T cells have important functions in development and disease processes through T cell receptor (TCR)-dependent activities. Many tools were developed to predict the binding between TCRs and antigens. However, one of the uncertainties is whether such tools can decipher how small changes in the TCRs or antigenic peptides contribute to binding. We develop a deep learning model, pMTnet-omni, which not only predicts the binding vs. non-binding of TCRs towards pMHCs, but also distinguishes the stronger vs. weaker binding of TCRs similar in sequence. We leverage this capability to interpret the biological rules that govern TCR-antigen pairing. This also enables pMTnet-omni to accurately predict variant TCRs with desired stronger or weaker binding to the antigen, in conjunction with a Lab-in-the-Loop (LiL) mechanism. We show that pMTnet-omni can also predict binding of TCRs towards similar pMHCs. Overall, we provide a flexible toolkit for research and translational applications involving antigens and TCRs. Computational prediction of peptide-MHC (pMHC) antigen binding to T cell receptors (TCR) is useful for T cell-based therapeutic development and predicting responses to immunotherapies. Here the authors use a deep learning model pMTnet-omni to predict the binding versus non-binding of TCRs to pMHC antigens and then use this model to tune the affinity of TCRs against antigens.
Han et al. (Sat,) studied this question.