Perspective: ML models see representations, not molecules, so demand accountability
- A Perspective (no new dataset or benchmark) examining what ML models actually receive across structural formulas, SMILES, fingerprints, graphs, conformers, simulated ensembles and language-model tokens.
- It argues that both the molecule and the bioactivity label are representations, so tautomer ambiguity, stereochemistry, activity cliffs, noisy labels and generative failures reflect this double reduction.
- The proposed standard is accountability: explicit entity definitions, assay provenance, calibrated uncertainty and prospective testing, with controlled perturbation tests to separate representational loss from model limits.
💡 Points Based on the abstract
Most of this will be familiar to anyone who has watched a model flip on a re-enumerated SMILES, but the perturbation-test idea is the usable part. On our own project I'd run it before trusting any benchmark gain: randomize SMILES, swap tautomer or protonation states, then see whether error tracks the representation or the model. Without prospective data, though, a Perspective can't tell us which fixes matter.
📄 Paper · Journal of Chemical Information and Modeling ↗ Ceci n’est pas le benzène: The Treachery of Chemical Representations in Machine Learning for Drug Discovery