Graph embedding on biomedical networks: methods, applications and evaluations.

Area of Science:

• Biomedical Informatics
• Network Science
• Machine Learning

Background:

• Graph embedding learning aims to create low-dimensional node representations for network analysis.
• Existing methods are primarily evaluated on social/information networks, with limited systematic study on biomedical networks.
• Traditional methods like matrix factorization show promise, highlighting the need to evaluate newer graph embedding techniques.

Purpose of the Study:

• To systematically evaluate recent graph embedding methods on biomedical network analysis tasks.
• To compare their usability and potential against state-of-the-art techniques.
• To provide guidelines for selecting and parameterizing graph embedding methods for biomedical applications.

Main Methods:

• Selected 11 representative graph embedding methods for evaluation.
• Conducted systematic comparisons on five key biomedical tasks: drug-disease association (DDA) prediction, drug-drug interaction (DDI) prediction, protein-protein interaction (PPI) prediction, medical term semantic type classification, and protein function prediction.
• Utilized the BioNEV Python package for implementation and data sharing.

Main Results:

• Recent graph embedding methods achieve promising results on biomedical tasks, warranting further attention.
• These methods demonstrate competitive performance compared to state-of-the-art approaches for DDA, DDI, and protein function prediction, even without biological features.
• Learned embeddings can serve as valuable complementary representations to existing biological features.

Conclusions:

• Recent graph embedding techniques are effective for various biomedical network analysis tasks.
• The BioNEV package facilitates the study and application of these methods in biomedical research.
• The study provides practical guidelines for method selection and hyperparameter tuning in biomedical graph analysis.