Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
Leash Bio’s Ian Quigley discusses DNA-encoded library data and the new Minos dataset for scientific agents.
A Latent Space interview with Ian Quigley of Leash Bio covers the company’s DNA-encoded library platform and its use for large-scale protein-ligand modeling. Topics include a Kaggle competition with about 4 billion measurements and roughly 2,000 teams, models that infer a molecule’s chemist from ChEMBL, and Hermes matching Boltz on public benchmarks when trained only on Leash data. Quigley also describes an agent sandbox for structure-activity reasoning over 6.5 million molecules. Leash is releasing the Minos preprint, blog post, and open dataset for lab-in-the-loop scientific reinforcement learning.
- Leash Bio builds models from DNA-encoded library protein-ligand measurements.
- A Kaggle challenge used about 4 billion measurements and roughly 2,000 teams.
- Hermes, trained only on Leash data, kept pace with Boltz on public benchmarks.
- Leash is releasing the Minos preprint, blog, and open dataset for agent reasoning.
- The sandbox lets agents search 6.5 million molecules with a 100-query budget.
From the original article
3 min read at latent.spaceHow Leash Bio tamed DELs, the assay that traumatizes everyone, and used the data to catch cheating models. And why that same data is perfect for AI agents learning to think like chemists.
By Brandon Anderson · latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.