Leash Bio, a biotech company specializing in DNA-encoded libraries (DELs), has turned a notoriously difficult assay into a reliable data engine for AI. In a recent Latent Space podcast episode, Ian Quigley explained how the company adds dozens of barcodes to each molecule to catch proteins that bind DNA instead of the drug, producing the kind of clean, high-throughput interaction data that large-scale modeling demands.
The data has already exposed embarrassing failure modes in existing models. A Kaggle competition built on about 4 billion measurements saw roughly 2,000 teams fail to beat chance on an unseen chemical library. Separately, Leash found that models trained on public data like ChEMBL could guess a molecule's chemist from its structure alone, and that 'who made it' predicted activity about as well as the structure did — a Clever Hans effect that undermines apparent predictive power.
Leash's own model, Hermes, trained only on internal data, keeps pace with Boltz on public benchmarks. The team also caught Hermes memorizing 'nicknames' for proteins rather than learning binding sites, prompting them to engineer protein mutants at scale to break the shortcut. Now, the same DEL data is being repurposed for Minos, an AI playground where LLMs reason about structure-activity relationships in plain English and hunt for the best binders among 6.5 million molecules on a budget of 100 experiments. The project has matured into a preprint, blog, and open dataset, offering tens of thousands of lab-in-the-loop simulations for scientific reinforcement learning.