Synthyra/ecoli_holdout_ppi_large

Clustered PPI datasets (BIOGRID + STRING) with sequence-disjoint splits

This dataset repo contains multiple dataset variants of protein–protein interactions (PPIs), built by clustering proteins by sequence similarity and then constructing train/valid/test splits that are intended to be disjoint at the protein level (and thus hard to memorize via near-identical sequences).

Artifacts are stored as compressed pickles (*.pkl.gz). A helper downloader exists in this repo:

• data_processing/download_ppi_data.py::download_clustered_ppi_data

What’s in each split dataframe?

Each split is a pandas.DataFrame with (at minimum):

• IdA / IdB: protein identifiers
• OrgA / OrgB: organism identifiers (STRING taxon id for STRING datasets; BIOGRID org id for BIOGRID datasets)
• labels: >0 indicates a positive interaction, 0 indicates a sampled negative

Some variants also include additional columns (e.g. cluster_a, cluster_b, confidences, org_a, org_b). When negatives are concatenated, some of these columns may be NaN for negative rows.

Dataset variants (index)

A machine-readable index is available at:

• tables/dataset_index.csv

variant	source	threshold	train rows	valid rows	test rows	train pos rate	protein overlap (max)
ecoli_holdout_st030	ecoli_holdout	st030	185132088	201460	976732	0.500	0

Per-variant deep dive (plots + stats)

Each variant has:

• plots/<variant>/...png (rendered below)
• tables/<variant>/summary.csv and tables/<variant>/schema.csv

ecoli_holdout_st030

Open report

Summary tables

• tables/ecoli_holdout_st030/summary.csv
• tables/ecoli_holdout_st030/schema.csv

Label balance

• train: plots/ecoli_holdout_st030/train_label_counts.png
• valid: plots/ecoli_holdout_st030/valid_label_counts.png
• test: plots/ecoli_holdout_st030/test_label_counts.png

Organism distributions (positives vs negatives)

• data: plots/ecoli_holdout_st030/train_organism_distribution.csv
• stats: plots/ecoli_holdout_st030/train_organism_distribution_stats.csv

• data: plots/ecoli_holdout_st030/valid_organism_distribution.csv
• stats: plots/ecoli_holdout_st030/valid_organism_distribution_stats.csv

• data: plots/ecoli_holdout_st030/test_organism_distribution.csv
• stats: plots/ecoli_holdout_st030/test_organism_distribution_stats.csv

Cross-split organism shift tests

• positives: plots/ecoli_holdout_st030/cross_split_pos_stats.csv
• negatives: plots/ecoli_holdout_st030/cross_split_neg_stats.csv

Sequence length distributions (unique proteins)

• stats: plots/ecoli_holdout_st030/train_seq_length_stats.csv

• stats: plots/ecoli_holdout_st030/valid_seq_length_stats.csv

• stats: plots/ecoli_holdout_st030/test_seq_length_stats.csv

Top organism pairs

• train positives: plots/ecoli_holdout_st030/train_top_org_pairs_pos.png

• train negatives: plots/ecoli_holdout_st030/train_top_org_pairs_neg.png

• valid positives: plots/ecoli_holdout_st030/valid_top_org_pairs_pos.png

• valid negatives: plots/ecoli_holdout_st030/valid_top_org_pairs_neg.png

• test positives: plots/ecoli_holdout_st030/test_top_org_pairs_pos.png

• test negatives: plots/ecoli_holdout_st030/test_top_org_pairs_neg.png

How to download and load

Use the helper in this codebase:

from data_processing.download_ppi_data import download_clustered_ppi_data

# BIOGRID example
train_df, valid_df, test_df, interaction_set, seq_dict = download_clustered_ppi_data(
    data_type='biogrid',
    cluster_percentage=0.5,
    hf_repo='Synthyra/ecoli_holdout_ppi_large',
)

# STRING example (descriptor must match the variant prefix: e.g. 'human' or 'model_orgs')
train_df, valid_df, test_df, interaction_set, seq_dict = download_clustered_ppi_data(
    data_type='string',
    descriptor='human',
    cluster_percentage=0.5,
    hf_repo='Synthyra/ecoli_holdout_ppi_large',
)