Artificial intelligence-driven biomarker discovery and response prediction in cancer immunotherapy: a pan-cancer perspective

Abstract

Cancer immunotherapy has transformed medical oncology, while individual’s therapeutic response to treatment is difficult because of the diverse nature and features of tumors and low precision and accuracy of markers, such as PD-L1, Tumor Mutational Burden (TMB), and Microsatellite Instability (MSI). Artificial intelligence integrates different types of data to cover the complexity of tumor-immune biological mechanisms. This review summarizes the recent studies in AI application in biomarker discovery and cancer immunotherapy prediction from a pan-cancer perspective, summarizing AI techniques, data types, clinical performance metrics, and implementation barriers. Selected AI-derived biomarkers have demonstrated promising performance compared with conventional biomarkers in specific studies. Indeed, selected AI-derived biomarkers demonstrated moderate-to-high discriminatory performance while reported estimates varied according to cancer type, prediction endpoint, and validation design. Deep Learning has shown promising results in Digital Pathology and Radiomics via providing tumor-infiltrating lymphocytes quantification and spatial immune phenotyping. Meanwhile, multi-omics integration frameworks, such as graph neural networks and transformer-based models can reveal complex associations among genomic, transcriptomic, and proteomic data. AI has indicated promising results in predicting both primary and acquired resistance to immune checkpoint inhibitors, stratifying potential patients for immunotherapy, and predicting immune-related adverse events. Post-hoc explainability methods, including SHAP and LIME, can provide feature-attribution information, although their outputs should not be interpreted as causal or mechanistic explanations. The intersection of artificial intelligence and cancer immunotherapy is opening new roads for precision-oncology. However, the clinical translation of AI-driven biomarkers requires robust external validation, standardized methodologies, and evaluation across diverse patient populations.