Learning over Molecules: Representations and Kernels

Abstract

In this paper, we tackle machine learning over molecular space by considering three representations for molecules: (1) a vector of molecular properties that we treat as predictor variables, (2) a graph that captures the relationship between individual atoms in a molecule, and (3) a cheminformatic fingerprint that “identifies” a molecule. We assess the viability of each representation by training a model to predict energy values. In particular, we look a class of models that use kernel methods, whereby the prediction algorithm relies on a similarity measure between training data. On a subset of the Harvard Clean Energy Project (CEP) database, we find a simple fingerprint similarity kernel to be the fastest and most accurate for predicting HOMO-LUMO energy gap values.