<scp>TinyHLAnet</scp> : A Light‐Weight <scp>3D</scp> Structure‐Aware Architecture for Rapid and Explainable Identification of <scp>CD8</scp> + T‐Cell Antigens

CD8+ T cell response is contingent on the surveillance and recognition of peptide antigens presented on HLA class I molecules. The highly polymorphic nature of the HLA loci combined with the vast cognate peptide space makes comprehensive experimental characterisation of the peptides presented by different individuals in different disease conditions intractable. Several computational methods for epitope prediction exist, but there is a necessity in the field for rapid, interpretable and personalised prediction of CD8+ T cell epitopes. We address this gap by developing a novel deep learning architecture, termed TinyHLAnet, that produces highly accurate and fast prediction of peptide-HLA (class I) complex formation. This architecture is developed using constraints derived from domain knowledge of contact formation in peptide-HLA (class I) complexes. Specifically, the estimation of binding affinity of the peptide-HLA complex is decomposed into subproblems relating to estimating the interaction strength of each intermolecular residue pair in the complex, which in itself is factorised into three different components based on knowledge of residue pair interactions. This compartmentalisation allows for the production of a model that uses a fraction of the parameters used by existing gold standard methods at no loss of prediction accuracy but provides much greater throughput. The TinyHLAnet prediction engine serves as a framework for several applications, with two examples shown here: TinyHLAnet-SCAN, for proteome-wide scanning of epitopes, and TinyHLAnet-ESCAPE, for prediction of likely immune escape mutants.