Ian Quigley

Ian Quigley is a co-founder and CEO of Leash Bio, where he works on a problem at the intersection of laboratory measurement and machine learning: gathering enough reliable examples of small molecules interacting with proteins to train and test predictive models. Before Leash, he worked at the Salk Institute and Recursion; he and Andrew Blevins started the company in a basement lab.[2][1]

In an authored post, Quigley describes a working culture that mixes laboratory and computational skills. At Leash, he has written, data engineers learn to pipette and bench scientists write SQL queries. The point is practical: modelers can ask for controls and replicates while experiments are being designed, and lab scientists can tell them which variations—such as reagent batches—need to be recorded for later analysis.[3]

That emphasis shows up in Leash’s use of DNA-encoded libraries. The company screens a mixture of small molecules, each carrying a DNA barcode, against a protein. Sequencing the barcodes recovered with the protein gives a pooled readout of which candidates were enriched; related molecules in the library let researchers compare small chemical changes.[1]

The barcode itself can create a false positive: some proteins bind DNA rather than the attached chemical. The team describes assigning dozens of different barcodes to the same chemical and checking whether enrichment follows particular DNA sequences. Negative controls, including a bead without the protein, help identify molecules that stick to the experimental support or broadly to proteins.[1]

Leash has applied the same skepticism to its training data. In a reported analysis of ChEMBL, the team first predicted which chemist was likely to have made a molecule, then used those predicted probabilities to represent it in an activity model. That representation performed about as well as one based on chemical formulas in the reported test. A model can pick up patterns in what chemists chose to make and test, making apparent prediction skill harder to interpret.[1]

The company’s Kaggle competition made a related generalization test concrete: competitors trained on one chemical library and were evaluated on an unfamiliar library with different components and chemistry. The team reported that no entrant did better than chance on that most distant split. The result describes that test, not every protein–molecule model or every new chemical library.[1]

By October 2026, Leash had also begun using its measured interactions as a lab-in-the-loop environment for agents: a model can select molecules, receive results already measured in the lab, and revise its next choices. The company’s approach connects experiment and evaluation: useful prediction depends on knowing what the measurements—and the model trained on them—actually represent.[1]