ProteinAffinityKD:基于知识蒸馏的蛋白质结合亲和力预测项目

可用于蛋白质-蛋白质结合亲和力预测,支持外部评估和实际应用。通过知识蒸馏技术,利用结构信息教师网络和序列学生网络,仅需蛋白质序列即可实现预测,无需结构信息。【此简介由AI生成】

Branch1Tags0
FilesLast commitLast update
10 months ago
10 months ago
10 months ago
10 months ago
10 months ago
10 months ago

Knowledge Distillation–Based Protein–Protein Binding Affinity Prediction

(Inference-Only Usage Guide)

This repository provides an inference-only implementation of a knowledge distillation–based regression model for predicting protein–protein binding affinities from protein sequence information alone. The model has been trained using a structure-informed teacher network and a sequence-based student network, enabling the transfer of structural knowledge during training while retaining sequence-only applicability at test time. This implementation is intended for external evaluation and practical use, consistent with the methodological standards of Bioinformatics and Briefings in Bioinformatics.


Method Summary

The released model corresponds to the best-performing distillation configuration identified under Leave-One-Complex-Out (LOCO) cross-validation, using:

  • Teacher network: structure-based descriptors (Moal features)
  • Student network: sequence-based descriptors (Propy features)

At inference time, only sequence-based features are required. No structural information is needed for the test complexes.


Repository Scope

This repository supports:

  • Testing the pre-trained distillation-based student model
  • Predicting binding affinities for new protein–protein complexes

Model training and hyperparameter optimization are not included, as these are fully described in the associated manuscript and supplementary material.


Directory Structure

  • model/ - Contains Pre-trained student model (PyTorch).
    • student_with_diss_weights.pth - trained model.
    • scaler_labels.pkl - StandarScaler for labels.
    • caler_propy.pkl - StandarScaler for features.
  • example/ - Example FASTA files (ligand & receptor).
  • src/ - python files.
    • feature_extraction.py - Propy feature extraction.
    • predict.py - Model inference script.
  • README.md - The file you are reading now.

Software Requirements

  • Python (version: 3.13.11)
  • PyTorch (version: 2.7.0)
  • Scikit-learn (version: 1.7.1)
  • NumPy (version: 2.3.5)
  • joblib
  • Propy (protein sequence feature extraction) (pip install propy3)

Input Format

Protein Sequences

  • FASTA format
  • Ligand and receptor chains provided separately

Feature Extraction (Sequence-Based)

  • Sequence features are extracted through Propy by using method define in feature_extraction.py.
  • All feature standardization is performed using training-set statistics only. The fitted normalization parameters were saved and reused for external validation to prevent information leakage.

Running Inference

  • To predict binding affinities for new complexes, use predict.py

Reproducibility Statement

  • The released model corresponds exactly to the configuration reported in the manuscript.
  • Feature normalization parameters are fixed and embedded.
  • No external data is accessed during inference.

Intended Use and Limitations

This model is intended for:

  • Benchmarking
  • External validation
  • Large-scale sequence-based affinity screening

Limitations:

  • Predictive performance is constrained by the size and diversity of available training data
  • Performance does not yet reach that of structure-based predictors, as discussed in the manuscript

Citation

If you use this model or code, please cite: Wajid et al., Investigating knowledge distillation through neural networks for protein binding affinity prediction. preprint, 2026.(http://arxiv.org/abs/2601.03704)


License

MIT License


Contact

For questions related to model usage or evaluation, please email at wajidarshad@uajk.edu.pk.

Introduction

可用于蛋白质-蛋白质结合亲和力预测,支持外部评估和实际应用。通过知识蒸馏技术,利用结构信息教师网络和序列学生网络,仅需蛋白质序列即可实现预测,无需结构信息。【此简介由AI生成】

Customize your domain