red module included Acidimicrobium ferrooxidans DSM 10331 (CCG) (cid:1)9), Synechococcus sp. (cid:1)10), Haliglobus japonicus (CEG) (r = 0.52, p = 2.66 3 10 (r = 0.53, p = 8.19 3 10 (cid:1)8), Candidatus Puniceispirillum marinum IMCC1322 (CCG)) (r = CC9605 (CCG) (r = 0.50, p = 2.28 3 10 0.47, p = 4.56 3 10 (cid:1)7), and Prochlorococcus marinus str. MIT 9301 (r = 0.40, p = 1.98 3 10 (cid:1)4) (Table 3). We analyzed the Hellinger-transformed enzyme level output from paprica to better understand the enzy- matic potential of those CEG and CCG that were correlated with ammonia. We found 35 enzymes associ- ated with the nitrogen cycle (Figure 6). The enzyme nitrogenase EC 1.18.6.1 had a mean value of 0.23 G 0.05 for the low disturbed sites, signiﬁcantly higher than that in the intermediate (0.17 G 0.06) and high (cid:1)10) (Figure S5). Nitrate reductase EC 1.7.99.4 had a mean value (0.17 G 0.05) disturbance sites (p = 1.2 3 10 of 0.23 G 0.05 for low disturbance site, 0.45 G 0.14 for intermediate disturbance site, and 0.44 G 0.13 for high disturbance site, and the nitrate reductase value was signiﬁcantly higher in the high disturbance sites (cid:1)14). The same was observed with nitrate reductase NADH EC 1.7.1.4 with a mean of 0.13 G (p = 8.7 3 10 0.05 for low disturbed sites, 0.23 G 0.14 for intermediate disturbance, and 0.23 G 0.13 for highly disturbed (cid:1)15) (Figures 6 andS5). The taxa that were associated with nitrogenase were Methylocella sites (p = 2 3 10 silvestris BL2, genus Calothrix, and Synechococcus sp. CC9605. For nitrate reductase members of the Be- taproteobacteria, Desulfococcus oleovorans Hxd3, and P. mikurensis NBRC 102666 were found to contribute to enzyme abundance. The taxa that were associated with nitrate reductase NADH were A. cellulolyticus 11B and members of the Rhodobacteraceae (Table 4). DISCUSSION Mangrove forests are experiencing a high degree of perturbation through nutrient enrichment, pollution, and deforestation. Shrimp aquaculture efﬂuent in particular is associated with the input of excess nutrients to mangrove forests. In this study we found that shrimp aquaculture efﬂuent is associated with changes in microbial community structure with likely consequences for biogeochemical cycles and mangrove forest health. Previous work suggests that for intensive shrimp farming, 2.22 km2 of mangrove forest is required iScience 24, 102204, March 19, 2021 7 ll OPEN ACCESS A B iScience Article Figure 6. Metabolic pathways and nitrogen cycle enzyme indicators for levels of disturbance (A) CCA ordination for metabolic pathways showing top four pathways with cos2 ranging from 0.6–0.8. Large value of cos2 shows a relatively large contribution to the total distance for bacterial metabolic prediction. (B) Heatmap of key nitrogen cycle enzymes (Bray-Curtis distance) for the bacterial community (see Figure S5). to remove efﬂuent from one pond of 0.01 km2, whereas 0.20 km2 is required for less-intensive farming from one pond of 0.01 km2 (Robertson and Phillips, 1995). As of 2014 in the Muisne region there were 20.47 km2 of shrimp farms and 12.06 km2 of mangrove forests, indicative of an intensive farming system. Cayapas- Mataje had 11.04 km2 of shrimp aquaculture farms and 302.05 km2 of mangrove forest, suggesting less intensive farming (Figure 1) (Hamilton, 2020). As the areal extent of shrimp aquaculture increases so does the volume of the efﬂuent, elevating the ﬂux of ammonia and nitrate to the surrounding ecosystem. Based on our observations we found that microbial communities in mangrove forests are signiﬁcantly altered by this perturbation. The bacterial communities in our mangrove systems were characterized by members of the Pelagibacter- aceae, Flavobacteriaceae, Rhodobacteraceae, Actinobacteria, and cyanobacteria (Figure 4, Figure S1). The archaeal community was dominated by members of the Thermoplasmata, Thaumarchaeota, and Methanococcales (Figure 4, Figure S1). This was in accordance with other studies that have identiﬁed Rhodobacteraceae, SAR86 clade, Actinobacteria, and Flavobacteriaceae, and Thaumarchaeota as the most abundant taxonomic groups (Dhal et al., 2020). Rhodobacteraceae has been found to be dominant in mangrove-dominated estuaries, and members of this family are associated with marine phytoplankton blooms where they play a role in transformations of derived phytoplankton organic matter (Ghosh et al., 2010; Simon et al., 2017). The presence of Actinobacteria has been documented previously in mangrove ecosystems (Azman et al., 2015; Gong et al., 2019), and it has been suggested that they could play a role in carbon cycling by decomposing the plant biomass including refractory lignins (Scott et al., 2010). Thau- marchaeota are the most abundant archaea in the surface ocean (Santoro et al., 2015), and Thermoplas- mata have been found in mangrove ecosystems (Zhang et al., 2019). Both these groups play an important role in the nitrogen cycle by carrying out the oxidation of ammonia in nitriﬁcation (Santoro et al., 2015; Zhang et al., 2019). Both bacterial and archaeal communities were less diverse at our more disturbed sites. This pattern extended to predicted metabolic diversity (Figure 3). We hypothesize that this reduction in diversity could cause reductions in ecosystem functions. This has been observed in previous mangrove forest studies, for example, where lower microbial diversity was associated with a reduction in microbial productivity in sites with high levels of deforestation, sewage, and ﬁshing activities (Carugati et al., 2018). Our results showed differences in biogeochemical parameters between sites at varying levels of distur- bance (Figure 2). In particular, nitrogen was a driver of the microbial community structure leading to segre- gation into three clusters of disturbance in the CA analysis based on our ANOSIM test and signiﬁcantly 8 iScience 24, 102204, March 19, 2021 iScience Article ll OPEN ACCESS Table 2. Signiﬁcant correlated taxa with ammonia and nitrate+nitrite result from WGCNA Taxon Thermogutta terrifontisf Acidothermus cellulolyticus 11B Oceanicola sp. D3 Moorea producens PAL-8-15-08-1 Moorea producens PAL-8-15-08-1 Actinobacteria bacterium IMCC26256 Steroidobacter denitriﬁcans Aureitalea sp. RR4-38 Pontimonas salivibrio Pontimonas salivibrio Acidothermus cellulolyticus 11Bf Candidatus Xiphinematobacter sp. Idaho grape Synechococcus sp. CB0101 Candidatus Cyclonatronum proteinivorum Candidatus Cyclonatronum proteinivorum Halioglobus paciﬁcus Synechococcus sp. WH 8101 Rhodoluna lacicola Actinobacteria bacterium IMCC26256 Thiohalobacter thiocyanaticus Actinobacteria bacterium IMCC26256 Chromatiaceae bacterium 2141T.STBD.0c.01a Wenzhouxiangella marina Thiolapillus brandeum Candidatus Puniceispirillum marinum IMCC1322 Thermogutta terrifontis Candidatus Pelagibacter sp. IMCC9063 Sulfurivermis fontisd Actinobacteria bacterium IMCC26256 Candidatus Methylopumilus planktonicus Moorea producens PAL-8-15-08-1 Owenweeksia hongkongensis DSM 17368 Phycisphaera mikurensis NBRC 102666e Thermogutta terrifontis Candidatus Planktophila vernalis Actinobacteria bacterium IMCC26256 Candidatus Xiphinematobacter sp. Idaho grape Candidatus Pelagibacter sp. IMCC9063 Owenweeksia hongkongensis DSM 17368 Syntrophus aciditrophicus SB Candidatus Pelagibacter sp. IMCC9063 Phycisphaera mikurensis NBRC 102666e Map IDa Module color GS.Nitrogenb 0.87 0.94 0.97 0.82 0.82 0.88 0.91 0.92 0.98 0.98 0.94 0.88 0.98 0.87 0.87 565 0.98 0.98 0.88 0.92 0.88 0.95 0.98 0.94 0.97 0.87 0.91 0.86 0.88 0.96 0.82 0.89 0.8 0.87 0.95 0.88 0.88 0.91 0.89 0.84 0.92 0.8 Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Blue Pink Pink Pink Pink Pink Pink Pink Pink Pink Pink Pink Pink Pink Pink Pink 0.74 0.69 0.69 0.69 0.67 0.62 0.61 0.61 0.59 0.59 0.58 0.58 0.58 0.57 0.56 0.56 0.54 0.53 0.53 0.51 0.50 0.49 0.48 0.45 0.45 0.44 0.40 0.90 0.90 0.89 0.89 0.87 0.85 0.85 0.85 0.84 0.83 0.82 0.81 0.80 0.79 0.79 p.GS.Nitrogenc 6.35 3 10 3.37 3 10 (cid:1)25 (cid:1)20 4.17 3 10 5.07 3 10 2.62 3 10 8.57 3 10 3.62 3 10 5.70 3 10 5.75 3 10 1.52 3 10 1.29 3 10 (cid:1)20 (cid:1)20 (cid:1)18 (cid:1)15 (cid:1)14 (cid:1)14 (cid:1)13 (cid:1)15 (cid:1)12 1.80 3 10 (cid:1)12 3.36 3 10 9.00 3 10 2.06 3 10 2.19 3 10 3.96 3 10 (cid:1)12 (cid:1)12 (cid:1)11 (cid:1)11 (cid:1)10 1.58 3 10 1.60 3 10 (cid:1)9 (cid:1)9 1.94 3 10 9.01 3 10 (cid:1)12 (cid:1)12 8.88 3 10 1.48 3 10 2.19 3 10 3.59 3 10 (cid:1)8 (cid:1)7 (cid:1)6 (cid:1)6 4.27 3 10 1.29 3 10 (cid:1)6 (cid:1)4 1.45 3 10 9.62 3 10 (cid:1)53 (cid:1)53 1.98 3 10 5.18 3 10 4.75 3 10 1.19 3 10 2.24 3 10 9.07 3 10 1.34 3 10 (cid:1)50 (cid:1)50 (cid:1)46 (cid:1)40 (cid:1)40 (cid:1)40 (cid:1)39 7.44 3 10 (cid:1)37 1.31 3 10 (cid:1)34 3.97 3 10 3.58 3 10 1.73 3 10 1.89 3 10 (cid:1)33 (cid:1)32 (cid:1)31 (cid:1)30 (Continued on next page) iScience 24, 102204, March 19, 2021 9 iScience Article ll OPEN ACCESS Table 2. Continued Taxon Steroidobacter denitriﬁcans Cyanobium gracile PCC 6307 Cyanobium gracile PCC 6307 Halioglobus japonicus Halomicronema hongdechloris C2206 Thermogutta terrifontis Mariniﬁlaceae bacterium SPP2 Steroidobacter denitriﬁcans Aureitalea sp. RR4-38 Halioglobus paciﬁcus Synechococcus sp. WH 7803 Candidatus Methylopumilus planktonicus Flavobacteriaceae bacterium Thiolapillus brandeum Owenweeksia hongkongensis DSM 17368 Thermogutta terrifontis Candidatus Pelagibacter sp. IMCC9063 Coraliomargarita akajimensis DSM 45221 Oscillatoria acuminata PCC 6304 Acidothermus cellulolyticus 11Bf Map IDa Module color GS.Nitrogenb 0.91 0.98 0.98 0.91 0.9 0.87 0.85 0.91 0.92 0.94 0.99 0.96 0.91 0.94 0.89 0.87 0.92 0.89 0.81 0.94 Pink Pink Pink Pink Pink Pink Pink Pink Yellow Yellow Yellow Yellow Yellow Yellow Yellow Yellow Yellow Yellow Yellow Yellow 0.79 0.78 0.72 0.56 0.53 0.53 0.47 0.40 (cid:1)0.57 (cid:1)0.50 (cid:1)0.49 (cid:1)0.47 (cid:1)0.45 (cid:1)0.42 (cid:1)0.39 (cid:1)0.38 (cid:1)0.37 (cid:1)0.36 (cid:1)0.34 (cid:1)0.34 p.GS.Nitrogenc 2.31 3 10 (cid:1)30 7.76 3 10 2.75 3 10 2.91 3 10 7.85 3 10 9.21 3 10 3.13 3 10 1.07 3 10 4.73 3 10 1.58 3 10 4.25 3 10 4.55 3 10 2.36 3 10 3.22 3 10 2.45 3 10 7.07 3 10 1.17 3 10 2.84 3 10 6.15 3 10 9.57 3 10 (cid:1)30 (cid:1)23 (cid:1)11 (cid:1)10 (cid:1)10 (cid:1)7 (cid:1)4 (cid:1)12 (cid:1)8 (cid:1)8 (cid:1)7 (cid:1)6 (cid:1)5 (cid:1)4 (cid:1)4 (cid:1)3 (cid:1)3 (cid:1)3 (cid:1)3 aMap ID phylogenetic classiﬁcation. Value = 1 represents a perfect placement on the tree. bGS = Pearson correlation to ammonia and nitrate + nitrite. cp.GS = p-adjusted value (Bonferroni correction) for correlation to ammonia and nitrate+nitrite. dRepresents presence of nitrogenase enzyme EC.1.18.61. eRepresents presence of nitrate reductase enzyme EC.1.7.99.4. fRepresents presence of nitrate reductase enzyme EC.1.7.1.4. correlated with ammonia concentrations (Figure 5, Figure S2). We note that the variance explained by the ﬁrst and second dimensions in our ordination analyses is relatively low (30.6% and 19.9% for the bacterial and archaeal communities, respectively). We attribute this to the complexity associated with the mangrove ecosystem and the large number of physical, chemical, and biological factors that could impact changes in the microbial community. We found a strong connection between N:P ratio and genome size among planktonic bacteria across study sites (Figure 2). Generally, smaller predicted genomes and lower 16S rRNA gene copy number was asso- ciated with higher N:P ratios, whereas larger predicted genomes and higher 16S rRNA gene copy number was associated with lower N:P ratios. The differences in genome sizes between communities associated with different levels of disturbance suggest differing ecological strategies. Studies suggest that generalists possess larger genomes in contrast to the smaller genomes in more specialized microbes (Sriswasdi et al., 2017; Willis and Woodhouse, 2020). This falls from the generalist requirement for a larger gene repertoire to boost activity in multiple environmental conditions and to cope with different stressors associated with a broad physicochemical niche (such as low levels of nitrogen and tidal ﬂuctuations in mangrove-dominated estuaries). The low disturbance sites showed a higher metabolic diversity and larger genomes, which we interpret as a more generalist microbial community. Taxa with larger genomes included Planctomycetes such as Singulisphaera acidiphila. This taxon has been found in other wetland ecosystems (Kulichevskaya et al., 2008; Dedysh and Ivanova, 2019), and it has been shown to play an important role in degradation of plant-derived polymers such as pectin and xylan (Dedysh and Ivanova, 2019). The S. acidiphila genome en- codes several dozen proteins that do not belong to any of the currently carbohydrate-active enzymes, but the enzymes display a distant relationship to glycosyltransferases and carbohydrate esterases, suggesting that this taxon has a diverse glycolytic and carbohydrate metabolic potential (Dedysh and Ivanova, 2019). Other taxa included Sandaracinus amylolyticus. This taxon has been found in association with plant 10 iScience 24, 102204, March 19, 2021 iScience Article ll OPEN ACCESS Table 3. Signiﬁcant correlated taxa with salinity result from WGCNA Taxon Acidimicrobium ferrooxidans DSM 10331 Kordia sp. SMS9 Halioglobus japonicus Synechococcus sp. CC9605 Synechococcus sp. RCC307 Candidatus Puniceispirillum marinum IMCC1322 Salipiger profundus Halioglobus paciﬁcus Roseovarius mucosus Acidimicrobium ferrooxidans DSM 10331 Candidatus Pelagibacter ubique HTCC1062 Owenweeksia hongkongensis DSM 17368 Prochlorococcus marinus str. MIT 9301 Synechococcus sp. KORDI-100 Sulfurivermis fontis Map IDa Module color GS.Salinityb 0.94 0.91 0.93 1.00 1.00 0.98 0.82 0.95 0.96 0.82 1.00 0.89 1.00 1.00 0.87 Red Red Red Red Red Red Red Red Red Red Red Red Red Red Red 0.53 0.53 0.52 0.50 0.47 0.47 0.46 0.46 0.44 0.41 0.40 0.40 0.40 0.39 0.39 aMap ID phylogenetic classiﬁcation. Value = 1 represents a perfect placement on the tree. bGS = Pearson correlation to salinity. cp.GS = p-adjusted value (Bonferroni correction) for correlation to salinity. p.GS.Salinityc 8.19 3 10 1.49 3 10 (cid:1)10 (cid:1)9 2.66 3 10 2.28 3 10 4.36 3 10 4.56 3 10 8.37 3 10 1.02 3 10 7.04 3 10 9.08 3 10 1.10 3 10 1.64 3 10 (cid:1)9 (cid:1)8 (cid:1)7 (cid:1)7 (cid:1)7 (cid:1)6 (cid:1)6 (cid:1)5 (cid:1)4 (cid:1)4 1.98 3 10 2.07 3 10 3.93 3 10 (cid:1)4 (cid:1)4 (cid:1)4 residues (Mohr et al., 2012), in coral ecosystems (Rubio-Portillo et al., 2016), and it is known to survive in poor nutrient conditions by developing desiccation-resistant spores (Mohr et al., 2012). We also observed larger genomes in cyanobacteria including members of the genus Calothrix, genus Os- cillatoria, and M. producens PAL-8-15-08-1. Cyanobacteria are known to have large genomes with low cod- ing density and a high level of gene duplication; it has been proposed that the large non-protein-coding sequences contribute to the genome expansion and metabolic ﬂexibility observed in diazotrophs (nitrogen ﬁxers) that are associated with nitrogen-limited environments (Sargent et al., 2016). The high diversity of cyanobacteria observed in mangrove ecosystems suggests that they play a key role in the ecosystem. Rele- vant functions associated with cyanobacteria include nitrogen and carbon ﬁxation and the production of herbivory-defense molecules and plant growth-promoting