splitplotnestedANOVA

Determination of significant differential expression was undertaken with analysis of variance (ANOVA) style models. From the ideas provided by Kerr and Churchill (2001) and Wolfinger, et al. (2001), and based on the specific design of this experiment, the following ANOVA models were constructed: Model (1) y_gijklr = \mu + L_i + A(L)_j(i) + T_k + LT_ik + D_l + AD_jl + \epsilon_gijklr Model (2) r_gijklr = GL_gi + GA(L)_gj(i) + GT_gk + GLT_gik + GD_gl + GS(A)_gr(j) + \epsilon_gijklr In model (1), y is the logarithm (base 2) of the intensity for a particular spot. L, T and D are global effects due to differences in lines (Mo17 and B73), treatments (control and UV exposure) and dyes, respectively, while LT represents interaction between lines and treatments. The AD interaction term is present as suggested by Wolfinger et al. to account for intensity scaling done on the two channels. The nested term, A(L), accounts for variation across replicate arrays in a slightly different fashion than that set forth in Kerr and Churchill's ANOVA models. Their approach considered connected designs; however, in this experiment, lines are not connected because no array has samples from two different lines hybridized to it. The output from model (1) is given in the file "standardized".Model (2) takes residuals from model (1) as normalized response values and includes gene specific effects. Additionally, it models replicate spot variation via the GS(A) term. Both models were fit using the SAS/STAT® software's MIXED procedure, with model (2) being fit gene by gene. Observed significance levels for all effects tests were adjusted for multiplicity of testing by use of the Sidak method. Significance was assessed based on an experiment-wise significance level of 0.05. The levels of expression of the SP10 negative controls were used to determine the log-ratio level for expression above background. All cDNAs with all levels below the SP10 median were set to background, thus diminishing their influence on the final list of genes. Ten genes, that may have been otherwise called significant, were removed from the 5366 total using this criterion. There are still 201 genes with at least one very low level; however, these were retained for significance testing as some may have significant expression level alterations in one line or condition. There were six cDNAs with significant differences between lines but not interaction or treatment; these genes were not analyzed further. A table of summary information for all genes is included as Table S1. Cluster analysis on gene expression levels was performed using the SAS v. 8 Enterprise Miner software Cluster node, using the default settings except for specifying input cluster number as equal to the total sample number. Clustering was performed separately on genes significant by interaction (Table S2) and genes significant by treatment (Table S3); the output cluster number is listed for each gene name in these tables. The entire model (1) output file can be obtained by within the citation below.