PABbind: A Hierarchical Geometric Framework for Mutation-Involved Antigen−Antibody Interaction Prediction and SARS-CoV-2 Immune Escape Assessment

Abstract

Antibody therapies exhibit remarkable specificity and therapeutic potential, yet their efficacy can be compromised by rapidly emerging antigen mutations that enable immune escape. Therefore, predicting how antigen variants alter antibody recognition is critical for assessing antibody durability, but experimental evaluation of large-scale mutation combinations remains challenging. Here, we develop PABbind, a hierarchical geometric deep-learning model that integrates sequence and structural information to predict antigen−antibody interaction changes using wild-type protein−protein interaction (PPI) structures and mutant sequences. By simultaneously incorporating sequence and structural information and leveraging pretraining on general mutant PPI datasets, PABbind accurately predicts immune escape effects of antigen variants against antibodies using wild-type PPI structures only, enabling large-scale screening of antigen mutations. PABbind demonstrates favorable performance on mutant PPI benchmarks and enables prediction of variant-specific immune escape scores for SARS-CoV-2 antibodies, therefore providing an effective computational framework for large-scale assessment of SARS-CoV-2 antigen variants.