Molecular machine learning with conformer ensembles

Abstract Virtual screening can accelerate drug discovery by identifying promising candidates for experimental evaluation. Machine learning is a powerful method screening, as it learn complex structure–property relationships from data and make rapid predictions over virtual libraries. Molecules inherently exist three-dimensional ensemble their biological action typically occurs through supramolecular recognition. However, most deep approaches to molecular property prediction use 2D graph representation input, in some cases single 3D conformation. Here we investigate how the information of multiple conformers, traditionally known 4D cheminformatics community, improve models. We introduce models that expand upon key architectures such ChemProp SchNet, adding elements multiple-conformer inputs conformer attention. then benchmark performance trade-offs these on 2D, representations activity using large training set geometrically resolved molecules. The new perform significantly better than models, but often just strong with many. also find interpretable attention weights each conformer.