PharmKG: a dedicated knowledge graph benchmark for bomedical data mining.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Data Science

Background:

• Biomedical knowledge graphs (KGs) are crucial for understanding complex biological systems and diseases.
• Challenges in KG construction include sparse data, insufficient modeling, and non-uniform evaluation metrics.

Purpose of the Study:

• To establish a comprehensive KG system for the biomedical field to address existing challenges.
• To introduce PharmKG, a multi-relational, attributed biomedical KG with extensive interconnections and relation types.

Main Methods:

• PharmKG integrates over 500,000 interconnections between genes, drugs, and diseases, featuring 29 relation types and ~8000 disambiguated entities.
• Entities are enriched with heterogeneous, domain-specific information from multi-omics data (gene expression, chemical structure, disease word embeddings).
• Evaluated nine state-of-the-art KG embedding (KGE) approaches and a novel graph neural network-based KGE method using a proposed benchmark and multiple metrics.

Main Results:

• Extensive experiments were conducted to assess the performance of various KGE models on the PharmKG benchmark.
• The study provides insights and guidelines for utilizing KGs in biomedical applications based on observed performance across downstream tasks.

Conclusions:

• PharmKG offers unprecedented quality and diversity, aiming to advance biomedical KG construction, embedding, and application.
• The developed KG system and evaluation framework facilitate better understanding and utilization of biomedical data.