Fig. 4

The proposed model begins by constructing a heterogeneous graph that integrates four types of biomedical entities–drugs, proteins, diseases, and side effects–along with their corresponding interactions. Initial node features, as illustrated in Fig. 1, are extracted and then combined with the constructed heterogeneous graph structure to serve as input to two parallel encoding modules: a two-layer NV encoder and a two-layer DV encoder. These modules capture complementary semantic perspectives from the heterogeneous network and generate enriched node-level representations. To align and enhance the expressiveness of the representations from both views, a contrastive learning mechanism is employed. Specifically, a similarity-based sampling strategy (Get Pos) is used to construct a dictionary of positive and negative samples (Pos_dict), which guides the contrastive optimization process between the NV and DV encoder outputs. Finally, the resulting node representations from both encoders are fused and passed into a DistMult decoder. This decoder incorporates multiple types of biomedical relations (e.g., drug–drug, protein–protein, drug–disease, etc.) through relation-specific diagonal matrices to reconstruct the full set of heterogeneous interactions. In particular, it focuses on computing the DTI matrix, which constitutes the primary prediction task of the model.