Unite and conquer: univariate and multivariate approaches for finding differentially expressed gene sets

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• Numerous univariate and multivariate methods exist for testing differential gene expression between phenotypes.
• Existing performance studies on simulated and real data highlight the need to quantify relative performance under different null hypotheses.

Purpose of the Study:

• To compare the performance of univariate and multivariate gene set testing approaches.
• To evaluate how correlations, dimensionality, and changing gene percentages affect test power.
• To assess the utility of different null hypotheses in identifying differentially expressed gene sets.

Main Methods:

• Performance evaluation using simulated data.
• Application of various test statistics to biological data.
• Comparison of three specific statistics: sum of squared t-tests, Hotelling's T(2), and N-statistic.

Main Results:

• High correlations equally reduce the statistical power of both univariate and multivariate tests.
• Gene set dimensionality and the proportion of changing genes similarly affect test power.
• Three distinct statistics, testing different null hypotheses, identify both common and complementary differentially expressed gene sets.

Conclusions:

• Different gene set testing statistics, by evaluating distinct null hypotheses, capture different facets of the data.
• Simultaneous application of multiple tests (sum of squared t-tests, Hotelling's T(2), N-statistic) provides a more complete analysis of biological data than relying on a single method.