enFoldX Turns AlphaFold3 Noise into TCR Recognition Signals

Researchers introduced enFoldX, a novel framework that transforms AlphaFold3's (AF3) structural prediction noise into predictive signals for T-cell receptor (TCR) recognition. The model achieves an AUC of 0.82 on human VDJdb data and 0.80 on mouse data, addressing the poor generalization of traditional sequence-based models to novel peptides or TCRs.

Confirmed

Core Mechanism: Instead of relying on a single predicted structure, enFoldX generates a full structural ensemble using 10 seeds and 5 samples for AF3 predictions of TCRα, TCRβ, MHC, and peptides. The model uses this ensemble directly as classification input, extracting 106 interface features to capture the true characteristics of TCR:pMHC interactions.

Hallucination Frequency as Signal: Taking the CMV epitope NLVPMVATV as an example, AF3 hallucinates for both binders and non-binders, but the frequency of these hallucinations is informative. AF3 exhibits greater structural divergence for non-cognate pairs, producing hallucinations in about 7 out of 10 predictions. enFoldX leverages this high-frequency noise characteristic specific to non-binders to effectively predict T-cell specificity.

Evaluation & Extensibility: In 10-fold cross-validation on VDJdb, enFoldX achieved an AUC of 0.82 on human TCR-split evaluations and 0.80 on mouse data, demonstrating robust cross-species transferability. It also delivered leading results on neoantigen tasks with single amino acid differences and unfamiliar peptide predictions. The authors emphasize that structural ensembles are the core of robust generalization, meaning this framework can theoretically extend beyond AF3 to other structural prediction models like RoseTTAFold and Boltz.

2026-07-22 ~ 2026-07-22 · 11 related posts

• enFoldX uses noisy AlphaFold3 ensembles to predict T cell specificity — quaidmorris · 2026-07-22
• [source] AF3 hallucinations appear in 7 of 10 seeds for a non-cognate TCR:pMHC pair — quaidmorris · 2026-07-22
• enFoldX gains accuracy as AF3 ensemble disagreement rises for non-binders — quaidmorris · 2026-07-22
• enFoldX classifies TCR–pMHC pairs from AF3 ensembles instead of one structure — quaidmorris · 2026-07-22
• [source] enFoldX extracts 106 interface and confidence features from AF3 ensembles — quaidmorris · 2026-07-22
• enFoldX reaches AUC 0.82 on human VDJdb and transfers to mouse at 0.76 — quaidmorris · 2026-07-22
• enFoldX tops 8 neoantigen scans and an unseen-peptide benchmark — quaidmorris · 2026-07-22
• [source] enFoldX turns AlphaFold3 ensemble noise into a TCR–peptide–MHC predictor — quaidmorris · 2026-07-22
• Structural ensembles beat single predictions in TCR:pMHC generalization study — quaidmorris · 2026-07-22
• Structural ensembles, not single predictions, drive robust TCR:pMHC generalization — quaidmorris · 2026-07-22
• Turning Noise into Signal: Predicting TCR Binding Using AlphaFold3 Hallucinations — quaidmorris · 2026-07-22