HAPIR: a refined Hallmark gene set-based machine learning approach for predicting immunotherapy response in cancer patients.

Immune checkpoint inhibitors (ICIs) have revolutionized cancer treatment, yet the response rate remains limited, with only about 30% of solid tumor patients benefiting. Identifying reliable biomarkers to predict ICIs response remains a significant challenge. In this study, we proposed a refined Hallmark gene set-based Approach for Predicting Immunotherapy Response (HAPIR). Through comprehensive multi-cohort analyses encompassing six TIGER cohorts (n = 352) and TCGA-SKCM (n = 472), we validated the optimal performance of HAPIR. Using transcriptomic data from a training cohort, we firstly refined seven Hallmark gene sets enriched with differentially expressed genes between responder and non-responder patients. Then, a logistic regression model trained based on the activities of these gene sets demonstrated superior predictive performance (AUROC = 0.778) in ten-fold cross-validation, significantly outperforming 13 existing biomarkers, including PD-1 (AUROC = 0.678) and PD-L1 (AUROC = 0.54). HAPIR's robustness was further validated in the validation set and four independent cohorts spanning multiple cancer types (melanoma, NSCLC, and STAD), consistently achieving average AUROC = 0.745. Beyond well-known biomarkers, HAPIR surpassed both gene-based and alternative gene set-based models. Importantly, HAPIR scores correlated significantly with patient survival and effectively recapitulated the immune microenvironment, enabling the prediction of potential drug targets and drug candidates to overcome immunotherapy resistance. In conclusion, HAPIR is a promising tool for predicting ICIs response and guiding the development of new immunotherapy strategies.