Leash Bio releases Minos, a lab simulation platform for AI agents

Summary

Leash Bio, a DNA-encoded library (DEL) specialist, has released Minos. The platform transforms the company's years of accumulated high-efficiency protein-ligand interaction data into tens of thousands of verifiable 'lab-loop' simulations, designed to enhance the scientific reasoning and reinforcement learning capabilities of AI agents. Alongside the release, a preprint, a blog post, and an open dataset were provided. Additionally, a Kaggle competition was held, involving approximately 4 billion measurements and 2,000 teams; in this competition, no model was able to exceed chance level on an unseen chemical library.

Why it matters

The core issue is this: AI models trained on open chemical data like ChEMBL are memorizing lab-specific artifacts rather than true binding physics. This poses a silent but serious obstacle to reliable drug discovery. Leash pairs epistemically rigorous DEL data with an agent sandbox that challenges verifiable reasoning over 6.5 million molecules within a limited budget. This approach offers a concrete pathway to reinforcement learning environments where an agent's scientific claims are validated not just by plausibility, but by comparison with real interaction data. If the 'Clever Hans' finding generalizes, the field needs to reframe how it audits existing structure-activity models.

Highlights

• Clever Hans in ChEMBL: A model can predict which lab produced a molecule based on its structure. The answer to 'who made it?' predicts activity as well as the molecule's structure does.
• Hermes, trained solely on Leash's own DEL data, competes with Boltz in open benchmarks. However, it emerged that the model memorized protein 'nicknames' rather than binding pockets. This necessitated a mutation engineering step known as 'pocket breaking.'
• Minos agent sandbox: Large language models (LLMs) handle structure-activity relationships (SAR) in plain English. They must select the best binders from 6.5 million molecules within a budget of 100 queries. This process provides a verifiable reinforcement learning signal.

Critical take

This text is not a technical review, but rather a note description of a podcast episode. Consequently, the actual mechanism of Minos's simulation loop, the reinforcement learning reward design, and the Kaggle evaluation protocol remain unclear. The claim that 'no one beat chance' and the Clever Hans result deserve independent replication before the community accepts them as definitive. Furthermore, while 'tens of thousands of simulations' sounds impressive, the question of how many distinct chemical spaces and protein families these simulations cover is critical. The article does not provide this detail.