SPRINT: Sequence-based Immunogenicity Prediction Networks

📄 RECOMB 2025 Supplementary Material: The file appendix_recomb.pdf contains the supplementary materials for our RECOMB submission.

A unified PyTorch-based benchmarking framework for deep learning methods in T-cell receptor (TCR) and peptide-MHC (pMHC) binding prediction.

Table of Contents

• Quick Start
• Available Resources
• Usage
• Citation

Quick Start

Installation

git clone https://github.com/Computational-Machine-Intelligence/SPRINT.git
cd SPRINT
pip install -r requirements.txt

Requirements: Python >= 3.8, PyTorch >= 2.0.0

Set HuggingFace Token (Required)

# Linux/Mac
export HF_TOKEN="your_huggingface_token"

# Windows PowerShell
$env:HF_TOKEN = "your_huggingface_token"

Available Resources

Methods (7)

• pmtnet: Multi-head attention for pMHC-TCR binding
• piste: Pre-trained immune system transformer encoder
• fusionpmt: Fusion model for pMHC-TCR interactions
• fusionpm: Peptide-MHC binding prediction
• transphla: Transformer for pHLA binding
• ergo2: LSTM/Autoencoder TCR specificity model
• nettcr: CNN-based TCR-pMHC prediction

Datasets (6)

• pmt: Large-scale pMHC-TCR training set
• pm: Peptide-MHC binding dataset
• pt: Peptide-TCR binding dataset
• allelic_ood: Out-of-distribution test (unseen alleles)
• modality_ood: Cross-modality test (BA/EL assays)
• temporal_ood: Temporal test (post-2021 data)

Modes (3)

• train: Train models from scratch
• eval: Evaluate models on test data
• both: Train then evaluate

Usage

1. Evaluate Pre-trained Models

Evaluate a pre-trained model on test data:

# Evaluate on standard dataset
python scripts/run_benchmark.py --method pmtnet --dataset pmt --mode eval --pretrain

# Evaluate on OOD datasets
python scripts/run_benchmark.py --method transphla --dataset temporal_ood --mode eval --pretrain
python scripts/run_benchmark.py --method fusionpm --dataset modality_ood --mode eval --pretrain

What happens:

• Downloads pre-trained model from HuggingFace (first time only)
• Loads test data automatically
• Evaluates and saves results to outputs/<method>/pre_train_results/evaluations/

2. Train Models from Scratch

Train a model on a dataset:

# Basic training
python scripts/run_benchmark.py --method pmtnet --dataset pmt --mode train

# Train with custom config
python scripts/run_benchmark.py --method ergo2 --dataset pt --mode train --config configs/methods/ergo2.yaml

Output: Trained model saved to outputs/<method>/<dataset>_<timestamp>/

3. Train and Evaluate

Train a model and immediately evaluate:

python scripts/run_benchmark.py --method fusionpmt --dataset pmt --mode both

4. List Available Resources

# List all methods
python scripts/run_benchmark.py --list-methods

# List all datasets
python scripts/run_benchmark.py --list-datasets

Command Options

Option	Description	Example
--method	Model to use	pmtnet, ergo2, piste
--dataset	Dataset to use	pmt, temporal_ood
--mode	Operation mode	train, eval, both
--pretrain	Use pre-trained model (eval only)	Flag, no value
--device	Computing device	cuda, cpu, auto
--seed	Random seed	42

Citation

If you use SPRINT in your research, please cite:

@software{yin2025sprint,
  title={SPRINT for Benchmarking Sequence-based Immunogenicity Prediction Networks},
  author={Yin, Yujia and Li, Hongzong and Ma, Jiahao and Chen, Weijia and Yu, Yingying and Zhang, Xiaoyuan and Qu, Tianyi and Wu, Xinhong and Li, Junyi and Huang, Jian-Dong and Hu, Ye-Fan and Chen, Yifan},
  year={2025},
  url={https://github.com/Computational-Machine-Intelligence/SPRINT}
}

License

This project is licensed under the MIT License.

Contact

• GitHub: https://github.com/Computational-Machine-Intelligence/SPRINT
• Issues: https://github.com/Computational-Machine-Intelligence/SPRINT/issues