substances (Alvarenga et al., 2015). In the disturbed sites the parasite C. Dependentiae accounted for much of the decrease in genome size. Studies have found that C. Dependentiae infects a wide range of protists, including heterotrophs and phytoplankton (Deeg et al., 2019). Other studies have shown that C. Dependentiae is associated with free-living ameba, suggesting that it could be an endosymbiont (Delafont et al., 2015). C. Dependentiae has very limited metabolic capability, lacks complete biosynthetic pathways for various essential cellular building blocks, and has protein motifs to facilitate eukaryotic host interactions (Yeoh et al., 2016; Deeg et al., 2019). C. Nasuia deltocephalinicola was also identiﬁed as having a small genome. C. Nasuia delto- cephalinicola is an obligate symbiont of plant phloem-feeding pest insects, and its main role is to provide essential amino acids that the host can neither synthesize nor obtain in sufﬁcient quantities from a plant diet (Bennett and Moran, 2013). The increase in the concentration of nitrogen species associated with aquaculture efﬂuent could further select for specialist microbes with reduced metabolic potential and lower diversity in nitrogen-processing enzymes. Our nitrogen isotope values are consistent with this, showing reduced variability for the highly disturbed sites (Table 1, Figure 2). Previous work in mangrove systems (Bernardino et al., 2018) associated reduced isotopic variability with a loss of trophic diversity. Higher variability in stable carbon isotopes has also been observed in salt marshes due to contribution dominated by allochthonous material derived from the phytoplankton community (Boschker et al., 1999). The larger variation in the isotopic signal observed in the low and intermediate disturbance sites suggests that these pristine systems contain a more diverse iScience 24, 102204, March 19, 2021 11 ll OPEN ACCESS iScience Article Table 4. Number of enzymes copies for nitrogenase and nitrate reductase enzymes and top 10 associated taxa Taxon Methylocella silvestris BL2 Genus Calothrix Synechococcus sp. CC9605 Family Rhodobacteraceae Oscillatoria nigroviridis PCC 7112 Acidothermus cellulolyticus 11B Phycisphaera mikurensis NBRC 102666 Rhodopirellula baltica SH 1 Desulfococcus oleovorans Hxd3 Class Betaproteobacteria Nitrogenase EC.1.18.6.1 Nitrate reductase EC.1.7.99.4 Nitrate reductase NADH EC.1.7.1.4 16,984 15,186 38,937 0 0 0 0 0 0 0 4,246 3,796 0 0 5,418 0 8,288 10,956 4,773 17,102 0 0 0 1,273 5,418 23,205 0 21,912 0 0 trophic food web as result of a wide range of metabolic and ﬁxation pathways, and environmental condi- tions in the mangrove-estuarine ecosystem (Boschker and Middelburg, 2002). The low N:P ratios we observed in the low disturbance sites suggest that the system is N limited. Pristine mangrove forests tend to be N limited, although nutrients are not uniformly distributed within the mangrove ecosystem and they can switch from N to P limitation. It has been shown that mangrove trees within fringe and tidally exposed zones tend to be N limited (Feller et al., 2003). One way mangrove trees cope with N limitation is through associations with diazotrophs that play a crucial role in N cycling within the mangrove forest (Holguin et al., 1992). Here we showed that the biological nitrogen ﬁxation signal, conﬁrmed by the N value (the linear combination of nitrate and phosphate that eliminates the effect of nitriﬁcation; thus, the remaining variability can be explained by nitrogen ﬁxation and denitriﬁcation) (Gruber and Sarmiento, 1997) and nitrogenase EC.1.18.6.1 abundance, were higher at low disturbance sites in contrast to high disturbance sites (Figures 2, 6, and S5). The microbial denitriﬁcation signal was further conﬁrmed by negative N* values in the highest disturbance sites (Figure 2) (Gruber and Sarmiento, 1997). Because excess nitrate is being introduced into the system via aquaculture efﬂuent, we expect denitriﬁca- tion rates to be high. Conversely, the lowest disturbance sites have a slight positive N* consistent with our identiﬁcation of putative diazotrophs such as genus Calothrix, genus Oscillatoria, and taxa of the order Rhi- zobiales such as M. silvestris (Essien et al., 2008; Liu et al., 2019). GBT degradation I was one of the major pathways contributing to the differences observed between low and high disturbed sites (Figures 6 and S3). GBT is an important source of nitrogen in oligotrophic systems, acts as an organic osmolyte, and plays an important role in phytoplankton-bacteria interactions (Becker et al., 2019; Jones et al., 2019; Zecher et al., 2020). The intertidal coastal mangrove ecosystem experiences daily ﬂuctuations in a range of environmental conditions, including water levels and salinity. Organisms living in this dynamic environment cope with changing environmental conditions by synthesizing a range of organic and inorganic osmolytes including GBT. The results from WGCNA showed that Pelagibactera- ceae taxa correlated with salinity (Table 3) and primary contributors of the GBT degradation I pathway. This suggests that osmolyte production is an important adaptation to salinity intrusions from oceanic waters into the mangrove environment, and GBT could be an additional pool of organic N within this system. Shrimp aquaculture impacts the water quality in adjacent mangrove forests by creating eutrophic condi- tions that can lead to anoxia. Eutrophic conditions were evident through high levels of nutrients and chlo- rophyll a (Figure 2, Table 1). Although we did not measure oxygen concentrations, we observed taxa indic- ative of hypoxic or anoxic conditions. These included purple sulfur bacteria (PSB), such as family Chromaticeae, and sulfur-oxidizing bacteria (SOB), such as genus Sulfurivermis (Figure 4, Table 2). PSB use sulﬁde, elemental sulfur, and thiosulfate as electron donors in anoxygenic photosynthesis and have been shown to play an important role in regime shifts from oxygenated to anoxic conditions (Diao et al., 2018). PSB ﬂourish in micro-aerobic conditions oxidizing sulﬁde into sulfate (Diao et al., 2018). As the oxy- gen inﬂux is reduced below a critical threshold, sulfate-reducing bacteria (SRB) and PSB can take over and outcompete the SOB. This suggests a more anoxic regime in the high disturbance site, allowing for PSB groups and SRB to become more abundant. 12 iScience 24, 102204, March 19, 2021 iScience Article ll OPEN ACCESS Based on our WGCNA analysis we also found nitrate-reducing bacteria (NRB)—indicative of reduced ox- ygen availability—that strongly correlated with the level of ammonia, nitrate, and nitrite. Putative NRB taxa included P. mikurensis and S. denitriﬁcans (Table 2). In addition, we also saw a microbial signature associated with dissimilatory nitrate reduction to ammonium (DNRA) with the presence of genus Acido- thermus, and anaerobic ammonium oxidation (annamox) with the presence of Planctomycetes (Thermo- gutta terrifontis) (Table 2); the presence of the genes involved in these pathways (denitriﬁcation, annamox, DNRA) were inferred by paprica, although further work is needed to conﬁrm the presence and activity of these enzymes. Overall, as nitrate and ammonia inputs increased with aquaculture efﬂuent the relative abundance of NRB increased. We identiﬁed speciﬁc microbes that can be used as sensitive indicators of aquaculture impacts. These included P. balearica (Figure 4), which has been associated with other contaminated wetland systems, sug- gesting that this taxon could be a potential bio-indicator of a disturbed mangrove ecosystem (Salva` -Serra et al., 2017). Similar studies have also identiﬁed aquaculture efﬂuent as a source of pathogens to the coastal ecosystem (Garren et al., 2009). In the disturbed site we saw the presence of members of the genus Arco- bacter (Figure 4). These bacteria have been identiﬁed in coral systems exposed to aquaculture efﬂuents, and have been associated with feces (human, porcine, and bovine) and with sewage-contaminated waters (Garren et al., 2009). PSB taxa such as family Chromaticeae have also been shown to be potential bio-in- dicators for anthropogenic contamination associated with other agriculture efﬂuent systems (Mohd-Nor et al., 2018). P. salivibrio in the order Micrococcales was elevated at the disturbed sites. This taxon has been isolated from high-salinity systems and aquaculture farms (Jang et al., 2013); high salinity levels have been associated with shrimp aquaculture efﬂuent due to high evaporation in the ponds (Barraza- Guardado et al., 2013). Previous studies have shown that taxa in the Micrococcales order are part of the core microbiome signal in shrimp ponds (Chen et al., 2017). Thus, P. salivibrio is a sensible indicator of shrimp aquaculture efﬂuent. Further work is needed to establish robust spatiotemporal baselines of micro- bial indicators of aquaculture to effectively monitor biogeochemical changes and health of the mangrove forests. Aquaculture could impact the health of mangrove ecosystems involving the direct loss of mangrove forests, efﬂuent associated with high levels of nutrients, and the development of anoxic and sulfuric water condi- tions (Robertson and Phillips, 1995). Aquaculture efﬂuent released into mangrove forests may be seques- tered and processed by bacteria. However, processing efﬁciency could change with increasing input. High organic loadings, for example, may shift the balance from aerobic to anaerobic systems (Lønborg et al., 2020). Anaerobic systems are less efﬁcient in nutrient cycling. The signals of SOB, SRB, denitriﬁers, and po- tential pathogenic taxa associated with the perturbed site suggest that aquaculture efﬂuent is playing a role in shifting the microbial community to a more pathogenic and less nutrient efﬁcient community that could impact the health of the mangrove forest. Conclusion In this study, we showed the impacts of aquaculture efﬂuent on the microbial community structure in mangrove forests and identiﬁed microbial signals associated with NRB, PSB, and SRB taxa that could have impacts in nutrient cycling. The high level of nutrients in the perturbed sites were associated with changes in microbial community structure that could impact ecosystem functions. In the low disturbance sites, we saw that the presence of Calothrix species and nitrogen ﬁxers could be important in increasing nitrogen inventories via nitrogen ﬁxation. Denitriﬁcation reduces excess inorganic nitrogen concentration, and in the highly disturbed sites we saw the presence of NRB-associated microbes. Nutrient cycling in mangrove habitats is a balance between nutrient inputs, availability, and internal cycling, and the changes in microbial community structure we see in disturbed sites could be indicators of biogeochemical changes. The results of the study highlight the sensitivity of the mangrove-estuarine microbial community to aqua- culture efﬂuent, and the impacts of land use changes could be ampliﬁed by climate change such as chang- ing precipitation patterns, heat, and rising sea level with severe consequences for the ecosystem. Limitations of the study Our analysis was based on comparison between sites of low, intermediate, and high disturbance in two mangrove systems in coastal Ecuador. Although ammonia concentration is a good proxy for disturbance from shrimp aquaculture efﬂuent, quantiﬁcation of land use changes, and the hydrological connections be- tween aquaculture facilities and our sampling sites was beyond the scope of the current work. We iScience 24, 102204, March 19, 2021 13 ll OPEN ACCESS iScience Article considered salinity, macronutrient concentrations, and isotopes in our analysis, but anticipate that other variables not considered here are contributing to differences in microbial community structure. These include physical processes such as tides and hydrology. The complexity of these environments is evident in our CCA and CA analyses, which capture a relatively small amount of variability in the ﬁrst two dimensions (see discussion section). Other limitations of note are typical of microbial community structure analyses. These include primer bias and a dependence on relative rather than absolute abundance. Resource availability Lead contact Further information and requests for resources should be directed to and will be fulﬁlled by the lead con- tact, Natalia Erazo ([email protected]). Materials availability This study did not generate new unique reagents. Data and code availability The data that support the ﬁndings of this study and sequences were submitted to the NCBI sequence read archive (SRA) under BioProject ID: PRJNA633714. Code for analysis is available on github repository: https://github.com/galud27. METHODS All methods can be found in the accompanying Transparent methods supplemental ﬁle. SUPPLEMENTAL INFORMATION Supplemental information can be found online at https://doi.org/10.1016/j.isci.2021.102204. ACKNOWLEDGMENTS N.G.E. was supported by Organization of American States and SENESCYT fellowships. J.S.B. was sup- ported by a grant from the Simons Foundation Early Career Investigator in Marine Microbiology program. We would like to thank Jesse Wilson and Avishek Dutta for helpful discussion on methods and interpreta- tion, Jesse Estacio for ﬁeld work assistance, and Fernando Rivera for assistance with logistics and permits. Samples were collected under the Ecuador environmental permit (MAE-UAFE-DPAE-2017-2009-E). We would like to thank the platform (mindthegraph.com) used here to create the graphical abstract. AUTHOR CONTRIBUTIONS Conceptualization, N.G.E. and J.S.B.; Methodology, N.G.E. and J.S.B; Investigation, N.G.E. and J.S.B; Writing – Original Draft, N.G.E.; Writing – Review & Editing, J.S.B.; Funding Acquisition and Supervision, J.S.B. DECLARATION OF INTERESTS The authors declare no competing interests. Received: July 30, 2020 Revised: September 25, 2020 Accepted: February 15, 2021 Published: March 19, 2021 REFERENCES Alongi, D.M. (1994). The role of bacteria in nutrient recycling in tropical mangrove and other coastal benthic ecosystems. Hydrobiologia 285, 19–32. Alvarenga, D.O., Rigonato, J., Branco, L.H.Z., and Fiore, M.F. (2015). Cyanobacteria in 14 iScience 24, 102204, March 19, 2021 mangrove ecosystems. Biodivers. Conserv. 24, 799–817. Azman, A.S., Othman, I., Velu, S.S., Chan, K.G., and Lee, L.H. (2015). Mangrove rare actinobacteria: taxonomy, natural compound, and discovery of bioactivity. Front.Microbiol. 6, 856. Barraza-Guardado, R.H., Arreola-Liza´ rraga, J.A., Lo´ pez-Torres, M.A., Casillas-Herna´ ndez, R., Miranda-Baeza, A., Magallo´ n-Barrajas, F., and Ibarra-Ga´ mez, C. (2013). Efﬂuents of shrimp farms and its inﬂuence on the coastal ecosystems of Bahı´a de Kino, Mexico. Sci.WorldJ. 2013, 306370. iScience Article Becker, J.W., Hogle, S.L., Rosendo, K., and Chisholm, S.W. (2019). Co-culture and biogeography of Prochlorococcus and SAR11. ISME J. 13, 1506–1519. Essien, J.P., Antai, S.P., and Benson, N.U. (2008). Microalgae biodiversity and biomass status in Qua Iboe Estuary mangrove swamp, Nigeria. Aquat. Ecol. 42, 71–81. Bennett, G.M., and Moran, N.A. (2013). Small, smaller, smallest: the origins and evolution of ancient dual symbioses in a phloem-feeding insect. Genome Biol. Evol. 5, 1675–1688. Ewel, K.C., Twilley, R.R., and Ong, J.E. (1998). Different kinds of mangrove forests provide different goods and services. Glob. Ecol. Biogeogr. Lett. 7, 83. Bernardino, A.F., Gomes, L.E.O., Hadlich, H.L., Andrades, R., and Correa, L.B. (2018). Mangrove clearing impacts on macrofaunal assemblages and benthic food webs in a tropical estuary. Mar. Pollut.Bull. 126, 228–235. Boschker, H.T.S., de Brouwer, J.F.C., and Cappenberg, T.E. (1999). The contribution of macrophyte-derived organic matter to microbial biomass in salt-marsh sediments: stable carbon isotope analysis of microbial biomarkers. Limnol. Oceanogr. 