AgentsAutonomy & tool use
Leash Bio releases Minos, replayable drug-screening environments that give LLM agents real binding results
Leash Bio has released Minos, which turns its archive of DNA-encoded library screens into replayable environments where LLM agents choose molecules and instantly receive real, previously measured binding results, letting drug-discovery search strategies be tested at compute speed rather than wet-lab speed.
From the Latent Space podcast episode of October 14, 2026 and cited sources: In Minos, an agent starts with no hints. [2] [7] [1]
SOURCES[2] Leash: Making reality replayable leashbio.substack.com[7] Minos preprint repo github.com[1] Latent Space: Lab-in-the-Loop at Compute Speed latent.space
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Latent Space podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 14, 2026 podcast episode published
- Covered
- Oct 15 edition
- Evidence
- 12 sources across 8 sites; no transcript
- 1Latent Space: Lab-in-the-Loop at Compute Speedlatent.space
- 2Leash: Making reality replayableleashbio.substack.com
- 3Leash: BELKA resultsleashbio.substack.com
- 4Leash: Good binding data is all you needleashbio.substack.com
- 5Clever Hans in Chemistry (arXiv)arxiv.org
- 6Hermes arXiv 2602.13503arxiv.org
RisingHow it’s measured
YouTube video · 3.6K views · 22 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
In Minos, an agent starts with no hints. It tests up to five molecules per round within a 100-molecule budget, searching a library of about 6.3 million molecules across hundreds of protein targets. Because every result was measured in advance, the agent never waits on a lab. Leash reports running about 100,000 trajectories in six weeks. The preprint and the dataset are public. [2] [7] [1]
Earlier Leash work suggests that data quality and how evaluations are split matter more than model design. In the BELKA Kaggle contest, about 2,000 teams trained on billions of measurements, yet the models failed on an unseen chemical library. A separate study found that ChEMBL activity can be predicted from who made a molecule about as well as from its structure. [3] [11] [5] [9]
Leash's binding model, Hermes, is much faster than Boltz-2, but its accuracy results are mixed. It scores higher on a split that holds out proteins (AUROC 0.683 vs 0.639) and lower on a split that holds out chemical libraries and on MF-PCBA. Its own authors link its performance to memorization. All of these figures come from Leash, and no independent replication has been reported. [6] [8]
Minos grounds agent testing in physical measurements, and Leash's studies suggest that random splits overstate generalization.
Read the full assessment
Implication: agent builders gain a low-cost benchmark built on real data, and drug-discovery teams should ask for evaluations split by author or by library.
Executive brief
Leash Bio has turned years of DNA-encoded library (DEL) screening data into Minos, a set of "replayable" chemistry worlds. An LLM agent picks molecules to test and gets real, already-measured binding results back immediately, so it never waits for a wet lab. Leash says it ran about 100,000 structure-activity trajectories in six weeks (Leash). The release follows two earlier Leash results that cut against hype. In a Kaggle contest of about 2,000 teams, models failed on unseen chemistry. And a molecule's author predicts its activity in public data about as well as its structure does. Latent Space's interview with Ian Quigley frames all of this (Latent Space).
What changed and event timeline
BELKA shows memorization, not generalization
About 2,000 Kaggle teams trained on 4 billion measurements. For two targets, none of the winning team's most confident predictions on a new library actually bound ().
"Clever Hans in Chemistry" preprint
Shows that ChEMBL activity models can exploit signals about who made a molecule rather than learning how it binds ().
Hermes announced
A lightweight model trained only on Leash's data, claimed to be 200–500× faster than Boltz-2 and comparable on proprietary validation ().
Hermes paper on arXiv
Nine Leash authors report benchmark results that favor Hermes on some tests and Boltz-2 on others ().
Minos released alongside the podcast
A preprint, blog and open dataset for lab-in-the-loop agent environments (;).
Capabilities and access
- Minos task: the agent starts with no hints, tests up to 5 molecules per round, and has a 100-molecule budget.
- The repository shows no license.
- Models tested: the blog does not name which LLMs it evaluated.
Read the full section
- Minos task: the agent starts with no hints, tests up to 5 molecules per round, and has a 100-molecule budget. It searches a library Leash puts at about 6.3M molecules; the podcast notes say 6.5M. Hundreds of protein targets are covered (Leash).
- Access: the preprint PDF is on GitHub and the data is at strelka.leash.bio. The repository shows no license.
- Models tested: the blog does not name which LLMs it evaluated.
- Hermes: described in a paper; no public weights are referenced in the reviewed sources.
Technical analysis for researchers and developers
- Hermes architecture: ESM2 protein embeddings and ChemBERTa SMILES embeddings feed a cross-attention module, attention pooling and an MLP that outputs a binding probability (arXiv).
