A novel candidate disease gene prioritization method using deep graph convolutional networks and semi-supervised...

Area of Science:

• Computational biology and bioinformatics
• Machine learning in genomics
• Network-based disease gene prioritization

Background:

• Prioritizing candidate disease genes is crucial but challenging due to the cost and time of laboratory validation.
• Existing machine learning methods face challenges with diverse data structures (e.g., graphs) and lack of negative data.
• Graph neural networks offer superior performance for complex data structures.

Purpose of the Study:

• To develop a novel semi-supervised learning method for prioritizing candidate disease genes.
• To address challenges in feature vector construction and model training for biological network data.
• To leverage graph convolutional networks for enhanced gene prioritization.

Main Methods:

• Constructed three novel feature vectors for each gene using Gene Ontology (GO) terms (molecular function, cellular component, biological process).
• Employed a graph convolutional network (GCN) trained on these feature vectors and protein-protein interaction (PPI) network data.
• The model integrates topological network information with gene features for robust candidate gene identification.

Main Results:

• The proposed method achieved superior performance in prioritizing candidate disease genes across 16 diseases.
• Evaluated metrics including precision, Area Under the ROC Curve (AUC), and F1-score demonstrated significant improvements.
• Outperformed eight state-of-the-art network and machine learning-based prioritization methods.

Conclusions:

• The developed semi-supervised GCN method effectively classifies and ranks candidate disease genes.
• The innovative feature vector construction based on GO terms enhances prioritization accuracy.
• This approach offers a powerful tool for accelerating disease gene discovery.