DOI: 10.1039/d6dd00309e ISSN: 2635-098X

Notation matters: cross-representation inconsistency in chemistry language models and its mechanistic origins

Animesh Mishra

Predictions from a chemistry language model shift by 1.69 log  S on average when the same molecule is rewritten in another notation, affecting 88% of ESOL. Descriptor models are unaffected, locating the failure in string-based models.