TLGP: a flexible transfer learning algorithm for gene prioritization based on heterogeneous source domain

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Gene prioritization is crucial for cancer diagnosis and therapy, identifying key genes as biomarkers or drug targets.
• Existing methods struggle with limited known disease-causing genes, particularly for rare cancers.
• There is a need for effective algorithms for gene ranking with scarce prior disease-causing gene information.

Purpose of the Study:

• To propose a transfer learning-based algorithm for gene prioritization (TLGP) in cancer.
• To address the challenge of limited disease-causing genes in the target cancer domain.
• To leverage knowledge from other cancers (source domains) to improve gene prioritization accuracy.

Main Methods:

• TLGP quantifies target-source domain similarity using gene affinity matrices.
• It learns a fusion network by integrating affinity matrices, pathogenic genes, and genomic data from source cancers.
• Genes in the target cancer are then prioritized based on the learned network.

Main Results:

• The learned fusion network is more reliable than traditional gene co-expression networks.
• Transferring knowledge from similar cancers significantly improves network construction accuracy.
• TLGP outperforms state-of-the-art methods, achieving at least a 5% accuracy improvement.

Conclusions:

• The developed TLGP model offers an effective and efficient strategy for gene ranking.
• It successfully integrates genomic data across various cancers.
• This approach provides a robust solution for gene prioritization, even with limited disease-specific data.