Evaluation of microarray data normalization procedures using spike-in experiments.

Background: Recently, a large number of methods for the analysis of microarray data have been proposed but there are few comparisons of their relative performances. By using so-called spike-in experiments, it is possible to characterize the analyzed data and thereby enable comparisons of different analysis methods.

Results: A spike-in experiment using eight in-house produced arrays was used to evaluate established and novel methods for filtration, background adjustment, scanning, channel adjustment, and censoring. The S-plus package EDMA, a stand-alone tool providing characterization of analyzed cDNA-microarray data obtained from spike-in experiments, was developed and used to evaluate 252 normalization methods. For all analyses, the sensitivities at low false positive rates were observed together with estimates of the overall bias and the standard deviation. In general, there was a trade-off between the ability of the analyses to identify differentially expressed genes (i.e. the analyses' sensitivities) and their ability to provide unbiased estimators of the desired ratios. Virtually all analysis underestimated the magnitude of the regulations; often less than 50% of the true regulations were observed. Moreover, the bias depended on the underlying mRNA-concentration; low concentration resulted in high bias. Many of the analyses had relatively low sensitivities, but analyses that used either the constrained model (i.e. a procedure that combines data from several scans) or partial filtration (a novel method for treating data from so-called not-found spots) had with few exceptions high sensitivities. These methods gave considerable higher sensitivities than some commonly used analysis methods.

Conclusion: The use of spike-in experiments is a powerful approach for evaluating microarray preprocessing procedures. Analyzed data are characterized by properties of the observed log-ratios and the analysis' ability to detect differentially expressed genes. If bias is not a major problem; we recommend the use of either the CM-procedure or partial filtration.