README.md

---
license: other
library_name: xgboost
tags:
  - biology
  - protein-protein-interaction
  - protein-language-model
  - xgboost
  - reclip
---

# ReCLIP Release Artifacts

This repository hosts release artifacts for **ReCLIP**:

> Learning residue-level context for modeling protein-protein interactions

The code repository is available at:

https://github.com/SiweiLab/ReCLIP

## Current Contents

```text
classifier_heads/
  mutation/
    reclip_mutation_xgb_best_fold5_by_ROC_AUC_macro_ovr.pkl
    reclip_mutation_xgb_best_fold5_by_ROC_AUC_macro_ovr.json
  ptm/               Placeholder for PTM classifier heads
  peptide_mhc/       Placeholder for peptide-MHC classifier heads

figure_data/
  mutation/          Placeholder for release figure tables
  ptm/               Placeholder for release figure tables
  peptide_mhc/       Placeholder for release figure tables
  interpretation/    Placeholder for interpretation analysis tables

metadata/            Model configuration, feature schema, checksums, and notes
```

## Mutation Classifier Head

The uploaded mutation artifact is the best-fold XGBoost classifier head selected
by macro one-vs-rest AUROC on the four-class mutation effect prediction task.
It expects features produced by the ReCLIP mutation pipeline with the validated
setting: ESM2 layer 32, top-k 5 residues per head, context features with layer
normalization, flattened target representation, and MINT partner
representations.

| Artifact | Description |
| --- | --- |
| `classifier_heads/mutation/reclip_mutation_xgb_best_fold5_by_ROC_AUC_macro_ovr.pkl` | XGBoost classifier head |
| `classifier_heads/mutation/reclip_mutation_xgb_best_fold5_by_ROC_AUC_macro_ovr.json` | Metadata, metrics, XGBoost parameters, and checksum |

Key metadata:

- Best fold: 5
- Selection metric: `ROC_AUC_macro_ovr`
- Best-fold AUROC: 0.973531
- Feature dimension: 7680
- SHA256: `8744392e55c067903ffefcb5fd0862a7b1f971043f211b8f26bb8c87f89cbb16`

## Planned Additions

Additional PTM and peptide-MHC classifier heads, together with larger
reproducibility artifacts, will be added after journal submission.

## Citation

The manuscript is currently in preparation. Citation details will be updated
after the preprint or final publication is available.

```bibtex
@misc{reclip2026,
  title = {Learning residue-level context for modeling protein-protein interactions},
  author = {ReCLIP authors},
  year = {2026},
  note = {Manuscript in preparation}
}
```

## License

License information will be finalized before public release.