Abstract

Join us on Zoom

While it is clear that 3D information is valuable to improve molecular property prediction, there are many molecules for which the geometry is not available. However, in principle, all information about a molecule’s possible 3D configurations is contained in its representation as a molecular graph. Can a Graph Neural Network’s (GNN’s) property predictions be improved if it understands a molecule’s geometry from only the molecular graph?

We show this to be the case. Using methods from self-supervised learning, we 3D pre-train a GNN to generate latent 3D information. During finetuning on molecules with unknown geometry, the GNN still generates implicit 3D information and uses it to inform downstream molecular property predictions, improving their accuracy.

Our 3D pre-training can provide large improvements for a wide range of molecular properties. Crucially, 3D pre-training always improves or performs on par, unlike many prior SSL methods for molecules that suffer from negative transfer for some tasks. Lastly, the learned representations are very generalizable and can be transferred between datasets with vastly different molecules. All these qualities of our method are essential for real-world applications and make it highly interesting for practice.