Endpoints News webinar — “Pilots to Production: AI across R&D, Safety, Regulatory, and Clinical” (IBM-sponsored). Panel: Sudhir Saxena (Co-Founder & CTO, GNQ Insilico), Deva Ramachandran (Partner, AI & Automation Practice Leader, IBM Consulting), Stuart Pyle (Partner, Life Sciences R&D, IBM).
GNQ highlights (Sudhir)
- Framed the regulatory tailwind first: recent FDA guidance that mechanistically justified evidence can reduce the number of required trials — i.e., the agency already accepts that “why it works” carries evidentiary weight.
- Laid out three levels of AI maturity: what (correlation), how (biological reasoning), why (counterfactual) — all three must be productionized in regulated settings, but the latter two are absent from most AI platforms today, which reach only the “what.”
- Four shifts required: mechanism over correlation, individual over population average, transparency/traceability, and automated navigation of regulatory bottlenecks.
- Multiple myeloma case study was the centerpiece — biomedical reasoning model built on a systems-biology graph of 200M+ edges, layered with drug binding and affinity data. A variant (HSD11B1 family) scored benign by SIFT/PolyPhen but was flagged pathogenic by GNQ’s quantum models. Mechanism: the variant inactivates the enzyme metabolizing dexamethasone in bone marrow, so immunosuppression fails where it’s needed, while the steroid still circulates and suppresses in liver and kidney. Predicted creatinine/eGFR toxicity in silico; observed it in the data. Clean articulation of “the answer wasn’t in the data, it was in the mechanism.”
- On quantum, notably candid: not ready broadly, but two deployable applications today — molecular binding energy prediction and combination therapy optimization.
- 90-day recommendation: before asking AI to predict your next success, ask it to explain your past failures. Retrospective causal analysis on archived failed trials — generates hypotheses, demonstrates auditable reasoning to regulatory teams, and produces board-ready proof points.
IBM highlights
- Deva’s production case: agentic AI for regulatory filings across IND/NDA/BLA. First go-live summer 2024 with a large biopharma (Azure); latest this summer with a rare disease company (Google Cloud).
- IBM Content Collaborator — ~40 document types autogenerated including tables, figures, charts; many documents drafted in under eight minutes; stakeholder-rated quality scores of 80–95%.
- Governance architecture worth noting: operational agents (write, critique, review) supervised by governance agents (security, privacy, quality), with every action fully traceable for 21 CFR Part 11 and GxP.
- Process IQ ingests 50–100 SOPs, work instructions and transcripts at a time to visualize current-state process, recommend eliminate/simplify/automate, and generate the agent prompts.
- Forward vision: domain expert agents operating at MD/PhD level, biomedical foundation models, causal reasoning, quantum algorithms, and digital twins against EMR/omics — explicitly citing GNQ’s causal + quantum layer as the complement to IBM’s agentic orchestration. Also referenced Moderna and Cleveland Clinic.
- Stuart’s framing: winners won’t be those running the most pilots but those that learn continuously; PTRS as the real R&D productivity metric. His 90-day ask: establish a baseline of where AI is actually deployed across your organization — most companies don’t know.
Watch here:
https://webinars.endpoints.news/from-pilots-to-production-ai-across-rd-safety-regulatory-and-clinical/