General Molecular Similarity Assessment Using Message Passing Neural Network

Graduate School of Artificial Intelligence ArtificiWe present a new way to express the similarity between molecules. The molecular similarity is subjectively defined for each field used for the chemical properties or structural aspect of a molecule. In order to know the chemical properties of molecules, high computational costs such as Density Functional Theory(DFT) calculation or real-world experiments are required. In addition, high domain knowledge is essential because understanding molecular structures and comparing structural differences between two molecules requires analysis in all aspects, such as 2D or 3D similarities or local and global similarities. Each similarity measure uses a completely different method and metric. Therefore, it is difficult to express and compare the two properties of chemical and structural properties simultaneously. However, in tasks that require molecular screening, such as drug discovery, both factors are important for comparing molecules. So to express the general molecular similarity, we consider both the chemical properties and the physical structure of the molecule. We convert the existing deterministic and costly method for similarity assessment into a learnable method through a deep learning approach and unify the similarity for each factor into one. In this study, the molecular structure is expressed as a graph, and the molecules are aligned in the latent space after embedding by reflecting the chemical and structural characteristics of the molecule using a Message Passing Neural Network(MPNN). The distance in the latent space between aligned molecules is directly related to the similarity between molecules, and adjacent molecules have similar structures and properties. In the experimental part, we validate the methods for mapping similarity to distance in latent space and qualitatively compare similar molecules in the real molecular population using the ZINC dataset.ope