New AI Predicts Epitopes by Analyzing 3D Molecular Surfaces

Understanding how antibodies recognize antigens is crucial for drug discovery and vaccine development, with epitope prediction at its core. Current methods for predicting these binding sites often rely on molecular sequences or backbone structures, which struggle to capture the complex, discontinuous patterns found on 3D molecular surfaces that are key to antibody-antigen interactions. A new framework called SurfBind addresses this limitation by directly operating on molecular surface representations. It employs a Transformer-based architecture that integrates both geometric and physicochemical information from the surface. Key components include patch-level surface modeling, cross-attention mechanisms aware of binder interactions, and a hierarchical prediction approach. Evaluations on challenging benchmarks like SAbDab and DB5.5 demonstrate that SurfBind achieves state-of-the-art performance. It shows strong generalization across previously unseen antibodies and various conformational states, underscoring the significant value of its interaction-aware surface modeling for deciphering protein-protein interaction mechanisms.