A Machine Learning-Optimized Immunogenic Cell Death Signature Reveals Tumor Immunogenicity and the Immunotherapy Response of Pancancer

Abstract

Tumor immunogenicity determines their response to immune checkpoint inhibitors (ICIs), but the mechanisms governing pancancer immunogenicity remain incompletely understood. A further critical barrier to developing reliable predictive biomarkers is data set shift, which undermines model generalizability. Here, we address these challenges by developing a novel adversarial validation (AV)-integrated machine learning framework, focusing on immunogenic cell death (ICD)-related gene signatures (ICDRSs). We designed three AV-based strategies to mitigate data set shift and validate the efficacies across multiple machine learning algorithms. Using dual-modal data from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO), four optimal AV-based classifiers (e.g., GradientBoosting, XGBoost, LGBM, and CatBoost) were screened, which effectively reduced inter-cohort shift, enhancing both accuracy and robustness of downstream analysis. We identified novel risk/protective ICDRSs that strongly predicted patient survival and tumor immunogenicity across cancers. High-risk ICDRSs correlated with immune-exclusive microenvironments marked by impaired antigen presentation and aberrant tumor-associated macrophage development, as revealed by single-cell RNA sequencing. Validation across 13 ICI-treated cohorts revealed the capacity of ICDRSs for anti-PD-1 nonresponse. Mechanistically, risk ICDRSs were linked to CD47-SIRPA-mediated immune evasion and proliferative macrophage subsets with terminal dysfunction. This study advances understanding of tumor immunogenicity, provides novel biomarker development tools, and supports personalized cancer immunotherapy decision-making.