The Neuron: AI Explained Can AI Really Design New Drugs? Google DeepMind Spin-out Isomorphic Labs Explains
17 snips
May 6, 2026 
Michael Schaarschmidt, foundational AI research lead building large-scale models for protein structure and binding. Rebecca Paul, head of medicinal drug design applying models to medicinal chemistry. They discuss why drug discovery is slow and costly. They explain limits of “generate a molecule and ship it.” They cover structure prediction, binding models, model confidence, and hopes for targeting previously undruggable proteins.
AI Snips
Chapters
Transcript
Episode notes
Drug Discovery Is A Multidecade Multifaceted Problem
- Drug discovery is costly, slow, and failure-prone, often taking over a decade with ~90% clinical failure and up to billions in capitalized cost.
- Rebecca Paul frames the problem as many interconnected challenges from target ID to clinic, not a single algorithmic fix.
Many Specialized Capabilities Create The Hardness
- Building useful AI for drug design requires many distinct capabilities, making it a Manhattan Project–style effort rather than a single-solution task.
- Michael Schaarschmidt emphasizes finding root problems whose solutions impact many programs to maximize ROI.
Trust Validated Structure Models To Speed Design
- Trust structure prediction models when they've been repeatedly validated and show high confidence scores to accelerate design decisions.
- Rebecca Paul describes co-fold predictions matching later crystal structures, letting chemists use models as experimental hypotheses.