Fig 1.
(A) Drug-drug interaction networks: This component illustrates the network of interactions between drugs. (B) Topology information learning module: This module extracts common H-Hop neighbor nodes for drug pairs to form a subgraph. Subsequently, the subgraph is passed through GCN to generate a global representation for the drug pair. (C) Structural information learning module: This module employs GCN with shared parameters to acquire representations for drugs within drug pairs. (D) Graph contrastive learning module and prediction module: In this module, graph contrastive learning and cross-entropy loss are employed to regulate the model’s iteration and predict the interaction probability between input drug pairs.
Table 1.
Dataset statistics.
Table 2.
Comparative evaluation (mean ± std).
Best performance in each metric is shown in bold font.
Fig 2.
Comparative performance of DeepGCL and its variants on multiple evaluation metrics.
Fig 3.
Metrics for models in edge attack scenarios.
(a) Performance in BioSNAP dataset, (b) Performance in AdverseDDI dataset, (c) Performance in DrugBank dataset.
Table 3.
Assessment of DeepGCL and competitive methods under the drug-wise and pairwise settings.
The best score is in bold.
Fig 4.
Red points indicate drug pairs without interactions and green points indicate drug pairs with interactions.
Fig 5.
Effect of convolutional neural networks of different depths.
(a) Performance in BioSNAP dataset, (b) Performance in AdverseDDI dataset, (c) Performance in DrugBank dataset.