Abstract

Predicting the properties/bioactivity of a molecule from its structure is a longstanding challenge in drug discovery, and in the past few years graph neural networks have found increasing popularity amongst the cheminformatics community. Such models operate on atomic and/or bond features over 2D graph descriptions of molecules, and they have been established as the current state-of-the-art.

However, many properties of interest (eg solvation, binding affinity) involve 3D interactions between the local regions of molecules, which might not be easily described by 2D descriptors and motivates the use of a model which captures the 3D shape of a molecule.

In this talk I will introduce a GP regression model which uses the SOAP rematch kernel, and show that it can beat state-of-the-art graph neural networks on benchmark datasets. In addition, I will describe some recently proposed metrics for quantifying the quality of the uncertainty predictions from ML models.