EPIC: Epitope–Paratope Interleaved Contact Learning for Antibody–Antigen Interface Prediction

Predicting antibody–antigen interfaces directly from sequence is challenging because residue–residue contacts are sparse, the interface depends on the specific antibody–antigen pair, and predictions must remain consistent between residue-pair contacts and residue-level binding sites. We present EPIC, a geometry-aware multi-task framework for predicting antibody–antigen contacts, paratopes, and epitopes from sequence inputs. EPIC combines pretrained protein language model representations with sequence-derived geometry obtained from predicted monomer structures, without requiring experimentally determined structures at inference time. Its geometry-aware interleaved interaction block integrates dense sequence self-attention with geometry-aware self-attention over local distance, direction, and relative-orientation features, and uses contact-gated cross-attention to enable information exchange conditioned on the specific antibody–antigen pair. A low-rank pair decoder models residue-pair interactions, while pairwise contact supervision and antibody- and antigen-side marginal supervision jointly constrain a shared representation of the interface. Task-specific readouts produce contact, paratope, and epitope predictions. We evaluate EPIC on SAbDab2 and AsEP for the interface-prediction tasks annotated in each benchmark. EPIC achieves a strong overall performance profile across interface-prediction levels and remains competitive with methods that use experimentally determined antibody or antigen structures. These results show that interleaving pretrained sequence representations with predicted geometric priors enables a unified framework for jointly predicting epitopes, paratopes, and residue-residue contacts from antibody-antigen sequences.