44, 309–319. Boschker, H.T., and Middelburg, J.J. (2002). Stable isotopes and biomarkers in microbial ecology. FEMS Microbiol. Ecol. 40, 85–95. Carugati, L., Gatto, B., Rastelli, E., Lo Martire, M., Coral, C., Greco, S., and Danovaro, R. (2018). Impact of mangrove forests degradation on biodiversity and ecosystem functioning. Sci. Rep. 8, 13298. Chen, W.-Y., Ng, T.H., Wu, J.H., Chen, J.W., and Wang, H.C. (2017). Microbiome dynamics in a shrimp grow-out pond with possible outbreak of acute hepatopancreatic necrosis disease. Sci. Rep. 7, 9395. Dedysh, S.N., and Ivanova, A.A. (2019). Planctomycetes in boreal and subarctic wetlands: diversity patterns and potential ecological functions. FEMS Microbiol. Ecol. 95, 227. Deeg, C.M., Zimmer, M.M., George, E.E., Husnik, F., Keeling, P.J., and Suttle, C.A. (2019). Chromulinavorax destructans, a pathogen of microzooplankton that provides a window into the enigmatic candidate phylum dependentiae. PLoS Pathog. 15, e1007801. Delafont, V., Samba-Louaka, A., Bouchon, D., Moulin, L., and He´ chard, Y. (2015). Shedding light on microbial dark matter: a TM6 bacterium as natural endosymbiont of a free-living amoeba. Environ.Microbiol. Rep. 7, 970–978. Dhal, P.K., Kopprio, G.A., and Ga¨ rdes, A. (2020). Insights on aquatic microbiome of the Indian Sundarbans mangrove areasJ.-S. Hwang, ed. 15, e0221543. Diao, M., Huisman, J., and Muyzer, G. (2018). Spatio-temporal dynamics of sulfur bacteria during oxic-anoxic regime shifts in a seasonally stratiﬁed lake. FEMS Microbiol. Ecol. 94, 40. Feller, I.C., McKee, K.L., Whigham, D.F., and O’Neill, J.P. (2003). Nitrogen vs. phosphorus limitation across an ecotonal gradient in a mangrove forest. Biogeochemistry 62, 145–175. Friess, D.A., Rogers, K., Lovelock, C.E., Krauss, K.W., Hamilton, S.E., Lee, S.Y., Lucas, R., Primavera, J., Rajkaran, A., and Shi, S. (2019). ‘The state of the world’s mangrove forests: past, present, and future’. Annu. Rev. Environ. Resour. 44, 89–115. Friess, D.A., Yando, E.S., Abuchahla, G.M.O., Adams, J.B., Cannicci, S., Canty, S.W.J., Cavanaugh, K.C., Connolly, R.M., Cormier, N., and Dahdouh-Guebas, F. (2020). Mangroves give cause for conservation optimism, for now. Curr. Biol. 30, R153–R154. Garren, M., Raymundo, L., Guest, J., Harvell, C.D., and Azam, F. (2009). Resilience of coral- associated bacterial communities exposed to ﬁsh farm efﬂuent. PLoS One 4, e7319. Ghosh, A., Dey, N., Bera, A., Tiwari, A., Sathyaniranjan, K.B., Chakrabarti, K., and Chattopadhyay, D. (2010). Culture independent molecular analysis of bacterial communities in the mangrove sediment of Sundarban, India. Saline Syst. 6, 1. Gomes, N.C., Cleary, D.F., Calado, R., and Costa, R. (2011). Mangrove bacterial richness. Commun.Integr. Biol. 4 (4), 419–423. Gong, B., Cao, H., Peng, C., Per(cid:1)culija, V., Tong, G., Fang, H., Wei, X., and Ouyang, S. (2019). High- throughput sequencing and analysis of microbial communities in the mangrove swamps along the coast of Beibu Gulf in Guangxi, China. Sci. Rep. 9, 9377. Gruber, N., and Sarmiento, J.L. (1997). Global patterns of marine nitrogen ﬁxation and denitriﬁcation. Glob.Biogeochem. Cycles 11, 235–266. Hamilton, S.E. (2020). Assessing 50 Years of Mangrove Forest Loss along the Paciﬁc Coast of Ecuador: A Remote Sensing Synthesis. Coastal Research Library (Springer), pp. 111–137. Hamilton, S.E., and Lovette, J. (2015). Ecuador’s mangrove forest carbon stocks: a spatiotemporal analysis of living carbon holdings and their depletion since the advent of commercial aquacultureB. Ruttenberg, ed. 10, e0118880. Dittmar, T., and Lara, R.J. (2001). Do mangroves rather than rivers provide nutrients to coastal environments south of the Amazon River? Evidence from long-term ﬂux measurements. Mar. Ecol. Prog.Ser. 213, 67–77. Holguin, G., Gonzalez-Zamorano, P., de-Bashan, L.E., Mendoza, R., Amador, E., and Bashan, Y. (2006). Mangrove health in an arid environment encroached by urban development-a case study. Sci. Total Environ. 363 (1–3), 260–274. Duke, N.C., Meynecke, J.-O., Dittmann, S., Ellison, A.M., Anger, K., Berger, U., Cannicci, S., Diele, K., Ewel, K.C., Field, C.D., et al. (2007). A world without mangroves? Science 317, 41b–42b. Holguin, G., Guzman, M.A., and Bashan, Y. (1992). Two new nitrogen-ﬁxing bacteria from the rhizosphere of mangrove trees: their isolation, identiﬁcation and in vitro interaction with ll OPEN ACCESS rhizosphere Staphylococcus sp. FEMS Microbiol. Lett. 101 (3), 207–216. Holguin, G., Vazquez, P., and Bashan, Y. (2001). The role of sediment microorganisms in the productivity, conservation, and rehabilitation of mangrove ecosystems: an overview. Biol. Fertil. Soils 33, 265–278. Imchen, M., Kumavath, R., Barh, D., Avezedo, V., Ghosh, P., Viana, M., and Wattam, A.R. (2017). Searching for signatures across microbial communities: metagenomic analysis of soil samples from mangrove and other ecosystems. Sci. Rep. 7, 1–13. Jang, G.I., Cho, Y., and Cho, B.C. (2013). Pontimonas salivibrio gen. nov., sp. nov., a new member of the family Microbacteriaceae isolated from a seawater reservoir of a solar saltern. Int. J. Syst. Evol.Microbiol. 63, 2124–2131. Jones, H.J., Kro¨ ber, E., Stephenson, J., Mausz, M.A., Jameson, E., Millard, A., Purdy, K.J., and Chen, Y. (2019). A new family of uncultivated bacteria involved in methanogenesis from the ubiquitous osmolyte glycine betaine in coastal saltmarsh sediments. Microbiome 7, 1–11. Kathiresan, K., and Bingham, B.L. (2001). Biology of mangroves and mangrove ecosystems. Adv. Mar. Biol. 40, 81–251. Kulichevskaya, I.S., Ivanova, A.O., Baulina, O.I., Bodelier, P.L., Damste´ , J.S., and Dedysh, S.N. (2008). Singulisphaera acidiphila gen. nov., sp. nov., a non-ﬁlamentous, Isosphaera-like planctomycete from acidic northern wetlands. Int. J. Syst. Evol.Microbiol. 58, 1186–1193. Liu, M., Huang, H., Bao, S., and Tong, Y. (2019). Microbial community structure of soils in Bamenwan mangrove wetland. Sci. Rep. 9, 8406. Lønborg, C., Carreira, C., Jickells, T., and A´ lvarez- Salgado, X.A. (2020). Impacts of global change on ocean dissolved organic carbon (DOC) cycling. Front. Mar. Sci. 7, 466. Lovelock, C.E., Feller, I.C., Mckee, K.L., Engelbrecht, B.M.J., and Ball, M.C. (2004). The effect of nutrient enrichment on growth, photosynthesis and hydraulic conductance of dwarf mangroves in Panama´ . Funct. Ecol. 18, 25–33. Maher, D.T., Sippo, J.Z., Tait, D.R., Holloway, C., and Santos, I.R. (2016). Pristine mangrove creek waters are a sink of nitrous oxide. Sci. Rep. 6, 25701. Mohd-Nor, D., Ramli, N., Sharuddin, S.S., Hassan, M.A., Mustapha, N.A., Amran, A., Sakai, K., Shirai, Y., and Maeda, T. (2018). Alcaligenaceae and Chromatiaceae as reliable bioindicators present in palm oil mill efﬂuent ﬁnal discharge treated by different biotreatment processes. Ecol. Indicators 95, 468–473. Mohr, K.I., Garcia, R.O., Gerth, K., Irschik, H., and Mu¨ ller, R. (2012). Sandaracinus amylolyticus gen. nov., sp. nov., a starch-degrading soil myxobacterium, and description of Sandaracinaceae fam. nov. Int. J. Syst. Evol.Microbiol. 62, 1191–1198. Reef, R., Feller, I.C., and Lovelock, C.E. (2010). Nutrition of mangroves. Tree Physiol. 30, 1148– 1160. iScience 24, 102204, March 19, 2021 15 ll OPEN ACCESS iScience Article Reis, C.R.G., Nardoto, G.B., and Oliveira, R.S. (2017). Global overview on nitrogen dynamics in mangroves and consequences of increasing nitrogen availability for these systems. Plant and Soil 410, 1–19. Robertson, A.I., Alongi, D.M., and Boto, K.G. (2011). Food Chains and Carbon Fluxes (American Geophysical Union (AGU)), pp. 293–326. Robertson, A.I., and Phillips, M.J. (1995). Mangroves as ﬁlters of shrimp pond efﬂuent: predictions and biogeochemical research needs. Hydrobiologia 295, 311–321. Rosentreter, J.A., Maher, D.T., Erler, D.V., Murray, R.H., and Eyre, B.D. (2018). ‘Methane emissions partially offset ‘‘blue carbon’’ burial in mangroves’. Sci. Adv. 4, eaao4985. Rubio-Portillo, E., Santos, F., Martı´nez-Garcı´a, M., de Los Rı´os, A., Ascaso, C., Souza-Egipsy, V., Ramos-Espla´ , A.A., and Anton, J. (2016). Structure and temporal dynamics of the bacterial communities associated to microhabitats of the coral O culina patagonica. Environ.Microbiol. 18, 4564–4578. Salva` -Serra, F., Jakobsson, H.E., Busquets, A., Gomila, M., Jae´ n-Luchoro, D., Seguı´, C., Aliaga- Lozano, F., Garcı´a-Valde´ s, E., Lalucat, J., Moore, E.R., and Bennasar-Figueras, A. (2017). Genome sequences of two naphthalene-degrading strains of Pseudomonas balearica, isolated from polluted marine sediment and from an oil reﬁnery site. Genome Announc 5, e00116–e00117. Santoro, A.E., Dupont, C.L., Richter, R.A., Craig, M.T., Carini, P., McIlvin, M.R., Yang, Y., Orsi, W.D., Moran, D.M., and Saito, M.A. (2015). ‘Genomic and proteomic characterization of ‘‘CandidatusNitrosopelagicus brevis’’: an ammonia-oxidizing archaeon from the open ocean’. Proc. Natl. Acad. Sci. USA 112, 1173– 1178. Sargent, E.C., Hitchcock, A., Johansson, S.A., Langlois, R., Moore, C.M., LaRoche, J., Poulton, A.J., and Bibby, T.S. (2016). Evidence for polyploidy in the globally important diazotroph Trichodesmium. FEMS Microbiol.Lett. 363, 244. Scott, J.J., Budsberg, K.J., Suen, G., Wixon, D.L., Balser, T.C., and Currie, C.R. (2010). Microbial community structure of leaf-cutter ant fungus gardens and refuse dumpsC.-H. Yang, ed. 5, e9922. Shiau, Y.J., and Chiu, C.Y. (2020). Biogeochemical processes of C and N in the soil of mangrove forest ecosystems. Forests 11, 492. Simon, M., Scheuner, C., Meier-Kolthoff, J.P., Brinkhoff, T., Wagner-Do¨ bler, I., Ulbrich, M., Klenk, H.P., Schomburg, D., Petersen, J., and Go¨ ker, M. (2017). Phylogenomics of Rhodobacteraceae reveals evolutionary adaptation to marine and non-marine habitats. ISME J. 11, 1483–1499. Sriswasdi, S., Yang, C.C., and Iwasaki, W. (2017). Generalist species drive microbial dispersion and evolution. Nat. Commun. 8, 1–8. Vazquez, P., Holguin, G., Puente, M.E., Lopez- Cortes, A., and Bashan, Y. (2000). Phosphate- solubilizing microorganisms associated with the rhizosphere of mangroves in a semiarid coastal lagoon. Biol. Fertil. Soils 30, 460–468. Willis, A., and Woodhouse, J.N. (2020). ‘Deﬁning cyanobacterial species: diversity and description through genomics. Crit. Rev. Plant Sci. 39, 101–124. Yeoh, Y.K., Sekiguchi, Y., Parks, D.H., and Hugenholtz, P. (2016). Comparative genomics of candidate phylum TM6 suggests that parasitism is widespread and ancestral in this lineage. Mol. Biol. Evol. 33, 915–927. Zecher, K., Hayes, K.R., and Philipp, B. (2020). Evidence of interdomain ammonium cross- feeding from methylamine- and Glycine betaine- degrading Rhodobacteraceae to diatoms as a widespread interaction in the marine phycosphere. Front.Microbiol. 11, 533894. Zhang, C.J., Pan, J., Duan, C.H., Wang, Y.M., Liu, Y., Sun, J., Zhou, H.C., Song, X., and Li, M. (2019). Prokaryotic diversity in mangrove sediments across southeastern China fundamentally differs from that in other biomes. mSystems 4, e00442- 19. Zhang, Y., Yang, Q., Ling, J., Van Nostrand, J.D., Shi, Z., Zhou, J., and Dong, J. (2017). Diversity and structure of diazotrophic communities in mangrove rhizosphere ,revealed by high- throughput sequencing. Front.Microbiol. 8, 2032. 16 iScience 24, 102204, March 19, 2021 iScience, Volume 24 Supplemental information Sensitivity of the mangrove-estuarine microbial community to aquaculture efﬂuent Natalia G. Erazo and Jeff S. Bowman A B Disturbance high intermediate low 1 0.8 0.6 0.4 0.2 0 Disturbance Coraliomargarita akajimensis DSM 45221 Candidatus Carsonella ruddii Candidatus Pelagibacter sp. IMCC9063 Thiolapillus brandeum Synechococcus sp. WH 7803 Synechococcaceae Candidatus Dependentiae Candidatus Pelagibacter ubique HTCC1062 Acidimicrobium ferrooxidans DSM 10331 Pelagibacteraceae Kordia sp. SMS9 Bacteria <prokaryotes> Flavobacteriaceae bacterium Candidatus Methylopumilus planktonicus Synechococcus sp. CC9605 Geminocystis Sulfurivermis fontis Cyanobacteria Cyanobacteria Candidatus Vesicomyosocius okutanii HA Chromatiaceae bacterium 2141T.STBD.0c.01a Actinobacteria bacterium IMCC26256 Thalassococcus Thalassococcus sp. S3 Celeribacter indicus Flavobacteriaceae Owenweeksia hongkongensis DSM 17368 Pelagibacteraceae Sulfitobacter sp. AM1−D1 Owenweeksia hongkongensis DSM 17368 Rhodoluna lacicola Pelagibacteraceae Candidatus Pelagibacter sp. IMCC9063 Candidatus Methylopumilus planktonicus Candidatus Puniceispirillum marinum IMCC1322 alpha proteobacterium HIMB59 Sulfurivermis fontis Roseovarius mucosus Halioglobus pacificus Acidimicrobium ferrooxidans DSM 10331 Disturbance high intermediate low 1 0.8 0.6 0.4 0.2 0 Disturbance (cid:38)(cid:68)(cid:81)(cid:71)(cid:76)(cid:71)(cid:68)(cid:87)(cid:88)(cid:86)(cid:3)(cid:49)(cid:76)(cid:87)(cid:85)(cid:82)(cid:86)(cid:82)(cid:83)(cid:88)(cid:80)(cid:76)(cid:79)(cid:88)(cid:86)(cid:3)(cid:68)(cid:71)(cid:85)(cid:76)(cid:68)(cid:87)(cid:76)(cid:70)(cid:88)(cid:86) (cid:55)(cid:75)(cid:68)(cid:88)(cid:80)(cid:68)(cid:85)(cid:70)(cid:75)(cid:68)(cid:72)(cid:82)(cid:87)(cid:68) (cid:39)(cid:76)(cid:68)(cid:73)(cid:82)(cid:85)(cid:68)(cid:85)(cid:70)(cid:75)(cid:68)(cid:72)(cid:68) (cid:55)(cid:75)(cid:68)(cid:88)(cid:80)(cid:68)(cid:85)(cid:70)(cid:75)(cid:68)(cid:72)(cid:82)(cid:87)(cid:68) (cid:55)(cid:36)(cid:38)(cid:46)(cid:3)(cid:74)(cid:85)(cid:82)(cid:88)(cid:83) (cid:48)(cid:72)(cid:87)(cid:75)(cid:68)(cid:81)(cid:82)(cid:80)(cid:76)(cid:70)(cid:85)(cid:82)(cid:69)(cid:76)(cid:68)(cid:79)(cid:72)(cid:86) (cid:55)(cid:36)(cid:38)(cid:46)(cid:3)(cid:74)(cid:85)(cid:82)(cid:88)(cid:83) (cid:55)(cid:36)(cid:38)(cid:46)(cid:3)(cid:74)(cid:85)(cid:82)(cid:88)(cid:83) (cid:55)(cid:36)(cid:38)(cid:46) (cid:74)(cid:85)(cid:82)(cid:88)(cid:83) (cid:55)(cid:36)(cid:38)(cid:46) (cid:74)(cid:85)(cid:82)(cid:88)(cid:83) (cid:55)(cid:36)(cid:38)(cid:46) (cid:74)(cid:85)(cid:82)(cid:88)(cid:83) (cid:55)(cid:75)(cid:68)(cid:88)(cid:80)(cid:68)(cid:85)(cid:70)(cid:75)(cid:68)(cid:72)(cid:82)(cid:87)(cid:68) (cid:55)(cid:36)(cid:38)(cid:46) (cid:74)(cid:85)(cid:82)(cid:88)(cid:83) (cid:38)(cid:68)(cid:81)(cid:71)(cid:76)(cid:71)(cid:68)(cid:87)(cid:88)(cid:86)(cid:3)(cid:49)(cid:76)(cid:87)(cid:85)(cid:82)(cid:86)(cid:82)(cid:83)(cid:88)(cid:80)(cid:76)(cid:79)(cid:88)(cid:86)(cid:3)(cid:68)(cid:71)(cid:85)(cid:76)(cid:68)(cid:87)(cid:76)(cid:70)(cid:88)(cid:86) (cid:38)(cid:68)(cid:81)(cid:71)(cid:76)(cid:71)(cid:68)(cid:87)(cid:88)(cid:86)(cid:3)(cid:48)(cid:72)(cid:87)(cid:75)(cid:68)(cid:81)(cid:82)(cid:80)(cid:68)(cid:86)(cid:86)(cid:76)(cid:79)(cid:76)(cid:76)(cid:70)(cid:82)(cid:70)(cid:70)(cid:88)(cid:86)(cid:3)(cid:76)(cid:81)(cid:87)(cid:72)(cid:86)(cid:87)(cid:76)(cid:81)(cid:68)(cid:79)(cid:76)(cid:86)(cid:3)(cid:44)(cid:86)(cid:86)(cid:82)(cid:76)(cid:85)(cid:72)(cid:16)(cid:48)(cid:91)(cid:20) (cid:55)(cid:36)(cid:38)(cid:46)(cid:3)(cid:74)(cid:85)(cid:82)(cid:88)(cid:83) (cid:38)(cid:68)(cid:81)(cid:71)(cid:76)(cid:71)(cid:68)(cid:87)(cid:88)(cid:86)(cid:3)(cid:49)(cid:76)(cid:87)(cid:85)(cid:82)(cid:86)(cid:82)(cid:83)(cid:88)(cid:80)(cid:76)(cid:79)(cid:88)(cid:86)(cid:3)(cid:68)(cid:71)(cid:85)(cid:76)(cid:68)(cid:87)(cid:76)(cid:70)(cid:88)(cid:86) (cid:3)(cid:48)(cid:72)(cid:87)(cid:75)(cid:68)(cid:81)(cid:82)(cid:80)(cid:76)(cid:70)(cid:85)(cid:82)(cid:69)(cid:76)(cid:68)(cid:79)(cid:72)(cid:86) (cid:38)(cid:68)(cid:81)(cid:71)(cid:76)(cid:71)(cid:68)(cid:87)(cid:88)(cid:86)(cid:3)(cid:49)(cid:76)(cid:87)(cid:85)(cid:82)(cid:86)(cid:82)(cid:83)(cid:88)(cid:80)(cid:76)(cid:79)(cid:88)(cid:86)(cid:3)(cid:68)(cid:71)(cid:85)(cid:76)(cid:68)(cid:87)(cid:76)(cid:70)(cid:88)(cid:86) (cid:38)(cid:68)(cid:81)(cid:71)(cid:76)(cid:71)(cid:68)(cid:87)(cid:88)(cid:86)(cid:17)(cid:49)(cid:76)(cid:87)(cid:85)(cid:82)(cid:86)(cid:82)(cid:83)(cid:88)(cid:80)(cid:76)(cid:79)(cid:88)(cid:86)(cid:3)(cid:68)(cid:71)(cid:85)(cid:76)(cid:68)(cid:87)(cid:76)(cid:70)(cid:88)(cid:86) (cid:55)(cid:36)(cid:38)(cid:46)(cid:3)(cid:74)(cid:85)(cid:82)(cid:88)(cid:83) (cid:38)(cid:68)(cid:81)(cid:71)(cid:76)(cid:71)(cid:68)(cid:87)(cid:88)(cid:86)(cid:17)(cid:48)(cid:72)(cid:87)(cid:75)(cid:68)(cid:81)(cid:82)(cid:80)(cid:68)(cid:86)(cid:86)(cid:76)(cid:79)(cid:76)(cid:76)(cid:70)(cid:82)(cid:70)(cid:70)(cid:88)(cid:86)(cid:17)(cid:76)(cid:81)(cid:87)(cid:72)(cid:86)(cid:87)(cid:76)(cid:81)(cid:68)(cid:79)(cid:76)(cid:86)(cid:17)(cid:44)(cid:86)(cid:86)(cid:82)(cid:76)(cid:85)(cid:72)(cid:17)(cid:48)(cid:91)(cid:20)(cid:17)(cid:20)(cid:3) (cid:48)(cid:72)(cid:87)(cid:75)(cid:68)(cid:81)(cid:82)(cid:70)(cid:82)(cid:70)(cid:70)(cid:68)(cid:79)(cid:72)(cid:86) (cid:38)(cid:68)(cid:81)(cid:71)(cid:76)(cid:71)(cid:68)(cid:87)(cid:88)(cid:86)(cid:17)(cid:48)(cid:72)(cid:87)(cid:75)(cid:68)(cid:81)(cid:82)(cid:80)(cid:68)(cid:86)(cid:86)(cid:76)(cid:79)(cid:76)(cid:76)(cid:70)(cid:82)(cid:70)(cid:70)(cid:88)(cid:86)(cid:17)(cid:76)(cid:81)(cid:87)(cid:72)(cid:86)(cid:87)(cid:76)(cid:81)(cid:68)(cid:79)(cid:76)(cid:86)(cid:17)(cid:44)(cid:86)(cid:86)(cid:82)(cid:76)(cid:85)(cid:72)(cid:17)(cid:48)(cid:91)1 