---
license: other
tags:
  - rna-splicing
  - splice-site-prediction
  - bioinformatics
---

# oncosplice model weights

Weight files for the splice-site engines used by
[oncosplice](https://github.com/nicolasalynn/oncosplice). One sub-folder per
model; oncosplice downloads them on first use.

## Provenance & licensing of the trained weights

| sub-folder | model | weights © / license |
|---|---|---|
| `openspliceai/` | OpenSpliceAI (MANE) | OpenSpliceAI authors — see upstream |
| `spliceai_pytorch/` | SpliceAI weights (Keras→PyTorch) | Jaganathan et al. 2019 (Illumina) |
| `spliceai_keras/` | SpliceAI original Keras `.h5` | Jaganathan et al. 2019 (Illumina) |
| `pangolin/` | Pangolin 40-model ensemble | Zeng & Li 2022 — **GPL-3.0** |
| `spliceformer/` | Spliceformer 45k base ensemble | Jónsson et al. 2024 — MIT |

These weights are redistributed for **non-commercial research use** under their
respective original terms. The Pangolin weights are GPL-3.0 (© Zeng & Li 2022);
the oncosplice architecture code that runs them is an independent MIT
re-implementation. The Spliceformer base ensemble (`transformer_encoder_45k_171022_*`)
is the general-purpose model — the blood-finetuned variants are intentionally
not included.