Immunolocalization and transcriptome data summary

To determine correlations between disparate methods of assessing gene expression levels, we selected a cohort of CD antigens as a reference gene set and compared transcript abundance measurements acquired by microarray hybridization with protein abundance measurements determined by immunohistochemistry in specific cell types found in the human prostate gland. The microarray and immunostaining results for each cell type are summarized in Figure 1. All annotations are deposited in a public database [14]. The sample data conform to the Minimum Information Specification For In Situ Hybridization and Immunohistochemistry Experiments (MISFISHIE) standards [15], and describe the tissue, distribution of reaction product in the tissue, localization pattern within cell types, and provide an assessment of the level of protein expression based on the immunostaining data. Our initial criteria for assessing expression centered on using 'present' vs. 'absent' call for both Affymetrix and Agilent array analyses. Based on this qualification, CD antigen expression for MACS-sorted cell populations agreed with the following frequency to IHC data by cell type: endothelial = 55%, luminal = 64%, stromal = 46%, basal epithelial = 63%, stem = 64%, cancer = 68%. CD antigen expression for LCM cells analyzed by Affymetrix array agreed 39% for stromal cells with IHC data, and analyzed by Agilent array agreed 25% for stromal cells and 36% for luminal cells.

We next sought to establish a more quantitative relationship between CD immunoreactivity of prostate cells and corresponding gene expression data (Figure 2). For example, staining for CD26 (DPP4) in normal (A) and cancer (B) prostate tissue shows clear luminal staining, which corresponds with intense gene expression measured by array in the luminal and cancer (05-179_CD26t) cell populations. The greyscale gradient indicates robust multi-array average (RMA) normalized Affymetrix signal intensity. Signals of 10 or less are represented as white and signals greater than or equal to 10,000 are represented as black. Higher Affymetrix signal (more black) indicates higher levels of gene expression. CD26 expression determined by Agilent array in LCM luminal cells was also in concordance (data shown in Figure 1). Immunostaining with CD56 (NCAM1) showed clear stromal staining and corresponding gene expression levels in both the MACS-sorted and LCM stromal cell populations. Agilent array analysis results for CD56 were 'absent' in luminal and 'present' in stromal cells as well (data shown in Figure 1).

Data type discordance is typified by CD44 (Figure 3). Immunostaining with CD44 (G44-26, BD PharMingen, San Diego, CA), was confined to the basal cells (A), however, measured gene expression was high in all cell types (B). This antibody recognizes the epitope 1 of the standard variant of CD44 antigen. Affymetrix has 11 probe sets for CD44, showing variable expression profiles, of which only probe set 217523_at showed agreement with IHC results. CD44 has a wide range of splice variants [16] which may account for the discrepancy. Although microarray experiments detect mRNA, and it is mRNA that has the direct relationship with protein, the methodology and algorithms for Affymetrix data analysis are gene-based. Gene level annotation fails to discriminate multiple mRNAs transcribed from the same gene [17] such as is the case with CD44 [18]. Affymetrix probesets for CD44 were analyzed with Basic Local Alignment Search Tool (BLAST). Probesets 204489_s_at, 204490_s_at, 209835_x_at, and 212063_at map directly to multiple variants of CD44, whereas the other probesets map to individual variants. It is likely that gene level annotation compared to protein immunostaining will be less straightforward in cases such as CD44, where multiple protein isoforms are generated from a single gene.

Further examples of data discordance for CD64 (FCGR1A), CD6 and CD24 expression are shown in Figure 4. For CD64, immunostaining identified the luminal cell population, but measured gene expression was low or absent for all cell types regardless of probeset (A). This was also the case for CD6 (B). CD6 shows equivocal staining of luminal cells. Examination of CD24 immunostaining identifies the cancer cells, however measured gene expression was high across all cell types for some probe sets (i.e. 209771_x_at) and absent for others (i.e. 1560395_at). Gene expression in LCM samples was also discordant (data shown in Figure 1).

