Biomedical Knowledge Graph for Drug Repurposing with 2.2 Million Edges

Functioning as a multimodal biomedical knowledge graph assembled from DrugBank, DisGeNET, and Hetionet, comprising more than 2.2 million edges. It was used to train and evaluate a heterogeneous attention-based meta-learning graph neural network (HAMGNN) for computational drug repurposing and biomarker discovery. The model was tested using a disjoint disease-based (cold-start) evaluation protocol.

Use Cases

• Train heterogeneous graph neural networks like HAMGNN on the 2.2 million edges for drug-disease link prediction.
• Benchmark new models for drug repurposing against the disjoint disease-based evaluation protocol used in the source study.
• Extract subgraphs focused on specific diseases like Alzheimer's or Long COVID from the knowledge graph for case study analysis.
• Use the multimodal knowledge graph, which integrates curated databases and literature, for multi-omics profile analysis in biomarker discovery.

Strengths

• The knowledge graph contains more than 2.2 million edges, providing a large-scale relational structure.
• Integrates heterogeneous evidence from multiple authoritative biomedical sources: DrugBank, DisGeNET, and Hetionet.
• Designed for and tested under a stringent disjoint disease-based (cold-start) evaluation protocol to assess generalization.

Limitations

• The dataset is small in file size (419.1 KB), suggesting it may contain processed graph indices or metadata rather than the full raw source data.
• Specific node and edge schemas, column names, and sample data are unavailable, limiting immediate analytical utility.
• The knowledge graph's construction and content are tied to the specific methodology (e.g., LLM-extracted literature data) of the source article, which may introduce specific biases.

Provenance

Source

Assembled from DrugBank, DisGeNET, and Hetionet.

Collection Method

Constructed via a literature-enhanced knowledge graph construction pipeline encoding LLM-extracted therapeutic information.

Freshness

Last updated 2026-03-20.