Generative transformer that writes candidate cognate epitope sequences from a TCR CDR3-beta input, annotating repertoires without functional assays.
No providers recorded yet. Browse all providers
Single-cell TCR sequencing produces repertoires of thousands of T-cell receptor sequences, but a repertoire on its own says nothing about what those T cells recognize. Determining the cognate epitope for a receptor requires functional assays that are impractical at repertoire scale, leaving most single-cell immune datasets functionally unannotated. Existing computational tools mostly score a supplied TCR-epitope pair or cluster receptors by sequence similarity — neither of which helps when no candidate epitopes are on hand.
EpitopeGen inverts the problem. It is a decoder-only generative transformer that takes a TCR CDR3β sequence and writes candidate cognate epitope sequences directly, so a repertoire can be annotated with predicted antigen targets without any epitope list to test against. This distinguishes it from binding predictors such as DecoderTCR, which need both halves of the pair supplied up front.
The model was developed by Minuk Ma, Wilson Tu, Carlos Vasquez-Rios and Jiarui Ding at the University of British Columbia and posted as a preprint in January 2025. Its central engineering problem is data: roughly 100,000 experimentally measured TCR-epitope pairs exist, far too few to train a generative language model, and mining pseudo-labels at enormous scale is what makes the approach work.
EpitopeGen is a GPT-2 small architecture with 12 transformer decoder layers and 124 million parameters, trained with a custom tokenizer fitted to recurring motifs in CDR3β and epitope sequences. The pipeline begins with the Robust Affinity Predictor, a BERT-based binding model built on TABR-BERT and trained with Triple Negative Sampling — negatives drawn by shuffling paired data and by pairing against external TCRs or epitopes — using data from VDJdb, IEDB, PIRD and McPAS-TCR; it reaches 0.78 AUROC against 0.69 for TEIM and 0.53 for PanPep. BINDSEARCH then scores 7,331,478 unique TCRs from TCRdb against samples drawn from 21,801,187 epitope sequences, retaining top predicted binders to yield an intermediate set of 17 million pairs, which the Antigen Category Filter rebalances from 37.7% eukaryotic-dominated to virus-dominated. Training ran 100 epochs on four NVIDIA L40S GPUs; the epoch-28 checkpoint was selected on validation next-token accuracy of 0.5026 to avoid overfitting. On held-out splits partitioned by whether the TCR, the epitope or both were seen in training, generated epitopes reached mean binding-affinity percentile ranks of 81.2–81.9 against 100 randomly sampled epitopes, essentially unchanged for entirely unseen receptors. Molecular dynamics simulations provided an orthogonal check on binding.
The immediate use is functional annotation of single-cell TCR sequencing data. In oncology, EpitopeGen identifies clonally expanded tumour-infiltrating lymphocytes whose predicted targets are tumour-associated antigens, which in the paper's analysis showed elevated cytotoxic markers and reduced exhaustion signatures — directly relevant to selecting receptors for adoptive cell therapy. Applied to COVID-19 patient repertoires, it separated T cells recognizing spike and non-structural proteins, and revealed that severe cases retained naive and memory phenotypes where mild and moderate cases acquired cytotoxic effector programs. The same workflow supports vaccine design by surfacing the epitopes a patient's repertoire is already primed against. A hosted web interface allows submission of receptor sequences without local installation.
EpitopeGen makes a case that generative modelling is the right frame for antigen specificity: rather than asking whether a given pair binds, it asks what a receptor is likely to see, which is the question a repertoire dataset actually poses. The BINDSEARCH strategy is a transferable answer to data scarcity in immunogenomics, and the Antigen Category Filter is an unusually explicit attempt to build immunological priors into a training distribution rather than a loss function. The caveats follow from the same design: training labels are model-generated, so systematic errors in the affinity predictor propagate into the generator, the scope is restricted to CD8+ T cells and class I MHC, and generated epitopes are hypotheses requiring experimental confirmation. The repository ships no license file, which constrains downstream reuse.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.