Disease Gene Classification with Metagraph Representations.

Area of Science:

• Bioinformatics
• Computational Biology
• Network Science

Background:

• Protein-protein interaction (PPI) networks are crucial for understanding protein functions in disease.
• Existing PPI networks lack sufficient biological context for accurate disease protein prediction.
• Integrating protein properties like molecular functions and biological processes is necessary to enhance PPI network analysis.

Purpose of the Study:

• To develop a novel method for identifying candidate disease-causing proteins.
• To enhance protein-protein interaction networks by integrating protein properties using keywords.
• To represent proteins using metagraphs for improved topological analysis.

Main Methods:

• Constructed a heterogeneous PPI-Keyword (PPIK) network integrating protein-protein interactions and protein-keyword associations.
• Utilized metagraphs to capture topological arrangements considering both protein interactions and properties.
• Fed metagraph representations into classifiers for disease protein prediction.

Main Results:

• The proposed method consistently improved disease protein prediction performance across various classifiers, achieving an average AUC increase of 15.3%.
• Outperformed diffusion-based (e.g., RWR) and module-based baselines by 13.8-32.9% in overall disease protein prediction.
• Demonstrated superior performance in breast cancer protein prediction (6.6-14.2% improvement) and showed better correlation with PubMed literature findings.

Conclusions:

• Integrating protein properties via keywords into PPI networks significantly enhances disease protein prediction.
• Metagraph representations effectively capture complex topological information for improved predictive power.
• The developed method offers a robust and accurate approach for identifying disease-associated proteins.