Correlation of immunolocalization with array data

LCM confirmation

Prostate tissue

The methods of tissue collection, immunostaining and expression data collection and analysis used in this study have been published previously [13, 20, 35]. Briefly, tissue samples for IHC and Affymetrix array analysis consisted of both cancer-enriched and cancer-free samples obtained from 55 radical prostatectomies. Upon receipt of prostatectomy specimens, 3-mm thick transverse sections were made of the prostate after inking the exterior surface (the surgical margin). The same approach was used for both cancer-free and cancer-enriched (where at least 85% of the cells in the corresponding frozen section were cancer cells) samples. Between 4 and 6 blocks of tissue from the posterior aspect (the peripheral zone) of each prostate were frozen. Sections from the frozen blocks of tissue provided a histological template for characterizing the cell composition of the adjacent non-frozen tissue blocks. Cancer-free samples, weighing between 2 and 10 g, were harvested primarily from the anterior aspect of the prostate (transition zone) as described [12, 13, 35]. Cancer-enriched samples, weighing at least 100 mg, were dissected from the opposing aspect of the non-fixed section adjacent to the block of tissue that had been frozen. Tissue specimens for cell sorting were digested by collagenase [36, 37].

As an external control, IHC and array data from MACS-sorted cells were compared to LCM cells from prostate stromal and luminal cells analyzed by hybridization to long-oligonucleotide microarrays (Agilent Corporation). Additionally, LCM stromal cells were analyzed by Affymetrix array to control for potential platform variation. Tissue specimens for LCM were obtained from prostate needle biopsies from 14 men identified from a biospecimen repository with no evidence of cancer as well as a second or third set of biopsies that also showed no evidence of carcinoma. Each biopsy specimen was immediately embedded in OCT and flash frozen in isopentane immersed in liquid nitrogen. All protocols for tissue acquisition and processing were under approval by the University of Washington Institutional Review Board following a standard protocol.

Immunohistochemistry

Blocks of unfixed prostate tissue were harvested at surgery. Serial 5-μm sections were prepared from randomly selected frozen blocks, fixed in cold acetone, and processed for immunohistochemistry. We used frozen sections since a majority of the anti-CD antibodies we used do not immunoreact with antigens in formalin-fixed, paraffin-embedded tissue. Immunohistochemistry was performed as described previously [20]. The primary antibodies used were mouse monoclonal CD antibodies (BD-PharMingen, San Diego, CA) diluted to 8 ng/μl or less. Antigen was localized using biotinylated anti-mouse IgG or IgM (Vector Labs, Burlingame, CA) as the secondary antibody and diaminobenzidine tetrahydrochloride as the chromogen. The sections were counterstained with hematoxylin.

The percentage of cells of a specific histological phenotype that expressed the antigen was estimated in five randomly selected fields at a final magnification of 40× (ocular 10×; objective 4×). Staining intensity was evaluated by two parameters (staining intensity and percentage of cells exhibiting each level of intensity). Intensity of reaction product was based on a 3-point scale – none, faint/equivocal and intense. These categories of staining intensity were defined as follows. Intense: immunoreaction deposit is distinctly more optically dense than background and than tissue that does not express the antigen. Equivocal: immunoreaction deposit is either similar enough in optical density to the background and/or to tissue that does not express the antigen, or is so focal, i.e. < 5% of cells, that there is reasonable uncertainty regarding whether the cells express the antigen. None: there is either no immunoreaction deposit or reaction product is no more optically dense than background.

WF = weighting factor, usually 0.2

The value of the weighting factor can be changed on our web site by independent observers of the images. We did this since we thought that observers of the images should be able to alter the value assigned to the weighting factor based on whether they thought that faint/equivocal immunostained cells either did or did not truly represent immunolocalization of each respective CD antigen. For the present study we used a weighting factor of 0.2.

where:

n = number of immunostains for each CD by cell type

A = staining value for each immunostain by cell type

Average scores in the range of 0.751 to 1.0 were categorized as intense, 0.251 to 0.750 as equivocal or 0 to 0.250 as none. Immunostained sections were imaged with an Olympus BX41 microscope (Olympus, Melville, NY) equipped with a MicroFire digital camera (Optronics, Goleta, CA). Composite images were constructed with Photoshop CS (Adobe Systems, San Jose, CA) and are publicly available online [38].

