OmniTCR

Repertoire sequencing yields hundreds of millions of TCRβ CDR3 sequences with no antigen attached to any of them, while antigen-specific assays yield a few tens of thousands of records that name the peptide, the HLA allele and the receptor together. Methods have had to choose a pile: binding predictors train on the small annotated set and overfit the handful of well-studied epitopes it contains, while repertoire classifiers train on the large unannotated set and never learn what any receptor recognises. OmniTCR removes the choice by writing both kinds of record in one format and pretraining a single model over all of them.

Each record is serialised with sequence-type tokens — [EPI], [HLA], [TRA], [TRB] — that bracket each component and mark its boundaries, so a record holding one component and a record holding four are the same kind of object to a causal decoder. Because such a decoder can only condition on what precedes a token, multicomponent records are also presented in alternative component orders: reversing peptide and receptor, or letting either chain open a four-component record. That augmentation is what lets one pretrained model both score a receptor given a pMHC and generate a receptor given a pMHC, rather than needing a separate model per direction. The work comes from the West China Biomedical Big Data Center at West China Hospital of Sichuan University with the College of Computer Science at Sichuan University.

The same pretrained backbone is then adapted to three uses — recognition prediction, cancer-associated repertoire classification, and pMHC-conditioned TCRβ generation. Even before any task-specific adaptation, the base model assigns lower perplexity to experimentally supported associations than to non-binding pairs across all three held-out recognition settings.

#Key Features

• Component-aware record format: Sequence-type tokens preserve the identity and boundaries of peptide, HLA pseudo-sequence, TRA and TRB components, so partial and complete records share one input representation without architectural changes.
• Complementary component orders: Multicomponent records are replayed in reversed or chain-first orders during pretraining, exposing the model to conditioning relationships a single canonical order would hide.
• Recognition prediction across combinations: One backbone is fine-tuned for peptide-MHC, peptide-TCRβ and peptide-MHC-TCRβ prediction, and absorbs a paired TRA chain when one is present.
• Repertoire-level cancer scoring: A cancer-associated enrichment score, the median receptor probability among up to 1,000 of the most abundant clonotypes, separates cancer from healthy repertoires with no cognate antigen information.
• Conditional generation with likelihood reranking: Ranking candidates by the log-likelihood gain under the target pMHC over an empty prompt cut cross-pMHC sequence sharing among 20 candidate sets by 62.9% without retraining.

#Technical Details

OmniTCR is a 113-million-parameter, 12-layer decoder-only transformer with 12 attention heads, a hidden dimension of 768, a 3,072-dimensional SwiGLU feed-forward block, RMSNorm and rotary positional embeddings, over an amino-acid vocabulary of 34 tokens. Pretraining used 328,232,215 formatted records assembled from TCRdb 2.0, iReceptor, IEDB, VDJdb and McPAS-TCR together with data released with BigMHC, NetMHCpan, TCRAI and pMTnet, covering 434,143 unique peptides, 203 HLA-I pseudo-sequences, 5,095,003 TRA CDR3s and 319,410,970 TRB CDR3s. On epitope-disjoint test sets the fine-tuned model reached AUPRCs of 0.7009 for peptide-TCRβ and 0.8235 for TCR-pMHC recognition, exceeding the strongest task-matched comparators by 0.3396 and 0.3451. Across 11 independent pan-cancer cohorts it reached a mean AUROC of 0.9436, roughly 0.02 above ESM-2. For seven pMHC targets absent from training, 27.6% of its generated CDR3β candidates met a composite AlphaFold 3 structural-confidence criterion, against 19.9% for the strongest specialised generator.

#Applications

The model suits immunologists and immuno-oncology groups who need to rank peptide-MHC and receptor pairings before committing to assays: prioritising neoantigen-reactive receptors, proposing CDR3β candidates for a chosen target as starting points for TCR-T engineering, and scoring bulk repertoires for blood-based immune monitoring. Repertoire scores tracked disease stage across early-stage and established cancer cohorts and shifted with time since neoadjuvant chemotherapy in a colorectal liver-metastasis cohort, which points at longitudinal monitoring as much as at one-off classification.

#Impact

OmniTCR shows that the abundant, unannotated half of immune sequencing data can be made to carry weight in recognition and design tasks, rather than sitting unused beside the small annotated sets those tasks normally rely on. Its scope is bounded in ways worth stating: the corpus is restricted to MHC class I, because class II records are sparser and its open-ended binding groove needs a different representation; paired-chain annotations and experimentally characterised negatives remain limited; and the structural support for generated receptors is AlphaFold 3 modelling, not measured binding. Inference code and all checkpoints are public, though neither release carries a license, and the work is a preprint awaiting peer review.