SC08: Adapting Generative AI for Biomedical Research and Drug Development

Large language models (LLMs) are rapidly being adopted in biomedical research and drug development for literature review, evidence synthesis, protocol and report drafting, and regulatory intelligence. However, general-purpose models can hallucinate; lack provenance; and produce unstructured outputs that are hard to audit, validate, or reuse. In regulated, collaborative research settings, these limitations can undermine scientific rigor and affect patient safety. This short course provides a practical, statistics-oriented introduction to how generative AI can be adapted into research-grade, auditable tools that fit biomedical and clinical workflows. We will introduce three complementary approaches---retrieval-augmented generation (RAG), supervised fine-tuning, and structured-output LLM agents---and show how they transform probabilistic text generators into grounded, reproducible, and machine-readable analytic tools. Using examples drawn from evidence synthesis and survival analysis workflows, participants will see how to: (1) retrieve and reason over validated sources; (2) generate statistically structured outputs; and (3) coordinate multi-step analytical pipelines such as extracting individual patient data from Kaplan–Meier curves. The short course emphasizes reproducibility, transparency, and statistical reliability, bridging modern AI with the requirements of regulatory review, clinical evidence synthesis, and pharmaceutical decision-making. Intended for statisticians with no prior AI background, this short course focuses on conceptual clarity, practical workflows, and safely deploying generative AI in high-stakes biomedical settings.