Figure S1. Top abundant microbial community (bacteria and archaea). Heatmap for the most abundant bacteria (A) and archaea (B) taxa. Samples were clustered using Bray-Curtis dissimilarity distance and normalized (Hellinger transformation) abundance. Related to Figure 4. A R = 0.56, p = 1.4e−10 B 1.5 R = 0.54, p = 1.2e−09 ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 1 CA dimension 1 2 3 4 ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ●● ● ● R = 0.2, p = 0.075 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 1 0 ] M µ [ 4 H N −1 ● ● ● ● ● ● C 1.5 1.0 0.5 0.0 ] M µ [ 4 H N −0.5 ● ● −1.0 1.0 0.5 0.0 −0.5 −1.0 1.5 1.0 0.5 0.0 ] M µ [ 4 H N D ] M µ [ 4 H N ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● −4 −2 0 CA dimension 2 R = 0.49, p = 7.2e−05 ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● −0.5 −1.0 ● ● ● ● −2.0 −1.5 −1.0 CA dimension 1 ● ● −0.5 0.0 −6 −4 ● ● −2 CA dimension 2 0 Figure S2. Microbial community structure and association to disturbance levels. CA dimension 1 and dimension 2 vs ammonia concentrations for bacteria (A, B) and for archaea (C, D) (Spearman correlation). Related to Figure 5. (cid:48)(cid:82)(cid:71)(cid:88)(cid:79)(cid:72)(cid:239)(cid:87)(cid:85)(cid:68)(cid:76)(cid:87)(cid:3)(cid:85)(cid:72)(cid:79)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:86)(cid:75)(cid:76)(cid:83)(cid:86) green(cid:3) (cid:239)(cid:19)(cid:17)(cid:23) (cid:11)(cid:22)(cid:72)(cid:239)(cid:19)(cid:26)(cid:12) (cid:19)(cid:17)(cid:19)(cid:21)(cid:25) (cid:11)(cid:19)(cid:17)(cid:26)(cid:12) (cid:239)(cid:19)(cid:17)(cid:19)(cid:26)(cid:23) (cid:11)(cid:19)(cid:17)(cid:23)(cid:12) (cid:19)(cid:17)(cid:19)(cid:26)(cid:28) (cid:11)(cid:19)(cid:17)(cid:22)(cid:12) (cid:19)(cid:17)(cid:19)(cid:19)(cid:25)(cid:25) (cid:11)(cid:19)(cid:17)(cid:28)(cid:12) (cid:19)(cid:17)(cid:22)(cid:25) (cid:11)(cid:23)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) (cid:239)(cid:19)(cid:17)(cid:19)(cid:23)(cid:25) (cid:11)(cid:19)(cid:17)(cid:25)(cid:12) (cid:239)(cid:19)(cid:17)(cid:21) (cid:11)(cid:19)(cid:17)(cid:19)(cid:21)(cid:12) (cid:239)(cid:19)(cid:17)(cid:21)(cid:22) (cid:11)(cid:19)(cid:17)(cid:19)(cid:19)(cid:23)(cid:12) (cid:19)(cid:17)(cid:22)(cid:27) (cid:11)(cid:20)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) (cid:20) turquoise(cid:3) (cid:239)(cid:19)(cid:17)(cid:24)(cid:22) (cid:11)(cid:22)(cid:72)(cid:239)(cid:20)(cid:21)(cid:12) (cid:239)(cid:19)(cid:17)(cid:23)(cid:26) (cid:11)(cid:21)(cid:72)(cid:239)(cid:19)(cid:28)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:26) (cid:11)(cid:23)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:25) (cid:11)(cid:25)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:28) (cid:11)(cid:26)(cid:72)(cid:239)(cid:19)(cid:26)(cid:12) (cid:19)(cid:17)(cid:20)(cid:21) (cid:11)(cid:19)(cid:17)(cid:20)(cid:12) (cid:19)(cid:17)(cid:23)(cid:23) (cid:11)(cid:20)(cid:72)(cid:239)(cid:19)(cid:27)(cid:12) (cid:19)(cid:17)(cid:19)(cid:25)(cid:24) (cid:11)(cid:19)(cid:17)(cid:23)(cid:12) (cid:19)(cid:17)(cid:24)(cid:21) (cid:11)(cid:26)(cid:72)(cid:239)(cid:20)(cid:21)(cid:12) (cid:19)(cid:17)(cid:22)(cid:25) (cid:11)(cid:25)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) yellow(cid:3) (cid:239)(cid:19)(cid:17)(cid:24)(cid:23) (cid:11)(cid:20)(cid:72)(cid:239)(cid:20)(cid:21)(cid:12) (cid:239)(cid:19)(cid:17)(cid:24)(cid:20) (cid:11)(cid:21)(cid:72)(cid:239)(cid:20)(cid:20)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:24) (cid:11)(cid:27)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) (cid:239)(cid:19)(cid:17)(cid:23)(cid:21) (cid:11)(cid:27)(cid:72)(cid:239)(cid:19)(cid:27)(cid:12) (cid:239)(cid:19)(cid:17)(cid:23)(cid:21) (cid:11)(cid:27)(cid:72)(cid:239)(cid:19)(cid:27)(cid:12) (cid:19)(cid:17)(cid:21)(cid:20) (cid:11)(cid:19)(cid:17)(cid:19)(cid:19)(cid:27)(cid:12) (cid:19)(cid:17)(cid:24) (cid:11)(cid:23)(cid:72)(cid:239)(cid:20)(cid:20)(cid:12) (cid:19)(cid:17)(cid:19)(cid:22)(cid:21) (cid:11)(cid:19)(cid:17)(cid:26)(cid:12) (cid:19)(cid:17)(cid:22)(cid:22) (cid:11)(cid:22)(cid:72)(cid:239)(cid:19)(cid:24)(cid:12) (cid:19)(cid:17)(cid:20)(cid:27) (cid:11)(cid:19)(cid:17)(cid:19)(cid:22)(cid:12) k r o w t e n b u s / e l u d o M blue(cid:3) pink(cid:3) (cid:19)(cid:17)(cid:25)(cid:23) (cid:11)(cid:25)(cid:72)(cid:239)(cid:20)(cid:28)(cid:12) (cid:19)(cid:17)(cid:26)(cid:20) (cid:11)(cid:21)(cid:72)(cid:239)(cid:21)(cid:23)(cid:12) (cid:19)(cid:17)(cid:24)(cid:24) (cid:11)(cid:23)(cid:72)(cid:239)(cid:20)(cid:22)(cid:12) (cid:19)(cid:17)(cid:25)(cid:23) (cid:11)(cid:21)(cid:72)(cid:239)(cid:20)(cid:27)(cid:12) (cid:19)(cid:17)(cid:25)(cid:23) (cid:11)(cid:27)(cid:72)(cid:239)(cid:20)(cid:28)(cid:12) (cid:239)(cid:19)(cid:17)(cid:20)(cid:21) (cid:11)(cid:19)(cid:17)(cid:20)(cid:12) (cid:239)(cid:19)(cid:17)(cid:25)(cid:21) (cid:11)(cid:22)(cid:72)(cid:239)(cid:20)(cid:26)(cid:12) (cid:239)(cid:19)(cid:17)(cid:20)(cid:23) (cid:11)(cid:19)(cid:17)(cid:19)(cid:27)(cid:12) (cid:239)(cid:19)(cid:17)(cid:23)(cid:28) (cid:11)(cid:20)(cid:72)(cid:239)(cid:20)(cid:19)(cid:12) (cid:239)(cid:19)(cid:17)(cid:19)(cid:28)(cid:25) (cid:11)(cid:19)(cid:17)(cid:21)(cid:12) (cid:19)(cid:17)(cid:27)(cid:27) (cid:11)(cid:21)(cid:72)(cid:239)(cid:23)(cid:28)(cid:12) (cid:19)(cid:17)(cid:26)(cid:26) (cid:11)(cid:24)(cid:72)(cid:239)(cid:22)(cid:20)(cid:12) (cid:19)(cid:17)(cid:28)(cid:23) (cid:11)(cid:22)(cid:72)(cid:239)(cid:25)(cid:28)(cid:12) (cid:19)(cid:17)(cid:25)(cid:26) (cid:11)(cid:26)(cid:72)(cid:239)(cid:21)(cid:20)(cid:12) (cid:19)(cid:17)(cid:27)(cid:25) (cid:11)(cid:21)(cid:72)(cid:239)(cid:23)(cid:24)(cid:12) (cid:19)(cid:17)(cid:20)(cid:25) (cid:11)(cid:19)(cid:17)(cid:19)(cid:24)(cid:12) (cid:239)(cid:19)(cid:17)(cid:23)(cid:25) (cid:11)(cid:23)(cid:72)(cid:239)(cid:19)(cid:28)(cid:12) (cid:239)(cid:19)(cid:17)(cid:20)(cid:26) (cid:11)(cid:19)(cid:17)(cid:19)(cid:23)(cid:12) (cid:239)(cid:19)(cid:17)(cid:25)(cid:25) (cid:11)(cid:23)(cid:72)(cid:239)(cid:21)(cid:19)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:25) (cid:11)(cid:24)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) brown(cid:3) (cid:19)(cid:17)(cid:20)(cid:22) (cid:11)(cid:19)(cid:17)(cid:20)(cid:12) (cid:239)(cid:19)(cid:17)(cid:21)(cid:20) (cid:11)(cid:19)(cid:17)(cid:19)(cid:19)(cid:28)(cid:12) (cid:239)(cid:19)(cid:17)(cid:20)(cid:27) (cid:11)(cid:19)(cid:17)(cid:19)(cid:22)(cid:12) (cid:239)(cid:19)(cid:17)(cid:21)(cid:26) (cid:11)(cid:26)(cid:72)(cid:239)(cid:19)(cid:23)(cid:12) (cid:239)(cid:19)(cid:17)(cid:21)(cid:24) (cid:11)(cid:19)(cid:17)(cid:19)(cid:19)(cid:21)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:21) (cid:11)(cid:26)(cid:72)(cid:239)(cid:19)(cid:24)(cid:12) (cid:19)(cid:17)(cid:20)(cid:20) (cid:11)(cid:19)(cid:17)(cid:21)(cid:12) (cid:19)(cid:17)(cid:21)(cid:21) (cid:11)(cid:19)(cid:17)(cid:19)(cid:19)(cid:25)(cid:12) (cid:19)(cid:17)(cid:20)(cid:27) (cid:11)(cid:19)(cid:17)(cid:19)(cid:22)(cid:12) (cid:239)(cid:19)(cid:17)(cid:25)(cid:28) (cid:11)(cid:21)(cid:72)(cid:239)(cid:21)(cid:21)(cid:12) black(cid:3) (cid:239)(cid:19)(cid:17)(cid:20)(cid:22) (cid:11)(cid:19)(cid:17)(cid:20)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:26) (cid:11)(cid:22)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:21) (cid:11)(cid:25)(cid:72)(cid:239)(cid:19)(cid:24)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:27) (cid:11)(cid:21)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:27) (cid:11)(cid:21)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:24) (cid:11)(cid:27)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) (cid:19)(cid:17)(cid:21)(cid:25) (cid:11)(cid:19)(cid:17)(cid:19)(cid:19)(cid:20)(cid:12) (cid:19)(cid:17)(cid:21)(cid:25) (cid:11)(cid:19)(cid:17)(cid:19)(cid:19)(cid:20)(cid:12) (cid:19)(cid:17)(cid:24)(cid:23) (cid:11)(cid:28)(cid:72)(cid:239)(cid:20)(cid:22)(cid:12) (cid:239)(cid:19)(cid:17)(cid:19)(cid:19)(cid:20)(cid:21) (cid:11)(cid:20)(cid:12) (cid:19)(cid:17)(cid:24) (cid:19) (cid:239)(cid:19)(cid:17)(cid:24) red (cid:239)(cid:19)(cid:17)(cid:21)(cid:25) (cid:11)(cid:19)(cid:17)(cid:19)(cid:19)(cid:20)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:21) (cid:11)(cid:26)(cid:72)(cid:239)(cid:19)(cid:24)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:21) (cid:11)(cid:24)(cid:72)(cid:239)(cid:19)(cid:24)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:20) (cid:11)(cid:20)(cid:72)(cid:239)(cid:19)(cid:23)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:23) (cid:11)(cid:21)(cid:72)(cid:239)(cid:19)(cid:24)(cid:12) (cid:239)(cid:19)(cid:17)(cid:22)(cid:25) (cid:11)(cid:24)(cid:72)(cid:239)(cid:19)(cid:25)(cid:12) (cid:19)(cid:17)(cid:21) (cid:11)(cid:19)(cid:17)(cid:19)(cid:20)(cid:12) (cid:19)(cid:17)(cid:20)(cid:24) (cid:11)(cid:19)(cid:17)(cid:19)(cid:25)(cid:12) (cid:19)(cid:17)(cid:24)(cid:24) (cid:11)(cid:22)(cid:72)(cid:239)(cid:20)(cid:22)(cid:12) (cid:19)(cid:17)(cid:23)(cid:26) (cid:11)(cid:28)(cid:72)(cid:239)(cid:20)(cid:19)(cid:12) (cid:239)(cid:20) Chl Phosphate Nitrate+Nitrite Total Nitrogen Ammonia N:P (cid:49)(cid:13) 15N 13C Salinity Figure S3. Microbial community and environmental variables. Weighted Gene Correlation Network Analysis (WGCNA) was used to identify subnetworks (or modules) of bacteria that correlated with environmental variables. Pearson correlation coefficients for subnetworks that were significantly correlated with environmental variables are shown in the top number (ρ value) and the number in the parentheses is the p-value for each relationship. Positive relationship is in red and negative relationship is in blue. Related to Table 2 & 3. Module membership vs. Taxa significance cor=0.73, p=5.6e−10 A n e g o r t i N r o f e c n a c i i f i n g s a x a T 8 . 0 6 . 0 4 . 0 2 . 0 0 . 0 0.2 0.4 0.6 0.8 Module Membership in blue module C Module membership vs. Taxa significance cor=0.53, p=6.3e−05 n e g o r t i N r o f e c n a c i f i n g i s a x a T 5 . 0 4 . 0 3 . 0 2 . 0 1 . 0 0 . 0 B n e g o r t i N r o f e c n a c i f i n g s i a x a T y t i n i l a S r o f e c n a c i f i n g s a x a T i Module membership vs. Taxa significance cor=0.87, p=4.9e−11 9 0 . 7 0 . 5 0 . 3 0 . 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Module Membership in pink module D Module membership vs. Taxa significance cor=0.38, p=0.013 5 . 0 4 . 0 3 . 0 2 . 0 1 . 0 0.2 0.4 0.6 0.8 0.4 0.5 0.6 0.7 0.8 0.9 Module Membership in yellow module Module Membership in red module Figure S4. WGCNA modules. Module membership of taxa in the blue (A), pink (B) and yellow (C) subnetworks (or modules) which strongly correlated with ammonia and nitrate + nitrite. Module membership of taxa in red subnetwork (D) which strongly correlated with salinity. Related to Table 2 & 3. A B Kruskal−Wallis, p = 1.2e−10 * * 0.4 0.3 0.2 . . . . . 1 6 8 1 1 E e s a n e g o r t i N 0.1 low intermediate high . . . . . 4 9 9 7 1 E e s a t c u d e r . e t a r t i N . 4 . 1 . 7 . 1 . E H D A N . e s a t c u d e r . e t i r t i N 0.8 0.6 0.4 0.2 0.5 0.4 0.3 0.2 0.1 0.0 Kruskal−Wallis, p = 8.7e−14 * * low intermediate high Kruskal−Wallis, p = 2e−15 * * low intermediate ● high Figure S5. Metabolic pathways. (A) Contribution of top taxa from CCA ordination analysis and cos2 values. (B) Nitrogenase EC 1.18.6.1, Nitrate reductase EC 1.7.99.4 and Nitrite reductase NADH EC 1.7.1.4 normalized (Hellinger transformation) abundance. Kruskal-Wallis test and p-values with Dunn post-test, **denotes p-value<0.001. Related to Figure 6. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Supplementary Information Transparent Methods Study sites, sample collection, and physiochemical parameter measurements The study was conducted in two ecological reserves along the coast of Ecuador (Fig. 1). The Cayapas-Mataje Ecological Reserve, located within Esmeraldas province along the Colombian border (1° 17’ 02.14’’ N, 78° 54’ 22.29’’ W), encompasses 302.05 km2 of largely non-disturbed mangrove forests. This reserve is located in the delta formed by the estuary of the Cayapas- Santiago-Mataje rivers, and it is part of what used to be a continuous mangrove belt that ranged from the central area of the Colombia Pacific coast to the south area of Esmeraldas in Ecuador. Cayapas-Mataje is considered one of the most pristine mangrove ecosystems along the Pacific coast of the Americas (Hamilton, 2020a). The dominant mangrove species is Rhizophora mangle, representing 98% of all the mangrove area (Hamilton, 2020a). Traditional uses, such as artisanal fishing and cockle picking are still practiced, and only 2% of mangrove forest area loss is attributed to aquaculture (Hamilton, 2020b). The Muisne Ecological Reserve, also located in the province of Esmeraldas (0° 36’ 41.81’’ N, −80° 1’ 14.36’’ W), is highly perturbed by aquaculture (Fig. 1). The site compromises the delta of the Muisne River and numerous smaller rivers and contains a total of 12.06 km2 of mangrove forests. The species composition is 71% Rhizophora mangle, 1% Avicennia germinans, and 28% Languncularia racemose (Hamilton, 2020a). Muisne has been severely impacted by shrimp aquaculture, accounting for 36% of mangrove loss. Only 1% of the remaining mangrove forest is protected (Hamilton, 2020b). 