Why Do Binding Models Learn the Chemist Instead of the Chemistry?

Ian Quigley says a model can predict ChEMBL activity from who probably made a molecule about as well as it can from the chemical formula, so it is learning the chemist rather than the chemistry.

At Recursion, a medicinal chemist walked past a screen Ian Quigley was looking at with his co-founder and called the molecule a Stuart Schreiber molecule. If a model can name the chemist from the structure alone, it can infer that chemist's intent and skip the physics. They took every ChEMBL chemist above a bar he remembers as more than 30 papers and 600 molecules. That was 1,815 chemists. A classifier that saw only the molecule picked the chemist with a top-five accuracy of 60%. Confusion showed up when two chemists shared papers and an institution. He reads that as trainees picking up part of a mentor's chemical style.

They then dropped the structure and stored each molecule as a vector of about 1,800 creator probabilities. That vector predicted activity as well as the chemical formula. Combining the two did not get meaningfully better. He calls it sociology, not ground truth: molecules that were never ordered, favorite blocks that once reached the clinic, and adding groups until a hit gets heavy. He names the failure Clever Hans, the horse that seemed to do arithmetic while reading its owner's face. The models are reading the face and not the math.