Variant-aware CDR-H3 design improves predicted binding across antigen variation

Computational antibody design aims to generate antibodies that bind to a fixed target antigen. However, a central challenge is that antigens are diverse and have the propensity to mutate, creating a moving target for design. To better model this variable-target problem, we focus on mutations in or near the epitope region, where substitutions are most directly positioned to change antibody binding. We propose a simple training-free method across computational design paradigms for designing variant-resistant antibodies. We achieve this by exposing the computational model to multiple antigen variants during design, allowing it to favor antibodies that bind across the variant panel rather than only to the original antigen. To evaluate this, we create AntivarBench, a benchmark of 200 antibody--antigen complexes with 30 antigen variants per target. Across three inverse-folding generators, accounting for antigen variation during design increases the fraction of targets that outperform the native antibody across all held-out variants by 18--25 percentage points when compared at the same sampling temperature. These designs generally improve binding to the original antigen while retaining an additional advantage after antigen mutation. These results show that variant-aware design can produce antibodies whose predicted binding generalizes beyond the antigen states seen during design.