Nano-AbNovo is a lightweight and educational version for reproducing the core concepts from the paper "AbNovo: Constrained Preference Optimization for Multi-objective Antibody Design". This project implements a diffusion-based model for antibody design, combined with constrained preference optimization (CPO) to meet multi-objective biophysical property requirements.
- Antibody Design with Diffusion Models: Joint design of antibody structures and sequences using diffusion generative models.
- Constrained Preference Optimization (CPO): Optimizes binding affinity while explicitly constraining key biophysical properties such as non-specific binding, self-association, and stability.
- Primal-Dual Optimization: Dynamically adjusts reward and constraint weights to improve training stability.
- Structure-aware Protein Language Model: Leverages large-scale protein structural data to mitigate overfitting caused by limited antibody-antigen training data.
- Reproducible Results: Includes code and configurations to reproduce the main experiments from the paper.
Working in progress....
@inproceedings{
ren2025multiobjective,
title={Multi-objective antibody design with constrained preference optimization},
author={Milong Ren and ZaiKai He and Haicang Zhang},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=4ktJJBvvUd}
}
This project is inspired by the following works:
- AbNovo: Multi-objective antibody design with constrained preference optimization
- DiffAb: Denoising Diffusion for Antibody Design
- AbX: Score-based Diffusion for Antibody Design
- ABDPO: Direct Preference Optimization for Antibody Binding
This project is licensed under the MIT License.