T-cell receptors that are k-binding have defined sequence features.

Previous studies have revealed that individual T cell receptors (TCRs) can recognize a diverse set of peptide targets displayed by Major Histocompatibility Complexes (MHCs) to enable effective adaptive immune surveillance. However, how TCR sequences encode their cross-reactivity remains poorly understood. Here, we used an in vitro assay to characterize the k-binding of 19⁶ (~47 million) different TCRs in the context of a single TCR framework for binding to seven related peptides displayed by HLA-A*02:01. We define k-binding to be the number of peptide-MHC targets recognized by a TCR within a specific universe of targets. We found a hierarchy of TCR complementarity-determining region 3 (CDR3) alpha and beta chain residue importance that determined k-binding for the seven targets. Our machine learning model that embedded TCR sequences using BLOSUM-50 provided an overall F1 score of 0.698 and an AUPRC of 0.745 for predicting TCR-pMHC binding, which was significantly superior to model results from VHSE-8 embedded or one hot encoded sequences. When we used our model to predict observed k-binding, we found that experimentally derived sequence motifs do not fully explain the relative importance of different CDR3 residues. We determined CDR3 residue importance by examining the reduction in machine learning model predictive ability by masking individual CDR3 residues. We found that the resulting residue importance ranking was significantly correlated to residue importance determined with a computational alanine scan using Rosetta. Our findings validate past theoretical predictions of TCR cross-reactivity and demonstrate that TCRs used in therapeutics must be carefully evaluated for their specificity.

Copyright © 2025 Park, Krog, Carter, Balivada, Pogue, Anand, Birnbaum and Gifford.