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 Water samples were taken from the surface (0.5 m depth) along a proximity gradient to the mangrove trees. Samples were grouped by level of disturbance based on the concentration of ammonia in the water column: Low = < 1 μM, Intermediate = 1 – 3 μM, High = > 3 μM. Similar ammonia ranges have been identified in previous studies exposed to aquaculture effluent (Robertson and Alongi, 1992); however, reported values in the literature can vary depending on spatial parameters and aquaculture land expansion (Cifuentes et al., 1996; Barraza-Guardado et al., 2013a, Samocha et al., 2004). Here we also take into account the area of shrimp aquaculture in the two ecological sites. Muisne was identified as highly disturbed, and all the samples were taken near shrimp aquaculture facilities (N = 29) with high levels of ammonia with the exception of two samples near the mouth of the estuarine. The site has 20.47 km2 of shrimp farms and 12.06 km2 of mangrove forests for an approximate 2:1 ratio of aquaculture to mangrove forest (Hamilton, 2020b). Cayapas-Mataje has 11.04 km2 of shrimp aquaculture farms and 302.05 km2 of mangrove forest for a 1:27 ratio of aquaculture to mangrove forest (Fig. 1) (Hamilton, 2020b). Thus, samples that were collected along mangrove forests in Cayapas-Mataje (no presence of aquaculture) were characterized as a low disturbance with lower levels of ammonia, and we included one sample from Muisne with low level of ammonia (N = 89). Within Cayapas-Mataje, there’s a smaller presence of shrimp aquaculture facilities and the samples that were collected near the shrimp facilities were classified as intermediate disturbance with intermediate levels of ammonia in addition to one sample from Muisne with intermediate ammonia (N = 34). For DNA samples, approximately 400 ml of water was filtered through a sterile 47 mm 0.2 µm Supor filter (Pall) directly from 0.5 m depth using a peristaltic pump. The filter was immediately stored on ice and transferred to long term storage at –80 °C within 8 hours. Chlorophyll a concentration was measured with an Aquaflash handheld active fluorometer (Turner 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 Designs) following the manufacturer’s instructions. Temperature, salinity, and turbidity were measured using a YSI ProDss (Xylem). For nutrient analysis, 50 ml of water was filtered through a combusted GF/F filter (Whatman), frozen immediately after collection, and stored at –80 °C. Samples were sent to the UC Santa Barbara Marine Institute and analyzed by flow injection analysis following standard protocols ( Lachat instrument methods: 31-107-04-1A, 31-107-06-5A, 31-115-01-3A). For CHN and isotope analysis, 50 ml of water was filtered through a combusted GF/F filter, and filters were wrapped into a tin envelope (Costech). Samples were analyzed by EA-IRMS at the Scripps Institute of Oceanography Isotope Facility yielding percent carbon and nitrogen by mass, as well as δ13C and δ15N following standard methods (Pestle, Crowley and Weirauch, 2014). The reference materials used were NBS-19 and NBS-18, and IAEA N1 and the analytical precisions were +/- 0.3 to 0.5 for C and 0.7 to 1.0 for N. The standards used for δ13C and δ15N calculations were the Pee Dee Belemnite and atmospheric N2, respectively. DNA extraction and sequencing DNA was extracted using the DNAeasy PowerWater DNA extraction kit (Qiagen). Extracted DNA was quantified using the Qubit HS DNA quantification kit (Invitrogen) and then quality checked by gel electrophoresis and PCR amplification of the 16S rRNA gene using primers 515F and 806R (Walters et al., 2015) for bacteria and archaea. High quality extracted DNA was submitted to the Argonne National Laboratory sequencing center for amplification and library preparation with the same primer set, followed by 2 x 151 paired-end sequenced on the Illumina Miseq platform. Sequences were submitted to NCBI Bio project accession number: PRJNA633714. 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 Sequence analysis Illumina Miseq reads were demultiplexed using the ‘iu-demultiplex’ command in Illumina utils. Demultiplexed reads were quality controlled and denoised using the ‘FilterandTrim’ and ‘dada’ commands within the R package dada2 (Benjamin J Callahan et al., 2016), and assembled with the ‘mergePairs’ command. The final merged reads had mean quality scores >30, and the non-redundant fasta files of the generated unique reads produced by dada2 were used as an input for the paprica pipeline for microbial community structure and metabolic inference (https://github.com/bowmanjeffs/paprica). The paprica method for determining microbial community structure differs from OTU clustering methods in that it relies on the placement of reads on a phylogenetic tree created from the 16S rRNA gene reads from all completed bacterial and archaeal genomes in Genbank (Bowman and Ducklow, 2015). Because the metabolic potential of each phylogenetic edge on the reference tree is known, paprica generates a reasonable estimate of genome sizes, gene content, and metabolic pathways for the organisms of origin of each read. To estimate metabolic potential, a phylogenetic tree of the 16S rRNA genes from each completed genome was generated. For each internal node on the reference tree we determined a “consensus genome”, defined as all genomes shared by all members of the clade originating from the node, and predict the metabolic pathways present in the consensus and complete genomes (Bowman and Ducklow, 2015). Unique sequences (referred to as amplicon sequence variants or ASVs) and estimated gene abundances were normalized according to predicted 16S rRNA gene copy number prior to downstream analysis. The paprica community structure results are described in terms of closest estimated genomes (CEGs; for phylogenetic placements to non-terminal edges) and closest completed genomes (CCGs; for placements to terminal edges). CCGs are names according to their 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 lowest consensus taxonomic ranking, while CEGs are named according to their closest relative on the phylogenetic reference tree. Diversity and statistics analysis The alpha diversity index, inverse of Simpson, for ASVs was calculated using the phyloseq package in R (McMurdie and Holmes, 2013) following methods described in Callahan (Ben J. Callahan et al., 2016). Kruskal-Wallis tests were performed to test differences among groups in the vegan package in R (Oksanen et al., 2019). For the biogeochemical parameters, we used the Kruskal-Wallis test to test differences among groups, and the Spearman correlation to evaluate relationships between N:P ratio, genome size, and 16S rRNA gene copy number. We determined N in disturbed and less disturbed sites; this is a measure of nitrogen vs. phosphorus availability based on the Redfield ratio (N:P = 16:1) (Gruber and Sarmiento, 1997), and we calculated based on nutrient concentrations using the following equation (Wilson, Abboud and Beman, 2017): (1): 𝑁 ∗ = (𝑁𝑂! " + 𝑁𝑂 $ " + 𝑁𝐻% &) − 16 × 𝑃𝑂% !" We used correspondence analysis (CA) to quantify taxon contributions to the sample ordination. This method allowed us to determine the degree of correspondence between sites and species, and which taxa were associated with gradients of disturbance. We performed CA on Hellinger- transformed data such that each value represents a contribution to the Pearson's χ2 (chi-squared) statistic computed for the data (Legendre P., 1998). We also calculated a cos2 value that describes the contribution of each taxa to the major axes of disturbance (Kuramae et al., 2012). Analysis of similarity (ANOSIM) was used to assess significant differences with respect to level of disturbance. This nonparametric permutation procedure uses the rank similarity matrix underlying 114 115 116 117 118 119 120 121 122 an ordination plot to calculate an R test statistic, and it was calculated using the vegan package in R (Oksanen et al., 2019). We examined association of levels of disturbance by a Spearman correlation between ammonia concentrations and dimensions 1 and 2 of CA analysis. To examine the impact of environmental variables associated to aquaculture outflow on the estimated metabolic pathways we performed a canonical correspondence analysis (CCA) to the metabolic output generated in paprica to restrict the sample ordination to nitrogen, phosphate, and isotopic signals to better understand the impact of aquaculture outflow on microbial metabolic potential. The cos2 value was used to determine the contribution of key metabolic pathways to the major axis. The ordinations were performed in R using the factoMiner and CA package (Husson et al., 123 2020). 124 125 126 127 128 129 130 131 132 133 134 135 136 To identify unique reads differentially present between disturbed and non-disturbed sites we used DESeq2 (Michael I Love, Huber and Anders, 2014), following the methods of Webb et al. (2019). DESeq2 performs differential abundance analysis based on the negative binomial/Gamma-Poisson distribution. The default settings were used, which estimates size factors with the median ratio method (Michael I. Love, Huber and Anders, 2014), followed by estimation of dispersion. Next, a Wald test for generalized linear model coefficients was used to test for significance of coefficients, considering size factors and dispersion. The p-values were attained by the Wald test and corrected for multiple testing using the Benjamini and Hochberg method (Michael I. Love, Huber and Anders, 2014). The most abundant bacterial and archaeal taxa that were significantly differentially present were further examined to identify potential microbial markers of shrimp aquaculture effluent. To determine the role of differentially abundant microbes in nutrient cycling, we utilized the BioCyc database (Karp et al., 2019) in combination with the paprica output to assess the potential for genes coding enzymes associated with nitrogen fixation and denitrification. Enzymes included with our assessment included: nitrogenase; EC 1.18.6.1, EC 1,19.6.1, nitrate reductase; EC 17.99.4, EC 1.7.1.1, EC 1.7.1.2, EC 1.9.6.1, EC 1.7.2.2, and nitrite reductase; EC1.7.2.1, EC 1.7.2.2, EC 1.7.1.4. To identify modules of highly correlated taxa we used Weighted Gene Correlation Network Analysis (WGCNA) (Langfelder and Horvath, 2008), following the methods of Wilson et al. (2018). A signed adjacency measure for each pair of features (unique reads) was calculated by raising the absolute value of their Pearson correlation coefficient to the power of parameter p. The value p = 8 was used for each global network to optimize the scale-free topology network fit. This power allows the weighted correlation network to show a scale free topology where key nodes are highly connected with others. The obtained adjacency matrix was then used to calculate the topological overlap measure (TOM), which, for each pair of features, considers their weighted pairwise correlation (direct relationships) and their weighted correlations with other features in the network (indirect relationships). For identifying subnetworks or ‘modules’ a hierarchical clustering was performed using a distance based on the TOM measure. This resulted in the definition of several subnetworks, each represented by its first principal component. A subnetwork is the association between the subnetworks and a given trait that is measured by the pairwise relationships (correlations) between the taxa. To find correlations between subnetworks and environmental factors, Pearson’s correlation coefficients were calculated between the considered environmental factor and the respective principal components. P-values were adjusted using Bonferroni method. All procedures were applied to Hellinger-transformed abundances. 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 References 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 Barraza-Guardado, R. H. et al. (2013) ‘Effluents of shrimp farms and its influence on the coastal ecosystems of Bahía de Kino, Mexico.’, TheScientificWorldJournal, 2013, p. 306370. Bowman, J. S. and Ducklow, H. W. (2015) ‘Microbial Communities Can Be Described by Metabolic Structure: A General Framework and Application to a Seasonally Variable, Depth- Stratified Microbial Community from the Coastal West Antarctic Peninsula’, PLOS ONE. Edited by C. Moissl-Eichinger, 10(8), p. e0135868. Callahan, Ben J. et al. (2016) ‘Bioconductor workflow for microbiome data analysis: From raw reads to community analyses [version 1; referees: 3 approved]’, F1000Research, 5. Callahan, Benjamin J et al. (2016) ‘DADA2: High-resolution sample inference from Illumina amplicon data.’, Nature methods, 13(7), pp. 581–3. Cifuentes, L. A. et al. (1996) ‘Isotopic and Elemental Variations of Carbon and Nitrogen in a Mangrove Estuary’, Estuarine, Coastal and Shelf Science, 43(6), pp. 781–800. Gruber, N. and Sarmiento, J. L. (1997) ‘Global patterns of marine nitrogen fixation and denitrification’, Global Biogeochemical Cycles, 11(2), pp. 235–266. Hamilton, S. E. (2020a) ‘Introduction to Coastal Ecuador’, in Coastal Research Library. Springer, pp. 69–110. Hamilton, S. E. (2020b) ‘Assessing 50 Years of Mangrove Forest Loss Along the Pacific Coast of Ecuador: A Remote Sensing Synthesis’, in Coastal Research Library. Springer, pp. 111–137. Husson, F. et al. (2020) Package ‘FactoMineR’ Title Multivariate Exploratory Data Analysis and Data Mining. [Online] Available at: http://factominer.free.fr Karp, P. D. et al. (2019) ‘The BioCyc collection of microbial genomes and metabolic pathways’, 20(4), pp. 1085–1093. Kuramae, E. E. et al. (2012) ‘Soil characteristics more strongly influence soil bacterial communities than land-use type’, FEMS Microbiology Ecology, 79(1), pp. 12–24. Langfelder, P. and Horvath, S. (2008) ‘WGCNA: an R package for weighted correlation network analysis.’, BMC bioinformatics, 9, p. 559. Legendre P., L. L. F. J. (1998) Numerical Ecology, Volume 24 - 2nd Edition, Elsevier Science. Love, Michael I, Huber, W. and Anders, S. (2014) ‘Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2’, Genome Biology, 15(12), p. 550. McMurdie, P. J. and Holmes, S. (2013) ‘phyloseq: An R Package for Reproducible Interactive Analysis and Graphics of Microbiome Census Data’, PLoS ONE. Edited by M. Watson, 8(4), p. e61217. Oksanen, J. et al. (2019) Package ‘vegan’ Title Community Ecology Package Version 2.0-8. [Online] Available at: http://CRAN.R-project.org/package=vegan Pestle, W. J., Crowley, B. E. and Weirauch, M. T. (2014) ‘Quantifying Inter-Laboratory Variability in Stable Isotope Analysis of Ancient Skeletal Remains’, PLoS ONE. Edited by L. Bondioli, 9(7), p. e102844. Robertson, A. I. and Alongi, D. M. (eds) (1992) Tropical Mangrove Ecosystems. Washington, D. C.: American Geophysical Union (Coastal and Estuarine Studies). Samocha, T. M. et al. (2004) ‘Characterization of intake and effluent waters from intensive and semi-intensive shrimp farms in Texas’, Aquaculture Research, 35(4), pp. 321–339. Walters, W. et al. (2015) ‘Transcribed Spacer Marker Gene Primers for Microbial Community Surveys’, mSystems, 1(1), pp. e0009-15. Webb, S. J. et al. (2019) ‘Impacts of Zostera eelgrasses on microbial community structure in San Diego coastal waters’, Elem Sci Anth, 7(1), p. 11. Wilson, J., Abboud, S. and Beman, J. M. (2017) ‘Primary Production, Community Respiration, and Net Community Production along Oxygen and Nutrient Gradients: Environmental Controls and Biogeochemical Feedbacks within and across “Marine Lakes”’, Frontiers in Marine Science, 4. Wilson, J. M., Litvin, S. Y. and Beman, J. M. (2018) ‘Microbial community networks associated with variations in community respiration rates during upwelling in nearshore Monterey Bay, California’, Environmental Microbiology Reports, 10(3), pp. 272–282. 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237	null
