abCAN: a Practical and Novel Attention Network for Predicting Mutant Antibody Affinity

Abstract

Accurate prediction of mutation effects on antibody-antigen interactions is critical for antibody engineering and drug design. In this study, we present abCAN, a practical and novel attention network designed to predict changes in binding affinity caused by mutations. abCAN requires only the pre-mutant antibody-antigen complex structure and mutation information to perform its predictions. abCAN introduces an innovative approach, Progressive Encoding, which progressively integrates structural, residue-level, and sequential information to construct the complex representation in a systematic manner, effectively capturing both the topological features of the structure and contextual features of the sequence. During which, extra weight to interface residues would also be applied through attention mechanisms. These learned representations are then transferred to a predictor that estimates changes in antibody-antigen binding affinity induced by mutations. On the benchmark dataset, abCAN achieved a root-mean-square error (RMSE) of 1.195 (kcal/mol-1) and a Pearson correlation coefficient (PCC) of 0.841, setting a new state-of-the-art (SOTA) benchmark for prediction accuracy in the field of antibody affinity prediction.