Fig. 2

When extracting protein node features from the constructed heterogeneous network, the protein interactions are first divided into four edge types according to connected node categories: protein–drug, protein–protein, protein–side effect, and protein–disease. Subsequently, an HGCN is applied independently to each edge type to extract relational features. These learned features, together with the original protein features reduced via PCA, are integrated using a multi-modal fusion module, referred to as Multiple, which performs element-wise multiplication to combine modalities. Finally, the four fused features are aggregated through mean pooling to yield the final protein feature matrix. Drug features are processed in the same manner.