Fig. 5

The operation of the graph convolution and GCN model architecture of binary classification is described in this paper. To include the surrounding environments of each atom (local environments), the multiplication of topology graph \(\tilde{L}\) and the compound features (either the input atomic features AF or the feature map F^(l−1) from the last layer) is required, which should be further multiplied by W_l to learn the information provided by the local environment.