Detecting disease genes based on semi-supervised learning and protein-protein interaction networks.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Genomics

Background:

• Predicting human disease genes is crucial for understanding genetic disorders.
• Network-based approaches often assume disease genes have similar neighbors.
• Existing methods primarily use supervised learning with known disease genes.

Purpose of the Study:

• To develop an effective method for disease gene prediction by leveraging network neighborhoods and integrating omics data.
• To enhance disease gene prioritization using semi-supervised learning (SSL).

Main Methods:

• Developed a novel SSL method utilizing protein-protein interaction networks.
• Integrated multi-omics data from six biological databases (e.g., UniProt, GO, Pfam) and gene expression data.
• Employed a 10-fold cross-validation strategy for performance evaluation.

Main Results:

• The SSL method outperformed k-nearest neighbors and support vector machines, achieving 85% sensitivity, 79% specificity, 81% precision, and 82% accuracy.
• Demonstrated advantages with limited labeled data, yielding 78% accuracy.
• Identified 572 putative disease genes with biological validation.

Conclusions:

• SSL significantly enhances disease gene study, particularly when known disease genes are scarce.
• The method provides computational improvements and aids in deciphering pathogenic mechanisms through predicted disease proteins.