Graph Neural Network for 3-Dimensional Structures Including Dihedral Angles for Molecular Property Prediction

Abstract

The prediction of molecular properties using graph neural network (GNN)- based approaches has attracted great attention in recent years. Topological molecular graphs are commonly used for representing molecules in machine learning (ML). However, the challenge is to utilize the complete geometry information, such as, bonds, angles, and dihedral angles while processing a molecular graph. In this work, we present predictive GNN accounting three-dimensional molecular structures including the dihedral angles (GNN3Dihed) in a systematic manner. Additionally, we demonstrate that the usage of autoencoders to generate latent space embeddings for usually sparse atomic and bond vectors reduces the number of parameters in the message passing stage while not reducing performance. We compare the performance of GNN3Dihed with state-of-the-art baselines on several tasks (regression and classification), for example, solubility prediction, toxicity prediction, binding affinity, and quantum mechanical property prediction, and showed that the present architecture often outperforms other models—demonstrating the importance of 3D structural information for ML in chemistry.