Lab in the Loop
The iterative architecture where data, AI, and experimental biology form a continuous cycle: experiment → model → predict → repeat. Articulated by Aviv Regev (Genentech) at Session II.
The Architecture
Four loops, one framework:
- Targets — hundreds of thousands of perturbations across cells and organoids; single-cell readouts; cross-screen foundation models (with Recursion)
- Indications — multimodal patient atlases train pathobiology foundation models; target-selection impact has more than doubled
- Medicines — AI delivered 50× better hit rate vs. medicinal chemist on same data; de novo antibody design, virtual screening, oracle scoring, immunogenicity cleanup all routine
- Patients — AI trained on 2,000+ OCTs would have caught a failed phase 3 from phase 2 data; personalized neoantigen vaccine autogene cevumeran in pancreatic cancer
Why It Matters
- The loop subsumes the AI-vs-data-vs-biology debate: you need all three
- Foundation models are bending toward generalization across screens, diseases, and modalities
- Patient-derived corpora are now training-grade assets, not just trial results
- Hard problems become tractable when computational and lab scientists share one workflow
Cross-References
- AI in drug discovery — the loop is the practical architecture for AI integration
- Productivity paradox — the loop is one proposed solution
- Jonathan Cohen’s multiplicative compute stack is the infrastructure layer beneath the loop
Comments