Comprehensive analysis of correlation coefficients estimated from pooling heterogeneous microarray data.

Area of Science:

• Bioinformatics
• Statistical Genetics
• Computational Biology

Background:

• Microarray data synthesis uses 'mosaic' (combining results) or 'melting pot' (pooling data) approaches.
• Data heterogeneity in microarray studies (lab differences, experimental conditions) can yield ambiguous results with the 'melting pot' method.

Purpose of the Study:

• To investigate the impact of mean differences and heteroskedasticity on gene-to-gene Pearson correlation coefficients in pooled microarray data.
• To compare the biases of pooled correlation coefficients with those from an effect-size model and classical meta-analysis.

Main Methods:

• Applied statistical theory to analyze 19 groups of microarray data.
• Quantified biases of pooled coefficients and compared them to an effect-size model.
• Used simulation studies to corroborate findings on mean differences and heteroskedasticity.

Main Results:

• Mean differences across microarray groups significantly influenced the magnitude and sign of pooled correlation coefficients, increasing bias.
• Heteroskedasticity led to less efficient correlation estimations compared to classical meta-analysis.
• Pooled coefficients showed the largest bias when approaching ±1.

Conclusions:

• Combining statistical results (mosaic approach) is the preferred method for synthesizing gene expression correlations across multiple microarray studies.
• The 'melting pot' approach is susceptible to biases introduced by data heterogeneity.