tcrga Evolving T-cell receptor sequences against a target peptide

TCR specificity predictors: what is actually usable offline

This page mirrors the repository's docs/predictors.md — read that file directly on GitHub for the canonical, always-current version. It records the findings from the Phase 0 risk-retirement probe, 2026-09-25.

ePytope-TCR is not on PyPI

pip install epytope gives epytope 4.0.1 with no TCR module. Its submodules are CleavagePrediction, Core, Data, EpitopeAssembly, EpitopePrediction, EpitopeSelection, HLAtyping, IO, TAPPrediction — no TCRSpecificityPrediction. The same is true of KohlbacherLab/epytope on both main and develop.

ePytope-TCR lives only in the fork used for the benchmark paper:

The 22 wrapped predictors

External.py (17) — Ergo1, Ergo2, pMTnet, EpiTCR, ATM_TCR, AttnTAP, TEIM, BERTrand, TEINet, PanPep, DLpTCR, TULIP, iTCep, NetTCR22, MixTCRpred, TCRGP

ML.py (5) — ImRex, TITAN, TCellMatch, STAPLER

The catch: none of them are self-contained

Every External predictor inherits ARepoTCRSpecificityPrediction, which requires the user to clone that predictor's own repository and pass its path: 

raise AttributeError("Please provide 'repository' as a input argument to predict ...
                      'git clone {self.repo}'")

Execution then shells out through conda run -n <env> <interpreter> <cmd> — each predictor runs in its own conda environment, which is why the benchmark repo ships docker/envs/. Model weights are likewise obtained per-predictor from the original publications; nothing is bundled.

This machine has no conda installed.

Consequences for tcrga

  1. The torch-free scorers are not a fallback, they are the foundation. TCRdist kNN, the per-peptide PWM and the GBDT classifier are the only scorers that will run here without a multi-day environment build. Phase 3 order stands.
  2. Subprocess-per-individual is architecturally impossible in a GA inner loop — conda activation plus process spawn per candidate, per generation. If a neural predictor is ever wired in, it must be through the cascade path: one batched call per generation over the top-K survivors only, not per individual. The BindingScorer batch signature already allows this; ensemble.py must enforce it.
  3. The dl extra stays optional and unimported. Do not add ePytope-TCR as a dependency. If needed later, vendor a thin adapter behind the existing BindingScorer protocol and document the per-predictor setup separately.