TRIOPS

TRIOPS is a deep learning model that predicts the human leukocyte antigen (HLA) allele restricting a given T-cell receptor (TCR) directly from amino acid sequence. It was developed by Nicholas R. Rose, Cynthia M. Ramirez, Lydia Mok, Christopher K. Wong, and Vanessa D. Jonsson at the University of California, Santa Cruz, and posted to bioRxiv in July 2026.

T cells recognize peptide antigens only when those peptides are presented by a specific MHC molecule, and a TCR's function is defined jointly by the receptor sequence and the restricting HLA allele. Establishing that restriction experimentally is laborious, and the combinatorial diversity of both TCR repertoires and HLA polymorphism means most receptor-allele pairings are never directly measured. TRIOPS reframes this as a sequence-to-sequence recognition problem: given a TCR sequence and a candidate HLA allele, it scores how likely that allele is to restrict the receptor, learning the biophysical grammar of restriction from paired data rather than requiring per-pair structural modeling.

The model sits alongside other TCR-MHC sequence models in the catalog, such as TULIP and DecoderTCR, but targets a distinct task: predicting the restricting HLA allele itself, and then transferring that predictor to bulk tumor data to study immune escape.

#Key Features

• HLA-restriction prediction from sequence: TRIOPS predicts a TCR's restricting HLA allele using amino acid sequence alone, without structural input, covering both TCR and HLA pseudosequence information.
• Dual-branch convolutional architecture: Separate convolutional branches encode the receptor and the HLA allele, and a cross-attention mechanism couples the two branches to model their joint recognition.
• Cross-reactivity-aware negative sampling: The training procedure constructs negatives in a way that accounts for TCR cross-reactivity, which lets the model generalize to receptors and to HLA alleles that were absent from the training set.
• Flexible chain input: TRIOPS scores both paired-chain receptors and inputs where only the TCR β-chain sequence is available, matching the coverage of real repertoire datasets.
• Transfer to bulk tumor data: The trained model is applied, without retraining, to bulk tumor RNA-seq to relate TCR repertoires to the HLA alleles they preferentially engage.

#Technical Details

TRIOPS is a dual-branch convolutional neural network with cross-attention between the receptor and HLA branches, trained as a supervised classifier of TCR-MHC restriction. On held-out data it reaches an AUC of 0.97 for paired-chain input and 0.92 when only the TCR β-chain sequence is provided. The cross-reactivity-aware negative sampling strategy is central to its generalization: the model maintains its accuracy on receptors and on HLA alleles that do not appear in training, rather than only interpolating within observed pairings. Because scoring requires only sequence, the same fixed checkpoint can be evaluated across large, independently collected cohorts. Applied to bulk tumor RNA-seq and HLA genotypes from The Cancer Genome Atlas (TCGA), TRIOPS recovers a pattern in which tumor TCR repertoires preferentially engage particular HLA alleles, a signal consistent with allele-specific antigen presentation and immune escape.

#Applications

TRIOPS is useful for computational immunologists and cancer-immunology researchers who need to assign HLA restriction to TCRs at scale. Because it operates from sequence, it can annotate large repertoire-sequencing datasets where experimental restriction assays are impractical, prioritize receptor-allele hypotheses for downstream validation, and support analyses of antigen presentation in tumor cohorts. The demonstrated TCGA analysis shows a concrete workflow: applying the frozen model to bulk transcriptomic and genotype data to study how HLA loss and allele-specific presentation shape the T-cell response, informing research into immune escape and immunotherapy resistance.

#Impact

TRIOPS extends sequence-based TCR-MHC modeling from binding and epitope prediction toward explicit HLA-restriction prediction, and pairs that predictor with a population-scale tumor analysis rather than leaving it as an in-silico benchmark. Its generalization to unseen alleles addresses a recurring weakness of receptor-specificity models, whose apparent accuracy often collapses outside the training distribution, and its cross-reactivity-aware negatives offer a reusable recipe for that problem. The work is a preprint awaiting peer review, and no public code or trained weights are linked from the paper, so reproduction and independent benchmarking are not yet possible. As a single-task restriction predictor rather than a broad multi-task representation model, its scope is deliberately narrow, but the reported held-out performance and the TCGA transfer make it a concrete tool for studying HLA-dependent immune recognition.