TCRfinder: Improved TCR virtual screening for novel antigenic peptides with tailored language models

Abstract

AbstractAccurate modeling of T-cell receptor (TCR) and peptide interactions is essential for immunoreaction elucidation and T-cell-based immunotherapeutic developments. We developed TCRfinder, a novel deep-learning architecture for TCR-peptide binding prediction and virtual screening. Large-scale benchmark experiments demonstrated a robust capability of TCRfinder in distinguishing interacting and non-interacting TCRs for unseen peptides, with accuracy significantly beyond current state-of-the-art methods. Furthermore, TCRfinder recognizes tumor neoantigen mutations from wild-type antigens of given TCRs, with a success rate nearly 50% higher than the best of existing methods. Detailed data analyses showed that the major advantage of TCRfinder lies in the specially trained TCR and peptide language models tailored with iterative attention network architecture, which can precisely reveal physical interaction patterns of cross-chain atoms and substantially enhance the precision of TCR-peptide interaction predictions. The open-source TCRfinder program can help facilitate large-scale deployment of high-quality TCR and neoantigen virtual screening, offering exciting potential for personalized TCR-based immunotherapies.