AI and the World's Hard Problems · Life sciences · Nature (machine learning) · 2026-10-26
Machine learning dataset and benchmark for accurate T cell receptor–pHLA binding prediction
Predicting how T cell receptors bind peptide–HLA complexes is a central challenge in immunology and therapeutic design, and existing tools are limited by incomplete sequence data and biased nonbinding examples. The authors built Hi-TPH, a large hierarchical dataset with an on-the-fly method for generating nonbinding data, and fine-tuned a family of protein language models (Hi-TPH-PLMs) on it. On an external wet-lab test set, these models achieved a 19.2% performance gain over state-of-the-art tools, and analyses showed that HLA's contribution to prediction depends on the presence of full T cell receptor chains.