metadata
license: mit
tags:
- protein
- binding-affinity
- deep-learning
- esm
- pytorch
language:
- en
🧬 Protein Binding Affinity Predictor
Dual-head model for predicting protein-protein binding affinity (ΔG) and mutation effects (ΔΔG).
Model Performance
| Metric | Validation Score |
|---|---|
| dG Pearson | 0.51 |
| ddG Pearson | 0.70 |
| Sum PCC | 1.21 |
Architecture
- Backbone: ESM-600M (frozen embeddings)
- Pooling: Sliced-Wasserstein Embedding (SWE)
- Heads: Dual-head (dG + ddG)
- Input: Protein sequences (1153-dim = 1152 ESM + 1 mutation channel)
Usage
from huggingface_hub import hf_hub_download
import torch
# Download checkpoint
ckpt = hf_hub_download(repo_id="supanthadey1/protein-binding-affinity", filename="best_model_checkpoint.pt")
checkpoint = torch.load(ckpt, map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
Predictions
- ΔG (kcal/mol): Binding free energy. More negative = stronger binding.
- ΔΔG (kcal/mol): Mutation effect. Negative = stabilizing, Positive = destabilizing.
Training Data
Trained on multiple datasets including SKEMPI, BindingGym, PDBbind, and others.
Citation
If you are using Seq2Bind / TwinPeaks please cite both papers as:
Ma, X., Dey, S., Zelinski, C., Li, Q. and Chowdhury, R., 2025. Seq2Bind webserver for binding site prediction from sequences using fine-tuned protein language models. NAR Genomics and Bioinformatics, 7(4), p.lqaf154.
Dey, S. and Chowdhury, R., 2025. Twin Peaks: Dual-Head Architecture for Structure-Free Prediction of Protein-Protein Binding Affinity and Mutation Effects. arXiv preprint arXiv:2509.22950.