Pharma AI Strategies: Who Is Doing What and Why
A cross-company comparison of AI approaches discussed at SDDS 2026. Every major company has an AI strategy — but they differ in where they place the bet, how they integrate it, and what they’ve learned from failure.
Company-by-Company Matrix
| Company | AI Strategy | Key Win | Key Lesson | Session |
|---|---|---|---|---|
| Sanofi | Agentic AI workflow for real-time portfolio decisions; constellation view with AI-driven indication expansion | Real-time capital allocation across pipeline, ventures, BD | AI is decision infrastructure, not molecule infrastructure | I |
| Schrödinger | Physics-based computation + AI; dynamic proteome modeling | SGR-1505: asset to clinic in 10 months | Public chemistry data saturated; dynamic biology data doesn’t exist yet | I |
| Genentech | Lab-in-the-loop: foundation models across targets, indications, medicines, patients | AI delivered 50× better hit rate than medicinal chemist on same data | The loop — not any single model — is the unit of work | II |
| NVIDIA | Multiplicative compute stack: instruments → models → agents → robotic labs | Parabricks 30× genome assembly; CodonFM mRNA model | Specialized sub-agents under generic orchestrators, not one super-agent | II |
| Takeda | AI-driven-by-default wet lab; pivoted after Nabla Bio rescue | In-silico-designed antibody sequence for myasthenia gravis | AI’s first real wins came from problems traditional methods had abandoned | II |
| BMS | Deep mutational scanning + AI target nomination; CELMoD design | AI broke months-long MPO impasse on sickle cell program | AI most valuable where conventional approaches fail | IX |
| Amgen | Convergent molecular innovation; antibody-peptide conjugate design | Maritide geometry (spatial heterodimerization) | The molecule’s geometry is the drug; AI helps optimize but doesn’t replace design insight | VII |
| Bayer | Cradle collaboration; goal 40% productivity gain by 2030 | Finerenone 25-year arc now accelerating | BD partnerships and AI must compound together | IX |
Where AI Is Already Load-Bearing
These applications are table stakes across large pharma (not differentiating):
- Regulatory submissions and dossier assembly
- Clinical trial site selection, enrollment optimization, supply chain
- Design-make-test-analyze (DMTA) cycle acceleration
- Single-cell immunology analysis
- Imaging-based readouts (pathology, OCT)
Where AI Is Differentiating (But Data-Constrained)
These applications separate leaders from followers — and are limited by data provenance, not algorithms:
- Target identification — Genentech’s pathobiology foundation models more than doubled target-selection impact
- De novo molecular design — Nabla Bio’s in-silico antibody, Schrödinger’s physics-discovered compounds
- Trial outcome prediction — Pande’s thesis: this is the biggest AI alpha in drug discovery
- Patient stratification — Genentech’s OCT-based AI would have flagged a failed Phase 3 from Phase 2 data
- Closed-loop experiment design — BMS’s automated DMTA, Genentech’s cross-screen foundation models
Where AI Hasn’t Delivered Yet
- End-to-end connectivity from target ID through clinical pharmacology (Quigley)
- Timeline compression — 10 years from start to approval has held steady for decades despite AI
- Predicting adverse effects in an integrated organism — “the clinical data is tiny, tiny, tiny” (Baker)
- First-principles drug design for truly novel biology — AI interpolates well, extrapolates poorly (Cohen)
The Data Problem
The consensus across sessions: algorithms are ahead of the data.
- Public chemistry and structure data are saturated
- Dynamic biology data (proteoform states, PPI dynamics, in vivo perturbation responses) mostly doesn’t exist
- An ML model learned which hospital a slide came from, not the biology (Akinsanya, Session I)
- 300,000 atomic-resolution structures exist; clinical data is orders of magnitude smaller
- Whoever builds closed-loop autonomous experiment datasets owns the next decade
The VC/Banking View
From Sessions III and X:
- Pande (VZ.VC): AI for trials is the asymmetric bet — pairs with China in-licensing arbitrage
- Gupta (Goldman): TechBio M&A awaits clinical proof of concept off an AI platform; Coefficient Bio/Anthropic ($400M) is notable but not the unlock
- Lee (J.P. Morgan): “There are companies branded as ML or computational biology that have been around for 30+ years” — the excitement is real; the specifics are still forming
Cross-References
- AI in Drug Discovery — concept hub page
- Lab in the Loop — Genentech’s architecture
- Dynamic Proteome — the data gap
- China Strategy — AI + China arbitrage thesis
- Productivity Paradox — AI is necessary but not sufficient
Comments