Hyperbolic hierarchical knowledge graph embeddings for biological entities.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Predicting biological relationships is crucial for biomedical research.
• Existing embedding methods in Euclidean space struggle to retain hierarchical information in large graph-structured data.
• This limitation hinders accurate semantic representation of biological entities and relationships.

Purpose of the Study:

• To propose HEM, a hyperbolic hierarchical knowledge graph embedding model.
• To generate accurate vector representations for biological entities.
• To capture latent hierarchical information effectively for improved biological predictions.

Main Methods:

• Developed HEM, a novel model utilizing hyperbolic spaces (Poincaré ball) for knowledge graph embedding.
• Encoded biological entities and relations within the hyperbolic geometry to leverage its negative curvature.
• Applied HEM to protein-protein interaction and gene-disease association prediction tasks.

Main Results:

• HEM successfully captures latent hierarchical information, outperforming Euclidean embedding methods.
• The model achieves superior accuracy in predicting protein-protein interactions and gene-disease associations.
• HEM preserves rich biological information in low-dimensional representations compared to Euclidean approaches.

Conclusions:

• Hyperbolic knowledge graph embedding offers a powerful approach for modeling biological data.
• HEM demonstrates significant improvements in accuracy and information preservation for biological relationship prediction.
• The proposed method provides a valuable tool for advancing biomedical research through enhanced bio-entity representation.