Comment from Rominder Singh

Introduction: I am submitting these comments as Professor of Practice in Regulatory Sciences and AI at Northeastern University. With over 30 years of scientific leadership at top biopharma companies, I’ve contributed to drug discovery, development, & global registration. I also served on the ICH Expert Working Group for the E17 guidelines and remain active in global regulatory harmonization. Background: I am commenting broadly across various sections of the draft guidelines on “Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products.” Comment 1. Definition of AI. The US FDA currently defines AI broadly, covering computer science, statistics, engineering, and decision sciences. This broad scope may cause ambiguity, scope creep, and resource dilution and hinder the ability to set focused regulatory goals for this emerging field. Recommendation: AI and ML should be redefined in the specific context of health regulatory sciences, as definitions vary by audience and use. Clarifying distinctions between generative and predictive AI would also help. A precise, unique definition tailored to human medicines would reduce confusion with definitions used by other US agencies. Comment 2. “Context of Use” (CoU) definition is vague. Unanticipated products and systems are bound to be presented for review to the FDA. Since the FDA is known to regulate through precedent or what is already established, the novel approach could create a conundrum for traditional modes of regulation. Does the FDA reject the application since there is no precedent, or does the FDA ask for an unexplainable human clarification on something AI created? How will this be put into CoU? Recommendation: To contextualize CoU in AI and drug development, a new term is proposed: AI-Enabled Ecosystems for Therapeutics (Ai2ET). These ecosystems encompass AI-driven tools, workflows, and outcomes in biotherapeutics, integrating systems, processes, platforms, and products to accelerate innovation, trials, and manufacturing while maintaining safety, efficacy, & quality. A Platform provides tools, infrastructure, and services to build and manage AI models. A System uses AI algorithms to perform tasks requiring human intelligence. A Process refers to structured steps like data collection, model building, and evaluation. Products are therapeutics designed, developed, or manufactured using AI. Ai2ET aligns well with FDA regulations—both AI and non-AI—across drug discovery, clinical trials, manufacturing, safety, regulatory operations, and lifecycle management. See Figure 1. Comment 3. Narrow Focus. The scope does not include early-stage drug discovery or operational efficiencies, such as regulatory operations, unless they impact patient safety, drug quality, or study reliability. Additionally, there is little or no cross-reference to existing FDA guidelines on Devices and SaMD. Recommendations: Referring to the FDA DHAC meeting in 2024—which focused on generative AI in medical devices—it would be helpful to include discussion on premarket evaluation, risk management, and postmarket monitoring. The committee’s recommendations on data bias, hallucination risks, and the need for transparency in AI development should be incorporated. Building on Comment 2, adding drug discovery guidelines or operational efficiencies within the Ai2ET framework would make the guideline more complete. Comment 4. Risk-based monitoring examples. These draft guidelines introduce a risk-based credibility framework for AI applications in drug and biologic regulation, requiring context-specific model evaluation. Without clear guidelines, how will such medicines be regulated? The FDA offers guidance on Generally Accepted Scientific Knowledge (GASK) for these applications. Due to the “black box” nature of AI, applying GASK will depend on human judgment to define what is considered “scientific” and “knowledge.” Recommendations: Include a decision tree (see Figure 2: “To Regulate or Not to Regulate AI-Enabled Ecosystem for Therapeutics?”). The focus is on AI-enabled therapeutic ecosystems—not traditional human therapeutics already addressed by regulations. This decision tree can aid regulatory choices where formal guidance is lacking, particularly using a risk-based approach and CoU. Conclusion: A clearer FDA definition of AI and examples for Context of Use (CoU) are needed. Regulatory uncertainty remains around AI technology, its components, and outcomes. The proposed AI-Enabled Ecosystems for Therapeutics (Ai2ET) spans tools, workflows, and AI-driven results in biotherapeutics, aiming to accelerate innovation while ensuring safety and quality. Including existing medical device and SaMD regulations could help align fast-evolving AI standards. However, even risk-based frameworks may struggle to address AI’s unknown risks without hindering innovation. A decision tree is suggested.