Chain-Aware Protein Language Modeling for Antibody Affinity and Mutational ∆∆G Prediction

Structure-free antibody affinity prediction requires chain-aware, partner-dependent representations. We present AbAffinity, a chain-aware, interpretable framework for structure-free antibody affinity and mutation prediction. Separate protein language model streams and gated interactions encode antibody/antigen relationships, calibrated affinity differences and a learned correction transfer the frozen model to mutational effects. AbAffinity outperforms early-fusion models under random and antigen-cold evaluation, reaching Pearson $r=0.86$ on SAAINT-DB dataset. On S1131, it surpasses evaluated frozen-ESM-2 and parameter-matched baselines on complex-/antigen-disjoint splits ($r=0.71/0.68$). Antigen interventions support partner dependence; explainability analysis recovers paratope residues, linking predictive gains to biologically plausible recognition.