Ian Quigley of Leash Bio has a quick way to test whether a binding model understands chemistry. Hand the model a protein sequence whose pocket has been altered on the computer. The pocket is the cavity where a drug-like molecule sits. Fill it in "like you're filling it with cement," delete it, or make it far larger. In the cases his team has seen so far, he said, a current state-of-the-art model "will shove the small molecule in where the pocket used to be because it just remembers where it is." He said a couple of different groups have shown the same thing.
Quigley described the test on an episode of Latent Space published on October 14, 2026. Brandon Anderson also took part in the technical discussion. The test explains one of Leash's more unusual lab habits. The company deliberately damages the pockets of real proteins, re-runs its screens, and uses the wreckage as training data.
What the broken-pocket test is probing
The models Quigley refers to are co-folding models. They predict the three-dimensional structure of a protein together with a small molecule sitting in it. Anderson named AlphaFold and Boltz as examples of the structure-prediction tools now available for this kind of work.
Quigley's hypothesis is that these models have not learned the forces involved. A pocket typically depends on a handful of amino acids. A few have hydrogen-bonding positions sticking into the cavity, and there may be "a little bit of pi stacking," a weak attraction between flat, ring-shaped chemical groups. Those contacts "really ought to be very, very important," he said, and they are important when binding is measured physically. A model that understood the physics should recognize that once they are gone, nothing can bind there.
He traced part of the problem to the data. Structural training sets are assembled from solved structures of molecules bound to proteins, so they contain few informative failures. As he put it, "there aren't great negatives, because what makes it into there" is the binding that worked. The sets lack the variation that disrupts binding, and no one is going to make crystal structures of a pile of disruptive mutants.
Manufacturing the failures
Leash's screening uses DNA-encoded libraries, or DELs. These are vast collections of small molecules, each carrying a DNA barcode. The protein is mixed with the library and pulled out, and sequencing the attached barcodes reveals which molecules came along. Quigley said Leash has a good understanding of how many copies of each molecule to expect in a baseline condition. Binding therefore shows up as enrichment, and a negative is a molecule that stops being enriched.
Quigley described the workflow in steps:
- Screen the normal (wild-type) protein and collect hits.
- If someone has already characterized a chemical known to occupy the pocket, buy it, add it at high concentration and screen again. The competitor displaces whatever normally binds there, so the hits that drop out belong to the pocket.
- Choose pocket residues to change. Quigley said the team uses conservation, structure-prediction models and a structural biologist on staff.
- Change the DNA construct that produces the protein so that one residue becomes, for instance, "a big bulky one," and screen once more. If the same chemical series disappear, he said, "we have really good evidence that they're all going into that pocket."
Protein chemists raise the obvious objection, which Quigley voiced himself: how do you know you didn't just break the whole protein? His answer is that only the competitor-sensitive series are lost. Molecules binding elsewhere on the protein remain. Leash's own write-up of these methods, How to screen a lot, says the agreement between blocker and mutant is imperfect, with differences between constructs offered as one explanation.
Anderson said the approach lets researchers treat activity cliffs, where a tiny change flips a result, "both in compound space and in protein space." He pointed to kinases, where around a hundred closely related enzymes can have binding sites that differ by a single residue at a time, as a case where that would be powerful. Quigley agreed. He said Leash has run "many, many hundreds" of such screens, with about a hundred protein constructs in a big week. Typically three to five residues drive the binding in a given pocket, and the team tries breaking each one independently.
Mutations do not only destroy binding. Changing the shape can encourage other molecules to bind. Quigley said Leash has definitely seen this and treats it as information.
He described a plot that shows both outcomes. Each dot is a molecule from a library Leash designed and made. One axis shows binding to the normal protein and the other shows binding to the mutant. Colored blobs mark chemical series that look nearly identical apart from where one group is attached. In that plot most of the series bound better after the mutation. A second mutation in the same pocket, Quigley said, kicked every one of those populations out entirely.
The cheat sheet in the attention maps
Leash's own binding model is called Hermes. Its introduction describes a roughly 50-million-parameter model built on pretrained protein and molecule encoders. It combines an amino-acid sequence with a molecule's chemical notation to predict whether they bind, and it neither needs nor produces a 3D structure. As Anderson noted, its cross-attention can be inspected to see which parts of a protein the model is using.
Quigley expected one of two answers. One was the binding pockets. The other, more modestly, was a spread across all residues because the model doesn't yet know much. Neither appeared. The model concentrated on short stretches of amino acids away from the functional parts of the protein. It also avoided conserved domains, the reusable modules that look similar across many proteins and that biologists use to infer families and functions. Instead it fixed on variable stretches not thought to do much.
"Hermes was probably coming up with a nickname for the protein," Quigley said. Rather than learning the functional part, "I can remember, you know, 10 or 12 amino acids that are different here than everyone else and I can just memorize it." He thought that was bad. He said it fit the co-folding behavior, where a model that "doesn't know what's going on" has simply made a nickname.
The mutant screens are the proposed cure. If two nearly identical sequences behave very differently, the nickname strategy fails, and the model is pushed back toward reading all the residues. Leash retrained with the added mutants and looked again. Quigley said Hermes now looks at all the residues "much more evenly."
The company's published analysis is more modest. It is a week-long exploratory investigation covering 20 proteins, and it uses learned residue-pooling weights as a proxy for which positions influence predictions. The newer model was trained with mutants and also with roughly 40% more ordinary Leash data. The result is therefore an expanded training regime, not an isolated mutant-only comparison. The authors propose collecting more targeted protein variation next.