Andrew D. Blevins and Ian K. Quigley, co-founders of Leash Bio, released a pre-print last year that I think deserves some more attention. Using a Bemis-Murcko scaffold split, the authors show that it is possible to train a machine learning model to identify which of 1,815 medicinal chemists produced a molecule from its structure alone, achieving a top 5 accuracy of 60% and a top 1 of 27%.

Let’s take a step back. Why aim to predict who made a molecule in the first place? The authors motivate this well, drawing on the story of Clever Hans. Owned and showcased by Wilhelm von Osten in 20th century Germany, Clever Hans was a horse purportedly able to solve mathematical problems, stamping out solutions with his hoof. However, it was quickly realized that Clever Hans could not actually perform arithmetic, but rather read the micro-expressions and posture of those posing the question, stopping his count when the faces of observers revealed the right answer had been reached. The moral? Clever Hans learned how to predict a correlate of a task, which was mistaken for mastery of the task itself.

Clever Hans predictors abound in machine learning: image classifiers that actually attend to the photographer’s watermark, self-driving cars that appear to follow lane markings but actually learn that roads are often adjacent to grass, and pinball-playing models that learn to achieve high scores by nudging the table rather than using the typical controls. The authors of this study set out to determine whether this affects activity prediction from data sourced from ChEMBL. That is, they set out to determine whether a model can predict the activity of a molecule by inferring who made it. They conduct two experiments, the first to determine whether it is possible to predict who created a molecule from its ECFP4 fingerprint, and the second to determine whether one can use a vector of author probabilities and the ID of a protein to predict activity. Critically, in this second experiment, no structural information is available at all, except insofar as it can be extracted from the author probability vector.

The conclusion of the second experiment is striking: using just a vector of predicted author probabilities and a protein ID, it is possible to rival the performance of an ECFP4 baseline at predicting binarized activity labels.

I recommend reading the full paper on Arxiv (Clever Hans in Chemistry: Chemist Style Signals Confound Activity Prediction on Public Benchmarks): https://arxiv.org/html/2512.20924v1#A1.SS1

Some caveats are due, however. The most significant is that for the author prediction task, the paper reports top-k accuracy, and not top-k balanced accuracy. The methods section of the paper does not mention drawing an equal number of molecules from each author, so we can assume that some authors will be represented by more molecules in the dataset. Without knowing the distribution of molecules per author, it is not possible to know whether the performance is due to a small handful of authors being particularly easy to predict.