Development of molecular-graph representations in graph neural networks for molecular-property prediction

Abstract

Molecular graph is one of the spotlighted molecular representations as deep learning models, especially graph neural network (GNN) models, have recently been proposed to predict molecular properties and intermolecular interactions. A Molecular graph is defined as graph G(V,E), composed of nodes (i.e., atoms) and edges (i.e., bonds). Many researchers have proposed three-dimensional GNN models, using basic two-dimensional molecular graphs and additional descriptors for three-dimensional molecular information, these days. Despite all the studies, three-dimensional molecular descriptors such as bond distance and relational positions between atoms are only considered as molecular representations, and other three-dimensional molecular structural information is ignored in the molecular graph. In this work, we proposed three new descriptors for molecular graphs: full-electron-configuration vector, interatomic overlap area, and noncovalent molecular graph based on noncovalent interactions. We introduced a new combination of atom and bond descriptors, including the proposed descriptors, to predict solvation free energy of organic light-emitting diode (OLED) molecules as solutes. We also proposed MolNet, a chemically intuitive covalent-noncovalent multimodal three-dimensional GNN model which learns both covalent and noncovalent interactions in molecules, to predict the binding affinity between a protein and ligands complex and aqueous solvation free energy. The MolNet model showed higher prediction performance than earlier GNN models. The descriptors and the noncovalent molecular graph introduced in this work could be applied and adopted for other types of molecular representations, and the MolNet model could predict various molecular properties.