Artificial intelligence for predicting immune checkpoint inhibitor outcomes through multi-omics integration

Immune checkpoint inhibitors (ICIs) have reshaped cancer therapy, yet durable benefit remains limited to a subset of patients. Current biomarkers, including PD-L1 expression and tumor mutational burden, provide only partial predictive value because response to ICIs arises from the interplay of tumor antigenicity, antigen presentation, immune infiltration, spatial organization of the tumor microenvironment, and host systemic factors. In this context, multi-omics profiling offers a more comprehensive framework for biomarker development by integrating genomic, transcriptomic, epigenomic, proteomic, metabolomic, microbiome, imaging, and longitudinal liquid-biopsy data. Artificial intelligence (AI) methods are increasingly used to fuse these heterogeneous modalities and model nonlinear interactions relevant to treatment benefit and resistance. This narrative review summarizes the biological basis for multimodal biomarker development, examines major AI-driven integration paradigms, including early, late, intermediate, hierarchical, graph-based, and latent-factor approaches, and highlights their respective strengths, limitations, and clinical use cases. We further discuss recurrent methodological challenges, particularly endpoint heterogeneity, missing modalities, information leakage, domain shift, inadequate external validation, and poor calibration. Finally, we propose practical standards for model evaluation and reporting to support clinical translation. Progress in this field will depend not only on increasingly sophisticated algorithms, but also on rigorous study design, harmonized outcome definitions, and multicenter validation.