MFDesign: Antibody Sequence and Structure Co-design

Repurposing AlphaFold3-like Protein Folding Models for Antibody Sequence and Structure Co-design

Accepted at NeurIPS 2025

Overview

MFDesign is a novel approach that repurposes AlphaFold3-style protein folding models for antibody sequence and structure co-design. By adapting state-of-the-art protein structure prediction models, we enable simultaneous optimization of both antibody sequence and structure, providing a powerful tool for antibody engineering and design.

Our method builds upon the open-source Boltz-1 framework, extending its capabilities to handle the unique challenges of antibody design, including CDR region optimization and antigen-antibody interaction modeling.

Installation

Prerequisites

MFDesign requires the same environment as Boltz-1. First, install the Boltz framework:

pip install boltz -U

Setup

Clone the MFDesign repository

git clone https://github.com/yangnianzu0515/MFDesign.git
cd MFDesign

Download Data and Models: We provide all processed data, raw source data, and pre-trained models in a separate repository on Hugging Face Hub. You can find them in the ./data and ./model directories within the corresponding repository. The model repository and data repository are available at MF-Design Model and MF-Design Data, respectively. After downloading the data, you will also need to unzip the compressed files.
Modify the system path in train.py and predict.py to point to your codebase:

import sys
sys.path.insert(0, '/$YOURPATH/MFDesign/src')

Usage

Training

For detailed training instructions, see docs/training.md.

Basic training command:

python scripts/train/train.py scripts/train/configs/stage_1.yaml

Inference

For detailed prediction instructions, see docs/predict.md.

Basic prediction command:

python scripts/predict.py --data <INPUT_PATH> --use_msa_server

Data Preprocessing

We provide both the pre-processed data and the original raw data. For users who wish to run the preprocessing pipeline themselves, please follow the comprehensive instructions in docs/preprocess.md.

Local MSA Generation

For local MSA data processing instructions, see scripts/process/local_msa/note.md.

Paper and Citation

If you use MFDesign in your research, please cite our paper:

@article{MFDesign,
  title={Repurposing AlphaFold3-like Protein Folding Models for Antibody Sequence and Structure Co-design},
  author={Nianzu Yang and Songlin Jiang and Jian Ma and Huaijin Wu and Shuangjia Zheng and Wengong Jin and Junchi Yan},
  journal={NeurIPS 2025},
  year={2025}
}

Acknowledgments

This work is built upon the excellent Boltz-1 framework. We thank the Boltz-1 team for their outstanding contributions to the protein structure prediction community and for making their code openly available.

Welcome to contact us via clorf6@sjtu.edu.cn or majian7@sjtu.edu.cn for any question (the first author Nianzu will not be able to respond to your questions as he is about to start working and will not have much time to continue research and answer questions).

MFDesign

Install / Use

README