Comparison of co-expression measures: mutual information, correlation, and model based indices

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Gene co-expression networks are crucial for understanding gene function and regulation.
• Mutual Information (MI) is a common metric for inferring gene-gene relationships, but its added value over correlation is unclear.
• Identifying biologically meaningful gene modules requires assessing co-expression measures and data transformations.

Purpose of the Study:

• To comprehensively compare Mutual Information (MI) with various correlation measures for gene co-expression network construction.
• To evaluate the impact of different adjacency matrix transformations, such as topological overlap, on module quality.
• To explore alternative methods for capturing non-linear relationships beyond MI.

Main Methods:

• Comparative analysis of MI and robust correlation measures (e.g., biweight midcorrelation) across eight empirical datasets and simulations.
• Assessment of topological overlap measure and other transformations on adjacency matrices.
• Evaluation of gene ontology enrichment for modules derived from different co-expression measures.

Main Results:

• A strong linear relationship exists between MI and correlation for most gene pairs, indicating MI offers limited advantage in stationary data.
• Biweight midcorrelation, transformed by topological overlap, yields superior gene ontology-enriched co-expression modules compared to MI and Maximal Information Coefficient (MIC).
• Polynomial or spline regression models are proposed as effective alternatives to MI for non-linear relationships.

Conclusions:

• Biweight midcorrelation is a more effective measure than MI for gene pairwise relationships.
• Topological overlap transformation combined with biweight midcorrelation significantly improves co-expression module enrichment.
• Correlation-based networks are a viable and often preferable alternative to MI networks for stationary co-expression data.