SimBinder-IF: Structure-Aware Antibody Affinity Optimization via Efficient Preference Learning

SimBinder-IF is a framework for antibody affinity maturation. It leverages an inverse folding protein model (ESM-IF) and an efficient preference learning strategy (SimPO) to optimize antibody binding affinity based on complex structures.

🛠️ Installation

Prerequisites

• Python 3.8+
• PyTorch 1.12+
• CUDA enabled GPU

Setup

1. Clone the repository:

   git clone https://github.com/MSBMI-SAFE/SimBinder-IF.git
   cd SimBinder-IF

2. Install dependencies:

   pip install torch torchvision torchaudio
   pip install biotite pandas scikit-learn
   # Install ESM-IF official library
   pip install git+https://github.com/facebookresearch/esm.git

📂 Project Structure

• train_esmif.py: Main training script for preference-based fine-tuning.
• data_utils.py: Data pipeline for PDB structure processing and mutation sequence mapping.
• util.py: Utility functions for loss calculation (SimPO/DPO) and correlation metrics.

📊 Data Preparation

The pipeline expects a metadata.json file pointing to PDB structures and affinity CSVs. Each CSV should contain:

• heavy_chain_seq: Mutated antibody heavy chain sequence.
• light_chain_seq: Mutated antibody light chain sequence.
• binding_score: Measured affinity value (e.g., $K_d$, $\Delta G$).

🏋️ Training

To start fine-tuning with SimPO loss:

python train_esmif.py \
    --meta_json path/to/metadata.json \
    --out_dir checkpoints \
    --batch_size 32 \
    --lr 1e-4 \
    --beta 0.1 \
    --margin 0.1 \
    --higher_is_better True

📥 Model Checkpoints

Fine-trained model checkpoints are available for download:

• Google Drive: Download Checkpoints

📝 Citation

If you find this work helpful for your research, please cite our paper:

@article{zhao2025simbinderif,
  title={SimBinder-IF: Structure-Aware Antibody Affinity Optimization via Efficient Preference Learning},
  author={Zhao, Xinyan and Tang, Yi-Ching and Monsia, Rivaaj and Cantu, Victor J and Ramesh, Ashwin Kumar and Liu, Xiaozhong and An, Zhiqiang and Jiang, Xiaoqian and Kim, Yejin},
  journal={arXiv preprint arXiv:2512.17815},
  year={2025}
}