A Gated Spatial-Channel Transformer Network for the Prediction of Molecular Properties

Although spatial features of molecules have been widely used for molecular property prediction, the importance of interactive features is coming to the surface these days. By using molecular voxel-based representation, each voxel contains the distribution of atoms, while each channel corresponds to one atomic type. Thus the interaction between multiple atoms is actually contained in the channel information. In this work, we propose a gated spatial-channel transformer (GatedSCT) network for molecular property prediction. We design a channel transformer to capture interactive features from channels, which indicates the relationship between multiple atoms. Also, a spatial transformer is used to extract spatial features of molecules. We apply a gated mechanism to merge these two parts efficiently. Since our proposed network takes advantage of channel information, the experiments show that it can predict molecular properties more accurately than other networks.