Graph Convolutional Networks for Molecules

Representation learning for molecules is important for molecular properties prediction, material design, drug screening, etc. In this work a graph convolutional network architecture for learning representations for molecules is presented. An operation for convolving k-neighbourhood of a specific node in graph is defined, which is corresponding to kernel size of k in convolutional neural networks. Besides, A module of adaptive filtering is defined to find the sampling locations based on graph connections and node features.