Interpretable Antibody-Antigen Structural Interface Prediction via Adaptive Graph Learning and Cyclic Transfer

Abstract

Experimental structural methods can identify antibody-antigen interfaces with high precision, but they remain time-consuming and resource-intensive, limiting their application across the rapidly expanding space of antibody and antigen sequences. Computational models capable of predicting these interfaces could therefore accelerate antibody discovery and provide insight into the principles governing immune recognition. However, this problem remains challenging due to limited structural datasets, severe class imbalance, and the complex, non-local nature of biomolecular interactions. Here we present VASCIF (Variable-domain Antibody-antigen Structural Complex Interface Finder), a structure-aware framework built on a Masked Graph Attention (MGA) architecture that represents protein complexes as residue graphs and captures long-range structural dependencies through attention-based message passing. The framework is straightforward to implement and enables efficient inference, allowing substantially faster predictions than other existing structure-based approaches. Evaluated on curated structural complexes across multiple benchmark datasets using rigorous cross-validation, VASCIF achieves state-of-the-art performance for residue-level interface prediction. Interpretability analyses reveal that the model recovers biophysically meaningful interaction patterns consistent with known principles of antibody recognition, and redefining interfaces using larger residue distance thresholds ([~]10 [A]) significantly improves predictive performance. Together, VASCIF provides a practical predictive framework and new insights into antibody-antigen molecular recognition.