Normalization Methods for Analysis of Microarray Gene-Expression Data

Abstract

This paper investigates subset normalization to adjust for location biases (e.g., splotches) combined with global normalization for intensity biases (e.g., saturation). A data set from a toxicogenomic experiment using the same control and the same treated sample hybridized to six different microarrays is used to contrast the different normalization methods. Simple t-tests were used to compare two samples for dye effects and for treatment effects. The numbers of genes that reproducibly showed significant p-values for the unnormalized data and normalized data from different methods were evaluated for assessment of different normalization methods. The one-sample t-statistic of the ratio of red to green samples was used to test for dye effects using only control data. For treatment effects, in addition to the one-sample t-test of the ratio of the treated to control samples, the two-sample t-test for testing the difference between treated and control samples was also used to compare the two approaches. The method that combines a subset approach (median or lowessfit) for location adjustment with a global lowess fit for intensity adjustment appears to perform well.

Keywords:

• Consistency
• Dye-swap experiment
• Global and subset normalizations
• Median and lowess adjustments
• One-sample and two-sample t-tests

Acknowledgments

We thank Drs. Robert Delongchamp, Cruz Velasco, James Fuscoe, and Sue-Jane Wang for comments on the manuscript. We are indebted to them for pointing out some important issues. The views presented in this paper are those of the authors and do not necessarily represent those of the U.S. Food and Drug Administration.