The microarray data were preprocessed utilizing gcrma (robust multi-array averaging with sequence composition correction; http://www.bioconductor.org) in 4 steps: background correction, normalization between chips, signal estimation, and adjustments for false hybridization with GC content as a surrogate. Genes with low expression values were excluded from further analysis. Differential expression for the log base 2 normalized ethanol and air time courses was assessed by using linear models and empirical Bayes methods with the limma package in the R statistical program (Smyth, 2004). For each probe set, a quadratic regression model was fit to the expression values for time and treatment. We used the p value from the moderated F statistic (limma interaction p value) to test for time-dependent differential expression between the ethanol- and air-treated samples. The false discovery rate (fdr) was also determined from the p value adjusted for multiple comparisons utilizing the Benjamini and Hochberg method in limma. Because specific genes were later validated by quantitative PCR (qPCR) and mutant analysis (see below), we did not use the fdr for further analysis (fdr values are included in Table S1). Genes