10.1371_journal.pone.0253387.pdf	Data Availability Statement: Raw genome sequencing data for 23 strains are available from the NCBI with BioProject PRJNA683613.	Raw genome sequencing data for 23 strains are available from the NCBI with BioProject PRJNA683613.	RESEARCH ARTICLE Classification of cannabis strains in the Canadian market with discriminant analysis of principal components using genome-wide single nucleotide polymorphisms Dan Jin1,2, Philippe HenryID 3,4, Jacqueline Shan2, Jie ChenID 1,5* a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Jin D, Henry P, Shan J, Chen J (2021) Classification of cannabis strains in the Canadian market with discriminant analysis of principal components using genome-wide single nucleotide polymorphisms. PLoS ONE 16(6): e0253387. https://doi.org/10.1371/journal.pone.0253387 Editor: Tzen-Yuh Chiang, National Cheng Kung University, TAIWAN Received: November 10, 2020 Accepted: June 3, 2021 Published: June 28, 2021 Copyright: © 2021 Jin et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: Raw genome sequencing data for 23 strains are available from the NCBI with BioProject PRJNA683613. Funding: PBG BioPharma Inc. (https:// pbgbiopharma.com/) provided funding support in the form of salaries for authors DJ and JS. PBG BioPharma Inc. also provided financial support for genome sequencing and travel expenses. JS is the founder and CEO of PBG BioPharma Inc., and reviewed the manuscript. Labs-Mart Inc. (http:// labs-mart.ca/) provided chemical standards and 1 Department of Biomedical Engineering, University of Alberta, Edmonton, Alberta, Canada, 2 PBG BioPharma Inc., Leduc, Alberta, Canada, 3 Egret Bioscience Ltd., West Kelowna, British Columbia, Canada, 4 Lighthouse Genomics Inc., Salt Spring Island, British Columbia, Canada, 5 Department of Electrical and Computer Engineering, University of Alberta, Edmonton, Alberta, Canada * [email protected] Abstract The cannabis community typically uses the terms “Sativa” and “Indica” to characterize drug strains with high tetrahydrocannabinol (THC) levels. Due to large scale, extensive, and unrecorded hybridization in the past 40 years, this vernacular naming convention has become unreliable and inadequate for identifying or selecting strains for clinical research and medicinal production. Additionally, cannabidiol (CBD) dominant strains and balanced strains (or intermediate strains, which have intermediate levels of THC and CBD), are not included in the current classification studies despite the increasing research interest in the therapeutic potential of CBD. This paper is the first in a series of studies proposing that a new classification system be established based on genome-wide variation and supple- mented by data on secondary metabolites and morphological characteristics. This study performed a whole-genome sequencing of 23 cannabis strains marketed in Canada, aligned sequences to a reference genome, and, after filtering for minor allele frequency of 10%, identified 137,858 single nucleotide polymorphisms (SNPs). Discriminant analysis of princi- pal components (DAPC) was applied to these SNPs and further identified 344 structural SNPs, which classified individual strains into five chemotype-aligned groups: one CBD dom- inant, one balanced, and three THC dominant clusters. These structural SNPs were all mul- tiallelic and were predominantly tri-allelic (339/344). The largest portion of these SNPs (37%) occurred on the same chromosome containing genes for CBD acid synthases (CBDAS) and THC acid synthases (THCAS). The remainder (63%) were located on the other nine chromosomes. These results showed that the genetic differences between mod- ern cannabis strains were at a whole-genome level and not limited to THC or CBD produc- tion. These SNPs contained enough genetic variation for classifying individual strains into corresponding chemotypes. In an effort to elucidate the confused genetic backgrounds of commercially available cannabis strains, this classification attempt investigated the utility of DAPC for classifying modern cannabis strains and for identifying structural SNPs. PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 1 / 14 PLOS ONEinstrumentation support for chemical testing, but did not have any additional role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Egret Bioscience Ltd. (https://egret.bio) and Lighthouse Genomics Inc. (https://lighthousegenomics.com/) provided support in the form of salaries for author PH, but did not have any additional role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The specific roles of these authors are articulated in the ‘author contributions’ section. Competing interests: The funder provided support in the form of salaries for authors DJ, PH, and JS. This does not alter our adherence to PLOS ONE policies on sharing data and materials. Genome-wide single nucleotide polymorphisms to classify cannabis strains Introduction Cannabis has a complex breeding history. Whether its botanical classification is monotypic (sativa) or polytypic (sativa and indica) remains controversial [1]. Since the 1980s, breeding for high psychoactive THC content has occurred very aggressively in North America [2]. Nearly all drug-type cannabis currently cultivated in the USA, Canada, and Europe are hybrid- ized, resulting in thousands of strains [3]. Recent genetic studies focused on validating the ver- nacular classification of “Sativa” and “Indica” [4–7]. However, this terminology is inadequate for identifying or selecting strains for clinical research and medicinal production due to the misuse of the botanical nomenclature, extensive cross-breeding, and unreliable labelling dur- ing unrecorded hybridization [2]. One genetic study found that the reported ancestry percent- age of “Sativa” vs. “Indica” for 81 drug stains is only moderately correlated with the calculated genetic structure (r2 = 0.36) [5]. In addition, CBD dominant strains and balanced strains (THC � CBD), which have gained increasing attention due to CBD’s use as a therapeutic [8– 12], have been omitted in recent classification studies. Cannabis has a diploid genome (2n = 20) with nine autosomal chromosomes and one pair of sex chromosomes [13]. The length of the haploid genome size is 818 Mbp for females and 843 Mbp for males [14]. An SNP is a variation of a single nucleotide at a specific position in the genome, and it is useful for understanding the genetic basis of diversity among populations [15]. SNPs are usually bi-allelic, with two alleles observed in the population [16]. Multiallelic SNPs have more than one alternative allele for that locus. Tri-allelic SNPs, which have three nucleotide substitution-based alleles at the same position, are relatively rare but are being con- sidered of great relevance in epidemiological studies [17], in disaster victim identification using mixed and/or degraded DNA samples [18], and in animals pedigree accuracy studies [19]. Tri-allelic SNPs are reported to have a higher power of discrimination than bi-allelic SNPs requiring fewer markers and lowering costs [18, 20]. However, tri-allelic SNPs have been excluded in cannabis population structural analysis in the current literature [6, 21]. Cannabis classification studies that employ SNPs generally used partial genome informa- tion with few or no overlap sequences between datasets [22]. Whole-genome sequencing is used less often in the literature, but is preferable despite its higher cost because it enables com- parison of genome datasets from different sources [22]. It also provides comprehensive genetic information [22], as studies showed that differences between fiber- and drug-type cannabis are at a genome-wide level and not necessarily limited to genes involved in THC production [5]. The recent release of the 10-chromosome map of the cannabis genome [23–27] may improve the understanding of the genetic architecture, identify a superior set of SNPs associated with interesting traits, and reduce future targeted genotyping costs by using fewer but more accu- rate SNPs [28]. Several approaches are now available for the analysis of population genetic structure. One of these approaches is the DAPC, which is a multivariate clustering method that combines the merits of both principal component analysis (PCA) and discriminant analysis (DA) [7, 29–31]. PCA is a multivariate analysis that can be applied to large datasets to reduce dimensions, but does not provide a group assessment, which is essential for investigating genetic structures of biological populations [32]. DA achieves the best classification of individuals into pre-defined groups by maximizing between-group variation and minimizing within-group variation, but the number of variables (alleles) needs to be fewer than the number of observations (individu- als), which is generally not the case for SNP data [29]. DAPC first uses PCA to transform raw data (genome-wide identified SNPs) into principal components (PC), which are mutually orthogonal linear combinations of the original variables. This ensures that variables submitted to DA are perfectly uncorrelated and that there are fewer variables than number of individuals. PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 2 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains Then, linear discriminant functions, which are synthetic variables of linear combinations of these SNPs, are constructed to maximize inter-cluster differences and minimize intra-cluster variation [29]. By combining the advantages of PCA and DA, DAPC can identify groups, assign individuals to groups, visualize between-population differentiation, and identify indi- vidual alleles that have contributed to population structuring. The objectives of this study are to: 1. investigate whether modern cannabis strains can be classified and differentiated at the whole-genome level, and 2. investigate the chromosomal location and putative functions of identified structural SNPs. This study is a part of an integrated cannabis strain classification project utilizing genetic, chemical, and morphological profiles, wherein plants were grown in a commercial greenhouse under the same condition. Materials and methods DNA extraction and whole genome sequencing This study included 23 commercially available cannabis strains, and the research was carried out under a cannabis research license issued by Health Canada. Where possible, the reported ancestry (“Sativa”, “Indica”, or “Sativa-dominant” and “Indica-dominant”) was obtained from the licensed producer providing the strain or from an online strain database (https://www. leafly.ca) (Table 1). Each strain was analyzed for chemical composition using methods estab- lished in our previous study [33] and labelled as “THC dominant”, “balanced”, and “CBD dominant”. DNA was extracted from 100 mg of fresh leaves for each strain using a Qiagen DNeasy Plant Mini Kit (QIAGEN, Canada). DNA concentrations were determined using a Qubit Fluorometer (Thermo Fisher Scientific, US). DNA integrity was tested by agarose gel electrophoresis. Library construction and sequencing were performed by BGI (USA) using DNBseq™ sequencing technology to a depth of 30x. DNBseq™ is a high-throughput sequencing solution, where DNA is fragmented into 100–300 bp and made into DNA nanoballs (DNB™), which are continuous DNA molecule with multiple head-to-tail copies of the same DNA frag- ment by linear isothermal rolling-circle replication. They are loaded onto high-density sequencing templates and sequenced by combinatorial probe-anchor synthesis (cPAS), where fluorescently tagged nucleotides complete for addition to the growing chain. After the addition of each nucleotide, high-resolution digital imaging is carried out where the DNB clusters are excited by a light source and a characteristic fluorescent signal is emitted. Hundreds of and thousands of clusters are sequenced in a massively parallel process. The emission wavelength, along with the signal intensity, determines the base call and the number of the cycles deter- mines the length of the read. Sequence reads were then aligned to the reference genome assem- bly ASM23057v4 of a drug type strain Purple Kush (PK) in the NCBI BioProject database under accession number PRJNA73819 [34] using Burrows-Wheeler Alignment (BWA) tool [35]. New assignments of chromosomes numbers (1–10) were used as in ASM23057v5 [36]. The first step of SNP calling is marking duplications in BAM format files, and selected duplica- tions are included in SNP calling by GATK (Genome Analysis Toolkit) (https://www. broadinstitute.org/gatk/). Local realignment around inDels is performed to avoid the bias of SNP calling, and the variation sites around inDel are identified as SNPs. A total of 235,334 SNPs was identified, including 225,046 bi-allelic and 10,288 multiallelic SNPs. After filtering for SNPs with no missingness by locus and a minor allele frequency less than 10% using PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 3 / 14 PLOS ONETable 1. Strain information of 23 strains and preassigned clusters by DAPC. Strain number Strain name Chemotypes Clusters (W-SNPs) Clusters (I-SNPs) "Sativa" or "Indica" Genome-wide single nucleotide polymorphisms to classify cannabis strains 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Lemon Garlic OG Royal Medic Blue Hawaiian Kandy Kush Special NN Dance World Treat High CB7 33˚ Banana Cake Bananium Burmese Blueberry Divine Banana Granddaddy Purple Lemon Love Lemon Sorbet MeatHead Nanitro Platinum Jelly Punch SBSK2 (Lemon Thai) Super sherbet 1-Balanced 2-Balanced 3-CBD 4-CBD 5-CBD 6-CBD 7-Balanced 8-CBD 9-Balanced 10-CBD 11-THC 12-THC 13-THC 14-THC 15-THC 16-THC 17-THC 18-THC 19-THC 20-THC 21-THC 22-THC 23-THC C1 C3 C3 C3 C3 C3 C3 C3 C3 C3 C1 C2 C3 C2 C2 C2 C1 C1 C2 C1 C1 C3 C1 C4 C2 C1 C1 C1 C1 C2 C1 C2 C1 C4 C5 C3 C5 C4 C5 C5 C4 C5 C4 C4 C3 C4 "Indica" dominant "Sativa" dominant "Sativa" dominant "Sativa" dominant Not provided Not provided "Sativa" dominant Not provided Not provided Not provided Not provided "Indica" dominant "Indica" dominant "Indica" dominant "Indica" dominant "Indica" dominant "Indica" dominant "Indica" dominant "Indica" dominant "Indica" dominant "Indica" dominant 50/50 hybrid "Indica" dominant �The column of clusters W-SNPs was obtained using the whole set of 137,858 filtered SNPs. The column of clusters I-SNPs was obtained using 344 structural SNPs. https://doi.org/10.1371/journal.pone.0253387.t001 VCFtools, 137,858 SNPs, including 128,810 bi-allelic and 9,048 multiallelic SNPs, remained for analysis. Analysis of population structure and identification of structural SNPs The population structure in this work was analyzed by DAPC using the adegenet package [37] in R software [38]. First, the find.clusters function ran successive