SDD Wiki

Session II — Drug Discovery II

Closing the Loop on Biology

Four leaders — Aviv Regev (Genentech), Jonathan Cohen (NVIDIA), Andy Plump (Takeda), and John Reed (J&J) — converge on a single thesis: drug discovery is becoming an iterative learning machine, and the companies that fail fastest and learn cleanest will own the next decade.


Aviv Regev — Lab in the Loop

Lab in the loop: experiment, model, predict, repeat — across targets, indications, medicines, and patients.

Four loops

  1. Targets — hundreds of thousands of perturbations; single-cell readouts; cross-screen foundation models with Recursion
  2. Indications — multimodal patient atlases; pathobiology foundation models; target-selection impact has more than doubled
  3. Medicines — AI delivered 50× better hit rate than a medicinal chemist using standard tools on the same data; de novo antibody design routine
  4. Patients — AI on 2,000+ OCTs would have flagged a failed phase 3 from phase 2 data; personalized neoantigen vaccine autogene cevumeran (with BioNTech) in pancreatic cancer

Jonathan Cohen — The Compute Stack Coming for Biology

Multiplicative stack

  1. Instruments → 2. Predictive models → 3. Autonomous agents → 4. Robotic wet labs — each layer multiplies the one beneath it.

Architecture: not one super-agent

Specialized sub-agents under generic reasoning orchestrators (like Claude Code), not a single oracle. Software vendors will ship agents that drive their tools.

Examples

  • Parabricks — 30× whole-genome assembly from ~half a day to ~10 minutes on GPU
  • CodonFM — mRNA foundation model tokenized at the codon level (~130M sequences, ~5B parameters)

Andy Plump — No Fear, No Ego

Two scars

  • TAK-994 (orexin 2 agonist for narcolepsy) — phase 2b liver toxicity; backup molecule from first principles about to be registered
  • Myasthenia gravis biologic — years of failed engineering; Nabla Bio returned an in-silico-designed sequence that worked; flipped Takeda’s stance on AI

Takeda’s new Cambridge wet lab will be AI-driven by default, not merely AI-enabled.

Company context

Founded 1781. Under Plump’s tenure: ~15K → ~50K employees; ~30B revenue; Japan from ~50% to ~10%.


John Reed — Pipelines Built on Persistence

Myeloma: average life expectancy from 2 to nearly 20 years

  • VelcadeDarzalexCarvykti (BCMA CAR-T, ~half show functional cure) → Tecvayli + Talvey (T-cell engagers)
  • Tecvayli: FDA Commissioner’s National Priority Review Voucher; 55-day approval

Prostate and bladder cancer

  • KLK2 — exquisitely prostate-specific, retained on ~90% of prostate cancers, cleaner than PSMA
  • Pasritamig — KLK2×CD3 bispecific; durable PSA responses, near-doubled PFS, only 15% all-grade-1 CRS
  • RIPTAC platform (Halda) — instead of PROTAC’s E3 ligase, engages an essential protein the cell can’t live without
  • Inlexzo — intravesical drug-eluting device (gemcitabine); >80% complete response in high-risk NMIBC

Immunology

Neuroscience

  • Spravato — intranasal esketamine; first rapid-acting antidepressant
  • Caplyta — balanced dopamine/serotonin; approved for both bipolar I and II depression

Cross-Cutting Themes

  1. The loop is the unit of work — Regev’s lab-in-the-loop, Cohen’s agent-plus-robot factory, Plump’s failure cycles, Reed’s persevere-and-redesign
  2. AI is now load-bearing, not decorative — Nabla’s sequence rescued Takeda’s mAb; Genentech’s models doubled target-selection impact
  3. Biology — not chemistry — is the bottleneck
  4. Data sharing is the unlock — pre-competitive consortia, multimodal atlases, Protein Data Bank precedent
  5. Necessity is forcing the change — cost pressure, pricing politics, unsustainable failure rate