PS04: Agentic AI in US Regulatory Science: Current Methodological Developments and Early Guidance for Sponsors

The FDA's December 2025 deployment of agentic AI-systems that can plan, analyze, cross-reference, and execute multi-step regulatory tasks-represents an emerging opportunity in how submissions, inspections, and reviews may be conducted. At this initial stage, agentic AI capabilities are being used for exploratory purposes. It is hoped that these tools will introduce time efficiencies and enhance review practices by accelerating routine analyses, enforcing consistency across documents, and flagging discrepancies that previously required manual or discretionary review. As a result, sponsors with clean, consistent, well-documented submissions may experience faster, more uniform regulatory scrutiny. This session will share preliminary agentic AI experiences from academia, government (FDA), and industry. Specifically, we will hear from three speakers who have valuable insight and expertise in this area. Speaker 1 (Yanxun Xu, Johns Hopkins University) will present current methodological advances in task automation for biomedical research, highlighting how large language models (LLMs) can streamline complex workflows and be tailored to the specialized needs of clinical studies. The talk will showcase practical applications, such as automating landscape analyses and extracting individual patient data from Kaplan-Meier plots using multimodal AI, and outline a methodological framework for responsible AI deployment. Best practices for data privacy will be emphasized, including strategies to minimize exposure of sensitive information, generate synthetic data, and deploy secure, locally hosted AI solutions. Next, Speaker 2 (Laura Thompson, FDA) will describe early experiences with agentic AI currently being explored by FDA, highlighting contributions from internal development efforts. These explorations span various aspects of regulatory review processes, from document analysis to data verification workflows. Examples in her talk may include proof-of-concept agents for clinical data analysis that investigate the potential to process CDISC submissions, extract endpoints from Clinical Study Reports, and create user-friendly datasets from ADAM data structures while generating reproducible verification code. The speaker will discuss how such tools might enhance reviewer capabilities by automating routine tasks, potentially allowing them to focus more attention on the complex analytical and interpretive work that requires human expertise and judgment. Finally, the Discussant (Justine Rochon, Takeda) will briefly summarize the first two presentations, and subsequently share the current state of generative AI and agentic AI from a sponsor's perspective, drawing on insights from the DISRUPT-DS (Data Science) industry roundtable and examples from her own organization. The remarks will include some of the generative AI tools and autonomous AI agents that pharma companies are already using. Referencing the DISRUPT-DS benchmarking across 14 major pharma companies and reflecting the evolving landscape, the speaker will suggest steps sponsors need to take now to prepare for a significant shift in US regulatory science. The proposed session is extremely timely, and should appeal to a large collection of biostatisticians and clinical trialists interested in learning about current developments and applications in this important and rapidly evolving area. Format of session: Chair, two speakers, and discussant