In Session 1 (Regulatory Perspectives & Opportunities), Dr. Mathew T. Thomas, ex-FDA and now Regulatory Advisor with InSilicoMinds, posed a two-part, operations-level question to his former colleagues: first, have AI-related inspections led to unusable data; and second, when and how should sponsors contact the Agency on AI models?
What the panel distilled to actions:
• Use existing FDA channels for generics: submit a Controlled Correspondence (CC) for written questions, for which FDA publishes scope, content, and timelines.
• If the approach is novel (including AI as a model): request the Model-Integrated Evidence (MIE) Industry Meeting Pilot under OGD to discuss feasibility, V&V expectations, and acceptance criteria for BE.
• For complex generics: leverage user-fee meeting programs (pre-ANDA and ANDA scientific meetings; Type A/B/C when applicable).
• To get input on the model artifact: consider a Model Master File (MMF), implemented via a Type V DMF, to centralize versioned model documentation and enable cross-reference in ANDAs.
• On inspections and unusable AI data: the panel noted more than 1,000 applications referencing AI but did not share inspection specifics. Regulators are treating AI as a modeling approach, not a novelty. The practical route is to use Controlled Correspondence, escalate to the MIE pilot when novel, show fit-for-purpose V&V, and document a governed process.