K-means [39] for a range of k values (where the number of clusters k = K), and identified the optimal number of clusters by comparing the Bayesian Information Criterion (BIC) [40] of the corresponding models. After groups were assigned, a cross-validation function (xvalDapc) was used to determine the opti- mal number of PCs to avoid over-sacrificing information or over-fitting in the subsequent DAPC. In cross-validation, the data were divided into a training set (90% of the data) and a validation set (10% of the data) by default. DAPC was carried out on the training set and the accuracy of predicting the membership of individuals in the validation set was used to identify the number of PCs. The sampling and DAPC were repeated 30 times by default at each level of PC retention. After assigning individuals to clusters, DA was carried out on the retained PCs and contributions of the alleles to each discriminant function were stored. An SNPZIP analysis (snpzip) in R was then used to provide objective delineation between structural and non-struc- tural SNPs, as identified by DAPC, to determine which SNPs contribute significantly to the between-population structure [41]. PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 4 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains First, the whole set of 137,858 SNPs were applied to DAPC to identify SNPs that contrib- uted most to the identified clusters. DAPC was carried out again using the identified SNPs to validate their differentiation efficiency by confirming the separation of the 23 strains into their preassigned clusters. A short sequence (about 600 nt) around each one of these identified SNP was searched using the BLAST software (https://blast.ncbi.nlm.nih.gov) against Cannabis sativa Annotation Release 100 [42]. In addition to DAPC, other clustering methods, including PCA, neighbor-joining (NJ) tree [43], and hierarchical dendrogram using Ward’s minimum variance method [44], were also employed to assess the robustness of the final inferred clusters. PCA and NJ tree were plotted using R. The hierarchical dendrogram was plotted using JMP 14.0.0. Results and discussions Discriminant analysis of principal components using 137,858 SNPs As indicated by the elbow in the curve of BIC values as a function of k in Fig 1(a), the optimal number of identified clusters was three, corresponding to the lowest BIC values. The number of PCs retained for DAPC analysis was four, as calculated by cross-validation in Fig 1(b), where it had 100% predictive success, and 0% associated root mean squared error (RMSE). In this study, the number of PCs associated with the highest mean success was also associated with the lowest MSE, which made it easier to choose the number of PCs to retain. For the sub- sequent DAPC analysis, four PCs and two discriminant functions were retained. The DAPC Fig 1. DAPC for 23 cannabis genotypes. (a) The x-axis is the number of clusters k and the y-axis is the corresponding value of BIC. (b) The plot of DAPC cross-validation. The x-axis is the number of PCA axes retained for DAPC, and the y-axis is the proportion of successful outcome prediction. Individual replicates appear as points, and the density of those points in different regions of the plot is displayed in blue. (c) DAPC plot for 23 cannabis genotypes along two linear discriminants (LD 1 and LD 2). https://doi.org/10.1371/journal.pone.0253387.g001 PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 5 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains plot of 23 cannabis genotypes is shown in Fig 1(c). The grouping assignment for individual strains by DAPC is listed in Table 1 (as W-SNPs). C1 is a THC dominant cluster and includes six THC dominant strains (11, 17, 18, 20, 21, and 23-THC) and one balanced strain (1-bal- anced). C2 is another THC dominant cluster and includes five THC dominant strains (12, 14, 15, 16, and 19-THC). C3 is a cluster dominated by CBD dominant and the balanced strains which includes six CBD dominant strains (3, 4, 5, 6, 8, and 10-CBD), three balanced strains (2, 7, and 9-balanced), and two THC dominant strains (13 and 22-THC). While C2 is closer to C3 and is more distant to C1, C1and C3 are clearly separated along linear discriminant 1 (LD1). While C1 and C3 are roughly at the same level with respect to linear discriminant 2 (LD2), C2 is separated from both. PCA was also carried out on the same set of SNPs and results are shown in S1 Fig. Twenty-three cannabis strains are plotted along pair-wise PCs of the first 4 PCs, which account for 18.4%, 11.5%, 9.5%, and 8.7% of the total variance, respectively. Simi- larly, the first PC suggests the existence of a relatively compact CBD & balanced clade on the left side of the plot and a more dispersed THC dominant clade on the right side of the plot. Bal- anced strains share a closer gene pool with CBD dominant strains, while the THC gene pool is more dispersed. Because THC is psychoactive and its potency can be readily assessed through consumption, selection for increasing THC content started early and widely for recreational purposes by traditional breeding [45]. In contrast, CBD is non-psychoactive and must be ana- lyzed in a laboratory for potency, and therefore breeding for high CBD concentrations began later [45]. A complete genome assembly implied that CBD dominant varieties were generated by integrating hemp-type CBD acid synthase gene clusters into a background of drug-type cannabis to elevate CBDA production [24]. These balanced strains may have been created by crossing purebred THC dominant types with CBD dominant types [46]. Therefore, there may be a relatively limited selection of CBD dominant strains for breeding balanced strains. Discriminant analysis of principal components using 344 structural SNPs DAPC was repeated using identified 344 structural SNPs. The optimal number of identified clusters was five, corresponding to the lowest BIC values (Fig 2(a)). Two PCs were retained for the following DAPC analysis in Fig 2(b), where it had 98.9% predictive success and 0.04% RMSE. For the subsequent DAPC analysis, two PCs and two discriminant functions were retained. The grouping assignment for individual strains by DAPC is listed in Table 1 (as I-SNPs). Within the five clusters (Fig 2(c)), C1 is a CBD dominant cluster that includes six strains (3, 4, 5, 6, 8, and 10-CBD), C2 includes three balanced strains (2, 7, and 9-balanced), and C3, C4, and C5 are THC dominant clusters that include two (13 and 22-THC), seven (1-balanced, 11, 15, 18, 20, 21, 23-THC), and five (12, 14, 16, 17, and 19-THC) strains, respectively. These multiallelic SNPs were also subjected to PCA, NJ tree, and hierarchical clustering analysis. In Fig 3, the 23 cannabis strains are plotted along PC1 and PC2, which account for 44.5% and 10.0% of the total variance, respectively. The proportions of explained variance are higher compared to the previous PCA results (18.4% and 11.5%) obtained using the whole set of SNPs. CBD dominant cluster C1 and balanced cluster C2 are on the left side of the scatter plot (PC1<0) and the THC dominant clusters C3, C4, and C5 are on the right side of the scat- ter plot (PC1>0). Notably, six CBD dominant strains are separated from three balanced strains, while they were previously combined in the analysis using the whole set of SNPs. In addition, two THC dominant strains 13-THC and 22-THC are separated from the CBD and balanced cluster, and instead placed closer to other THC dominant strains. Strain 1-balanced is closer to THC dominant strain regardless of whether the whole set of SNPs or 344 identified SNPs were used. PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 6 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains Fig 2. DAPC of 23 cannabis genotypes using 344 multiallelic structural SNPs. Clusters indicated as C1, C2, C3, C4, and C5 corresponds to the I-SNPs in Table 1. https://doi.org/10.1371/journal.pone.0253387.g002 The genetic structure from NJ-tree and hierarchical clustering using the 344 multiallelic are displayed in Fig 4, mostly congruent with that of DAPC. In the NJ-tree, all six CBD dominant strains are clustered together, with three balanced strains clustered closer on the same branch (Fig 4(a)). Most THC dominant strains are also clustered adjacent to strains within their own clusters. The dendrogram using hierarchical clustering by Ward’s method reveals two major groups, where one group is comprised of CBD dominant & balanced strains, and the other of THC dominant strains (Fig 4(b)). They are further separated into five subclusters, where CBD dominant and balanced clusters are consistent with the DAPC grouping results, and several THC dominant strains clustered differently. Two strains, 15-THC and 18-THC, were assigned to C4 using DAPC but are assigned closer to C5 in the dendrogram. Two other strains, 14-THC and 16-THC, were assigned to C5 in DAPC but are assigned closer to C3 in the den- drogram. The clustering results are congruent between DAPC and hierarchical clustering with an assignment agreement rate of 83% (19/23). Allele frequencies for 344 multiallelic SNPs in three chemotypes DAPC identified 344 highly contributing SNPs (S1 Table). All the structural SNPs are multial- lelic, among which 98.5% (339/344) are tri-allelic and the remainder 1.5% (5/344) are tetra- allelic. The dendrogram of 23 strains using hierarchical clustering based on the allele counts in the 344 structural SNPs (S2 Table) separated the strains into CBD dominant, balanced, and THC dominant strains, mostly corresponding to the grouping results of DAPC (Fig 5). The allele frequency was calculated by dividing the counts of that allele for all strains within the tar- geted group by the sum of the counts for each allele for that SNP within the targeted group. Allele frequencies of the structural SNPs were calculated for three major branches, each corre- sponding one of three chemotypes. (S1 Table). If 1-balanced strain was assigned to the THC PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 7 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains Fig 3. Scatter plot of 23 cannabis strains on PC1 & PC2 using 344 structural SNPs. Clusters indicated as C1, C2, C3, C4 and C5 correspond to I-SNPs in Table 1. https://doi.org/10.1371/journal.pone.0253387.g003 dominant group as indicated by DAPC for allele frequency calculation, there are 87% (300/ 344) SNPs in CBD dominant clusters, 46% (157/344) SNPs in balanced clusters, and 11% (39/ 344) SNPs in THC dominant clusters that have one allele with allele frequencies > 80% (S1 Table). Among them, 140 SNPs shared same alleles with allele frequencies > 80% in CBD dominant strains (140/300) and balanced strains (140/157), which further indicated that CBD dominant strains and balanced strains closely share a gene pool. There are 38 SNPs that have one allele present in CBD dominant strains with allele frequencies > 80% and are not detected in THC dominant strains. There are 322 SNPs whose alleles that are present in THC dominant strains but were not detected in CBD dominant strains. If the 1-balanced strain is assigned to the balanced group for allele frequency calculation, there are 87% (300/344) SNPs in CBD dominant clusters, 10% (36/344) SNPs in balanced clus- ters, and 13% (44/344) SNPs in THC dominant clusters that have one allele with allele frequencies > 80% (S2 Table). Among them, 32 SNPs shared same alleles with allele frequencies > 80% in CBD dominant strains (32/300) and balanced strains (32/36). There are 38 SNPs that have one allele present in CBD dominant strains with allele frequencies > 80% and are not detected in THC dominant strains. There are 321 SNPs whose alleles are present in THC dominant strains but were not detected in CBD dominant strains. Assigning the 1-bal- anced strain to the balanced group added more genetic diversity to the balanced group, and the effect of adding or deleting this strain for the THC dominant group in terms of allele fre- quency is small and can be neglected. PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 8 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains Fig 4. NJ-tree and hierarchical clustering using the 344 multiallelic SNPs (a) NJ-tree and (b) The dendrogram using hierarchical clustering by Ward’s method for 23 cannabis genotypes. Clusters indicated as C1, C2, C3, C4, and C5 corresponds to I-SNPs in Table 1. https://doi.org/10.1371/journal.pone.0253387.g004 Fig 5. Hierarchical clustering of 23 strains based on the allele counts for 344 structural SNPs identified by DAPC. https://doi.org/10.1371/journal.pone.0253387.g005 PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 9 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains Fig 6. Features of 344 multiallelic SNPs (a) Distribution of structural SNPs on chromosome 1–10 and unplaced scaffolds. (b) BLAST results for structural SNPs against a fully annotated genome. https://doi.org/10.1371/journal.pone.0253387.g006 BLAST analysis of 344 multiallelic SNPs These 344 SNPs were spread across all 10 chromosomes (Fig 6(a)), indicating that commer- cially available cannabis strains in North America are significantly differentiated at a genome- wide level. The number of identified SNPs ranged from 7 to 127 on each genome, with 37% of the genetic variation occurring (127 SNPs) on chromosome 6, where CBDAS and THCAS are located [13]. The rest SNPs were spread over the remaining nine chromosomes. All ten chro- mosomes have genes related to the biochemical pathways of secondary metabolites, including cannabinoids, monoterpenes, and sesquiterpenes [13, 24, 47–51]. BLAST results showed that 90% (310/344) of these structural SNPs had no feature, 7% (24/344) are uncharacterized loci with unknown functions, and 3% (10/344) are predicted for certain functions (Fig 6(b)). Conclusions Although the cannabis industry is rapidly advancing after the relaxation of legal restrictions in North America, the increasing number of THC dominant strains, CBD dominant strains, and balanced strains only adds confusion to the currently poorly understood genetic background of the thousands of varieties already in existence. Although there were only 23 strains included in this study, they covered the three typical chemotypes of cannabis strains currently available in the market. Leveraging as much genetic variation as possible using whole-genome sequenc- ing, we identified 344 multiallelic SNPs that were used to investigate the genetic structure of 23 cannabis genotypes using DAPC, PCA, NJ tree, and hierarchical clustering, which provided consistent observations and groupings despite the differences in algorithms. The clustering results revealed that these 23 strains could be separated into five clusters, with one cluster PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 10 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains containing six CBD dominant strains, another cluster containing three balanced strains, and the remaining three clusters containing 13 THC dominant strains and one balanced strain. CBD dominant strains and the balanced strains are closer genetically. This may be attributed to how medical interest in breeding for non-psychoactive, CBD-elevated strains (CBD domi- nant and balanced strains) has only recently been in vogue, resulting in an overlapping and less