Fig. 2. The architecture of the Bayesian GCN used in this work. (a) The entire model is composed of three augmented graph convolutional layers, readout layers and two linear layers and takes inputs as a molecular graph G(H⁽⁰⁾, A), where H⁰ is a node feature and A is an adjacency matrix. (b) Details of each graph convolution layer augmented with attention and gate mechanisms. The l-th graph convolutional layer updates node features and produces H^(l+1).