PHbinder-PSGM

PHbinder & PSGM: A Peptide-HLA Interaction Framework

Overview

This is a cascaded framework for peptide-HLA binding prediction (PHBinder) and HLA pseudo-sequence generation (PSGM). The PHbinder model is designed to predict the binding affinity between peptides and their corresponding MHC molecules (HLA). The PSGM model generates candidate HLA-I pseudo-sequences based on a given peptide and maps them to a corresponding list of HLA-I alleles.

Model Replication

First, clone this repository to your local machine:

git clone https://github.com/Zekkenwang/PHbinder-PSGM.git
cd PHbinder-PSGM

Setting Up the Python Environment

It is highly recommended to use Conda or venv to create an isolated Python virtual environment to avoid package conflicts.

1. Create a virtual environment using Conda:

   conda create -n ph_binding_env python=3.9  # Python 3.9 or higher is recommended
   conda activate ph_binding_env
   

2. Install dependencies:

   pip install -r requirements.txt
   

Model Training and Usage

Downloading Pre-trained Models

To skip the training process and use our pre-trained models directly for prediction, please download the model weights from our Hugging Face Hub repository:

https://huggingface.co/Zekkenwang/PHbinder-PSGM

After downloading, please organize the files into the models/ directory of your cloned repository, following this structure:

PHbinder-PSGM/
└── models/
    ├── esm2_t30_150_UR50D/
    ├── phbinder/
    ├── psgm/
    └── phbinder_pretrain/

Note: If the models directory or its subdirectories (phbinder, psgm, phbinder_pretrain) do not exist, please create them manually before placing the downloaded files.

PHbinder: Training and Usage

This section provides detailed instructions on how to replicate and use the PHbinder model. PHbinder includes a pre-training phase using ESM2+LoRA, which significantly improves training efficiency and model performance.

Training the PHbinder Model

To start the training process, run the following command:

python scripts/phbinder.py

Expected Output:

• The pre-trained model weights will be saved to PHbinder-PSGM/models/phbinder_pretrain.
• The final PHbinder model weights will be saved to PHbinder-PSGM/models/phbinder.
• After the script finishes, the final evaluation metrics (Accuracy, F1 Score, Recall, Precision, MCC, AUC) on the test set will be printed to the console.

Predicting Epitopes with the PHbinder Model

To predict whether a peptide is an epitope using the pre-trained model, you must first download the model weights and place them in the specified models/phbinder/ and models/phbinder_pretrain/ directories.

Then, run:

python scripts/predict_phbinder.py

This will evaluate the model on an external dataset. We have provided an example evaluation using the dataset in external_data. If you wish to use your own dataset, please update the EXTERNAL_DATA_PATH and PREDICTION_OUTPUT_PATH in the phbinder_config to your desired input and output paths.

PSGM: Training and Usage

Training the PSGM Model

To train the PSGM model, execute the following command:

python scripts/train_psgm.py

Expected Output:

• The PSGM model weights will be saved to PHbinder-PSGM/models/psgm.

Generating HLA-I Pseudo-Sequences with the PSGM Model

Use the pre-trained PSGM model to generate HLA-I pseudo-sequences from a given list of peptides and map them to a corresponding list of HLA Alleles. You must first download the model weights and place them in the specified models/psgm/ directory.

Run the following script for generation:

python scripts/predict_psgm.py

This will generate predictions on an external dataset, with an example provided in external_data. To evaluate on your own data, modify the EXTERNAL_DATA_PATH and PREDICTION_OUTPUT_PATH in the psgm_config file to your specific input and output file paths.