SC03: Statistical and Machine Learning Methods with Trial Design Strategies for Precision Medicine

Biomarkers are integral to modern health care and drug discovery. Across a drug's life cycle, they deepen disease understanding, guide compound selection, enable population targeting, and accelerate trials - foundations of precision medicine. Precision medicine, often termed "personalized medicine," aims to tailor treatments to individual biology rather than apply a "one size fits all" approach. This shift is propelled by advances in omics, bioinformatics, and data science, where machine learning and AI increasingly complement statistical methods to extract signal from complex, high-dimensional data. Since policy initiatives in the 2010s accelerated personalized medicine, approvals of biomarker-guided therapies have steadily grown (for example, hormone-receptor (HR)-guided approval of tamoxifen for breast cancer). The area has become more complex with multi-modal data and limited subjects causing a shift in the statistical approaches of biomarker-driven clinical trials. This momentum underscores the need for study designs and statistical approaches that can robustly identify, validate, and apply predictive biomarkers within trials. In clinical development, many therapies benefit only specific subpopulations. Biomarkers and model-based approaches help determine enrichment strategies, select responsive patients, and improve trial success. For continuous biomarkers, choosing cut-offs - or moving beyond fixed thresholds with probabilistic risk scoring - requires careful statistical reasoning. Key challenges include limited sample sizes, single arm trials, missing data, combining markers, high dimensionality, multiple testing, bias, and generalizability across cohorts. ML/AI methods - regularization, ensemble learning, causal ML, and Bayesian modeling - can augment classical statistics to address these challenges while maintaining interpretability and regulatory credibility. Decisions often rely on limited early-phase data. Model-informed decision-making, including synthetic control, digital/virtual twins, and simulation, can improve patient enrichment choices and optimize go/no-go criteria. Efficient designs - adaptive enrichment, model-based subgroup identification, and seamless phase designs - benefit from integrating statistical inference with ML-driven feature selection and prediction, provided rigorous validation, calibration, and bias reduction are in place. As precision medicine expands, biomarkers are central to patient stratification and trial optimization, and regulatory considerations remain pivotal. This short course emphasizes how statisticians can fuse ML/AI with statistical principles to strengthen biomarker discovery, validation, and clinical utility assessment - independent of any specific diagnostic pathway.