New model · Sichuan University
OmniTCR
A 113M-parameter foundation model unifying TCR recognition prediction and receptor generation
· 8 claims checked against 2 sources
What it is
OmniTCR is a 113-million-parameter autoregressive foundation model developed by researchers at Sichuan University. It unifies two tasks usually handled by separate models: predicting whether a T cell receptor recognizes a given peptide-MHC complex, and generating new TCR sequences designed to bind a target.
How it works
TCR recognition prediction and receptor generation are normally built as separate models, so the large collections of unpaired TCR sequences never inform the smaller TCR-peptide-MHC datasets used for specificity prediction. OmniTCR is pretrained on 328 million formatted human immune-sequence records. Sequence-type tokens and flexible ordering of components let the same architecture learn jointly from single TCR chains and from partial or complete TCR-pMHC associations.
What is new
On epitopes absent from training, OmniTCR reaches AUPRC 0.7009 for peptide-TCRβ recognition and 0.8235 for full TCR-pMHC interaction prediction, exceeding the strongest evaluated comparators by 0.3396 and 0.3451 respectively. Across 11 independent pan-cancer cohorts, it separates cancer from healthy repertoires with mean AUROC 0.9436. It achieves the highest sequence recovery on internal and external TCR generation benchmarks, and structural modelling supports the plausibility of selected pMHC-conditioned CDR3β candidates the model generated.
Limits
Generalization to epitopes absent from training data remains the main bottleneck for TCR specificity models generally. OmniTCR's gains are measured against that persistent difficulty, not a fully solved problem.
Availability
The preprint was posted on bioRxiv on September 10, 2026, under a CC-BY 4.0 license. The paper does not state a code or model weights release.
The evidence
Each statement above was written with the sentence from the source that supports it, and a program (not a model) checked that every quote and every figure appears in the pages it read. Here they are.
- 01
OmniTCR is a 113-million-parameter autoregressive foundation model pretrained on 328 million formatted human immune-sequence records
“a 113-million-parameter autoregressive foundation model pretrained on 328 million formatted human immune-sequence records” S1
- 02
On unseen epitopes, OmniTCR achieved AUPRCs of 0.7009 for peptide-TCRβ recognition and 0.8235 for TCR-pMHC interaction prediction
“OmniTCR achieved AUPRCs of 0.7009 for peptide-TCRβ; recognition and 0.8235 for TCR-pMHC interaction prediction” S1
- 03
These results exceed the strongest evaluated comparators by 0.3396 and 0.3451 respectively
“exceeding the strongest evaluated comparators by 0.3396 and 0.3451, respectively” S1
- 04
OmniTCR distinguishes cancer from healthy repertoires across 11 independent pan-cancer cohorts with mean AUROC 0.9436
“It distinguishes cancer from healthy repertoires across 11 independent pan-cancer cohorts (mean AUROC, 0.9436)” S1
- 05
The model achieved the highest sequence recovery on internal and external generation benchmarks
“The model achieved the highest sequence recovery on internal and external generation benchmarks” S1
- 06
Structural modelling supported the plausibility of selected pMHC-conditioned CDR3β candidates
“Structural modelling supported the plausibility of selected pMHC-conditioned CDR3β; candidates” S1
- 07
The work is affiliated with Sichuan University, funded in part by Sichuan Provincial Natural Science Foundation and West China Hospital of Sichuan University
“1.3.5 project for disciplines of excellence from West China Hospital of Sichuan University, ZYYC23007” S1
- 08
The preprint is made available under a CC-BY 4.0 International license
“It is made available under a CC-BY 4.0 International license” S1
