AntibodyDesignBFN: High-Fidelity Fixed-Backbone Antibody Design via Discrete Bayesian Flow Networks

TL;DR

AntibodyDesignBFN tackles fixed-backbone antibody design by modeling sequence logits with Discrete Bayesian Flow Networks conditioned on full 3D backbone geometry and antigen context. The framework uses a Geometric Transformer with Invariant Point Attention and a continuous-time loss, enabling differentiable, geometry-aware generation and Bayesian sampling for inference. Compared to a amino-acid recovery baseline, it achieves higher overall AAR on a rigorous 2025 temporal test set and shows strong gains in several CDR regions, indicating better handling of canonical loop geometries and potential for novel binders. This approach reduces reliance on expensive energy-based refinement and demonstrates practical potential for efficient, high-fidelity antibody design with accessible code and checkpoints.

Abstract

The computational design of antibodies with high specificity and affinity is a cornerstone of modern therapeutic development. While deep generative models, particularly Denoising Diffusion Probabilistic Models (DDPMs), have demonstrated the ability to generate realistic antibody structures, they often suffer from high computational costs and the difficulty of modeling discrete variables like amino acid sequences. In this work, we present AntibodyDesignBFN, a novel framework for fixed-backbone antibody design based on Discrete Bayesian Flow Networks(BFN). Unlike standard diffusion models that rely on Gaussian noise removal or complex discrete corruption processes, BFNs operate directly on the parameters of the data distribution, enabling a continuous-time, fully differentiable generative process on the probability simplex. While recent pioneering works like IgCraft and AbBFN have introduced BFNs to the domain of antibody sequence generation and inpainting, our work focuses specifically on the inverse folding task-designing sequences that fold into a fixed 3D backbone. By integrating a lightweight Geometric Transformer utilizing Invariant Point Attention (IPA) and a resource-efficient training strategy with gradient accumulation, our model achieves superior performance. Evaluations on a rigorous 2025 temporal test set reveal that AntibodyDesignBFN achieves a remarkable 48.1% Amino Acid Recovery (AAR) on H-CDR3, demonstrating that BFNs, when conditioned on 3D geometric constraints, offer a robust mathematical framework for high-fidelity antibody design$.$Code and model checkpoints are available at https://github.com/YueHuLab/AntibodyDesignBFN and https://huggingface.co/YueHuLab/AntibodyDesignBFN, respectively.