Prediction of Molecular Properties Using Graph Kernel and Active Learning

In this talk, we present a new machine learning method for training predictive models of extensive molecular properties through the application of a similarity kernel on graphical representations of molecules, which is intuitive and can adapt to molecules of arbitrary size and topology. The pairwise similarity matrices between 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. The method also uses a much smaller number of training samples as compared to other methods to achieve the same level of accuracy.