Leash Bio, led by Ian Quigley, has spent years refining DNA‑encoded libraries (DELs) into clean, reproducible interaction datasets suitable for large‑scale modeling. The team’s work has attracted attention across the community — discussed in their Substack and in a deep dive by Owl Posting — and recently culminated in a preprint, a blog post, and an open dataset tied to a new project called Minos.
The DEL problem and Leash’s approach
DELs use DNA barcodes attached to small molecules so that enormous numbers of protein–small molecule interactions can be read out by Next Generation Sequencing. In practice DELs can be noisy: non‑specific binding, proteins that interact with DNA rather than the small molecule, and other artefacts often confound results.
Leash Bio addresses these issues experimentally and methodically. They place dozens of barcodes on each molecule to detect cases where proteins bind the DNA barcode instead of the druglike moiety, and they emphasize epistemic rigor and repeated, controlled measurements to produce reproducible datasets that are useful for modeling.
The Kaggle challenge and hard numbers
Leash ran a Kaggle competition built on roughly 4 billion measurements and attracted about 2,000 teams. On the held‑out chemical library used for final testing, no team beat chance. The competition reinforced that DEL‑derived prediction remains challenging, and the group had previously worried that co‑folding effects might make the problem less tractable.
Model memorization and surprising artefacts
Analyses revealed Clever Hans‑type effects: models could infer the chemist or source of a molecule from its structure, and the information about “who made it” predicted activity almost as well as the structure itself. That implies models sometimes learn source or batch signals rather than true binding biology.
Leash’s Hermes model, trained solely on the company’s internal data, held its own against public models such as Boltz on benchmarks. The team also noted instances where Boltz had already seen a BRD4‑associated molecule in the PDB. To counter memorization of protein “nicknames” instead of binding sites, Leash engineered protein mutants at scale to test whether models were attending to genuine binding pockets.
Another observation: Lumacaftor‑like hits occasionally acted as detectors of misfolded proteins, indicating that particular chemical profiles can reflect off‑target or artefactual readouts rather than intended target engagement.
Minos: an agent sandbox and lab‑in‑the‑loop simulations
Near the end of an interview, Ian Quigley described an experimental direction: an AI playground where agent systems (LLM‑based) would perform verifiable scientific reasoning. The team realized their accumulated DEL data could provide tens of thousands of lab‑in‑the‑loop simulations — precisely the kind of data required to train scientific reinforcement learning agents.
Two months after the initial discussion that inspired the idea, the effort matured into Minos. Alongside the preprint and blog post, an open dataset was released. A showcased task had LLM‑agents reason about structure‑activity relationships (SAR) in natural language to select the best binders from a set of 6.5 million molecules under a budget of 100 — a concrete demonstration of agent‑driven, looped experimental design.
Translational and industry notes
The team reports that, in one inflammation program, they progressed from a DEL hit to favorable mouse PK/PD within six months. The conversation also emphasized cases where oral small molecules might be preferred over antibodies for cytokine targets.
When asked what would stop a better‑funded lab from outcompeting them at scale, Ian highlighted the lack of massively parallel purification as a bottleneck and expressed a desire (by fiat) for vastly parallel purification capabilities.
Community and conceptual takeaways
The discussion touched on the so‑called Recursion alumni network and why many spinouts from those teams have similar profiles. It raised the “Kosuri paradox”: if you can build the ability to collect massive datasets, do you still need more models, and what is the role of modeling when rich data exist?
Conclusion
Leash Bio demonstrates that, with careful experimental design and rigorous controls, DELs can yield high‑quality data for large‑scale machine learning. The Minos project and the accompanying open dataset create opportunities for training agents with lab‑in‑the‑loop simulations, moving agent‑assisted experimental planning and drug discovery forward. The broader lesson is practical and epistemic: measure carefully, control for artefacts, and design systems that avoid learning source or batch signals instead of true biological interactions.
(The above summarizes a podcast conversation with Ian Quigley and the subsequent public release of Minos‑related materials: preprint, blog, and dataset.)
