Permutation of class labels is a common approach to build null distributions for significance analyis of microarray data.
It is assumed to produce random score distributions, which are not affected by biological differences between samples. We
argue that this assumption is questionable and show that basic requirements for null distributions are not met.
We propose a novel approach to the significance analysis of microarray data, called permutation filtering. We show that it
leads to a more accurate screening, and to more precise estimates of false discovery rates. The method is implemented in the
Bioconductor package twilight available on http://www.bioconductor.org.