Leash Bio releases Minos: an open dataset and preprint for chemistry reasoning agents

On a podcast, Leash Bio revealed that its DNA-encoded library data from the past two years can generate tens of thousands of lab-in-the-loop simulations, exactly what agents need for scientific reinforcement learning. At recording time they were still experimenting with the idea; two months later it matured into Minos, released alongside a full preprint, blog, and open dataset. Previously, on Kaggle they used roughly 4 billion measurements and about 2,000 teams to show that binding prediction on unseen chemical libraries is still barely better than random, and found that in ChEMBL, who made a molecule predicts activity almost as well as its structure. The Hermes model matched Boltz on public benchmarks using only in-house data. Now this open data gives chemistry reasoning agent teams a simulated environment with experimental verification signals, searching for the best binder among 6.5 million molecules within a budget of 100 queries.