Leash Bio releases Minos and open dataset for agent science

Leash Bio's path to Minos started with DELs, short for DNA Encoded Libraries, a way of collecting high-throughput interaction data between proteins and small molecules. For the company, the harder problem was that DEL users generally had horror stories: Leash's answer involved putting dozens of barcodes on each molecule to catch proteins that bind the DNA rather than the drug, along with epistemic rigor.

The Minos release follows a string of earlier Leash work. A Kaggle competition produced about 4B measurements from a basement and about 2,000 teams, with nobody beating chance on an unseen chemical library. Leash also found that a molecule's author predicts activity about as well as its structure, and trained Hermes only on its own data to keep pace with Boltz on public benchmarks. Catching Hermes memorizing protein "nicknames" instead of binding sites, the team moved to engineering protein mutants at scale. Minos appears to be the next step from those datasets toward agent training.