Abstract
<jats:p>The configuration of small molecules plays a key role in their measurable properties. This includes their spatial arrangement in 3-dimension space, giving rise to stereoisomers. As different stereoisomers can have vastly different biological effects, its determination is of upmost importance to the pharmaceutical, agrochemical, and cosmetics industries. The easiest property to measure of stereoisomers is optical rotation, a non-destructive measurement that takes mere minutes. However, there is no method by which the stereoisomer can be determined from solely this empirical value. Modern machine learning methods have seen great success in linking chemical structure to property (e.g., ADMET, ligand binding, reaction outcomes, etc.). However, explicit stereochemical modelling using graph neural networks, a wildly-used, structure-based, highly accurate deep learning architecture, is limited due to the loss of 3-dimensional information in these models. Herein, we develop a stereodiscerning graph neural network which achieved unrivaled accuracy in determining the specific rotation sign from chemical structures containing one or more stereogenic units. This work lays the foundation for rapidly repurposing models previously unsuitable for stereogenic property prediction to state-of-the-art accuracy.</jats:p>