One of the most important tasks for artificial intelligence-assisted molecular design is the prediction of physicochemical qualities from molecular structures. To meet this problem, an increasing number of Graph Neural Networks (GNNs) have been proposed. By including more information in molecules, these models expand their expressive power while unavoidably increasing their computational complexity. In this work, we seek to create a powerful and effective GNN for molecular structures. By first representing each molecule as a two-layer multiplex graph, one layer of which only contains local connections that primarily capture covalent interactions and the other layer of which contains global connections that can simulate non-covalent interactions, we propose a molecular mechanics-driven approach to accomplishing this goal. Then, in order to balance the trade-off between expression strength and computing complexity, a corresponding message passing module is proposed for each layer. We construct the Multiplex Molecular Graph Neural Network based on these two modules. When it was verified using a dataset for big protein-ligand complexes and tiny compounds

Neural network, GNN, AI, deep learning, drug discovery, molecules.