VitTCR: A deep learning method for peptide recognition prediction

The identification of the interaction between T-cell receptors (TCRs) and immunogenic peptides is important for the development of novel cancer immunotherapies and vaccines. However, experimentally determining whether a TCR recognizes a peptide is still time– and labour-consuming. In this study, we introduced VitTCR, a predictive model based on the architecture of the vision transformer (ViT), designed to forecast TCR-peptide interactions. Prior to prediction, VitTCR converts the TCR-peptide interactions into a numerical tensor named AtchleyMaps using Atchley factors. Subsequently, VitTCR takes AtchleyMaps as inputs and predicts whether an interaction between a TCR and a peptide exists. Through comprehensive evaluations, we demonstrate that VitTCR surpasses other published methods in classifying TCR-peptide pairs, exhibiting superior performance in terms of the area under the receiver operating characteristic (AUROC) and the area under the precision-recall curve (AUPR). To determine the focal contact point between TCRs and peptides, we obtained a positional bias weight matrix (PBWM) from the empirical amino acid (AA) contact probabilities derived from 83 structurally resolved pMHC-TCR complexes. The comparison of VitTCR with and without the integration of the PBWM revealed significant enhancements in the performance of the model. Moreover, the predicted probabilities generated by VitTCR exhibit significant correlations with immunological factors such as the clonal expansion and activation percentages of T cells. This further supports the efficacy of VitTCR in capturing biologically meaningful TCR-peptide interactions. In conclusion, VitTCR provides a useful computational tool for the prediction of TCR-peptide interactions, thereby contributing to our understanding in this field.