A few years ago, Ian Quigley and his co-founder Andrew Blevins were at the drug-discovery company Recursion, looking at a small molecule on a screen. A medicinal chemist named Michael Cuccarese walked past, glanced at it and said it looked like a Stuart Schreiber molecule.
Quigley said he and Blevins reacted with alarm: "you can tell just by walking by?" and "That's horrifying."
Quigley, co-founder of Leash Bio, told the story on Latent Space, in an episode published October 14, 2026, because of the worry it raised. If a person can spot the chemist behind a molecule, a machine-learning model probably can too. In that case, a model trained to predict how well molecules work might be learning about the people who made them rather than the chemistry. Quigley said the idea "haunted" Blevins for six or seven years, and the two kept joking over beers that they should find out how bad the problem was. They eventually did, in work they published in December 2025.
Why the chemist matters
In drug discovery, a small molecule is a candidate drug, and the main question is whether it binds to a target protein strongly enough to affect it. To build models that predict this, researchers need many measured examples. Quigley said the temptation is to collect data from many groups. The largest public pool is ChEMBL, a database of molecules and their measured activities drawn from published papers.
The risk, Quigley said, is that if models "can identify the chemist by looking at the molecule alone, then maybe the models can infer chemist intent, which means maybe they don't learn anything at all." Brandon Anderson, who joined the conversation, objected that the models "did learn something." "They don't learn physics," Quigley replied. "They're not learning the thing you're trying to get them to learn."
Anderson stressed how much rides on ChEMBL: many academic projects and PhDs have trained on it. Quigley said that groups without their own data are using it, and that groups with their own data are probably affected anyway, because any organization employing several chemists has several chemical styles, each applied to different kinds of projects. The conversation also turned to the point that chemists with a given style tend to work on a particular class of problems, so a model can learn the style instead of the problem.
The experiment
Quigley and Blevins downloaded ChEMBL and kept every author with more than about 30 papers and 600 molecules, leaving 1,815 chemists. They trained a classifier to guess the chemist from the molecule's structure alone. Its correct answer was among its top five guesses 60% of the time. "It's a very, very easy problem," Quigley said. Their preprint confirms the 60% top-five figure and puts top-one accuracy at 27%. It used a scaffold split, which keeps molecules with the same core structure out of both the training and test sets.
Quigley said the model got confused where two chemists published together at the same institution. He takes that as a sign that individuals have a distinctive chemical style and that their students and trainees pick up part of it. Leash's essay on the work gives one example: the model mixed up Timothy Harrison with his frequent collaborator Jose L. Castro. The essay also describes a public leaderboard and an online game, "Name That Chemist."
Then came the step that worried them most. They stopped showing the second model the chemical structure at all. Instead, each molecule was represented only by the first model's guesses: a list of about 1,800 probabilities, one for each chemist who might have made it. Quigley said a model trained on that list to predict activity in ChEMBL "does as well as if you use chemical formulas," and combining the two was not meaningfully better.
The preprint adds detail and some limits. The activity model also received the identity of the target protein. Across five splits, the author-plus-protein version scored a mean AUROC of 0.652, against 0.656 for a version using a standard structural fingerprint plus protein. AUROC is a ranking score on which 0.5 is chance and 1.0 is perfect. On a second measure, average precision, structure did somewhat better: 0.430 against 0.399. Combining both inputs produced only modest gains. Authorship in the study comes from who is listed on a paper, so it stands in for a lab's style rather than proving who designed each molecule. The authors also found that the probability list still contains a lot of structural information: simple probes could recover much of the chemical fingerprint from it. So "sociology only" simplifies the result. The preprint's practical advice is to split data by its source and to test whether a source-only baseline matches a model before trusting it.
What the models pick up
Quigley's explanation is that public chemistry data records human decisions as much as physics. Ordering a molecule to be made and tested "is kind of a pain," he said, so chemists rule out many options and never record them. Anderson listed reasons from his own experience: a molecule costs too much, the chemistry team cannot make it, or the building blocks are not available. Not making a choice, Quigley said, is "a fingerprint" that shows up in what does get made.
Then there is habit. Quigley called medicinal chemists, "bless them," superstitious people. Anderson agreed that they are "some of the most superstitious people I've ever met." A chemist who saw a building block work in a program that reached clinical trials may keep putting it into new molecules. A model then learns that the chemist's favorite piece goes with success.
Quigley said none of this is a criticism of chemists, who are trying to reach drugs with as few chemical series as possible. "But when you train models on the exhaust, you might be in trouble." Asked whether that makes ChEMBL a poor machine-learning training set, he said: "We do not believe so."
The heavier-is-better shortcut
Anderson raised a cruder example: molecular weight. In most public datasets, he said, simply sorting molecules from heaviest to lightest is one of the strongest predictors of binding. He and Quigley recalled that this baseline tied for first, or came very close, in CASP16, a community-wide blind prediction exercise. Leash's essay identifies the test as CASP16's affinity challenge on two proteins, where the weight comparison tied the leading predictions.
Anderson described how this happens. A chemist finds a hit and adds a group that might bind a nearby part of the protein. If that works, they add another. If the improved molecule does not dissolve well, they attach groups that help with solubility, on the side away from the binding pocket. "You just keep tacking these things on and next thing you know you get this big good binder," he said, "but that has nothing to do with chemistry." Anderson called it "a property of the sociology of what gets made, what gets tested and like the way medicinal chemists work," a property of the dataset rather than of chemistry.
Quigley's shorthand was that "if a thing's working they're going to do more stuff to it." Molecules in a promising program become more baroque, with expensive reactions and hard-to-get building blocks that nobody would bother with otherwise. Weight is not a universal shortcut, though. The preprint and essay say Leash's chosen ChEMBL subset carries less weight bias, and there molecular weight plus protein scored an AUROC of 0.517, close to chance.
The horse that read faces
Anderson asked Quigley to tell the story of Clever Hans, which gave the preprint its title. Hedging that he should remember the details better, Quigley described a horse in Germany, "like the late 1800s," that seemed to solve arithmetic by stamping its hoof and stopping at the right number. The trick fell apart when the questioner did not know the answer. The horse, he said, was reading the posture and tiny expressions of its owner, who grew visibly excited as the count neared the answer.
Oskar Pfungst's account, by the German psychologist Quigley mentioned, places the controversy in 1904 and broadens the point. Hans responded to small head and body movements from whichever questioner knew the answer, not only from his owner. Pfungst showed this by producing the signals on purpose and controlling the horse's answers.
The analogy carries over to chemistry, Quigley said: the models "do not understand chemistry at all," and are "just remembering the sociology of the creators."