Fig. 1

HOLIgraph encompasses a ligand graph (in orange, where the nodes are the atoms, and the edges are the intra-ligand bonds) and a protein graph (in blue, where the nodes are amino acids) connected by protein-ligand interaction edges (magenta). For example, protein-ligand interactions are shown for an OATP1B1 complex with Simeprevir, a known OATP1B1 inhibitor (A, left). The simplified HOLIgraph representation (A, right) corresponds to the interactions involving the 4-isopropylthiazole group of Simeprevir. The HOLIgraph data representation includes detailed features (B) of each atom and amino acid (AA) node, intra-ligand bond edge, and protein-ligand interaction edge. HOLIgraph features are further detailed in Tables S6-S9. The HOLIgraph architecture (C) involves processing input feature data through multilayer perceptron (MLP), neural message passing convolution (NNConv), heterogeneous graph convolution, projection, pooling, and graph classifier layers to predict the probability that a given ligand is an inhibitor or noninhibitor of OATP1B1