Session I — Drug Discovery I
The Productivity Paradox, the Dynamic Proteome, and What Molecules Teach Us
Three R&D leaders — Mike Quigley (Sanofi), Karen Akinsanya (Schrödinger), and Alistair Henry (UCB) — circle a single uncomfortable truth: better science alone hasn’t moved the productivity needle. The fix is upstream of the bench.
Mike Quigley — Beyond Biology: Decision Systems for Scalable Discovery
Core argument: Pharma’s productivity paradox won’t be solved by more science. It requires treating discovery as a continuous decision ecosystem — where to invest, how to source, when to advance.
Three pillars of the go-forward operating model
- Where to invest — asset optionality, downstream value, portfolio-level risk
- How to source — assess every external lever on equal footing with internal execution
- When to advance — rigorous stage-gating, strict budgets, fast pivots on evidence
Constellation view
Sanofi’s constellation view maps biologies across (not within) therapeutic areas, supported by an AI-driven indication-expansion engine. Core node: immunoscience — overlapping immunologic pathways with broad cross-TA applicability.
Worked example — the complement pathway: multiple intervention nodes, over-activated across rare blood disorders, ophthalmology, and neurologic disease.
Capital allocation
Sanofi has built an agentic AI workflow proposing real-time decisions across internal pipeline, Sanofi Ventures, Sanofi Capital, and BD/M&A — collapsing decision lag.
Karen Akinsanya — From Edges to Atoms to Drugs
Core argument: We’ve built a trillion-dollar industry on ~900 protein targets out of ~20,000 proteins and millions of proteoforms. The next leap is dynamic biology at scale.
SARM1 case study
- Cryptic pocket: a tryptophan rotated and opened a previously invisible pocket once a designed molecule engaged
- Biomarker reversal at low dose: sub-stoichiometric coverage left catalytic sites uncovered, reversing the biomarker at 25 mg/kg — a clinical liability exposed only by dynamics
Three blind spots
- Edges: ~92% of the protein–protein interaction network is empty; only ~4% of an estimated 650,000 PPIs have resolved structures
- Atoms: The cytoplasm is crowded (~10 binding partners per protein); biochemistry runs at ~300× dilution
- Drugs: Across 234 cancer drug approvals (2003–2021), median PFS gain is ~3.3 months
SGR-1505
Physics-discovered compound, asset to clinic-ready in 10 months. 100% response rate in Waldenström’s macroglobulinemia.
Alistair Henry — Drug Discovery By Design, Not By Chance
Core argument: Every molecule tells a story. Listen to humans first, design at the atom level, and back-translate relentlessly.
Three molecules, three lessons
- Romosozumab (Evenity) — born from ultra-rare genetics; cloning SOST in ~60 patients worldwide gave sclerostin and a path into osteoporosis
- Rozanolixizumab (Rystiggo) — driven by patient need; led to FcRn inhibition
- Bimekizumab (Bimzelx) — the headliner; lessons in why mice aren’t men
F: human pathobiology wins
In peripheral blood under chronic inflammation: IL-17A initiates, IL-17F drives the late phase. Run the same in mouse — no IL-17F signature. Human tissue confirmed the human pattern.
Atom-level design for dual specificity
One antibody, selective for A and F only out of five isoforms. Pre-AI: >28,000 variants → ~20 sequences → 5 made → 1 drug.
Back-translation
Reverse expression profiling on psoriasis transcriptome with bimekizumab showed:
- Biology was monocyte-driven more than appreciated
- IL-23 as an upstream regulator of IL-17A/F — complicating the textbook “waterfall” model
Panel Highlights
- AI’s leverage today: Table stakes in regulatory, enrollment, supply chain, DMTA. Unsolved: end-to-end connectivity from target ID through clinical pharmacology.
- AI-native biotech thesis: Three-plus years on, timelines haven’t moved. 10 years from start to approval has held steady for decades.
- Training data problem: An ML model learned which hospital a slide came from, not the biology. Data provenance is the constraint.
- Talent: “Bilingual” tech-and-biology scientists needed. Hire for perpetual curiosity.
- NAMs: Organoids, tissue explants, patient-derived material are routine at UCB. The question is always what is this model actually modeling?
Cross-Cutting Themes
- The productivity paradox is an operating-model problem, not a molecule problem
- Static views of biology are the bottleneck — dynamics is where the next decade’s drugs hide
- Human pathobiology beats animal models, and the field is willing to say so
- AI’s binding constraint has shifted from algorithms to data provenance
- Back-translation from approved drugs is an under-exploited engine
Comments