DualKG-DC: A Drug-Centric Dual-Layer Knowledge Graph Framework for Drug Combination Prediction

Area of Science:

• Computational biology
• Pharmacology
• Biomedical informatics

Background:

• Current drug combination discovery is primarily disease-centered.
• A drug-centered approach can accelerate translational applications by repurposing known combinations.
• Leveraging established drug safety profiles is crucial for efficient therapeutic development.

Purpose of the Study:

• To introduce DualKG-DC, a novel drug-centered computational framework.
• To identify potential disease indications for given drug combinations.
• To complement existing disease-centered drug discovery methods.

Main Methods:

• Developed a dual-layer knowledge graph architecture (DualKG-DC).
• Pretrained the knowledge graph on a foundation biomedical knowledge graph.
• Refined the graph on a task-specific drug combination subgraph.
• Utilized existing knowledge on drug targets, pathways, and phenotypes.

Main Results:

• DualKG-DC outperformed three state-of-the-art models in systematic benchmarking.
• Achieved high performance metrics: Hits@10=0.48, MRR=0.30, AUROC=0.99, AUPRC=0.31.
• Demonstrated superior performance in cold start scenarios for unseen drug combinations (Hits@10=0.32, MRR=0.18, AUROC=0.98, AUPRC=0.23).

Conclusions:

• DualKG-DC is an effective platform for discovering therapeutic opportunities of drug combinations.
• The dual-layer architecture facilitates knowledge transfer, improving predictive performance and robustness.
• The framework shows particular promise for identifying indications for novel or previously unseen drug combinations.