Graph Neural Networks for Property-guided Molecule Generation

Abstract

Organic Photovoltaic Molecules (OPVs) have attracted chemists’ attention in improving their electricity production efficiency while maintaining a low production cost. Designing more efficient OPVs is slow and challenging due to the larger and more complex structures of OPVs compared to other molecules. Moreover, there are currently no large-scale datasets of inefficient and efficient versions of OPVs for the supervised generation of more efficient molecules. Hence, we formulate this molecule design task as an unsupervised, propertyguided molecule optimization task, such that the chemical properties of the generated OPV candidate molecules closely match the desired property values for high efficiency. Specifically, we propose a motif-based graph-to-graph generative model to address the large molecule size of OPVs. The generated OPV candidates are, according to chemists, rated promising with minimal modifications.