Three-Dimensionally Embedded Graph Convolutional Network (3DGCN) for Molecule Interpretation
share
Graph convolutional networks (GCNs) for learning graph representation of molecules widely expand their scope on chemical properties to biological activities, but learning the three-dimensional topology of molecules has not been explored. Many GCNs that have achieved state-of-the-art performances rely on node distances only, limiting the spatial information of molecules. In this work, we propose a novel model called three-dimensionally embedded graph convolutional network (3DGCN), which takes a molecular graph embedded in three-dimensional Euclidean space as an input and recursively updates scalar and vector features based on the relative positions of nodes. We demonstrate the capabilities of 3DGCN with tasks on physical and biophysical predictions.
READ FULL TEXT
