Add MixTCRpred pMHC:TCR binding predictor wrapper

[iskandr]

Wrap MixTCRpred as an mhctools pMHC:TCR binding predictor (pMHC_TCR_binding).

Upstream: https://github.com/GfellerLab/MixTCRpred — a transformer predicting TCR binding to a given pMHC; one model per pMHC.

Pretrained weights: on Zenodo (record 7930623, CC-BY-4.0) — 146 per-pMHC models (~32 MB each), auto-downloaded on demand by the code; 2 are also in-repo. No retraining needed.

Framework / deps: PyTorch + PyTorch-Lightning, modern (Python 3.9–3.11) — the most current stack of the benchmark set. Colab provided.

License: code is Ludwig Institute academic / non-commercial; the weights are separately CC-BY-4.0 (freely redistributable with attribution).

Inputs / output: paired αβ CDR3 + V/J genes → binding score. The presenting pMHC is selected by which model you load (mhc_dependence effectively per-model / single_allele). Note 43/146 models are flagged "high-confidence".

Ease: Easy–moderate. Modern install; the main wrinkles are the paired-chain + V/J input requirement (richer than mhctools.TCR's CDR3-centric default — the TCR type does carry all six CDRs, but V/J genes would need to be supplied) and the one-model-per-pMHC design (map requested allele+peptide → model name).

Suggested wrapping approach: in-process via Lightning as an optional extra, or subprocess; auto-fetch the requested pMHC model from Zenodo in a setup step. Expose the model catalog so callers can see which pMHCs are covered.

---

Surfaced from a survey of TCREpitopeBenchmark for methods with public pretrained weights that run easily for inference.

https://claude.ai/code/session_01LZahFhBSCiehXTESCYQ7wG

[iskandr]

⚠️ Read before investing effort: the generalization ceiling

Every TCR–epitope binding predictor in this backlog — this one included — shares a well-documented limitation: they do only modestly on epitopes seen in training and drop to near-random on unseen/novel epitopes. Independent benchmarks find that no current method reliably generalizes; the "best" scores on unseen data sit around the random floor (e.g. mean AUPRC ~0.52–0.53). A tool can look decent on a test set dominated by common epitopes (GILGFVFTL, NLVPMVATV, …) purely because those TCRs saturate the public databases, then fail on anything new.

References

• Moris et al. 2021, Briefings in Bioinformatics — Current challenges for unseen-epitope TCR interaction prediction (foundational: the unseen-epitope challenge, negative-data bias, imbalanced-epitope distribution)
• Meysman et al. 2023, ImmunoInformatics — Benchmarking solutions to the TCR epitope prediction problem: IMMREP22 workshop report (first public benchmark, 23 models) and the IMMREP23 follow-up
• Nature Methods 2025 — Assessment of computational methods in predicting TCR–epitope binding recognition

Implication for this wrapper: worthwhile for reproducibility, cross-method benchmarking, and the well-characterized "seen-epitope" regime — not as a trustworthy predictor for novel epitopes / neoantigens. When we build it, the wrapper's docstring should state this ceiling so users don't over-trust the output.