Biophysical modeling for accurate and efficient predictions of T cell-antigen specificity

Despite the central role that T cell receptor (TCR) recognition plays in eliminating cancer and viral nonself-signatures, we currently lack robust computational methods to reliably predict the specificity of TCRs against relevant target antigens. Such predictive modeling would immediately enable the development of optimized vaccines and cancer immunotherapies. Here, we describe insights derived from our random energy models focused on mapping the TCR-antigen contact interface. From this, we have developed a probabilistic method of characterizing the variance in TCR-antigen binding energy, which we use to cluster the expected recognition behavior of T cells across a variety of distinct Human Leukocyte Antigen (HLA)-I-restricted systems. We also discuss the development and implementation of a supervised residue-specific machine learning model trained on publicly available structural data and TCR-antigen primary sequences. In the setting of HLA-A*02:01-restricted TCRs, we demonstrate the utility of this model to simultaneously predict the antigenicity of disparate peptides, as well as closely related (point-mutated) variants of a preidentified antigen.