EpiCLIP: Learning Antibody-Antigen Interactions from Approximate Interfaces

Modeling antibody-antigen interactions in therapeutic discovery remains challenging because structurally resolved complexes are scarce, exact interfaces are unavailable at inference, and pairwise structure-based prediction does not scale. We introduce EpiCLIP, a contrastive dual-encoder framework that casts antibody-antigen interaction modeling as dense retrieval over approximate epitopes and paratopes. EpiCLIP trains on synthetic epitope candidates designed to match the noisy search space encountered at deployment, enabling both binder retrieval and epitope mapping in a shared embedding space. Empirically, EpiCLIP achieves strong performance on both tasks while operating directly on approximate interface candidates. These results suggest that retrieval over approximate interfaces is a scalable alternative to exact-interface supervision or exhaustive pairwise structural modeling for antibody-antigen interaction prediction.