MKGE: Knowledge graph embedding with molecular structure information.

Area of Science:

• Biomedical informatics
• Computational biology
• Data science

Background:

• Knowledge graphs (KGs) are crucial for organizing biomedical data, but their utility is limited by incomplete relations and data sparsity.
• Existing knowledge graph embedding (KGE) methods struggle with sparse biomedical KGs due to limitations in knowledge extraction.
• Domain-specific knowledge integration into KGE methods, particularly molecular structure information, is underexplored in biomedicine.

Purpose of the Study:

• To develop a novel approach for knowledge graph embedding (KGE) that incorporates molecular structure information to overcome data sparsity in biomedical KGs.
• To improve the reliability and performance of KGE models in the biomedical domain by leveraging entity structure.
• To introduce a new method for enhancing KGE by fusing text-structure and graph-structure based entity representations.

Main Methods:

• Developed two strategies to obtain vector representations of entities: text-structure-based and graph-structure-based.
• Spliced these dual representations to serve as input for KGE models.
• Constructed a novel KCCR knowledge graph for validation purposes.

Main Results:

• Demonstrated the superiority of the proposed model in entity prediction tasks.
• Validated the model's effectiveness in relation prediction tasks.
• Showcased improved performance in drug-drug interaction prediction tasks using the integrated molecular structure information.

Conclusions:

• This research presents the first integration of molecular structure information into knowledge graph embedding (KGE) methods for the biomedical field.
• The proposed approach effectively addresses the challenge of data sparsity in biomedical KGs, leading to improved KGE performance.
• Future work can explore fusing other feature annotations like Gene Ontology and protein structure with KGE for further advancements.