General prediction of T cell receptor antigen specificity from sequence using AlphaFold 3

Abstract

The Major Histocompatibility Complex (MHC):peptide:T Cell Receptor (TCR) complex is the most diverse trimolecular interface known in nature and a key trigger for adaptive immunity. TCRs are now routinely sequenced at scale, however decoding their specificity for antigens remains a bottleneck. While in silico approaches have advanced considerably, none enables prediction against "unseen" epitopes, limiting their applicability to a tiny fraction of cases. Here, we show that AlphaFold 3 (AF3) predicts the structures of MHC:peptide:TCR triads with unprecedented accuracy. By applying AF3 to >9,000 TCRs mapped to >1,000 distinct epitopes restricted by >70 MHC class I / II alleles, we identify features that distinguish cognate triads from controls. The resulting model achieves median AUCs of 0.81-0.92 on validation triads unseen by AF3 or during feature selection. Our results reveal that generalized prediction of TCR specificity from sequence is possible, with the potential to greatly accelerate the decoding of immune responses.