Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
- ID
- 35534
- Status
- summarized
- Published
- 15 Oct 2026, 4:35 AM
- Fetched
- 15 Oct 2026, 5:26 AM
- Provider
- Latent Space
- Category
- developer-ai
- Original URL
- https://www.latent.space/p/lab-in-the-loop-at-compute-speed
- Source URL
- https://www.latent.space/feed
Summary
- Score
- 7.0
- Created
- 15 Oct 2026, 5:26 AM
- Tags
- Audience
- ai_ml_learnersai_agent_usersdevelopers
What happened
Latent Space interviews Ian Quigley of Leash Bio about DNA Encoded Libraries (DELs), the high-throughput protein-small molecule interaction data they generate, and a Kaggle competition with about 4B measurements from a basement and about 2,000 teams where nobody beat chance on an unseen chemical library. Leash is releasing a preprint, blog, and open dataset for Minos, an AI playground for agents doing verifiable scientific reasoning via tens of thousands of lab-in-the-loop simulations. The episode also covers ChEMBL 'Clever Hans' effects, where models can infer a molecule's chemist from its structure.
Why it matters
If you build AI agents for science, the open Minos dataset/preprint is a concrete environment to test verifiable scientific reasoning and RL, not just a product announcement. The Kaggle result is the practical warning: ~4B DEL measurements and ~2,000 teams still failed to beat chance on an unseen chemical library, so don't assume high-volume data or co-folding solved protein-ligand binding; evaluate on held-out chemical libraries and check for artifacts like chemist identity leaking into ChEMBL models. If you're outside AI-for-science or agent evaluation, there is no direct Malaysia/SEA or general product-decision impact stated in the text.
Discussion angle
If you're building agent evaluations, would a lab-in-the-loop simulation environment like Minos be a useful benchmark, or does the Kaggle result show the real bottleneck is generalization to unseen chemical libraries rather than more simulation data?