Abstract / Summary
Biomarkers are central to modern diagnostics and therapeutics, yet traditional discovery approaches suffer from single-modality analyses, weak mechanistic foundations, and low reproducibility. Artificial intelligence (AI) enables the integration of complex multimodal biomedical data, but the translation of AI-derived biomarkers into clinical applications remains inconsistent. This review examines how AI reshapes biomarker discovery across diseases, focusing on biological mechanisms, validation requirements, and therapeutic integration. We synthesize evidence across molecular, cellular, imaging, and digital biomarker approaches, evaluating AI methodologies including classical machine learning, deep learning, graph-based models, foundation models, and causal inference. Findings are organized using a pipeline framework encompassing discovery, external validation, robustness testing, clinical utility, and deployment. AI enables identification of multiscale signatures reflecting cellular programs, tissue remodeling, and disease trajectories. However, many biomarkers fail external validation. Biologically grounded approaches using network-based modeling, spatial profiling, and mechanistic constraints improve interpretability and therapeutic relevance in oncology, cardiovascular disease, neurodegeneration, and metabolic diseases. Validation frameworks with decision curve analysis and prospective evaluation are necessary to demonstrate clinical utility beyond predictive accuracy. AI's impact will depend on integration into rigorous validation and therapeutic pathways. By transitioning from correlational pattern recognition to mechanistically informed, clinically evaluated systems, AI can support next-generation precision diagnostics and targeted interventions.