The genes MRPL19, PSMC4, SF3A1, PUM1, ACTB and GAPD were analyzed by real-time quantitative RT-PCR. Starting copy numbers for the six candidate housekeeping genes were measured across 80 primary breast tumor samples. The data are available as Additional data file 1 with the online version of this article. Plots of the raw and log-scaled expression levels (all logarithms in this paper are natural base (e) logarithms) are shown in Figure 1. The breast tumor samples are ordered according to the mean of the log-expression levels of all the genes. It is evident from the plot that for the raw data the variability of within-sample measurements increases with the mean expression, whereas the variability stays approximately the same for all the samples with the log-transformation. In addition, the log-transformation allows us to model fold changes in expression levels in an additive way.

To select the best housekeepers for normalizing data across a single tissue type, we tested three variations of a model (Model 1, a-c) with real-time quantitative RT-PCR data generated from primary breast samples (see Materials and methods for details).

Gene(s) with minimal variation in expression across different cell types serve as good 'universal' housekeepers. A universal control may be a single gene or combination of genes. While the former should display both low variability within a given tissue type and consistent basal levels of expression across tissue types, the latter may comprise a gene set with individually different, but complementary, basal expression levels across tissue types.

To test our models for selecting universal housekeepers, we used published data from Vandesompele et al. 13. They measured the expression level of 10 genes in neuroblastoma cell lines (NB), cultured normal fibroblasts (FIB), normal leukocytes (LEU) and cells from normal bone marrow (BM). In addition, normal tissues from pooled organs (breast, brain, fetal brain, heart, kidney, uterus, lung, trachea and small intestine) were also profiled. A plot of these housekeepers across the different tissues is shown in Figure 2. It is notable that a gene can have stable expression within a given tissue type but can change rank position compared to other housekeepers across tissues. For example, GAPD has relatively high expression in fibroblasts compared to other housekeepers but low expression in leukocytes. Thus, GAPD may be a good single housekeeper within certain tissue types but may not be an optimal universal housekeeper unless it is used within a complementary gene set.

Four candidate housekeepers (PSMC4, MRPL19, PUM1 and SF3A1) were selected from a microarray dataset containing 40 different breast tumors, three normal breast samples and 19 cell lines representing 17 different cell lines of diverse nature including lymphocytes, fibroblasts and epithelial cells 8. All experiments were done using a common reference strategy in which all experimental samples are compared to the same reference comprised of a pool of RNAs isolated from 11 diverse human cell lines 19.

To select housekeepers, we first 'filtered' the microarray data to select genes with Cy3 and Cy5 signal intensities greater than 500 units across at least 75% of the experiments. This requirement ensures that the gene is well expressed not only in the experimental samples, but also in the common reference sample. Next, we used the SAS/STAT Analysis Package Version 8 (SAS Institute Inc., Cary, NC) to identify a set of genes that showed a small range of expression across sample types and the least variance of the array-mean normalized log-ratios. For real-time RT-PCR, we selected four of the top six genes - PUM1, PSMC4, MRPL19 and SF3A. The two other low-variability genes identified in the data were IER3 (immediate early response 3) and SRY ((sex determining region Y)-box 2). We did not select these genes because of their potential for being differentially regulated under other conditions. However, we did include GAPD and ACTB, which are commonly used reference genes 20, in the set of candidate genes for comparison to the microarray selection.

Breast samples were acquired under informed consent and received at the Huntsman Cancer Institute (Salt Lake City, UT) for gene expression analysis (University of Utah, IRB #8533). All specimens were expediently processed in pathology upon arrival from surgery. Samples were grossly dissected, procured by flash freezing in liquid nitrogen, and stored at -80°C until RNA extraction. Approximately 50-100 mg cancer tissue was homogenized from each sample, and total RNA was prepared using the RNeasy midi kit (Qiagen). The integrity of RNA was determined using the RNA 6000 Nano LabChip kit (Agilent Technologies) and an Agilent 2100 Bioanalyzer. Two microliters of total RNA (50 ng/μl) were heated to 70°C and 1 μl was loaded on the column. Degradation was evaluated using the signal of the 18S and 28S ribosomal peaks 21.

