AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio

Collected Oct 23, 2026

Leash Bio has released a preprint, a blog post and an open dataset for Minos, an AI environment where agents perform verifiable scientific reasoning against simulated lab experiments. The release follows a conversation with Leash's Ian Quigley on the Latent Space podcast, during which he described the work as something the team was only beginning to explore at recording time. The interview material says the effort matured into Minos over roughly two months, and that a short follow-up podcast with an update is expected within days.

The starting point is Leash's core business: DNA-encoded libraries, usually shortened to DELs. The technique packages enormous numbers of small molecules together with DNA barcodes, so that binding between a protein and a compound can be read out at very high throughput by sequencing the barcodes instead of testing compounds one at a time. The interview describes DELs as widely painful to work with, and credits Leash with cleaning the signal into reproducible measurements suitable for large-scale modeling.

One concrete fix the episode covers is barcode redundancy. Leash attaches dozens of barcodes to each molecule, which lets the team catch proteins that bind the DNA tag itself rather than the drug candidate. That is a familiar failure mode in encoded libraries: a hit that looks real is actually an affinity for the barcode, and it will not survive follow-up chemistry. Redundant tagging gives a way to flag and discount such artifacts.

Quigley's stated rule is that biology only scales in two ways: taking pictures and sequencing DNA. That conviction is why the company invested in sequencing-based readouts rather than conventional plate-based screening.

The public evidence behind the modeling claims includes a Kaggle competition run with roughly 4 billion measurements and around 2,000 participating teams. On an unseen chemical library, no entrant beat chance. The result is a useful calibration point for anyone assuming protein-ligand binding is a solved prediction problem: models can fit a known chemical neighborhood and still fail to generalize to new scaffolds.

The episode also revisits a finding about ChEMBL, a large public database of compound bioactivity. Models can infer which chemist authored a molecule from its structure alone, and knowing the author predicts activity about as well as the structure does. This is a data-leakage problem in disguise. If a model learns lab-specific or chemist-specific signatures rather than binding physics, benchmark scores say less than they appear to.

Leash's own model, Hermes, is trained only on the company's internal data and, per the interview, keeps pace with Boltz on public benchmarks. Boltz is an open structure-prediction model in the AlphaFold lineage. The episode notes one caveat: a BRD4 molecule used in evaluation had already appeared in the Protein Data Bank, so that comparison may reward memorization. BRD4 is a common cancer target and therefore heavily represented in public structural data.

A related problem emerged when the team found Hermes memorizing protein "nicknames" rather than learning binding sites. The response was to engineer protein mutants at scale, altering the pocket so that a model cannot rely on a familiar protein identity. Separately, Lumacaftor-like hits turned up as an accidental detector for misfolded proteins.

Minos builds on this corpus with a budgeted search task. As described, LLM agents reason about structure-activity relationships in plain English while hunting for the best binders among 6.5 million molecules on a budget of 100. Budgeted search is attractive for agent evaluation because it produces a verifiable score rather than a plausible-sounding narrative, which is the usual weak point of language-model benchmarks. In practice, a scoring loop grounded in real assay data is a stronger reinforcement-learning signal than model-graded judgments, and tens of thousands of simulated lab iterations is the kind of volume agent training needs.

The interview also touches Leash's drug-discovery pipeline, citing a path from a DEL hit to good mouse PKPD in six months on an inflammation program, and the case for small-molecule pills over antibodies against cytokine targets. PKPD refers to pharmacokinetics and pharmacodynamics, the absorption, distribution and exposure behavior of a compound in animals.

Broader themes include the "Kosuri paradox" — if you can collect the data, why do you need the model — and the observation that alumni of Recursion tend to start companies that rhyme with one another. On competition, Quigley points to massively parallel purification as the capability he would most want, rather than more compute or capital.

Why it matters: Teams building scientific agents get an open dataset and a scored environment where reasoning can be checked against assay data, not just judged by another model. Model builders should also read the Kaggle result and the ChEMBL author-signal finding as warnings that benchmark wins on public chemistry data can reflect leakage or memorized compounds rather than real generalization. My inference is that lab-in-the-loop simulation will become a standard evaluation pattern for scientific agents, and that data quality, not model architecture, will decide who can build it.

Read at Latent Space

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.