Fig. 3

The proposed DGN architecture for sensitivity prediction. Given a PPIN subgraph \(\mathcal {G}_{\mathscr {S}}\), its node features \(\mathscr {X}_{\mathcal {G}}\) and connectivity functions \(\overrightarrow{\mathcal {N}_{\mathcal {G}}}\) and \(\overleftarrow{\mathcal {N}_{\mathcal {G}}}\) are provided as input to the message passing module, which computes the node embeddings by applying L graph convolutional layers to the node features. After each convolutional layer, the embeddings are passed through a ReLU non-linearity and a dropout layer. Then, the final embeddings are aggregated into a single graph representation via the pooling module. Lastly, the readout module takes in the aggregated graph representation and computes a sensitivity prediction \(\hat{y}_{\mathscr {S}}\)