First-strand cDNA was synthesized from 1 μg total RNA using oligo(dT) primers and Superscript III reverse transcriptase following manufacturer's instructions (Superscript III First-Strand Synthesis System, Invitrogen Life Technologies). Briefly, the reaction was held at 48°C for 50 min, followed by a 15 min step at 70°C. The cDNA was washed on QIAquick PCR purification column (Qiagen) and eluted in 2 × 50 μl of elution buffer. The cDNA was then diluted in TE' (10 mM Tris, 0.1 mM EDTA, pH 8.0), aliquoted and stored at -80°C for further use.

All PCR reactions were performed on the LightCycler. Each 20 μl reaction included 1 × PCR buffer with 3 mM MgCl₂ (Idaho Technology), 0.2 mM each of dATP, dCTP, and dGTP (Roche), 0.1 mM dTTP (Roche), 0.3 mM dUTP (Roche), 1 U of Platinum taq (Invitrogen Life Technologies), 1/40000 SYBR Green I (Molecular Probes), approximately 5 ng cDNA, and 0.4 μM of each primer. The primers used for the RNA control genes are shown in Table 5.

PCR was done using the following protocol: initial denaturation 95°C for 1 min 30 sec, then 50 cycles at 94°C for 1 sec for denaturation, 60°C for 5 sec (20°C/sec transition) for annealing, 72°C for 8 sec (2°C/sec transition) for extension. Fluorescence emission of SYBR Green I (channel 1, 530 nm) was acquired each cycle after the extension step. A melting step was performed after PCR to determine product purity. For melting curve analysis, the reactions were rapidly (20°C/sec) cooled from 95°C to 60°C and then slowly heated (0.1°C/sec) back to 95°C while continuously monitoring fluorescence.

Copy number was determined using the crossing point (Cp) value, which is automatically calculated using the LightCycler 3.5 software (Roche Molecular Biochemicals). The Cp value is reported as a fractional cycle number that is determined from the second derivative maximum (point of maximum acceleration) on the PCR amplification curve (fluorescence versus cycle number) 22. A relative starting copy number was determined for each housekeeper using a calibration curve done with the same batch of master mix. Efficiency (E) of PCR was calculated from a plot of Cp versus log ng cDNA 22.

E = 10^-1/slope

As the effects of interest are fold changes, we modeled the log-transformed expression Model 1a.

log y_ij = μ + T_i + G_j + ε_ij,

where independent

where μ denotes the overall mean (log) expression, T_i is the difference of the ith tissue sample from the overall average and G_j is the difference of the jth gene from the overall average. The key feature of this model that makes it different from a traditional ANOVA model is that it allows heteroscedastic errors: the variability of the genes is different.

We fitted the model using the gls routine of the nlme library for R, however other commonly available software such as PROC MIXED from SAS could have been used.

Based on the model, the variability of the logarithm of the geometric mean

of a gene-set S was estimated as

Vandesompele et al.'s M-value is the average of relative standard deviations of the log-expression levels. Under Model 1, the M-value of the gene is closely related to its variance (under Models 2 and 3 below, the similar relationships can be derived):

We tested the assumption of unequal variances by fitting Model 1b that forces all the genes to have the same variability (this is the classical ANOVA model).

log y_ij = μ + T_i + G_j + ε_ij,

where independent

Model 1c with a correlated error structure can be used to assess the assumption of (conditional) independence of the genes given the sample mean. If warranted, a more complicated correlation structure can be imposed.

log y_ij = μ + T_i + G_j + ε_ij,

where

For the multiple tissue-type set-up the notation and the model need to be extended. We will denote the expression level of gene j of in the ith sample of type k by y_i(k)j, i = 1,...n_k, j = 1,...,g and k = 1,...,m. The best-fitting model for the data, which we call Model 2, had the form

log y_i(k)j = μ + C_k + T_i(k) + G_j + (CG)_kj + ε_i(k)j,

where

Thus the errors are independent and their variability is decomposed into a gene-specific and tissue-type specific multiplicative components. The last restriction ensures the uniqueness of the solution. Simpler models that we considered used uniform error variance, equal error variance for tissue types, and equal error variance for genes. We also considered more complex models that used exchangeable correlation structure for the errors and unstructured error variance (each gene-tissue-type combination has a variance parameter). The BIC was used as a basis for model selection.