Motivation: Microarrays are becoming an increasingly common tool for observing changes in gene expression over a large cross section of the genome. This experimental tool is particularly valuable for understanding the genome-wide changes in gene transcription in response to thiazolidinedione (TZD) treatment. The TZD class of drugs is known to improve insulin-sensitivity in diabetic patients, and is clinically used in treatment regimens. In cells, TZDs bind to and activate the transcriptional activity of peroxisome proliferator-activated receptor gamma (PPAR-γ). Large-scale array analyses will provide some insight into the mechanisms of TZD-mediated insulin sensitization. Unfortunately, a theoretical basis for analyzing array data has not kept pace with the rapid adoption of this tool. The methods that are commonly used, particularly the fold-change approach and the standard t-test, either lack statistical rigor or resort to generalized statistical models that do not accurately estimate variability at low replicate numbers.

Results: We introduce a statistical framework that models the dependence of measurement variance on the level of gene expression in the context of a Bayesian hierarchical model. We compare several methods of parameter estimation and subsequently apply these to determine a set of genes in 3T3-L1 adipocytes that are differentially regulated in response to TZD treatment. When the number of experimental replicates is low (n = 2–3), this approach appears to qualitatively preserve an equivalent degree of specificity, while vastly improving sensitivity over other comparable methods. In addition, the statistical framework developed here can be readily applied to understand the implicit assumptions made in traditional fold-change approaches to array analysis.

Availability: Our statistical approach is implemented as a set of Java-based tools called VAMPIRE, accessible from our website at http://genome.ucsd.edu/VAMPIRE