Leash Bio has been widely followed for its work on DNA-Encoded Libraries (DELs), a technique for generating extremely high-throughput measurements of protein–small molecule interactions. The company has focused on converting DEL data — which are often noisy and hard to reproduce — into clean, scalable datasets suitable for large-scale machine learning.
The Minos project, announced today, grew out of Leash’s multi-year DEL collection efforts: the team released a preprint, a blog post, and an open dataset designed to support "lab-in-the-loop" simulations and agent-based scientific reasoning for reinforcement learning and other agent frameworks.
How Leash produces usable DEL data
In a conversation with Ian Quigley (Leash Bio) the team highlighted several practical elements that enabled high-quality data:
- Epistemic rigor: strict attention to data quality and reproducibility, backed by engineering practices.
- Multiple DNA barcodes per molecule: Leash places dozens of barcodes on each molecule, which helps detect and discard cases where the protein binds to the DNA tag rather than to the small molecule.
- Scaling biology through two readouts: Quigley emphasizes that biology scales reliably in two ways — imaging and DNA sequencing — and Leash’s approaches rely on sequencing-based readouts.
A large Kaggle challenge and modelling surprises
Leash ran a Kaggle competition built on roughly 4 billion measurements collected from a basement lab; about 2,000 teams participated. The outcome demonstrated that for unseen chemical libraries, competitors were unable to beat chance in a meaningful way. That result underscored the difficulty of generalizing across chemical space and prompted concern at Leash that co-folding effects might quickly render some approaches obsolete.
Another notable finding was a "Clever Hans" phenomenon in public datasets such as ChEMBL: models can infer a molecule’s chemist from its structure, and the meta-information of "who made it" can predict activity about as well as the molecular structure itself. This shows how dataset artifacts and metadata can mislead models if not accounted for.
Hermes, Boltz, and combating memorization
Leash’s in-house model Hermes, trained on their own data, kept pace with Boltz on public benchmarks. During development the team found Hermes sometimes memorized nicknames or identifiers for proteins rather than learning the physics of binding sites; to counteract that they engineered protein mutants at scale to force the model toward genuine generalization.
They also noted that some hits resembling Lumacaftor turned out to be accidental detectors of misfolded or improperly folded proteins, indicating that not all DEL hits reflect canonical target binding.
Minos: an agent-ready laboratory sandbox
Near the end of the interview Quigley mentioned an experimental idea: the DEL data Leash had amassed could be used to create tens of thousands of "lab-in-the-loop" simulations, providing the kind of environment agents need for scientific reinforcement learning. Two months after the interview that idea matured into Minos. The current release includes a preprint, blog, and an open dataset.
Minos aims to be a sandbox where large language models and agent systems can reason about SAR (structure–activity relationships) in plain English and search among large libraries. The team describes scenarios in which agents evaluate up to 6.5 million molecules under a constrained budget (for example, a budget of 100 queries).
Practical workflows and outcomes
Leash’s materials and the associated podcast discuss how a DEL hit can be progressed to a mouse PK/PD profile within six months in an inflammation program, and they lay out arguments for pursuing oral small molecules rather than antibodies against certain cytokine targets.
Quigley also addressed what could prevent better-funded labs from simply out-scaling Leash, and he expressed an aspirational desire for massively parallel purification workflows — a capability that, if implemented at scale, could change throughput dynamics.
Conclusion
Leash Bio’s work illustrates that DEL-derived datasets can be made reliable enough to support large-scale modeling, but doing so requires engineering controls, careful experimental design, and attention to dataset artifacts. The Minos release offers an open resource for developing and evaluating lab-aware agents and reinforcement learning approaches using real, laboratory-derived simulation data. The accompanying podcast interview with Ian Quigley covers these topics in depth for both newcomers to chemistry and experienced AI×Bio practitioners.
