Leash Bio releases Minos, an AI agent sandbox for verifiable scientific reasoning in chemistry

Leash Bio has released Minos, an AI agent sandbox intended for verifiable scientific reasoning in chemistry. The system is grounded in tens of thousands of lab-in-the-loop simulations derived from the company’s DNA-encoded library (DEL) screening data. These simulations provide a controlled environment where AI agents can propose hypotheses, design experiments, and receive feedback that mirrors real laboratory constraints and outcomes.

The release includes a full preprint, a project blog, and an open dataset to support reproducibility and downstream research. The authors describe how their DEL pipeline—historically challenging to standardize—was refined to produce clean, reproducible interaction data suitable for large-scale modeling. This rigor enabled the construction of the agent sandbox, which the team argues is uniquely suited for training agents to "think like chemists."

The work builds on prior Leash Bio contributions to the open scientific community, including a 2024 Kaggle competition based on approximately 4 billion measurements from a basement-scale DEL run. That competition drew about 2,000 teams, with none surpassing chance-level performance on an unseen chemical library, underscoring the difficulty of modeling DEL data. Leash’s internal model, Hermes, trained only on their proprietary data, reportedly kept pace with Boltz on public benchmarks, despite not having seen certain molecules in external datasets like the PDB.

The team also discusses the risk of "Clever Hans" behavior in biological datasets, where models exploit spurious correlations such as predicting a molecule’s activity based on its chemist’s identity rather than its structure. To mitigate this, Leash introduced barcoding strategies to detect proteins binding DNA instead of the intended small-molecule drug, and engineered protein mutants at scale to break memorization of protein "nicknames." These measures improved the reliability of the underlying data and, by extension, the agent sandbox’s training environment.

In the agent sandbox, LLMs reason about structure–activity relationships (SAR) in plain English, searching for optimal binders among 6.5 million molecules with a budget of 100 experimental steps. The authors position this as a step toward autonomous, lab-in-the-loop scientific discovery, where agents can iterate under realistic constraints of cost, time, and biological plausibility.