XGDAG: explainable gene-disease associations via graph neural networks

Area of Science:

• Computational biology
• Genetics
• Machine learning

Background:

• Disease gene prioritization identifies genes linked to diseases using computational methods.
• Deep learning approaches have shown superior results but lack interpretability.
• Existing methods struggle with reliability when identifying a large number of associated genes.

Purpose of the Study:

• To develop a novel, interpretable methodology for disease gene discovery.
• To leverage graph neural networks (GNNs) for enhanced gene prioritization.
• To provide an explainability phase for understanding model outputs.

Main Methods:

• Utilized graph-structured data and graph neural networks (GNNs).
• Implemented a positive-unlabeled learning strategy.
• Integrated an explainability phase for model interpretation.

Main Results:

• The proposed methodology outperforms existing gene discovery methods.
• Achieved superior performance in gene prioritization, even for large gene sets.
• Enabled non-black-box exploitation of GNNs for disease gene discovery.

Conclusions:

• The new approach offers an effective and interpretable solution for disease gene discovery.
• It enhances the reliability of gene prioritization, particularly in complex scenarios.
• The methodology advances the application of GNNs in biomedical research.