The a priori expectation is that not all genes are expressed in any given tissue at levels detectable by Affymetrix GeneChip® arrays (1–3 transcripts per cell) [3]. It therefore makes both biological and statistical sense to avoid analyzing differences in the apparent expression of genes that are not truly expressed. To compare two methods of filtering out such probe sets, we examined a dataset that compares 10 arrays from cells without interferon to 10 arrays from cells treated with interferon alpha (IFN dataset [16]). Of the 445,660 individual probe sets (20 arrays with 22,283 probe sets per array) in the IFN data set, 54% were called Absent by the MAS5 algorithm. There is a strong suggestion of bimodality in the distribution of log2(signal) for all probe sets (Fig. 1A), with a large number of Absent probe sets forming a shoulder at low signal values. Removing probe sets in which fewer than half of the arrays in at least one of the experimental groups was called Present eliminates the large shoulder of Absent probe sets, leaving a distribution that is more
