Deep learning identifies a T-cell exhaustion-dependent transcriptional signature for predicting clinical outcomes and response to immune checkpoint blockade.

Abstract

Immune checkpoint blockade (ICB) therapies have brought unprecedented advances in cancer treatment, but responses are limited to a fraction of patients. Therefore, sustained and substantial efforts are required to advance clinical and translational investigation on managing patients receiving ICB. In this study, we investigated the dynamic changes in molecular profiles of T-cell exhaustion (TEX) during ICB treatment using single-cell and bulk transcriptome analysis, and demonstrated distinct exhaustion molecular profiles associated with ICB response. By applying an ensemble deep-learning computational framework, we identified an ICB-associated transcriptional signature consisting of 16 TEX-related genes, termed ITGs. Incorporating 16 ITGs into a machine-learning model called MLTIP achieved reliable predictive power for clinical ICB response with an average AUC of 0.778, and overall survival (pooled HR = 0.093, 95% CI, 0.031-0.28, P < 0.001) across multiple ICB-treated cohorts. Furthermore, the MLTIP consistently demonstrated superior predictive performance compared to other well-established markers and signatures, with an average increase in AUC of 21.5%. In summary, our results highlight the potential of this TEX-dependent transcriptional signature as a tool for precise patient stratification and personalized immunotherapy, with clinical translation in precision medicine.