Lightweight Chain-Aware Modeling for Antibody Affinity and Mutational Effects

Sequence-based antibody affinity prediction requires chain-aware, antigen-dependent representations. We present AbAffinity, a chain-aware framework that transfers calibrated affinity differences to mutational effects through a lightweight gated residual head. AbAffinity outperforms early-fusion models on random and antigen-cold SAAINT-DB splits, reaching Pearson $r=0.84$ on the random split. On S1131, it surpasses evaluated frozen-feature and parameter-matched baselines under complex-disjoint ($r=0.71$) and antigen-disjoint ($r=0.68$) evaluation. Antigen interventions and paratope attribution support biologically plausible partner dependence. Frozen protein language model features enable sequence-only prediction with modest computational demands.