Supplemental Results Yeang CH*, Mak HC*, McCuine S, Workman C, Jaakola T, Ideker TI Genome Biology (2005) 6:R62   [fulltext] [PDF]  Computer Science and Artificial Intelligence Laboratory /     Laboratory for Integrative Network Biology MIT / UC San Diego  ·  Home Methods  ·  Experimental Protocols  ·  Internal Validation of Models  ·  Knockout Data Reproducibility ·  Regulated Gene Selection  ·  Building Physical Network Models ·  Protein-DNA Data ·  Protein-Protein Data ·  Knockout Data  ·  Model Inference  ·  Software Download  ·  Evaluating for New Experiments Data  ·  Inferred Network Models  ·  Download Network Model Data  ·  References Observations Of The Reproducibility Of Knockout Data We assessed the reproducibility of gene expression experiments by comparing swi4Δ and gcn4Δ data from Hughes at al. (Rosetta) to the same knockout data generated from our experimental settings. We measured of overlap of two data sets with several metrics, including the Pearson correlation coefficient between two expression profiles, the permutation p-value of the correlation coefficient, the rank correlation coefficient and its permutation p-value between the two expression profiles, and the hyper-geometric p-value of the overlap of up and down regulated genes. The rank correlation coefficient was computed by replacing each expression profile - a vector of expression log10 ratios - with a vector of their ranks. The permutation p-value of the correlation coefficient was computed by fixing one profile and randomly permuting the entries of the other profile. The hyper-geometric p-value evaluated the significance of overlap of significant changes between two data sets. Suppose N1 out of N genes are up or down regulated in Rosetta swi4Δ, N2 genes are up or down regulated in our new swi4Δ, and n genes are significantly changed in both data sets (in the same direction). The p-value calculated the probability that by randomly selecting N2 genes from a pool of N genes with N1 significant changes, ≥n selected genes had significant changes. N1 and N2 were then swapped to obtain a second p-value; the maximum of those two values was reported. Table S2: Correlations between swi4Δ and gcn4Δ data from Rosetta and the new experiments.   swi4Δ total gcn4Δ total correlation -0.013 0.34 correlation p-value 0.84 < 10-4 rank correlation -0.058 0.145 rank correlation p-value 1.0 < 10-4 hyper-geometric p-value 0.188 4.88×10-21         swi4Δ subset gcn4Δ subset correlation 0.344 0.831 correlation p-value 8×10-5 < 10-5 rank correlation 0.353 0.705 rank correlation p-value 9×10-5 < 10-5 hyper-geometric p-value 1.05×10-5 < 1.8×10-13 Thus, the two Swi4 profiles showed no correlation, while the Gcn4 profiles had significant correlations. Clearly, this lack of correlation for the swi4Δ experiments could be due to uncontrollable differences in our array procedure versus Rosetta's, such as the microarray fabrication process. However, we suspected comparing the expression profiles across the entire genome was also problematic, since a majority of genes were unaffected by a single deletion. Measuring the log10 ratios with respect to wild type in this case would be analogous to measuring the correlation of the background noise, leading to a low correlation between experiments. Since Gcn4 is a master regulator for many genes involved in amino acid synthesis, a gcn4Δ induced many changes which are not buried under random fluctuations, leading to a higher correlation than that of Swi4. We then restricted the expression profiles of each experiment to the subset of genes which are putatively regulated by Swi4 or Gcn4. For Swi4, we chose the genes which are directly bound by it according to the CHIP-chip data (Lee et al. 2002) or appeared in model #1, and for Gcn4 the genes which are bound by it and are significantly down regulated in the Rosetta gcn4Δ data; this list of selected gene sets is shown in Table S3, and the lower part of Table S2 shows the comparison results on the restricted gene sets. Both correlation coefficients and permutation p-values were much more significant in this restricted set compared to the genome-wide comparisons. Table S3: Restricted subsets used to evaluate the reproducibility. Swi4 subset AAP1 ATR1 BAT2 CCP1 CLB1 CLB2 CLB6 COX9 CPA2 CSI2 CUP9 ECM13 ECM33 ERG6 EXG1 HAP1 HAP4 HSC82 HTA1 HTB1 ISU2 LYS20 MNN1 MSN4 NCE102 PCL7 PDR12 PRY2 QCR2 RNR1 RPI1 RPL19B SAT2 SCW10 SGA1 SNQ2 SOK2 SRL1 SVS1 SWI4 TRP4 UTR2 YAP6 YBL029W YBL112C YBL113C YDR451C YDR545W YEL045C YEL047C YEL077C YER045C YER189W YER190W YFR006W YGR086C YGR153W YGR296W YHL049C YHR048W YIL056W YIL177C YJL217W YJL225C YJR078W YKL051W YLR084C YLR463C YLR465C YLR466W YLR467W YML133C YNL339C YNR067C YOL011W YOR248W YOR315W YOX1 YPL267W YPL283C YPR203W YPS4 Gcn4 subset ADH5 APG1 ARG1 ARG11 ARG3 ARG7 ARG8 ARO1 ARO3 CPA2 GCN4 HAD1 HAL2 HIS1 HIS4 HOM3 LEU4 MET16 PCL5 RIB5 TRP2 TRP3 UGA3 YER069W YER073W YGL184C YHM1 YHR122W YHR162W YJL200C YLR152C YMC1 YOL119C