Lab-in-the-loop drug discovery is an operating model problem, not a model problem

TL;DR

Lab-in-the-loop gets talked about like it’s a model problem. The Roche/Novartis/Microsoft panel at health.tech | global summit Basel, alongside Thomas Clozel’s OWKIN presentation, made the real point clearer: it’s an operating model problem. A simple maturity ladder — Level 1 (model proposes, humans curate) → Level 4 (portfolio-grade loops with QA, monitoring, defensible economics). Most orgs sit at Level 1–2, and that’s already useful. The first big value isn’t “more hits.” It’s faster, more confident kill decisions.

“Lab-in-the-loop” gets talked about like it’s a model problem.

The Roche/Novartis/Microsoft panel at health.tech | global summit Basel right now, as well as Thomas Clozel’s OWKIN presentation prior to that, made the real point clearer: it’s an operating model problem.

A simple maturity ladder:

• Level 1: model proposes, humans curate, lab runs

• Level 2: partial automation + active learning, still fragile

• Level 3: a closed loop that runs reliably for one domain (one assay family, one modality)

• Level 4: portfolio-grade loops with QA, monitoring, and economics you can defend

Most orgs are somewhere around Level 1–2, and that’s already useful. The first big value is not “more hits”. It’s faster, more confident kill decisions.

I really liked the insight from Finton Sirockin at Novartis that we still have an essential data capture problem in the actual wet lab experiments that needs to be overcome.

Marwin Segler from Microsoft highlighted the massive change management and cultural aspects needed beyond the technical implementation.

Anna Vangone from Roche listed some key KPIs being used at Roche: Cycle time, and the ability to be able to fail faster. It is not about cost cutting per se but about higher quality outputs.

Question for anyone building this: where does the loop break most often for you, assay noise, lab uptime, data provenance, or decision governance?

Key takeaways

• Lab-in-the-loop is an operating model problem, not a model quality problem.
• A simple maturity ladder: Level 1 (model proposes, humans curate) → Level 4 (portfolio-grade loops with QA, monitoring, and defensible economics).
• Most pharma orgs sit around Level 1–2. That’s already useful, but the gap to Level 3–4 is operating model investment, not model performance.
• The first big value unlock isn’t more hits. It’s faster, more confident kill decisions.
• Wet lab data capture is still an unsolved foundational problem (Finton Sirockin, Novartis).
• Change management and cultural transformation matter as much as the technical implementation (Marwin Segler, Microsoft).
• Roche tracks cycle time and ability to fail faster as KPIs. The frame is higher quality outputs, not cost cutting (Anna Vangone, Roche).