Decoding the specificity of the T cell repertoire: From random energy models to coarse-grained structure-based prediction

The adaptive immune system enlists millions of T cells, each with a unique T cell receptor (TCR), to recognize and eliminate foreign and tumor associated antigens presented on the cell surface. Despite substantial interest, high-throughput and reliable identification of relevant TCR-antigen pairs remains a highly desired research objective. Predicting TCR-antigen specificity pairs on the level of the human T cell repertoire would have far-reaching applications in immunology and cancer immunotherapy, including improved design of tumor-specific T cells, optimized donor transplant selection, and characterization of T cell-mediated autoimmunity. The immense complexity required to understand how large (~10⁸) TCR repertoires interact with thousands of potential antigens, together with increasingly available data on known TCR-antigen pairs, provides an attractive modeling problem well-suited to quantitative-based theoretical and computational investigation. Motivated by a statistical random energy model of TCR-antigen interactions, we describe the development of a structure-based, supervised machine learning model to resolve relevant TCR-antigen pairs from large immune repertoires and putative antigen lists. We utilize high affinity TCR-antigen pairs with their solved crystal structures to train an optimized energy model. Our results suggest that TCR-antigen pairs can be reliably predicted when training and testing pairs are restricted to the same MHC allele. We demonstrate the power of this approach by simulating the immunogenicity of a post-thymic selection immune system consisting of 10⁹ TCR-antigen pairs. Lastly, we discuss recent applications trained on HLA A*02:01-restricted systems and the resulting transferability of our model to distinct TCRs and antigens, arguing for the utility of incorporating known structural information into the learning procedure for TCR-antigen interaction.