Peptide-independent deep learning with TRIOPS predicts TCR–HLA restriction and recovers locus-biased HLA loss in tumors
T cell receptor (TCR) antigen recognition is MHC-restricted, yet the restricting HLA allele is unknown for nearly all sequenced receptors. We present TRIOPS, a dual-branch convolutional model that predicts TCR–HLA restriction from amino acid sequence alone, trained with cross-reactivity-aware sampling. TRIOPS reaches an AUC of 0.97 on held-out TCRαβ pairs, 0.92 on HLA alleles withheld from training, and 0.81 on unseen TCRs. We tested whether predicted engagement tracks allele-specific HLA loss in 8,906 TCGA tumors across 23 cancer types, assigning each clonotype to its likeliest restricting allele within the corresponding tumor genotype. At HLA-A and HLA-B, engagement was higher at the lower-expressed allele of each heterozygous pair (concordance 0.60 and 0.65; 23/23 cancers). TRIOPS resolves this locus-dependent pattern from receptor sequences and HLA genotypes alone, establishing HLA-restricted TCR engagement as a relevant measurable feature of the T cell receptor repertoire.
Authors
- Nicholas R. Rose (ORCID: https://orcid.org/0000-0002-8285-4215)
- Vanessa D. Jönsson (ORCID: https://orcid.org/0000-0002-6732-9607)
- Lydia Mok (ORCID: https://orcid.org/0000-0002-4029-5793)
- Christopher K. Wong (ORCID: https://orcid.org/0000-0001-6012-7001)
- Roberta Zappasodi (ORCID: https://orcid.org/0000-0002-7031-6671)
- Cynthia Ramirez
Institutions
- University of California, Santa Cruz (US)
- Weill Cornell Medicine (US)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11143133/v1
- Primary Topic
- vaccines and immunoinformatics approaches
- Type
- preprint