GeoPep

Geometric-aware Peptide-Protein Binding Site Prediction

GeoPep is a per-residue binding site predictor that combines the ESM3 protein foundation model with Kolmogorov-Arnold Network (KAN) heads. Given a peptide and a protein, it predicts which residues of the protein bind the peptide and which residues of the peptide make contact.

• 📦 Code: https://github.com/Dian0212/GeoPep
• 📄 Checkpoint: model_distanceLoss.ckpt (~16 GB)

Model Summary

Backbone	ESM3 (sm-open-v1, 1.4B params)
Head	5 stacked FastKAN layers, 1536 → 1153 → 770 → 387 → 3 → 3
Granularity	Per-residue (one prediction per amino acid)
Classes	3 — 0: non-interface, 1: interface (binding), 2: padding
Input	Peptide (≤50 residues) `
Output	Logits [B, 3, 551] → softmax binding probability per residue
Training loss	Weighted cross-entropy + differentiable geometric distance loss

How to Use

1. Install the GeoPep code

git clone https://github.com/Dian0212/GeoPep.git
cd GeoPep
conda create -n geopep python=3.10 -y
conda activate geopep
pip install -r requirements.txt

2. Download the checkpoint from this repo

huggingface-cli download dchenqwer/GeoPep model_distanceLoss.ckpt \
    --local-dir model_weights/ --local-dir-use-symlinks False

This places the file at model_weights/model_distanceLoss.ckpt.

3. Run inference on your PDBs

Put PDB files (named <PDBID>_<peptide_chain>_<protein_chain>.pdb) into a folder, then:

cd scripts
python inference_pipeline.py \
    --pdb-dir /path/to/pdb \
    --checkpoint ../model_weights/model_distanceLoss.ckpt

Output: result/predictions.json with per-residue binding probabilities.

4. Result format

{
  "1a1r_C_A": {
    "peptide_chain": "GSVVIVGRIVLSGKPA",
    "protein_chain": "VEGEVQIVSTATQTFLAT...",
    "peptide_bindingProbability": "0.99 0.99 0.99 ...",
    "protein_bindingProbability": "0.35 0.97 0.12 ..."
  }
}

• peptide_chain / protein_chain — actual residue sequences (padding stripped).
• *_bindingProbability — space-separated 2-decimal floats, one per residue. Value is the raw class-1 (interface) softmax probability.

Architecture

input tokens [B, 553]                  BOS + 551 residues + EOS
       │
   ESM3 encoder
       ↓ embeddings[:, 1:552, :]       drop BOS / EOS
   [B, 551, 1536]                      per-residue embeddings
       ↓ reshape
   [B * 551, 1536]                     each residue independent
       ↓
   KAN_1: 1536 → 1153
   KAN_2: 1153 → 770
   KAN_3:  770 → 387
   KAN_4:  387 → 3
   KAN_5:    3 → 3
       ↓ reshape + permute
   logits [B, 3, 551]
       ↓ softmax(dim=1)
   binding probability = softmax[:, 1, :]

The per-residue head is what makes GeoPep different from the original CLS-token approach — every residue's embedding flows independently through the KAN stack, giving sharper position-level predictions.

Training

• Dataset: peptide–protein complexes from the PDB, encoded as paired complex/ (full structure) + interface/ (binding residues only) PDB files.
• Length filters: peptide ∈ [10, 50] residues, protein ∈ [10, 500] residues.
• Loss:
  • Per-half cross-entropy with class weights [0.2, 0.8, 0.0] (padding contributes zero gradient).
  • Differentiable distance loss L_dist = Σᵢ P_binding(i) · dist(i) / num_valid_residues, which penalizes high binding probability at residues far from the true interface (distance computed from residue centers of mass).
• Backbone: ESM3 backbone is fine-tuned together with the KAN head.
• Optimizer: AdamW, learning rate 1e-4, weight decay 1e-4.
• Mixed precision: FP16.

Input Format

The model expects a single concatenated string of length 551 tokens:

PEPTIDESEQ<pad><pad>...|PROTEINSEQ<pad><pad>...
|<------- 50 ------->|<-------- 500 ----------->|

Position	Content
0 .. 49	Peptide (left-padded with <pad> to 50)
50	Separator `
51 .. 550	Protein (left-padded with <pad> to 500)

The inference_pipeline.py script handles padding and tokenization for you — you only provide raw PDB files.

Limitations

• Sequence-length cap: peptide must be ≤ 50 residues, protein ≤ 500 residues. Sequences longer than this are silently skipped at the preprocessing stage.
• Requires per-PDB chain annotation: filenames must follow the <PDBID>_<peptide_chain>_<protein_chain>.pdb convention so the pipeline knows which chain is the peptide.
• No uncertainty calibration: raw softmax probabilities are not calibrated; use them as relative scores rather than absolute confidences.
• Single peptide-protein pair per call: multi-peptide or multi-chain contexts are not supported in the standard pipeline.

Citation

If you use GeoPep in your work, please cite:

@misc{geopep2026,
  title  = {GeoPep: Geometric-aware Peptide-Protein Binding Site Prediction},
  author = {Chen, Dian},
  year   = {2026},
  howpublished = {\url{https://github.com/Dian0212/GeoPep}}
}

License

MIT