TACO: An open, interpretable model for microsecond-scale TCR specificity prediction

TACO, an open statistical model for HLA-A*02:01-presented peptides, predicts T-cell activation with AUROC 0.67 on a large held-out dataset and 0.68 on IMMREP25.

• Why it matters: Antigen-induced T-cell activation prediction is hard for TCR–peptide pairs not represented in training, and near-neighbor peptide similarity can create shortcut learning.
• What they did: TACO uses peptide sequence plus six TCR CDR loop sequences, fixes a single canonical CDR orientation, and learns five contact-potential matrices and two peptide-exposure profiles from binary activation labels. It is evaluated on held-out peptides with Levenshtein distance at least four from training/validation.
• The result: With 2,019 fitted parameters, it preserves most within-peptide ranking signal using TCRβ sequence alone and runs at 406,000 TCR–peptide predictions per second on CPU, enabling large search-space prioritization and interpretable physicochemical patterns.

AI summary of the abstract · may contain errors · not medical advice · read the paper.