DNA microarrays allow researchers to measure the coexpression of thousands of genes, and identify changes across experimental condition. Recently, many studies have shifted from tabulating the effects of individual genes to the effects of groups of genes that share biological features. We define a general framework for gene category testing, and show that most existing methods can be presented as a contrast of the differential expression within a category to that of the complementary set of genes on the array. Our framework includes post hoc tests that look for overrepresentation of the category in a list of significant associations, and methods that consider quantitative measures of differential expression for all genes.

We divide existing gene category tests into two classes. Class 1 tests are most commonly used, and assume gene-specific measures of differential expression are independent, despite overwhelming evidence of positive correlation. We provide analytic results and simulations based on real microarray data to demonstrate that Class 1 tests are strongly anti-conservative. Class 2 tests use array permutation to account for correlation in expression, and by construction have proper Type I error control. We have previously introduced a general framework for Class 2 procedures called Significance Analysis of Function and Expression (SAFE). Both classes of tests assume or induce a null where all genes have the same degree of differential expression, which may not be biologically reasonable.

We introduce a more sensible and general (Class 3) null hypothesis which states that the profile of differential expression is the same within the category as for the entire array. Under this broader null, we show that Class 2 tests tend to be conservative. We present a bootstrap approach to test for departures from the Class 3 null, and use simulations and real microarray data to demonstrate that it provides valid Type I error control and more power than Class 2 tests. If time permits, we will discuss extensions of our testing approaches to groups of genes with shared transcription factor motifs.