Benjamin D. Greenbaum

Ensembles of in silico structures enable T cell peptide-MHC binding prediction 1. The paper introduces enFoldX, a structure-ensemble method that predicts whether a TCR and a peptide-MHC (pMHC) are cognate binders by featurizing many AlphaFold3 (AF3) co-folded complex samples, rather than trusting a single “best” predicted structure. 2. Core idea: AF3 can generate plausible-looking false positives for non-cognate TCR:pMHC pairs, especially when sequences differ subtly. enFoldX turns this weakness into signal by measuring how confidence and interface geometry vary across an ensemble (multiple seeds, multiple samples per seed). 3. enFoldX focuses feature engineering on the binding interface, especially peptide–CDR3 contacts, using AF3 confidence outputs such as pLDDT, PAE/pAE, pTM, and chain-pair/interface variants (iPAE, ipTM), plus inter-chain contact probabilities. It summarizes each feature by mean and standard deviation across the ensemble, capturing both “typical” behavior and uncertainty/diversity. 4. A motivating example is a CMV epitope differing by one amino acid (wild-type NLVPMVATV vs mutant NLVPKVATV): both cognate and non-cognate ensembles can contain “hallucinated” misdocked structures, but the non-cognate shows a much higher hallucination rate across seeds—an ensemble-level signature that single-structure methods can miss. 5. Benchmarking on VDJdb (positives only) required synthetic negatives. The authors generate non-cognate pairs by swapping in “decoy” peptides predicted to bind the same MHC (NetMHC/NetMHCpan), aiming to make negatives fail at TCR recognition rather than MHC presentation. They also test a simpler “permuted” negative scheme (random TCR/peptide shuffling within an MHC allele). 6. With ridge regression and careful cross-validation splits to reduce leakage, enFoldX achieves (human VDJdb) mean AUC ~0.82 for TCR-wise CV and ~0.73 for epitope-wise CV; (mouse VDJdb) ~0.98 and ~0.95, respectively. A notable result is cross-species transfer: a model trained on human data can classify mouse TCR:pMHCs. 7. Feature attribution (SHAP) highlights interface-focused signals: iPAE and ipTM at TCR:pMHC and pMHC interfaces dominate, and both means and standard deviations matter—supporting the claim that “distribution over structures” carries non-redundant information. 8. Generalization tests include: new VDJdb release data (AUC ~0.86 on new TCRs, ~0.77 on new epitopes), a functionally validated TCRvdb benchmark (AUC ~0.75), promiscuous TCRs such as 1E6 (AUC ~0.92 on a 185-epitope evaluation), and MEL5/MEL8 where predicted probabilities separate validated cognate vs non-cognate sets. 9. On the IMMREP25 “unseen peptides” challenge setting (high epitope divergence), enFoldX attains AUROC ~0.65; an ensemble of ridge regression + XGBoost + random forest improves macro-AUROC to ~0.66 and performs strongly under low-FPR metrics compared to reported IMMREP25 baselines, despite label noise concerns in that dataset. 10. For the neoantigen-relevant regime of single amino-acid peptide substitutions, enFoldX is evaluated on eight mutational-scan datasets (macro-AUC ~0.71; range ~0.59–0.82). Additional leave-one-position and leave-one-TCR analyses on a multi-concentration activation dataset show consistent above-random performance across all positions/TCRs, supporting sensitivity to point mutations. 11. The paper also introduces BERTie, a sequence-only BERT-style baseline trained on CDR3β + peptide, showing strong performance when test sequences resemble training data but substantial degradation as epitope distance grows. In contrast, enFoldX performance is reported to be comparatively independent of epitope sequence similarity, aligning with its improved out-of-distribution behavior. 12. Limitations and practical notes: ensembling increases compute cost; performance drops for underrepresented HLA groups (notably some HLA-C alleles), pointing to data scarcity. enFoldX remains informative even when AF3 predictions degrade (e.g., shallow/no MSA, no templates), implying systematic ensemble differences persist even with imperfect structures. 💻Code: github.com/jonlevi/enFoldX 📜Paper: biorxiv.org/content/10.64898… #ComputationalImmunology #TCR #pMHC #AlphaFold3 #ProteinStructure #MachineLearning #Bioinformatics #CancerVaccines #Neoantigens #Immunotherapy