Learning from B Cell Evolution: Adaptive Multi-Expert Diffusion for Antibody Design via Online Optimization

Recent diffusion models show strong potential for antibody design, but most use uniform strategies that ignore antigen-specific needs. Inspired by B-cell affinity maturation—balancing affinity, stability, and self-avoidance—we present a biologically motivated framework that brings physics-based priors into an online meta-learning loop. Specialized experts (van der Waals, molecular recognition, energy balance, interface geometry) adapt their weights during generation via iterative feedback, mimicking natural refinement. Instead of fixed protocols, the system discovers target-specific guidance. In experiments, it: (1) learns SE(3)-equivariant guidance for different antigen classes without pretraining, preserving molecular symmetries; (2) improves hotspot coverage and interface quality through target-specific adaptation; (3) enables iterative refinement where each complex learns its own optimization profile via online evaluation; and (4) generalizes from small epitopes to large interfaces, supporting precision campaigns for individual targets.