Application and evaluation of knowledge graph embeddings in biomedical data.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Linked data and bio-ontologies are crucial for biological and biomedical databases, ensuring data integrity, organization, and searchability.
• These technologies are increasingly used to build knowledge graphs, representing biological information as complex networks.
• Knowledge graph embeddings enable relationship prediction between biological entities, advancing data analytics and machine learning for decision support.

Purpose of the Study:

• To provide a comparative assessment and standard benchmark for knowledge graph-based representation learning methods.
• To focus on the link prediction task for biological relations.
• To identify limitations and suggest guidelines for developing improved knowledge graph embedding methods.

Main Methods:

• Systematic investigation and comparison of state-of-the-art embedding methods.
• Evaluation based on training and testing design settings.
• Assessment of strategies for controlling relational information and their impact on performance.
• Analysis of knowledge graph feature quality using clustering and visualization.
• Application of multiple evaluation metrics to compare methods.

Main Results:

• Identified limitations of current knowledge graph-based representation learning methods.
• Provided insights into the effects of different training and evaluation strategies.
• Demonstrated the utility of various evaluation metrics for assessing link prediction performance.
• Highlighted the importance of controlling information related to each relation.

Conclusions:

• Knowledge graph embeddings offer a powerful approach for biological link prediction.
• Systematic benchmarking is essential for understanding method performance.
• Further research is needed to overcome current limitations and enhance predictive accuracy in biological knowledge graphs.