Fig. 3

To comprehensively extract deep-level protein node features, this study employs a heterogeneous network to mine three key protein multi-hop pathways: direct protein–protein interactions (P–P), first-order indirect protein–drug–protein associations (P–D–P), and second-order mediated protein–drug–drug–protein relationships (P–D–D–P). First, we construct adjacency matrices for the three pathways separately and feed them, together with the PCA-reduced protein features, into the GWT module for multi-scale feature extraction. This enables simultaneous capture of both local and global topological information and yields multi-view protein feature representations. During the feature fusion stage, an attention mechanism is introduced to adaptively assign weights to features from the three pathways, followed by element-wise scalar product fusion (i.e., inner product of tensor vectors) to generate the final unified protein feature matrix. Drug features are processed in the same manner.