CIPHER-SC: Disease-Gene Association Inference Using Graph Convolution on a Context-Aware Network With Single-Cell...

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Accurate inference of disease-gene associations is crucial for understanding disease mechanisms and developing treatments.
• Current machine learning methods face challenges due to inaccurate feature selection and multi-stage training architectures.
• Existing approaches often lack cell-type-specific resolution, limiting the detailed study of gene functions.

Purpose of the Study:

• To develop a novel computational approach for accurate disease-gene association inference.
• To integrate single-cell transcriptome data for higher-resolution gene function analysis.
• To create an end-to-end learning architecture that minimizes error accumulation.

Main Methods:

• Utilized single-cell transcriptome data and constructed a context-aware network for unbiased data integration.
• Developed a graph convolution-based approach named CIPHER-SC with a complete end-to-end learning architecture.
• Performed five-fold cross-validations on three distinct test sets to evaluate performance.

Main Results:

• CIPHER-SC achieved a maximum AUC of 0.9501, outperforming four state-of-the-art methods.
• The end-to-end design and unbiased data integration improved AUC from 0.8727 to 0.9443.
• Incorporating single-cell data enhanced prediction accuracy and enriched results with cell-type-specific genes.

Conclusions:

• CIPHER-SC demonstrates a robust ability to discover reliable disease genes, including novel and genetically based associations.
• The method's end-to-end architecture and integration of single-cell data significantly improve prediction performance.
• CIPHER-SC offers a powerful tool for advancing disease gene discovery with enhanced resolution.