Leash Bio CEO says chemistry AI can learn chemists’ habits instead of biology

An AI model predicting a molecule’s activity may be recognizing the habits of its creator, rather than learning the biology behind it. In an October 14 Latent Space interview, Leash Bio co-founder and CEO Ian Quigley described experiments that he says expose that shortcut—and explained why his company collects its own binding data.

A chemist’s signature becomes a prediction tool

Quigley traced the concern to an encounter at Recursion. He and co-founder Andrew Blevins were looking at a molecule on a screen when a passing medicinal chemist identified it as looking like a Stuart Schreiber molecule. That raised an uncomfortable possibility: if a person could recognize its maker by sight, a model might do the same.

His team assembled molecules and activity data from a chemical repository, selecting chemists with substantial publication and molecule records. The resulting group included 1,815 chemists. A classifier—a model trained to assign an input to a category—then tried to identify each molecule’s creator from the molecule alone.

Quigley reported 60% top-five accuracy: the correct creator appeared among the model’s five highest-ranked guesses in 60% of cases. The more consequential experiment came next. Instead of representing a molecule with its chemical structure or formula, the team used a list of probabilities indicating who had made it.

They used those creator probabilities to train another model to predict molecular activity. According to Quigley, it performed as well as one using chemical formulas. The comparison put two very different representations side by side: one describing chemistry, the other describing likely authorship. In his account, both supported equally good activity predictions.

Quigley’s interpretation is that these models may learn researchers’ practices and intentions, rather than the underlying physical principles. He described that as learning “sociology.” The distinction is between making a useful prediction within those records and learning the ground truth he wants a chemistry model to capture. These are his reported results and interpretation, not independently verified findings.

Collecting binding data instead of inheriting it

That concern connects directly to Leash Bio’s founding rationale. Quigley said the team found existing data inadequate for predicting whether small molecules would bind to protein targets. It built a data-collection operation around DNA-encoded libraries: pools of small molecules, each carrying an attached DNA barcode that identifies it.

Rather than testing chemicals one tube at a time, researchers expose a protein target to a pooled library. Leash commonly holds the protein on a magnetic bead, Quigley said. After retrieving and rinsing it, researchers sequence the DNA barcodes attached to recovered molecules to identify what bound.

The libraries combine a smaller set of chemical building blocks into many more molecules. Quigley acknowledged the restriction: researchers are limited by the blocks they start with. But the resulting molecules can differ by only one or two changes, allowing comparisons that help identify which part drives binding and which changes leave it largely unaffected.

Collecting measurements still requires separating the desired signal from other reasons a molecule gets recovered. Quigley described controls for molecules sticking to the bead itself or sticking broadly to proteins. Those comparisons help distinguish binding to the intended target from general stickiness—a different data problem from recognizing a chemist, but one the experiment must address.