Many successful machine learning models for molecular property prediction rely on Lewis structure representations, commonly encoded as SMILES strings. However, a key limitation arises with molecules exhibiting resonance, where multiple valid Lewis structures represent the same species. This causes inconsistent predictions for the same molecule based on the chosen resonance form in common property prediction frameworks such as Chemprop, which implements a directed message-passing neural network (D-MPNN) architecture on the input molecular graph. To address this issue of resonance variance, we introduce the resonance-invariant graph representation (RIGR) of molecules that ensures, by construction, that all resonance structures are mapped to a single representation, eliminating the need to choose from or generate multiple resonance structures. Implemented with the D-MPNN architecture, RIGR is evaluated on a large data set with resonance-exhibiting radicals and closed-shell molecules, comparing it against the Chemprop featurizer. Using 60% fewer features, RIGR demonstrates comparable or superior prediction performance. Alternative approaches, such as data augmentation with resonance forms, are assessed, and their limitations are explored. Available open-source as an optional featurization scheme in Chemprop, RIGR is benchmarked across a wide range of property prediction tasks, showcasing its potential as a general graph featurizer beyond resonance handling.
Zalte et al. (Wed,) studied this question.