- Hermes training data: 239 targets, about two-thirds of them kinases, with roughly 147k positive and 0.83M negative examples drawn from the Kin0 library (arXiv).
- Data quality in DEL screens: Leash tags each molecule with dozens of barcodes.
Read the full section
- Hermes architecture: ESM2 protein embeddings and ChemBERTa SMILES embeddings feed a cross-attention module, attention pooling and an MLP that outputs a binding probability (arXiv).
- Hermes training data: 239 targets, about two-thirds of them kinases, with roughly 147k positive and 0.83M negative examples drawn from the Kin0 library (arXiv).
- Data quality in DEL screens: Leash tags each molecule with dozens of barcodes. This helps flag proteins that stick to the DNA tag instead of the drug (Latent Space).
- Reusable lesson: split evaluations by author or by chemical library, not at random (arXiv 2512.20924).
- Minos design: it works like an RL environment, but agents can only choose molecules that have already been screened (Leash).
Claims and evidence
All of the figures below come from Leash itself. No independent replication was found. Part of MF-PCBA was in Boltz-2's training data.
Read the full section
All of the figures below come from Leash itself. No independent replication was found.
- Hermes vs Boltz-2 by test set (mean AUROC) (arXiv):
| Test set | Hermes | Boltz-2 |
|---|---|---|
| DEL protein split | 0.683 | 0.639 |
| DEL chemical-library split | 0.562 | 0.608 |
| MF-PCBA | 0.554 | 0.789 |
Part of MF-PCBA was in Boltz-2's training data.
- Inconsistent Hermes figures across sources:
- Owl Posting reports 0.761 vs 0.577 on Leash's private set.
- The July 2025 blog reports a best run of 0.669.
- Speed: claimed as 200–500× faster in the blog and 500–700× faster after hardware adjustment in the paper (28.2 vs about 0.04 samples per second per GPU).
- Clever Hans: an 1,815-class model guesses a molecule's author with 60% top-5 accuracy under scaffold splits. An activity model using only author probabilities matches a structure baseline (arXiv).
- Minos agents: the blog says only that "stronger" models succeed "a little bit" and are more willing to leave unproductive regions of chemistry. It gives no named models and no numbers (Leash).
Context and prior work
- BELKA: released 133M screened molecules against three targets (BRD4, sEH, HSA) and was presented at NeurIPS 2024 (OpenReview; NeurIPS slides).
- Data leakage: in the podcast's framing, Boltz had already seen a BRD4 molecule in the PDB, which illustrates benchmark leakage (Latent Space).
- Company context: Leash raised a $9.3M seed round in April 2024 (Utah Business).
Read the full section
- BELKA: released 133M screened molecules against three targets (BRD4, sEH, HSA) and was presented at NeurIPS 2024 (OpenReview; NeurIPS slides).
- Data leakage: in the podcast's framing, Boltz had already seen a BRD4 molecule in the PDB, which illustrates benchmark leakage (Latent Space).
- Company context: Leash raised a $9.3M seed round in April 2024 (Utah Business).
- Independent analysis: Owl Posting's December 2025 deep dive is the main outside write-up (Owl Posting).
Limitations, safety and contested findings
- Hermes:
- The authors say its performance is "strongly correlated with memorization."
- They say cross-assay generalization is probably capped by label noise.
Read the full section
- Hermes:
- The authors say its performance is "strongly correlated with memorization."
- They say cross-assay generalization is probably capped by label noise.
- Hermes is weak on new chemical libraries (arXiv).
- DEL chemistry bias: Owl Posting notes Hermes may do well only within DEL-style chemistry, which differs systematically from natural products and other unexplored chemistry (Owl Posting).
- Minos scope: it covers single-target binding optimization only, depends on heavy upfront wet-lab measurement, and only allows molecules already screened (Leash).
- Inconsistent headline numbers: Hermes AUROC and speed figures differ across Leash's own materials.
Business and practitioner implications
- Data is the bottleneck. BELKA and Clever Hans both suggest that model architecture matters less than unbiased, well-split data.
- Agent builders: Minos is one of few science environments grounded in physical measurements rather than simulation, so agent policies can be compared at compute cost.
- Drug-discovery teams: treat public ChEMBL benchmarks with suspicion. Ask for evaluations split by author or by library.
Read the full section
- Data is the bottleneck. BELKA and Clever Hans both suggest that model architecture matters less than unbiased, well-split data.
- Agent builders: Minos is one of few science environments grounded in physical measurements rather than simulation, so agent policies can be compared at compute cost.
- Drug-discovery teams: treat public ChEMBL benchmarks with suspicion. Ask for evaluations split by author or by library.
- Leash's strategy: the company has turned its proprietary data into an advantage and also runs drug programs. It claims progress from a DEL hit to good mouse PK/PD in six months on an inflammation program, an unverified claim (Latent Space).