AbFlowNet: Optimizing Antibody-Antigen Binding Energy via Diffusion-GFlowNet Fusion

[Download Paper] [Source Code]

Antibody CDR design methods typically optimize reconstruction losses and leave binding energy to a separate, expensive reinforcement-learning stage. AbFlowNet folds GFlowNet training into diffusion: each denoising step is a GFlowNet state, so reconstruction and binding-energy rewards are optimized together. Relative to a diffusion baseline, AbFlowNet improves amino-acid recovery, geometric reconstruction, and the fraction of designs with better binding energy, without online RL or pseudo-labeling the test set.