Benchmarking network propagation methods for disease gene identification.

Area of Science:

• Computational biology
• Bioinformatics
• Pharmacology

Background:

• In-silico identification of potential drug targets is crucial for efficient drug discovery.
• Leveraging genetic, genomic, and protein interaction data aids in finding successful drug targets.

Purpose of the Study:

• To systematically test 12 network propagation-based algorithms for identifying drug-targeted genes.
• To evaluate algorithm performance using gene-disease data from 22 common non-cancerous diseases.
• To assess the impact of network properties and validation strategies on performance.

Main Methods:

• Systematic testing of 12 algorithms on gene-disease data from OpenTargets.
• Utilized two biological networks, six performance metrics, and two types of gene-disease association scores.
• Introduced novel protein complex-aware cross-validation schemes to mitigate over-optimistic estimates.

Main Results:

• Machine learning and diffusion-based methods identified 2-4 true drug targets within the top 20 suggestions when seeding with known targets.
• Performance decreased significantly when seeding with genetically associated disease genes.
• Larger, albeit noisier, biological networks improved overall performance.

Conclusions:

• Diffusion-based prioritisers and machine learning on diffusion-based features are effective for practical drug discovery.
• These methods outperform simpler neighbor-voting approaches.
• The choice of validation strategy and seed disease genes definition significantly impacts results.