Which normalization method is best? A platform-independent biologically inspired quantitative comparison of normalization methods

Abstract

Since the introduction of microarray technology, several different normalization techniques have been introduced, but it is still unclear which normalization method is best. We present the first comparative study of normalization methods for both cDNA as well as oligonucleotide arrays that is based on their overall performance on five complementary performance measures. The presented comparison is unique in that it 1) compares normalization methods with very different outcomes, 2) is applied to two different array platforms, 3) introduces several different (biologically inspired) performance measures and 4) can be applied to any data set. The results show amongst others that, for cDNA arrays, the well-established lowest-compensation of logratio is not biologically beneficial and that a novel ratio-based normalization (without logarithm) performs best overall. For Affymetrix arrays, we found that Rosetta's Experiment Builder is generally to be preferred.