Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
Leash Bio has released a full preprint, a blog post, and an open dataset for Minos, an AI playground that lets agents perform verifiable scientific reasoning. The release arrives roughly two months after Ian Quigley, of Leash Bio, first described the project on a podcast, where he said his team had found that data collected over several years could supply tens of thousands of lab-in-the-loop simulations. That combination — an environment that can score an agent's proposed experiment against real measurements — is the raw material scientific reinforcement learning needs. Quigley is the guest in the accompanying interview episode.
Leash's core technology is DNA-encoded libraries, or DELs. The method attaches DNA barcodes to small molecules so that binding experiments between proteins and huge numbers of compounds can be read out by sequencing the barcodes of whatever sticks. It is a way of running binding assays at very high throughput. DELs have a rough reputation among the people who use them; Quigley's team instead extracted clean, reproducible measurements that suit large-scale modeling. One specific hazard they address: proteins that bind the DNA tag rather than the drug, which the company counters by putting dozens of barcodes on each molecule.
Some of the firm's results have circulated outside chemistry circles. In a Kaggle competition, roughly 4 billion measurements were generated, described as coming from a basement, and about 2,000 teams took part. On an unseen chemical library, no entrant beat chance — a result that argues protein-ligand binding is far from solved by current methods. Separately, work on ChEMBL showed models can infer which chemist made a molecule from its structure, and that knowing the author predicts activity about as well as the structure itself. The team's Hermes model, trained only on Leash's own data, has kept pace with Boltz on public benchmarks. They also caught Hermes memorizing protein "nicknames" instead of binding sites, and responded by engineering protein mutants at scale to break the pocket. Lumacaftor-like hits turned up as an accidental detector for misfolded proteins.
For the agent work, the environment described is a search over 6.5 million molecules on a budget of 100, with LLMs reasoning about structure-activity relationships in plain English. In practice, a budgeted search is what makes this interesting to developers: the agent must choose which compounds to test, and the lab result supplies a reward signal that resists the reward hacking that plagues purely textual benchmarks. Whether that signal transfers beyond Leash's own assay conditions is the open question I would watch.
Other threads from the conversation include going from a DEL hit to good mouse PKPD in six months on an inflammation program, an argument for pills over antibodies on cytokine targets, and what would stop a better-funded lab from outscaling Leash. Quigley also framed a rule that biology only scales two ways — taking pictures and sequencing DNA.
Why it matters: teams working on AI for science get a concrete, scored environment instead of a static benchmark, and a public dataset to test whether agents can choose experiments rather than just predict labels. If lab-in-the-loop simulation proves faithful to real assays, it could shorten the loop between a model's hypothesis and a wet-lab result. Expect scrutiny of how well Minos rewards correlate with actual binding, since that, not the model size, determines whether the approach generalizes.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
How Leash Bio tamed DELs, the assay that traumatizes everyone, and used the data to catch cheating models. And why that same data is perfect for AI agents learning to think like chemists.