Feature selection in molecular graph neural networks based on quantum chemical approaches

Feature selection is one of the important topics and it has been widely studied in data science. Recently, graph neural networks (GNNs) and graph convolutional networks (GCNs) have been also employed in chemistry. To enhance the performance of the GNN and GCN in the chemistry field, the feature selection should be also discussed in detail from the view point of chemistry. In this study, the authors proposed a new feature in molecular GNNs based on the quantum chemical approaches and discussed the accuracy, the overcorrelation between features, and interpretability. From the overcorrelation and the accuracy, the important graph convolution (IGC) with molecular atomic properties (MAPs), which was proposed in this study, showed good performance. Moreover, the integrated gradients analysis showed that the machine leaning model with the IGC(MAP) gave reasonable explanation to the prediction outputs.