Ceci n ’ est pas le benzène : The Treachery of Chemical Representations in Machine Learning for Drug Discovery

Chemists do not draw molecules; they draw agreements about molecules. A hexagon with three alternating lines is not benzene, and, read literally, it is not even an adequate model, since it localizes bonds that experiment treats as equivalent. Chemists repair this discrepancy automatically and never write the repair down; a machine receives the convention, plus whatever further theory its software compiles into it. This Perspective asks what our models actually see, across structural formulas, SMILES, fingerprints, graphs, conformers, simulated ensembles and the token sequences on which chemical language models are trained. We argue that both members of the training pair are representations─one of a chemical entity, the other of an assay-conditioned observation─and interpret tautomeric ambiguity, stereochemical and solid-form identity, activity cliffs, heterogeneous bioactivity labels and generative failure as consequences of this double reduction. We propose accountability, not realism, as the operative standard: explicit entity definitions, complementary representations, assay provenance, calibrated uncertainty, prospective testing. We then make the standard operational, setting out the controlled comparisons that separate representational loss from model limitation, sparse training coverage and experimental heterogeneity, and specifying for each perturbation test the control that makes it interpretable, the criterion for an actionable failure, and the corrective action it implies.