diverse gene pool for CBD dominant and balanced strains compared to the longer breed- ing history for THC strains. Some alleles are only present in CBD dominant strains or in THC dominant strains. More alleles present in balanced strains are shared with CBD dominant strains. One third of these structural SNPs are located on the chromosome containing THCAS and CBDAS. The remaining SNPs are located on the other nine chromosomes. An area of potential investigation is how the identified structural SNPs are associated with the production of other cannabinoids, mono- and sesquiterpenes, flavonoids, other compounds, or morpho- logical characteristics. Since the late 20th century, genetic methodologies have been developed for separating industrial hemp from drug-type cannabis for forensic purposes, thus differentiating CBD dominant and THC dominant strains [52–56]. For the past 20 years, with the extensive hybrid- ization of THC dominant strains, many classification studies have focused on separating “Sativa” and “Indica” strains and many have suggested abolishing this vernacular [5–7]. The genotyping results of this study indicate that modern, extensively hybridized strains can still be separated using genome-wide information. As a powerful multivariate approach that investi- gates population structures based solely on genetic information, DAPC separated strains into clusters aligned with their chemotypes. Additionally, DAPC has the potential to sort the disor- dered genetic background of thousands of THC dominant strains by identifying the number of genetic clusters within THC dominant strains, describing clusters by interpreting group memberships, and identifying the contributing SNPs that have the potential to be used as genetic markers for strain classification and identification. This would require a concerted effort from the cannabis industry by contributing whole genome sequence data to public data- bases and by building a common taxonomy based on genomics. Optimally, the identified genetic markers can be used as genomic fingerprints in combination with chemical finger- prints and morphological characteristics for strain identification. These markers can be lever- aged for strain selection in clinical trials and for manufacturing cannabis-based products and medicines. Supporting information S1 Fig. PCA of 23 strains using whole set of SNPs. (PDF) S1 Table. 344 multiallelic SNPs identified by DAPC. (XLSX) S2 Table. Allele counts for 344 structural SNPs identified by DAPC. (XLSX) Acknowledgments The authors are grateful to licensed grower Emerald Flower Farm who provided commercial greenhouse to cultivate cannabis. The authors are also grateful to Dr. Limin Wu for assisting DNA extraction, Dr. Jie Zeng for assisting BLAST analysis, and Shengxi Jin for proofreading the manuscript. PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 11 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains Author Contributions Conceptualization: Dan Jin. Data curation: Dan Jin. Formal analysis: Dan Jin, Philippe Henry. Funding acquisition: Dan Jin, Jacqueline Shan. Investigation: Dan Jin. Methodology: Dan Jin, Philippe Henry. Project administration: Dan Jin. Resources: Dan Jin. Software: Dan Jin, Philippe Henry. Supervision: Jie Chen. Validation: Dan Jin, Philippe Henry. Visualization: Dan Jin, Philippe Henry. Writing – original draft: Dan Jin. Writing – review & editing: Philippe Henry, Jacqueline Shan, Jie Chen. References 1. Hillig KW (2005) A systematic investigation of Cannabis, PhD thesis, Indiana University. PhD Thesis 2. McPartland JM (2017) Cannabis sativa and Cannabis indica versus “Sativa” and “Indica”. In: Cannabis sativa L.-Botany and Biotechnology. Springer, pp 101–121 3. McPartland JM, Guy GW (2017) Models of Cannabis Taxonomy, Cultural Bias, and Conflicts between Scientific and Vernacular Names. The Botanical Review 4:327–381 4. Knight G, Hansen S, Connor M, Poulsen H, McGovern C, Stacey J (2010) The results of an experimen- tal indoor hydroponic Cannabis growing study, using the ‘Screen of Green’(ScrOG) method—Yield, tetrahydrocannabinol (THC) and DNA analysis. Forensic science international 202:36–44 https://doi. org/10.1016/j.forsciint.2010.04.022 PMID: 20462712 5. Sawler J, Stout JM, Gardner KM, Hudson D, Vidmar J, Butler L, et al. (2015) The genetic structure of marijuana and hemp. PloS one 10:e0133292 https://doi.org/10.1371/journal.pone.0133292 PMID: 26308334 6. Lynch RC, Vergara D, Tittes S, White K, Schwartz CJ, Gibbs MJ, et al. (2016) Genomic and chemical diversity in Cannabis. Critical Reviews in Plant Sciences 35:349–363 7. Henry P (2015) Genome-wide analyses reveal clustering in Cannabis cultivars: the ancient domestica- tion trilogy of a panacea. https://doi.org/10.7287/peerj.preprints.1553v2 8. McGuire P, Robson P, Cubala WJ, Vasile D, Morrison PD, Barron R, et al. (2018) Cannabidiol (CBD) as an adjunctive therapy in schizophrenia: a multicenter randomized controlled trial. Am J Psychiatry 175:225–231 https://doi.org/10.1176/appi.ajp.2017.17030325 PMID: 29241357 9. French J, Thiele E, Mazurkiewicz-Beldzinska M, Benbadis S, Marsh E, Joshi C, et al. (2017) Cannabi- diol (CBD) significantly reduces drop seizure frequency in Lennox-Gastaut syndrome (LGS): results of a multi-center, randomized, double-blind, placebo controlled trial (GWPCARE4)(S21.001). Neurology 88:S21–001 10. Avraham Y, Grigoriadis NC, Poutahidis T, Vorobiev’ L, Magen I, Ilan Y, et al. (2011) Cannabidiol improves brain and liver function in a fulminant hepatic failure-induced model of hepatic encephalopathy in mice. British journal of pharmacology 162:1650–1658 https://doi.org/10.1111/j.1476-5381.2010. 01179.x PMID: 21182490 11. Bloomfield MAP, Green SF, Hindocha C, et al (2020) The effects of acute cannabidiol on cerebral blood flow and its relationship to memory: An arterial spin labelling magnetic resonance imaging study. Jour- nal of psychopharmacology (Oxford, England) 269881120936419 https://doi.org/10.1177/ 0269881120936419 PMID: 32762272 PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 12 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains 12. Upton R, Craker L, ElSohly M, Romm A, Russo E, Sexton M (2014) Cannabis inflorescence: cannabis spp.; standards of identity, analysis, and quality control. American Herbal Pharmacopoeia, Scotts Val- ley, CA https://doi.org/10.4081/ijfs.2020.8581 PMID: 32913724 13. Kovalchuk I, Pellino M, Rigault P, et al (2020) The genomics of cannabis and its close relatives. Annu Rev Plant Biol 71:713–739 https://doi.org/10.1146/annurev-arplant-081519-040203 PMID: 32155342 14. Sakamoto K, Akiyama Y, Fukui K, Kamada H, Satoh S (1998) Characterization; Genome Sizes and Morphology of Sex Chromosomes in Hemp (Cannabis sativa L.). Cytologia 63:459–464 15. Shastry BS (2007) SNPs in disease gene mapping, medicinal drug development and evolution. Journal of Human Genetics 52:871–880 https://doi.org/10.1007/s10038-007-0200-z PMID: 17928948 16. Sachidanandam R, Weissman D, Schmidt SC, et al (2001) A map of human genome sequence varia- tion containing 1.42 million single nucleotide polymorphisms. Nature 409:928–933 https://doi.org/10. 1038/35057149 PMID: 11237013 17. Casci T (2010) SNPs that come in threes. Nature Reviews Genetics 11:8–8 https://doi.org/10.1038/ nrg2725 PMID: 20050277 18. Westen AA, Matai AS, Laros JFJ, Meiland HC, Jasper M, de Leeuw WJF, et al. (2009) Tri-allelic SNP markers enable analysis of mixed and degraded DNA samples. Forensic Science International: Genet- ics 3:233–241 19. Kalbfleisch T, Petersen JL, Tait RG Jr., Qiu J, Basnayake V, Hackett PH, et al. (2020) Using triallelic SNPs for determining parentage in North American yak (Bos grunniens) and estimating cattle (B. tau- rus) introgression. F1000Res 9:1096 https://doi.org/10.12688/f1000research.25803.2 PMID: 33163159 20. Phillips C, Amigo J, Tillmar AO, et al (2020) A compilation of tri-allelic SNPs from 1000 Genomes and use of the most polymorphic loci for a large-scale human identification panel. Forensic Science Interna- tional: Genetics 46:102232 https://doi.org/10.1016/j.fsigen.2020.102232 PMID: 31986343 21. Henry P, Khatodia S, Kapoor K, et al (2020) A single nucleotide polymorphism assay sheds light on the extent and distribution of genetic diversity, population structure and functional basis of key traits in culti- vated north American cannabis. Journal of Cannabis Research 2:26 https://doi.org/10.1186/s42238- 020-00036-y PMID: 33526123 22. Vergara D, Baker H, Clancy K, Keepers KG, Mendieta JP, Pauli CS, et al. (2016) Genetic and genomic tools for Cannabis sativa. Critical Reviews in Plant Sciences 35:364–377 23. Laverty KU, Stout JM, Sullivan MJ, et al (2019) A physical and genetic map of Cannabis sativa identifies extensive rearrangements at the THC/CBD acid synthase loci. Genome Res 29:146–156 https://doi. org/10.1101/gr.242594.118 PMID: 30409771 24. Grassa CJ, Wenger JP, Dabney C, Poplawski SG, Motley ST, Michael TP, et al. (2018) A complete Cannabis chromosome assembly and adaptive admixture for elevated cannabidiol (CBD) content. bioRxiv 458083 25. 26. 27. Jenkins C, Orsburn B (2019) Constructing a Draft Map of the Cannabis Proteome. bioRxiv 577635 Jenkins C, Orsburn B (2019) The Cannabis Multi-Omics Draft Map Project. bioRxiv 753400 Jenkins C, Orsburn B (2019) The First Publicly Available Annotated Genome for Cannabis plants. bioR- xiv 786186 28. Benevenuto J, Ferrão LFV, Amadeu RR, Munoz P (2019) How can a high-quality genome assembly help plant breeders? Gigascience. https://doi.org/10.1093/gigascience/giz068 PMID: 31184361 29. Jombart T, Devillard S, Balloux F (2010) Discriminant analysis of principal components: a new method for the analysis of genetically structured populations. BMC Genetics 11:94 https://doi.org/10.1186/ 1471-2156-11-94 PMID: 20950446 30. Henry P (2017) Cannabis chemovar classification: terpenes hyper-classes and targeted genetic mark- ers for accurate discrimination of flavours and effects. PeerJ Preprints 5:e3307v1 31. Henry P (2018) The genetic basis of the human-cannabis relationship. bioRxiv 287938 32. 33. Jin D, Jin S, Yu Y, Lee C, Chen J (2017) Classification of Cannabis Cultivars Marketed in Canada for Medical Purposes by Quantification of Cannabinoids and Terpenes Using HPLC-DAD and GC-MS. J Anal Bioanal Tech 8:2 Jin D, Dai K, Xie Z, Chen J (2020) Secondary Metabolites Profiled in Cannabis Inflorescences, Leaves, Stem Barks, and Roots for Medicinal Purposes. Scientific Reports 10:3309 https://doi.org/10.1038/ s41598-020-60172-6 PMID: 32094454 34. ASM23057v4—Genome—Assembly—NCBI. https://www.ncbi.nlm.nih.gov/assembly/GCA_ 000230575.4/. 35. Li H, Durbin R (2010) Fast and accurate long-read alignment with Burrows–Wheeler transform. Bioinfor- matics 26:589–595 https://doi.org/10.1093/bioinformatics/btp698 PMID: 20080505 PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 13 / 14 PLOS ONEGenome-wide single nucleotide polymorphisms to classify cannabis strains 36. ASM23057v5—Genome—Assembly—NCBI. https://www.ncbi.nlm.nih.gov/assembly/GCA_ 000230575.5. 37. Jombart T (2008) adegenet: a R package for the multivariate analysis of genetic markers. Bioinformat- ics 24:1403–1405 https://doi.org/10.1093/bioinformatics/btn129 PMID: 18397895 38. R Core Team (2018) R: A language and environment for statistical computing. R Foundation for Statis- tical Computing, Vienna, Austria. https://www.R-project.org/. 39. Liu N, Zhao H (2006) A non-parametric approach to population structure inference using multilocus genotypes. Human Genomics 2:353 https://doi.org/10.1186/1479-7364-2-6-353 PMID: 16848973 40. Konishi S, Kitagawa G (2008) Information Criteria and Statistical Modeling. https://doi.org/10.1007/978- 0-387-71887-3 41. Deperi SI, Tagliotti ME, Bedogni MC, Manrique-Carpintero NC, Coombs J, Zhang R, et al. (2018) Dis- criminant analysis of principal components and pedigree assessment of genetic diversity and population structure in a tetraploid potato panel using SNPs. PLoS One 13:e0194398 https://doi.org/10.1371/ journal.pone.0194398 PMID: 29547652 42. Cannabis sativa Annotation Report. https://www.ncbi.nlm.nih.gov/genome/annotation_euk/Cannabis_ sativa/100/. 43. Saitou N, Nei M (1987) The neighbor-joining method: a new method for reconstructing phylogenetic trees. Mol Biol Evol 4:406–425 https://doi.org/10.1093/oxfordjournals.molbev.a040454 PMID: 3447015 44. Ward JH Jr (1963) Hierarchical Grouping to Optimize an Objective Function. Journal of the American Statistical Association 58:236–244 45. Clarke RC, Merlin MD (2016) Cannabis domestication, breeding history, present-day genetic diversity, and future prospects. Critical Reviews in Plant Sciences 35:293–327 46. de Meijer EPM, Bagatta M, Carboni A, Crucitti P, Moliterni VMC, Ranalli P, et al. (2003) The inheritance of chemical phenotype in Cannabis sativa L. Genetics 163:335–346 PMID: 12586720 47. Stout JM, Boubakir Z, Ambrose SJ, Purves RW, Page JE (2012) The hexanoyl-CoA precursor for can- nabinoid biosynthesis is formed by an acyl-activating enzyme in Cannabis sativa trichomes. The Plant Journal. https://doi.org/10.1111/j.1365-313x.2012.04949.x PMID: 22353623 48. Gagne SJ, Stout JM, Liu E, Boubakir Z, Clark SM, Page JE (2012) Identification of olivetolic acid cyclase from Cannabis sativa reveals a unique catalytic route to plant polyketides. Proceedings of the National Academy of Sciences 109:12811–12816 https://doi.org/10.1073/pnas.1200330109 PMID: 22802619 49. Van Bakel H, Stout JM, Cote AG, Tallon CM, Sharpe AG, Hughes TR, et al. (2011) The draft genome and transcriptome of Cannabis sativa. Genome biology 12:R102 https://doi.org/10.1186/gb-2011-12- 10-r102 PMID: 22014239 50. Zager JJ, Lange I, Srividya N, Smith A, Lange BM (2019) Gene Networks Underlying Cannabinoid and Terpenoid Accumulation in Cannabis. Plant Physiology 180:1877–1897 https://doi.org/10.1104/pp.18. 01506 PMID: 31138625 51. Booth JK, Page JE, Bohlmann J (2017) Terpene synthases from Cannabis sativa. PLOS ONE 12: e0173911 https://doi.org/10.1371/journal.pone.0173911 PMID: 28355238 52. Piluzza G, Delogu G, Cabras A, Marceddu S, Bullitta S (2013) Differentiation between fiber and drug types of hemp (Cannabis sativa L.) from a collection of wild and domesticated accessions. Genetic resources and crop evolution 60:2331–2342 53. Datwyler SL, Weiblen GD (2006) Genetic variation in hemp and marijuana (Cannabis sativa L.) accord- ing to amplified fragment length polymorphisms. J Forensic Sci 51:371–375 https://doi.org/10.1111/j. 1556-4029.2006.00061.x PMID: 16566773 54. Gilmore S, Peakall R, Robertson J (2003) Short tandem repeat (STR) DNA markers are hypervariable and informative in Cannabis sativa: implications for forensic investigations. Forensic Sci Int 131:65–74 https://doi.org/10.1016/s0379-0738(02)00397-3 PMID: 12505473 55. Hakki EE, Kayis SA, Pinarkara E, Sag A (2007) Inter simple sequence repeats separate efficiently hemp from marijuana (Cannabis sativa L.). Electronic Journal of Biotechnology 10:570–581 56. Hilyard A, Lewin S, Johnson S, Henry P, Orser C (2019) Application of a Simple Genetic Assay to Discriminate Hemp from Drug-Type Cannabis. In: Cannabis Science Tech. https://www. cannabissciencetech.com/view/application-simple-genetic-assay-discriminate-hemp-drug-type- cannabis. PLOS ONE \| https://doi.org/10.1371/journal.pone.0253387 June 28, 2021 14 / 14 PLOS ONE	null