Algorithms 1 The pseudo-code of GNN-DDI

input: Molecular graph ${\mathrm{G}}_{\mathrm{x}}$ of drug x and its original features ${\mathrm{H}}_{\mathrm{i}}^{\left(0\right)}$ of atomic nodes                Molecular graph ${\mathrm{G}}_{\mathrm{y}}$ of drug y and its original features ${\mathrm{H}}_{\mathrm{i}}^{\left(0\right)}$ of atomic nodes output: Probability score $p\left({\mathrm{G}}_{\mathrm{x}},{\mathrm{G}}_{\mathrm{y}}\right)$ of drug pair $\left(\mathrm{x},\mathrm{y}\right)$ 1: Initialize parameter sets in GNN-DDI. 2: for k in K: 3: Compute ${\mathrm{h}}_{\mathrm{x}}^{\left(\mathrm{k}+1\right)}$ and ${\mathrm{h}}_{\mathrm{y}}^{\left(\mathrm{k}+1\right)}$ based on Equations (1) to (3). 4: SAGPooling based on Equation (4) to obtain ${\mathrm{H}}_{{\mathrm{G}}_{\mathrm{x}}}^{\left(\mathrm{k}+1\right)}$ and ${\mathrm{H}}_{\mathrm{y}}^{\left(\mathrm{k}+1\right)}$ in layer k. 5: end for 6: Concatenate k-hops ${\mathrm{H}}_{{\mathrm{G}}_{\mathrm{x}}}^{\left(\mathrm{k}+1\right)}$ and ${\mathrm{H}}_{\mathrm{y}}^{\left(\mathrm{k}+1\right)}$ based on Equation (5) to obtain ${\mathrm{H}}_{{\mathrm{G}}_{\mathrm{x}}}$ and ${\mathrm{H}}_{{\mathrm{G}}_{\mathrm{y}}}$. 7: Concatenate ${\mathrm{H}}_{{\mathrm{G}}_{\mathrm{x}}}$ and ${\mathrm{H}}_{{\mathrm{G}}_{\mathrm{y}}}$ to obtain the latent feature vector of a drug pair $\mathrm{H}({\mathrm{G}}_{\mathrm{x}},{\mathrm{G}}_{\mathrm{y}})$ 8: Feed feature vector $\mathrm{H}({\mathrm{G}}_{\mathrm{x}},{\mathrm{G}}_{\mathrm{y}})$ into the predictor to get probability score $p\left({\mathrm{G}}_{\mathrm{x}},{\mathrm{G}}_{\mathrm{y}}\right)$.