bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelTermsPrivacyContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / protein / tcrgen
Protein
Arizona State UniversityGoogle DeepMindReleased January 2025

TCRGen

Epitope-conditioned T cell receptor generator that writes its own in-context examples, so receptors can be designed for targets with no known binders.

The short version

  • —Designs candidate receptors for epitopes that have no experimentally known binding TCRs
  • —Writes its own prompt: drafts candidates, scores them, then feeds the best back as context
  • —Training instances mirror the few-shot prompt format used at inference
  • —Screens output on predicted binding affinity and on resemblance to natural repertoires
72Openness

Where to run it

No providers recorded yet. Browse all providers

Engineered T cell therapies need receptors that recognise a specific epitope, and the epitopes that matter most clinically — a newly emerged viral strain, a patient's private neoantigen — are exactly the ones with no catalogued cognate TCRs. A generative model must therefore produce binders for a target it has never seen, with no examples to imitate.

TCRGen, from Heewook Lee's group at Arizona State University with Seojin Bang at Google DeepMind, imports in-context learning from large language models and adapts it in two steps. First, in-context training (ICT): instead of fine-tuning on isolated epitope-TCR pairs, the model is trained on instances that bundle an epitope with several of its cognate receptors, so the prompt format it learns from matches the format it will be given at inference. Second, self-contemplation prompting (SCP): because a novel epitope has no known binders to supply as context, the model generates its own candidates from the epitope alone, scores them, and reuses the best as the in-context examples for a final generation pass.

The paper's central finding is that the two steps are inseparable. A conventionally fine-tuned model handed context receptors does not merely fail to benefit — its success rate collapses from 80% to 3%, because the prompt shape is unlike anything it saw in training. With training and inference distributions aligned, self-generated context works about as well as ground-truth context, and with scored selection it works better.

#Key Features

  • Format-aligned training: Training instances group k+1 receptors known to bind the same epitope, so few-shot prompting at inference is in-distribution rather than an unfamiliar input shape.
  • Self-generated prompts: For a novel epitope the model drafts 300 zero-shot candidates, ranks them, and promotes the top handful into the prompt — removing the dependency on prior binding data entirely.
  • Scored prompt selection: Choosing context receptors by predicted binding affinity and sequence authenticity beats choosing them at random, and beats iteratively chaining single best outputs, which lets early errors accumulate.
  • Built-in candidate triage: Generated receptors are filtered on two independent axes — a binding predictor and a likelihood model of natural repertoires — with thresholds set by maximising Youden's index.

#Technical Details

The generator is a fine-tune of RITA_m, a 300-million-parameter, 24-layer autoregressive protein language model pretrained on over 280 million sequences. Instances are serialised as an epitope, a delimiter, and one or more receptor sequences; vanilla fine-tuning used 96.7K unique TCR-epitope pairs, while in-context training grouped receptors sharing an epitope, with a context width of k = 4 for the main model and up to 10 in the longer-context variants. Evaluation rests on three metrics: BAP, a catELMo-based binding predictor that separates known binders from non-binders at AUC 0.97; TCRMatch, a k-mer similarity score against ground-truth receptors; and GPT-LL, a protein language model fine-tuned on four million healthy-repertoire receptors from ImmunoSeq whose log-likelihood scores authenticity at AUC 0.87.

Across 29 novel epitopes with 300 generated receptors each, the strongest configuration — in-context training plus scored self-contemplation prompting at 10 shots — reaches a 91.91% good-TCR rate under BAP, against 87.17% for the oracle that is handed real cognate receptors and 88.14% for zero-shot generation. Under the authenticity metric it reaches 99.17%, and under TCRMatch 99.77%. Gains saturate quickly: BAP performance plateaus after about six context receptors, and the similarity-based metrics after one.

#Applications

The intended workflow is pre-screening for personalised immunotherapy. Given an epitope from a new pathogen or a tumour neoantigen, the model proposes receptor sequences and filters them to a short list whose members are both predicted to bind and statistically indistinguishable from natural repertoire sequences, shrinking what goes to wet-lab binding assays and T cell engineering. The recipe generalises beyond immunology wherever a generative model must extrapolate to targets absent from a small training set: train in the shape you will prompt in, and manufacture the context you lack. Catalogued epitope-conditioned generators such as TCRDiff approach the same design problem from a diffusion rather than a prompting angle.

#Impact

TCRGen isolates a failure mode that is easy to miss — in-context learning is not a free capability of a pretrained language model, but something a fine-tune destroys when its training format diverges from its inference format — and shows that closing that gap is what makes self-generated context viable. That synthesised context matches or exceeds ground-truth context is the load-bearing claim, since it is what makes the method applicable to the novel-epitope case it was built for. The limits are acknowledged by the authors: every result is scored by computational proxies, no generated receptor has been expressed or assayed, and the binding predictor that ranks candidates also grades them. The repository publishes training code, generation and evaluation scripts, and the k-shot data splits under CC BY 4.0; pretrained weights are announced but not yet released. The work remains a preprint.

At a glance

Parameters
300 Million
Released
January 2025
Category
Protein
Organizations
Arizona State University / Google DeepMind

Links

GitHub RepositorybioRxiv Preprint

Tags

immunologyin_context_learninglanguage_modelprotein_designsequence_generationtransformer

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.