Leash Bio: DNA-encoded libraries and an RL sandbox for 10,000 drug-discovery campaigns a weekend
Latent Space · youtube · 2026-10-15
Latent Space interviews Leash Bio co-founder and CEO Ian Quigley on how DNA-encoded libraries (DELs) produce clean protein–small-molecule binding data for AI drug discovery, and their new agent sandbox Minos.
- How DELs work: barcodes + sequencing screen millions of molecules; multiple barcodes per molecule catch proteins binding the DNA instead of the drug
- BELKA Kaggle competition: 4B measurements, 2,000 teams, nobody beat chance on an unseen chemical library — protein-ligand binding remains unsolved
- Clever Hans traps: models guess a molecule's chemist from structure in ChEMBL ("who made it" predicts activity about as well as structure); Hermes memorized protein "nicknames," prompting engineered mutants at scale
- Hermes: trained only on Leash's own data, keeps pace with Boltz on public benchmarks (which had already seen one BRD4 molecule in the PDB)
- Minos agent sandbox: years of data yield tens of thousands of lab-in-the-loop RL environments; LLMs reason about SAR in plain English, hunting best binders among 6.5M molecules on a $100 budget; 10,000+ med-chem campaigns in a weekend; preprint and open dataset released
- Translation to medicine: DEL hit to mouse PK/PD in six months on an inflammation program; the case for pills over antibodies on cytokine targets; economics and purification bottlenecks
- The Kosuri paradox: if you can collect the data, why do you need the model?
More from Embodied
- RAI Institute's single-track bicycle robot UMV lands in Science Robotics — joemeno · 2026-10-15
- Scoble visits Bucket Robotics, a startup using computer vision for factory quality control — Scobleizer · 2026-10-15
- fdotinc demos a consumer robot arm you can control with basic prompting — Scobleizer · 2026-10-15
- 1X, DeepMind, Physical Intelligence Gather for Humanoid Robot Demos — lukas_m_ziegler · 2026-10-15
- NVIDIA recaps GTC Berlin: open models to physical AI — NVIDIA Developer · 2026-10-15
- AlphaNIST: one RL policy solves all 20 NIST assembly tasks with 98.2% success in simulation — hbouammar · 2026-10-15