Finding Significantly Differentially Expressed Genes

Scientists are interested in how gene expression levels may vary under different treatments. They want to know which genes are expressed differently under treatment conditions and control conditions. Gene expression microarray data are modeled to find significant changes in gene expression levels between a control and treatment tissue co-hybridized on one slide. The model used is the Bayesian hierarchical model proposed by Newton et al. (2001). This Gamma-Gamma-Bernoulli model leads to a criterion for classifying genes as significantly differentially expressed based on the posterior odds of any real differential expression. It also generates an estimate of the proportion of genes in the sample that are differentially expressed. Several data sets of open frames from E. coli bacteria are examined. The variability and accuracy of Newton et al.’s classification method are explored through simulation from the posterior predictive distributions of the model. The specificity of the model for this data is examined over a range of possible true expression levels. Also, a whole data set is simulated to compare Newton et al.’s method for determining significant differential expression with the currently popular method. 1. OVERVIEW OF THE NEWTON