An integrative approach to modeling biological networks

Area of Science:

• Computational Biology
• Network Science
• Systems Biology

Background:

• Protein-protein interaction (PPI) networks and residue interaction graphs (RIGs) are crucial for understanding biological systems and protein structures.
• Selecting appropriate network models is essential for data analysis and prediction but challenging due to computational limitations.
• Existing methods often rely on single network properties, which may not reliably assess model fit.

Purpose of the Study:

• To develop and validate an integrative approach for identifying the best-fitting network models for PPI networks and RIGs.
• To evaluate the performance and robustness of machine learning classifiers in network model selection.
• To determine the most suitable network model for both RIGs and PPI networks.

Main Methods:

• Utilized five machine learning classifiers fed with diverse network properties.
• Applied the approach to residue interaction graphs (RIGs) and protein-protein interaction (PPI) networks.
• Tested the robustness of classifiers against noise in network data.

Main Results:

• Geometric random graphs (GEO) were confirmed as the best-fitting model for RIGs, validating the approach.
• High-coverage, high-confidence PPI data also showed consistency with GEO models.
• Classifiers exhibited varying robustness to noise; noisy GEO networks could be misclassified as scale-free (SF).

Conclusions:

• The integrative machine learning approach effectively identifies optimal network models for biological networks.
• Geometric random graphs (GEO) are a suitable model for residue interaction graphs (RIGs).
• Protein-protein interaction (PPI) network structures are most consistent with noisy geometric random graph (GEO) models, accounting for data imperfections.