PMBind

PMBind is a deep learning framework for predicting peptide-MHC binding interactions using a multi-task transformer architecture.

⚠️ Note: This project is under active development. Features and documentation may change.

Overview

PMBind predicts peptide binding to Major Histocompatibility Complex (MHC) molecules using:

• Multi-task learning (binding prediction + sequence reconstruction)
• 2D-masked cross-attention transformer
• ESM-C600m protein language model embeddings
• Physicochemical peptide encoding

Installation

git clone https://github.com/Amirreza-0/PMBind.git
cd PMBind/
pip install -r requirements.txt

Requirements: Python 3.10+, TensorFlow, CUDA-compatible GPU (recommended)

Data

Download the dataset from the link in the data/ folder:

• Main Dataset

Extract to the data/ directory.

Quick Start

1. Prepare Data

cd src/
python create_dataset.py
python preprocessing/run_ESM.py --input ../data/PMDb_alleles_sequences_aligned.csv \
                                --model esmc_600m --device cuda:0
python create_tfrecords.py

2. Train Model

python run_training.py --fold 1

3. Run Inference

python run_inference.py --model_weights best_model.weights.keras \
                       --test_data ../data/test.parquet

Project Structure

PMBind/
├── src/              # Source code
│   ├── models.py     # Model architecture
│   ├── run_training.py
│   ├── run_inference.py
│   └── preprocessing/
├── data/             # Dataset directory
├── tests/            # Tests
└── requirements.txt

Key Scripts

• run_training.py - Train the model
• run_inference.py - Run predictions

Model Architecture

PMBind uses a transformer with:

• Peptide encoding: BLOSUM62 + physicochemical properties (14D)
• MHC encoding: ESM-C embeddings (1152D)
• 2D-masked cross-attention (prevents self-attention, enables cross-attention)
• Multi-task heads: binding prediction, peptide reconstruction, MHC reconstruction

Citation

@software{pmBind2025,
  author = {Aleyasin, Amirreza},
  title = {PMBind: Multi-Task Deep Learning for Peptide-MHC Binding Prediction},
  year = {2025},
  url = {https://github.com/Amirreza-0/PMBind},
  orcid = {0000-0003-2742-7138}
}

Author: Amirreza Aleyasin (ORCID: 0000-0003-2742-7138)

License

Apache License 2.0 - see LICENSE file

Contact

• Email: amirreza.alise@gmail.com
• GitHub: @Amirreza-0