Affymetrix gene expression profiling of MACS-sorted cell populations and laser captured stromal cells

Transcriptomes of the following prostate cell types have been previously determined: CD104⁺ basal epithelial cells, CD26⁺ luminal epithelial cells, CD49a⁺ stromal fibromuscular cells, and CD31⁺ endothelial cells [13], ABCG2⁺ stem cells [39] and CD26⁺ cancer cells [in preparation]. The CD molecules cited above were the ones used for sorting each specific cell type for gene expression profiling. The expression of genes documented in the literature as cell-type specific was compared to Affymetrix datasets; analyses showed that the transcriptomes were representative of their respective cell type (p < 0.05) [13]. Five separate biological replicates of each MACS-sorted cell populations (except the CD26⁺ cancer cells which comprised of one biological sample) and three separate biological replicates of laser captured stromal cells were assayed to produce a dataset using the Human Genome U133 Plus 2.0 GeneChips (Affymetrix, Santa Clara, CA). The U-133 Plus 2.0 array contains probesets representing 54,675 genes, splice variants, and ESTs. The GeneChips were prepared, hybridized, and scanned according to the protocols provided by Affymetrix. Briefly, 200 ng of RNA was reverse transcribed with poly (dT) primer containing a T7 promoter, and the cDNA was made double-stranded. In vitro transcription was performed to produce unlabeled cRNA. Next, first-strand cDNA was produced with random primer. cDNA was made double-stranded with poly (dT) primer/T7 promoter. Finally, in vitro transcription was performed with biotinylated ribonucleotides. The biotin-labeled cRNA was hybridized to the GeneChips. The chips were washed and stained with streptavidin-PE using an Affymetrix FS-450 fluidics station. Data was collected with an Affymetrix GeneChip Scanner 3000.

LCM and RNA isolation of luminal and stromal cells for Agilent array analysis

Approximately two thousand luminal epithelial or stromal cells from histologically benign glands were captured using the Veritas Microdissection System (Molecular Devices Corporation, Sunnyvale, CA) as previously described [40]. Frozen sections (8 μm) were cut from tissue embedded in OCT (Miles, Inc. Diagonostic Division, Elkhart, IN) and fixed in cold 95% ethanol. After brief (5–10 seconds) staining with hematoxylin using the HistoGene staining solution (Arcturus Engineering, Mountain View, CA), the sections were dehydrated in 100% ethanol, followed by sequential ethanol-xylene solutions. Three stromal cell samples and 14 luminal cell samples were captured. Digital photographs were taken of tissue sections before, during, and after LCM and assessed independently by two investigators to confirm that the laser-captured cells were stromal and luminal cells, respectively.

Captured individual specimens were lysed and RNA was isolated using Arcturus PicoPure RNA Isolation. Samples were treated with RNase-free DNase (Qiagen, Inc., Valencia, CA). This RNA was converted to a cDNA library and amplified using the TransPlex WTA (Rubicon Genomics, Ann Arbor, MI) using Titanium Taq polymerase (Clontech Laboratories, Inc., Mountain View, CA) in the presence of amino-allyl deoxyuridine triphosphate (dUTP) for postamplification labeling. Real-time PCR amplifications were terminated at plateau phase, as measured by fluorescence incorporation, to preserve maximum representation. Concentration and purity was determined using the ND-1000 spectrophotometer (Nanodrop Technologies, Wilmington, DE). To provide a reference standard RNA for use on 2-color microarrays, we isolated total RNA from LNCaP, DU145, PC3, and CWR22 cell lines (American Type Culture Collection, Manassas, VA) growing at log phase in dye-free RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS; Life Technologies, Rockville, MD). RNA was purified using Trizol (Life Technologies). The RNA was then further purified by Qiagen RNeasy maxi and treated with Qiagen RNase-free DNase. This total RNA was then amplified using the TransPlex WTA similar to the experimental samples.

