In this talk, we will discuss how graphs, which are variably-sized non-linear data structures, can be used as intuitive and flexible representations of chemical entities such as molecules and atom clusters in the context of data-driven modeling. Through the use of a kernel function that evaluates and quantifies the similarity between the graphical representations, we are able to construct a machine learning pipeline for training predictive models of molecular properties such as atomization energy and charge distribution. We will talk about how the pairwise similarity matrices between the molecules as computed by the graph kernel can be used to construct Gaussian process regression models that can predict extensive properties with provable size scaling. Using an active learning procedure, we demonstrate that models created by our method can achieve a state-of-the-art accuracy of less than 1 kcal/mol on predicting atomization energies for molecules in the QM7 dataset without using any explicit energy decomposition/localization scheme. In constrat to common belief, the method also requires much less training timing to achieve the mentioned level of accuracy thanks to an optimized and GPU-accelerated open-source implementation.

Graph-based representations and kernel methods for the prediction of molecular properties

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.