GeOKG: geometry-aware knowledge graph embedding for Gene Ontology and genes.

Area of Science:

• Bioinformatics
• Computational Biology
• Network Science

Background:

• Deep learning for Gene Ontology (GO) and Gene Ontology Annotation (GOA) representation learning aids biological tasks like protein-protein interaction prediction.
• Existing methods embed GO and GOA in single geometric spaces, which are insufficient for GO's complex, nonmonotonic hierarchy.

Purpose of the Study:

• To address limitations in modeling GO's hierarchical structure.
• To develop a novel method for enhanced representation learning of GO and GOA.

Main Methods:

• Proposed Geometry-Aware Knowledge Graph Embeddings (GeOKG) method.
• Leveraged geometric interactions among various geometric representations during training.
• Modeled the complex hierarchy of GO more effectively.

Main Results:

• GeOKG effectively models the intricate hierarchical structure of GO.
• Experiments at the GO level demonstrated the benefits of geometric interactions.
• GeOKG outperformed existing methods in protein-protein interaction prediction at the gene level.

Conclusions:

• Geometric interaction is a promising approach for embedding heterogeneous biomedical networks.
• GeOKG enhances the representation learning of GO and GOA for downstream biological tasks.
• The findings highlight a new direction for improving biological network analysis.