Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
Latent Space Brandon Anderson
Leash Bio says its new AI lab sandbox can test science ideas against real data. That could make agents reason about chemistry instead of just sounding smart.
Based on reporting by Latent Space, Brandon Anderson — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Leash Bio has spent years doing something most chemists seem to dread: turning DNA-encoded libraries into data that machines can actually learn from. Ian Quigley’s pitch is simple and a little defiant. Biology, he argues, really only scales in two ways: by taking pictures or by sequencing DNA. Leash picked the second route and kept pushing until the noise became useful.
That matters because DELs have a nasty reputation. They can produce confusing results, and in the past Quigley says users told plenty of horror stories. Leash appears to have beaten that back with epistemic rigor and a stubborn refusal to treat the mess as fate. The result is clean, reproducible interaction data between proteins and small molecules, the kind that can support large-scale modeling instead of just another lab headache.
The company’s Kaggle competition made the point brutally. About 4 billion measurements came from a basement, around 2,000 teams took a shot, and nobody beat chance on an unseen chemical library. That was the warning shot: protein-ligand binding is still hard, even when the data pile looks enormous. It also explains why Leash worried quietly about co-folding models making them obsolete.
Then the story turns from model-building to model-testing. Quigley says the team realized their existing data could power tens of thousands of “lab-in-the-loop” simulations, enough to give an agent something more honest than vibes. That work has now become Minos, with a preprint, a blog post, and an open dataset arriving together. The point is not just to generate answers, but to verify reasoning about SAR in plain English against a real chemical search space.
There’s also a sharper edge here. Leash’s Hermes model, trained only on its own data, kept pace with Boltz on public benchmarks, while a BRD4 molecule in the PDB turned out to be one of the gotchas. The team even found that Hermes could memorize protein nicknames instead of binding sites, which led them to engineer protein mutants at scale and break the pocket. That is the whole AI-bio story in miniature: the models look clever until the data punches back, and then the real work starts.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of AI-bio story: less “the model knows chemistry,” more “show your work and let the data argue back.” Too much of the field still confuses fluent text with scientific reasoning, which is how you end up with expensive poetry. Open datasets and verifiable loops are the antidote, even if they’re less glamorous than the hype crowd would like.
Read more about this at: Latent Space