Leash Bio Minos: Measured Chemistry Worlds Where AI Agents Hunt for Drug Binders

Leash Bio has released Minos, a set of "premeasured" lab worlds where AI agents search for drug binders and receive real measured outcomes instantly. The company, known for large DNA-encoded-library (DEL) datasets, announced it on October 14 with a preprint, an open data release and thousands of agent reasoning traces. The idea matters because scientific reinforcement learning needs verifiable feedback, and wet-lab experiments are slow and expensive. Minos replays measurements already taken. For more on how the field is wiring agents into biology, see our posts on Claude and protein design and NVIDIA's BioNeMo agent toolkit.

Quick reference

What Minos is, in plain terms

Drug discovery often looks like a search problem. You have a protein linked to a disease and a huge space of candidate molecules, and you want to find ones that stick to the protein. Testing each in a lab is costly, so chemists choose a few at a time, see results, and adjust. That loop is called structure-activity relationship (SAR) work.

Leash's data collection method, DNA-encoded libraries, measures how a very large number of molecules interact with a protein in one experiment. Minos reuses those measurements. Per Leash's post, because the whole landscape is measured, the outcome of any molecule the agent did not pick is also known. An agent can play the same world many times, at compute speed, with real physical results rather than a simulator's guess.

A rounded ribbon folded into a compact shape, standing for the protein targets in the Leash Bio Minos binder search

The flagship task and its numbers

Per the post, the flagship task hands the agent a target and a library of a million or more molecules with no starting points. The budget is 100 molecules, tested up to five at a time over 20 iterations. With a library of roughly 6.3 million molecules, Leash estimates around 10 to the power 638 possible search trajectories for a single target. The flagship SAR task covers 33 targets and about 6 million molecules. Leash says it adds roughly two new environments a day, aiming for about 20 in the near term.

Agents have run about 100,000 independent SAR trajectories in the six weeks before the post. Performance versus cost is shown in the paper with 95% confidence intervals over 10 trials per target. We could not read exact values from the text we reviewed, so check the preprint PDF for the figures.

What the agents did wrong

The most useful part of the release is the failure analysis. Leash reports that stronger models find binders better than weaker ones, but the gain is modest. Weaker models stay in unproductive chemical neighborhoods too long: too much exploitation, too little exploration. Stronger models leave poor neighborhoods more often, and campaigns that leave unproductive areas have a better chance of success.

Observed failure modes include:

• not using the full molecule budget
• rerunning the same molecules
• poor belief updating after contradictory results
• failing to advance promising candidates

These are general agent weaknesses, not chemistry-specific: spending a budget well, remembering what you tried, and changing your mind. That makes Minos a possible testbed for agent behavior beyond biology.

Where this came from: the Latent Space interview

The announcement landed with a long interview of Ian Quigley of Leash Bio on the Latent Space podcast. The host recorded the conversation while the team was just starting to play with the agent idea, and notes it matured into Minos two months later. The episode covers why DEL data is notoriously noisy and how Leash put many barcodes on each molecule to catch proteins binding the DNA instead of the drug, a Kaggle competition built on about 4 billion measurements in which nobody beat chance on an unseen chemical library, and how models can sometimes guess a molecule's chemist from its structure. It also describes a Leash-data model named Hermes that keeps pace with Boltz on public benchmarks and a case where Hermes appeared to memorize protein "nicknames" rather than binding sites. The point of the history is trust: this team has spent years worrying about models that cheat, and Minos is built from that data.

For the broader picture of how data-first AI science is being pursued, see our coverage of the Periodic Labs science model and the clinical evidence review of AI drug discovery.

Why premeasured worlds are different from a simulator

Most agent training environments are either code sandboxes, where the world is software and feedback is a test result, or physics simulators, where the world is a model of reality that can be wrong. Minos sits between them. The world is a lookup of real measurements, so feedback is grounded in experiment, but the agent does not wait days for a lab. Three properties follow.

1. Counterfactuals are free. In a real campaign you never learn what would have happened had you picked a different molecule. Here, the entire landscape is measured, so you can score a choice against all alternatives and train on that gap.
2. Evaluation is repeatable. Two agents can face the identical world, which makes comparisons fairer than ad hoc lab runs.
3. Overfitting is a risk. A fixed set of libraries and targets can be memorized. Leash's own history with models that memorized protein names rather than binding sites suggests the team knows this; the open question is how well skills transfer to new targets and measurement types.

The Leash post's line that "playing the game changes the player" is a hint at the larger bet: environments, not only models, are where science progress will come from, and scientists may spend more time building worlds for agents. That is a claim about direction, not a result.

How to try it

Start with the three official pieces Leash points to.

• Read the preprint: Minos_Leash.pdf in the project repository.
• Get the data: the release at strelka.leash.bio, which Leash links from its post. Check the license on the site before use.
• Run an environment: a small number of environments are available as Harbor packages. If you already run an agent harness that supports Harbor-style tasks, start with the SAR task and a fixed model so you can compare trials.

A reasonable first experiment is a budget-discipline test. Give a model the flagship task and log how many of the 100 molecules it actually uses, how often it repeats a molecule, and whether it changes direction after poor results. Those map directly to the failure modes Leash lists, and they do not require any chemistry knowledge to measure. Compare two models on ten trials each, since the paper itself uses ten trials per target to get confidence intervals.

Questions readers are asking

Is this an AI scientist that finds drugs?

No. It is an environment for training and testing agents on one narrow step, choosing which molecules to test next. A binder is also not a drug: it still needs selectivity, safety and delivery work that Minos does not measure yet.

Why not just train on the data directly?

You can, and Leash did that with Hermes, a model trained on its own data. Minos adds sequential decision making: a model that predicts binding is not the same as an agent that decides what to measure given a budget.

Why does a 100-molecule budget matter?

Real programs cannot test millions of compounds. A tight budget forces the exploration versus exploitation trade-off that separates strong from weak agents, and it mirrors the cost pressure of wet-lab work.

What it means for builders

• If you train agents with reinforcement learning, Minos is a source of verifiable, physically grounded rewards. Leash released some environments as Harbor packages, the same packaging used in recent RL environment releases, such as the OpenAI Lean math tasks on Hugging Face.
• If you evaluate agents, the failure modes above are cheap to test for in any budgeted search task.
• If you work in biotech, treat this as a research environment, not a drug pipeline. A good score in Minos is not a drug.
• If you follow AI for science, note the pattern: companies with proprietary measurement pipelines turning them into training worlds.

Limits and what we could not verify

1. Agents can only choose molecules already screened. Picking others requires ordering or synthesis, which is slower and costlier.
2. Current worlds cover binding enrichment for single targets, not ADMET (absorption, toxicity and similar) or selectivity.
3. The landscape is fixed to specific libraries and targets, so the search space is bounded.
4. The GitHub repository we saw holds only the PDF, with no README or license shown, so reuse terms are unclear to us.
5. We reviewed Leash's own post and summaries, not the full preprint, and the results are the company's own and not independently reproduced.
6. Whether similar worlds can be built for other measurement types, such as cells or tissues, is listed by Leash as an open question.

Related reading

• Claude protein design and analytical chemistry
• NVIDIA BioNeMo agent toolkit
• Periodic Labs Neon science model
• AI drug discovery clinical evidence
• OpenAI Lean RL environments on Hugging Face
• Hugging Face Hub RL environments guide
• Novo Nordisk and Claude for science

Details are accurate as of October 14, 2026 and come from Leash Bio's announcement and the Latent Space episode page.