Implications of topological imbalance for representation learning on biomedical knowledge graphs

Area of Science:

• Computational biology
• Machine learning in drug discovery
• Bioinformatics

Background:

• Drug discovery knowledge graphs (KGs) leverage interconnected data for tasks like predicting gene-disease associations.
• Graph embedding methods (KGE) offer intuitive representations and inference capabilities.
• Ensuring biological meaningfulness of KG predictions is critical.

Purpose of the Study:

• To investigate biases in KG construction and their impact on gene ranking for disease association.
• To demonstrate how topological overrepresentation affects entity ranking in KGE models.
• To highlight the influence of entity frequency versus biological information in KGE predictions.

Main Methods:

• Utilized various datasets, KGE models, and predictive tasks to assess graph imbalances.
• Performed graph perturbation experiments to analyze model behavior.
• Evaluated the impact of structural imbalances on the ranking of genes for diseases.

Main Results:

• Demonstrated that densely connected entities are consistently ranked highly, irrespective of context.
• Provided evidence across diverse datasets and models that KGEs can prioritize entity frequency over biological relationships.
• Graph perturbation experiments confirmed the susceptibility of KGE models to structural biases.

Conclusions:

• Inherent structural imbalances in KGs can lead to biased predictions, overranking hubs.
• KGE models may rely more on entity frequency than biological relevance, impacting drug discovery.
• Practitioners must be aware of these modeling choices and biases when composing KGs and interpreting results.