TRIOPS: A deep learning framework for prediction of T cell receptor-MHC binding specificity

Abstract

T cell receptor (TCR) recognition is MHC-restricted, yet accurately predicting a TCRs restricting HLA allele remains an open problem. We present TRIOPS, a dual-branch convolutional model with soft cross-attention that predicts TCR-MHC restriction from amino acid sequence alone. TRIOPS uses cross-reactivity-aware negative sampling by HLA pseudosequence similarity to reduce allele-boundary label noise, extending prediction to alleles absent from training. TRIOPS reaches a held-out AUC of 0.97 for paired TCR{beta} and 0.92 for TCR{beta}-only inputs, generalizes to unseen receptors and HLA alleles, and after locus-specific calibration, assigns TCR clonotypes to their likeliest restricting allele across an individuals HLA genotype. In TCGA tumors, TCR repertoires preferentially engage the expression-lost allele at HLA-A and HLA-B and the retained allele at HLA-C, recapitulating from bulk tumor RNA-seq the allele specific HLA loss previously linked to immune escape.