OmniTCR preprint: one model predicts T-cell antigen recognition and proposes candidate receptors

OmniTCR preprint: one model predicts T-cell antigen recognition and proposes candidate receptors
A preprint posted on bioRxiv on September 13 describes a model with 113 million parameters. It was trained on 328 million immune sequence records, ranging from individual receptors to rare complete combinations of a receptor, a peptide, and HLA. In separate tests, the authors evaluated predictions of recognition for peptides excluded from training. In another task, the model generated candidate receptors for a specified peptide and HLA pair.

A T-cell receptor recognizes a short protein fragment called a peptide, which another cell displays on its surface together with HLA, a protein that holds the peptide. Data on complete peptide, HLA, and receptor combinations are scarce, while individual receptor sequences are abundant.

Predicting recognition and designing receptors for a target have generally required separate models. In OmniTCR records, boundary markers separate the peptide, HLA, and the receptor’s α and β chains. The model learns to predict the next sequence token. The authors present the components in different orders: when the receptor chains follow the peptide and HLA pair, the model learns to propose a receptor for that target; the reverse order helps assess whether a known receptor matches a target. This allows rare complete combinations and a large collection of individual sequences to teach the model different parts of the same task.

To test generalization, the authors selected peptides whose records had been entirely excluded from both pretraining and fine-tuning. OmniTCR achieved an AUPRC of 0,7009 on 642 peptide and receptor β-chain pairs and 0,8235 on 141 complete combinations that included HLA. AUPRC measures how well a model ranks correct associations above incorrect ones in a dataset where correct pairs are rare. In these two tests, the scores exceeded those of the strongest comparison models by 0,3396 and 0,3451. These tests assessed generalization to peptides the model had never encountered during training.

For generation, OmniTCR produces CDR3β, a variable region of the receptor’s β chain that often contacts the peptide. The authors compared each candidate’s score given a specified peptide and HLA pair with its score given an empty input, then ranked candidates with the largest difference more highly. Across lists of the top one hundred candidates for 20 targets, the mean proportion of shared sequences fell from 0,0870 to 0,0323: specifying the target changed the composition of the list. For seven complexes containing peptides absent from the training data, the authors used AlphaFold 3 to predict candidate structures and compared confidence in the structural models.

The combined dataset teaches the model relationships among the components; tests with new peptides assess generalization; conditional generation and structure prediction produce a list of candidates for subsequent experimental testing.