Skip to main content
. 2022 Jan 21;12:1124. doi: 10.1038/s41598-022-04967-9

Figure 1.

Figure 1

A high-level sketch of our method. In Step 1, a GNN is trained with property prediction tasks. The black and red arrows indicate the forward pass and backpropagation, respectively. In Step 2, the trained GNN is used to generate molecular vectors of candidate molecules, and then molecules are selected via submodular function maximization.