TCRcube

A T cell decides whether to attack by matching its receptor against a short peptide displayed on an MHC molecule. Predicting which T-cell receptors (TCRs) bind which peptide-MHC complexes (pMHCs) would shortcut neoantigen vaccine design and TCR-T cell therapy, and many machine learning models report strong accuracy on the task. TCRcube tests how much of that accuracy survives peptides the model has never seen, and builds a predictor that holds up better when it does not.

TCRcube is a family of binary classifiers developed by Martin Culka, Nicolas Lounsbury and colleagues in the AlQuraishi lab at Columbia University with a team at Genentech, posted to bioRxiv in April 2025 and not yet peer reviewed. Rather than learning receptor and peptide representations from scratch on the roughly 13,000 public binding measurements available, it reads representations out of frozen protein foundation models — ESM-2 embeddings and AlphaFold 2 evoformer or structure-module activations — and trains only a thin scoring head on top. The checkpoints released with the paper are fixed, and predict.py scores new TCR-pMHC pairs without retraining.

The accompanying evaluation is as much of the contribution as the model. On a leakage-controlled pan-peptide split, three published generalist predictors — TULIP, PanPep and TABR-BERT — perform at or below random, while physics-based scoring functions on predicted structures outperform them. TCRcube matches those biophysical baselines out of distribution and leads in distribution, but the paper is explicit that peptide generalization remains unsolved — a candour that distinguishes it from catalog neighbours such as TridentTCR and LoFT-TCR, which target the peptide-specific regime.

#Key Features

• Three-body interaction tensor: Peptide, MHC and CDR3 representations are projected to a shared 384-dimensional space and combined by outer product into a third-order tensor, so the score accounts for peptide-assisted TCR-MHC and MHC-assisted peptide-TCR contacts, not just peptide-TCR contacts.
• Energy-style read-out: Each residue triplet is mapped to a scalar energy through a ReLU and linear layer; energies are summed, length-normalised over the alpha and beta stacks, and passed through a sigmoid to give a binding probability.
• Six interchangeable input representations: One architecture accepts amino acid identity, identity plus position, ESM-2 embeddings of isolated fragments or full complexes, and AlphaFold2 evoformer or structure-module representations, isolating what foundation-model features contribute.
• Leakage-controlled benchmarking: Negatives are generated by mismatching TCRs to pMHCs within each split separately, and every complex with an experimental structure in the PDB is excluded, closing two routes by which earlier models inflated out-of-distribution scores.
• Distance-stratified generalization analysis: Accuracy is reported as a function of Levenshtein distance between test and training epitopes, a reusable protocol for quantifying how far any TCR-pMHC model extrapolates.

#Technical Details

Representations are precomputed outside the model: ESM-2 embeddings come from the 3-billion-parameter esm2_t36_3B_UR50D checkpoint (2,560 dimensions per residue, final layer), and AlphaFold2 representations from the TCRdock template-based pipeline, which needs no MSA and yields 384-dimensional single representations. Training uses binary cross-entropy with positives upweighted, Adam at a learning rate of 1e-4, for 500 epochs. The peptide-specific benchmark is IMMREP_2022 (14,719 training and 4,312 test points across 17 epitopes); the pan-peptide split draws on VDJdb, McPAS-TCR, IEDB and 10x Genomics data restricted to human MHC-I with paired CDR3 alpha and beta chains, giving 79,086 training and 1,207 test points.

On IMMREP_2022 the ESM2-full variant is the strongest of the family and competitive with the best published peptide-specific methods, with full-complex embeddings beating embeddings of isolated fragments. On the pan-peptide split every variant falls back to the level of structure-based scoring functions: AlphaFold2 PAE at the interface, CHARMM36 interaction energies after NAMD minimisation, and HADDOCK re-docking scores. On a blind Genentech/Adaptive Biotechnologies set of 1,446 points built from neoantigen-specific TCRs in 30 non-small-cell lung cancer patients, all methods including TCRcube predict poorly.

#Applications

The peptide-specific setting is where the released checkpoints earn their keep: screening a patient's repertoire for receptors against a well-characterised epitope for which training data already exists. The pan-peptide checkpoints are best read as a calibrated baseline — useful for ranking candidates when nothing better is available, and a warning against trusting confident predictions on novel cancer neoantigens. Pan-peptide weights also serve as initialisation for peptide-specific training, boosting accuracy on the HPVTKYIM epitope, a practical transfer recipe for groups with small epitope-specific datasets.

#Impact

TCRcube's main effect on the field is corrective: it shows that reported out-of-distribution success in TCR specificity prediction largely reflects negative-sampling choices and train-test epitope overlap, and it supplies both a cleaner benchmark and a competitive model against which to measure progress. Code, weights and public data splits are released under Apache-2.0. Two limitations are worth stating plainly. The AlphaFold2 and ESM-2 variants require precomputing representations through external TCRdock and ESM pipelines, so inference is not a single self-contained call. And the lung-cancer neoantigen test set is licensed from a third party and cannot be redistributed, so that headline zero-shot evaluation is not independently reproducible.