Gene expression profiling on Agilent microarrays

Two μg of WTA-amplified amino-allyl cDNA from each of 14 luminal and 3 stromal samples were labeled with Cy3 fluorescent dye (cell line reference cDNA was labeled with Cy5) as described previously [40], and hybridized to Agilent 44 K whole human genome expression oligonucleotide microarray slides (Agilent Technologies, Inc., Santa Clara, CA) following the manufacturer's suggested protocols. Fluorescence array images were collected for both Cy3 and Cy5 using an Agilent fluorescent scanner, and Feature Extraction software (Agilent Technologies, Santa Clara, CA, USA) was used to grid, extract image intensities, and normalize data. Spots of poor quality, as determined by the software, were removed from further analysis. Data from the 14 luminal or 3 stromal samples were combined and the average signal values were used for comparative analyses. For genes with multiple probes on the array, the data from the probe with the highest sequence specificity was chosen.

Data comparison and statistical analysis

Immunohistochemistry data was compared to Affymetrix array data for the following prostate cell types: MACS-sorted basal cells, luminal cells, stromal fibromuscular cells, endothelial cells, stem cells and cancer cells from a primary tumor of Gleason 3+3 = 6. Additionally IHC data was compared to Affymetrix array data for LCM stromal and Agilent array data for LCM stromal and epithelial cells. Data was analyzed for concordance and correlation of expression levels. Data concordance was defined as agreement between data types based on 'present' or 'absent' calls, i.e., if both techniques detected a gene, they were considered to be in agreement. Genes in the Affymetrix array dataset with an average raw fluorescence signal < 50 were considered to be undetected by the experiment and were classified as 'absent' by GCOS version 1.0 software (Affymetrix) [13]. Genes were also classified as undetected if the majority (> 50%) of individual probe sets were classified as 'absent'. All other hybridization patterns were classified as 'present'. Genes with signal 51–500 were considered to have 'low' expression, those with signal 501–5000 were considered to have 'medium' expression, and those with signal > 5000 were considered to have 'high' expression. Genes in the Agilent array dataset with an average raw fluorescence signal < 100 were considered to be 'absent', those with signal 100–999 were considered to have 'low' expression, those with signal 1000–9999 were considered to have 'medium' expression, and those with signal > 9999 were considered to have 'high' expression. Immunohistochemistry data having a staining level of intense (score range 0.751 – 1.0) or equivocal (score range 0.251 – 0.750) were classified as in agreement with genes classified as 'present'. Immunostaining results of none (score range 0 – 0.250) were considered equivalent to 'absent'.

The correlation between mRNA expression levels and immunostaining intensity for each CD antibody was described by Pearson product-moment and Spearman nonparametric correlation measures and tests. Correlations were calculated for mean and median values. Median was included as a more robust measurement in the presence of outlier or skewed values especially for the array data. Individual Affymetrix probes within a probe set were originally designated to hybridize with the same unique mRNA transcript. However, increasing knowledge of genomic sequences has shown that a substantial number of the manufacturer's original probe groupings and mappings are inaccurate and must be corrected [41]. To address the possibility that disparity in expression level was due to probe set definition, correlation was additionally determined using updated custom CDF definitions [19]. Dai et al., have reorganized probe set definitions based on the latest genome and transcriptome information [17]. These updated definitions based on Version 8 Entrez Gene mappings were downloaded and used in addition to the Affymetrix probe set definitions. To control for possible gene expression changes induced by the cell sorting process, LCM data for stromal cells were also compared for correlation. Additionally, LCM data for luminal and stromal cells assayed by Agilent array analysis was also used for correlation calculations. Correlations were calculated using both original and updated probe set definitions and were categorized as follows: no positive correlation (0.00 to 0.30), moderate positive correlation (0.31 to 0.79), and strong positive correlation (0.80 to 0.99).