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

Abstract

Predicting antigen-induced T-cell activation from T-cell receptor (TCR) and peptide sequence remains difficult for TCRs and peptides not represented in training data. We present TACO (T-Cell Activation from Chemistry and Orientation), an open statistical model for predicting TCR specificity against HLA-A*02:01-presented peptides. TACO takes as input the peptide sequence and six TCR CDR loop sequences. Learning from binary activation labels, TACO selects a single canonical orientation of the CDR loops over the peptide and learns five amino-acid contact-potential matrices as well as two peptide-exposure profiles.The orientation is fixed and does not require experimentally determined or predicted structures. We evaluate TACO on held-out peptides separated by Levenshtein distance at least four from all training and validation peptides, limiting near-neighbor shortcut learning. TACO achieves AUROC 0.67 on a large held-out activation dataset and 0.68 on IMMREP25 HLA-A*02:01, exceeding the open structure-based models evaluated here and matching the reported AlphaFold3-TCRdock result. The model has 2,019 fitted parameters, retains most within-peptide ranking signal with TCR$\beta$ sequence alone, and reaches 406,000 TCR--peptide predictions per second on CPU. This throughput enables computational evaluation of large TCR-peptide search spaces. We show its potential use for prioritizing candidate target peptides from large peptide libraries, reducing the search space for downstream experimental investigation. TACO fitted parameters also expose candidate physicochemical patterns that can be inspected directly. We release TACO's parameters and complete training, evaluation, and inference code.