(197ak) Towards the Development of Explainable Graph Neural Network Model for Predicting Viral Antibody-Antigen Binding Interfaces

The phenomenon of antigen epitopes being specifically recognized by antibodies is not only a critical aspect of our immune response, but is also imperative for vaccine and therapeutic-antibody design. Consequently, detailed characterization and prediction of antigen-antibody interaction is essential for immune system research and treatment strategies. Experimental techniques for capturing the structural details of antigen-antibody complexes, however, can be extremely challenging and resource intensive. Computational modeling is a viable alternative, but current methods for computationally predicting such antigen-antibody docking are based on general protein-protein interaction studies, and only provide global scores of relative performances, without explicit structural rationales. We will demonstrate prototypical examples of how an explainable AI (XAI) based deep learning algorithm can be applied to determine antigen-antibody binding interfaces and interaction parameters. With viral envelope proteins as our study case, we have curated an extensive structural dataset of epitope-paratope complexes and associated binding affinities, and extensively teased out clustering features underpinning these interactions. We have deployed this dataset to enable our machine learning driven predictive models, applying an efficient graph neural network-based method to model docking interfaces in atomistic level. Detailed treatment of covalent bonds, non-bonded interactions, and secondary structural features in defining the interfacial graph extraction preserves the core invariants of antigen-antibody specific conformational information, leading to improved performance. Structural predictions are being further interpreted through XAI methods to understand the underlying sequence-structure-binding relations. This methodology not only gives us a handle over the structural bases of antigen-antibody recognition, but also enables predictive engineering of the interface towards improved therapeutic design.