Machine learning dataset and benchmark for accurate T cell receptor–pHLA binding prediction

A central challenge in immunology and therapeutic design is accurately predicting the diverse interactions between T cell receptors and peptide-human leukocyte antigen (pHLA) complexes. Existing machine learning tools are hindered by incomplete sequence data and biased nonbinding examples. Here we present Hi-TPH, a large-scale hierarchical dataset featuring an on-the-fly selection method for generating nonbinding data. We further develop Hi-TPH-PLMs, a collection of protein language models (PLMs) with varied architectures and scales, fine-tuned on Hi-TPH. These models achieve a 19.2% performance gain over state-of-the-art tools on an external wet-lab test set. Leveraging the hierarchical structure of Hi-TPH, detailed analyses dissect the contribution of different molecular components to binding prediction and reveal their synergistic interplay; for instance, the prediction contribution of HLA relies on the presence of full T cell receptor chains. Hi-TPH and Hi-TPH-PLMs are publicly released to support the development of more reliable tools for advanced immunoinformatics research and personalized immunotherapy.