Label Noise Limits TCR-pMHC Specificity Prediction: Improved Performance Through AlphaFold3-Based Structural Modeling and Data Denoising
The study argues that label noise in TCR-pMHC specificity datasets is a major driver of weak performance for unseen peptide prediction, where structural modeling has so far shown the most predictive power. Using an AlphaFold3-based pipeline adapted for TCR-pMHC structural modeling, the authors report state-of-the-art specificity prediction that outperforms AlphaFold2.3-based and sequence-based methods and performs at par with the leading Immrep2025 competition submission. A cluster-based denoising step that removes mislabeled points from a large specificity dataset increased binder ranking accuracy by more than 70% relative to the full dataset, implying that data quality improvements could materially affect TCR-based immunotherapy and vaccine design.