Gene expression profiles of prostate cancer reveal involvement of multiple molecular pathways in the metastatic process

All tissue samples were acquired from the Health Sciences Tissue Bank of the University of Pittsburgh Medical Center under stringent Institutional Review Board guidelines with appropriate informed consent. The 18 donor and 64 primary prostate tumor samples have been described previously [7]. Specimens were received directly from the operating room. Samples (>500 mg) were excised and snap frozen in liquid nitrogen within 30 min of excision and stored at -80°C until extraction of RNA. Metastatic tumor samples were obtained from a warm autopsy program and processed similarly to primary tumors. An H&E stained frozen section of each sample was evaluated by a pathologist, to determine epithelial and stromal content and verify the presence of tumor in the sample. Dissection of the frozen tissue block was performed with the guidance of a marked H & E slide to minimize the presence of host tissue in the metastatic samples. All samples used in the study contained >80% tumor. Metastatic tumor samples were minced and divided into two equal portions to be extracted with the sample protocol used for each set of primary tumors.

The clinical characteristics of the 64 primary tumor samples used in the Affymetrix portion of our study have been previously described [6, 7]. These cases have a mean follow-up time of 3 years. The metastatic samples consisted of 24 tissues derived from 4 patients (Table 1). All patients with metastatic disease had received androgen ablation therapy and had shown progression of disease while on androgen ablation. The clinical characteristics for the additional 10 primary prostate tumor cases used in the CodeLink study are shown in Table 1.

RNA purification for the 64 primary samples has been previously described [6]. The set of metastatic samples analyzed with the Affymetrix platform was extracted with the same methodology. The set of metastatic samples and primary tumors analyzed with the CodeLink platform were extracted using the RNeasy kit (Qiagen, San Diego, CA). For the metastatic samples, one sample did not have enough for extraction with the Qiagen method, only 23 metastatic samples are included in the CodeLink assays. The concentration of each total RNA sample was measured with a Nanodrop ND-1000 spectrophotometer (Nanodrop Technologies,Wilmington, DE). RNA integrity was determined by capillary electrophoresis using an Agilent 2100 Bioanalyzer (Agilent, Willmington, DE).

cRNA was prepared and hybridized to Affymetrix GeneChip HGU95av2, HGU95b and HGU95c arrays (Affymetrix, Santa Clara, CA) as previously described [6]. For gene expression profiling with the CodeLink Gene Expression System (GE Healthcare, Piscataway, NJ), biotin-labeled cRNA was prepared as previously described [16]. Ten micrograms of biotin-labeled cRNA product from each sample were then fragmented with RNA fragmentation buffer at 94°C for 20 minutes. Hybridization mix was prepared according to the manufacturer's instructions and the final volume was adjusted to 260ul using nuclease-free water. The hybridization mix was heat denatured at 90°C for 5 minutes, cooled on ice and then applied to Human Uniset 20 K arrays (GE Healthcare, Piscataway, NJ). Arrays were incubated at 37°C for 18 h with shaking at 60 rpm in an Innova hybridization oven (New Brunswick, Edison, PA).

After hybridization, arrays were placed in a pre-heated (46°C) chamber filled with 0.75 × TNT (0.75 M Tris-HCL, pH 7.6, 3.75 M NaCl, Tween-20, and milli-Q water) and incubated at 46°C for 1 hour. Arrays were then stained with Streptavidin-Alexa Fluor 647 (Molecular Probes, Grand Island, NY) for 30 minutes at room temperature. Upon the completion of staining, the arrays were washed three sequential times in fresh 1 × TNT (1 M Tris-HCL, pH 7.6, 5 M NaCl, Tween-20, and milli-Q water) and then washed two final times in fresh solutions of 0.05% Tween-20 and 0.1 × SSC with gentle agitation. All arrays were dried by centrifugation at 2000 rpm for 3 minutes.

Affymetrix arrays were scanned in an Affymetrix GCS3000 Scanner (Affymetrix, Santa Clara, CA). CodeLink arrays were scanned with the GenePix 4000B scanner using GenePix Pro 4.1 software (Molecular Devices, Sunnyvale, CA).

The raw scanned array images from the Affymetrix GeneChip U95 arrays were processed using GCOS 1.1 software using the MAS5 algorithm (Affymetrix Corporation, Santa Clara, Ca) to generate probe cel intensity (*.cel) files. Data normalization to remove variation in overall chip intensities was performed by global scaling to a chip mean target intensity of 200 (MAS 5.0). Data for U95Av2, B and C arrays were combined for further analyses.

To identify differentially regulated genes in both datasets, these were analyzed with the Significance Analysis of Microarrays software (SAM v 1.2) [17]. Prior to analysis, genes that showed low variation across all samples were removed by using the filtering option in the Avadis 3.3 Pride Software (Strand Life Sciences, Bangalore, India) data analysis tool. To avoid false results due to difference in the tissue composition of metastatic and primary tumors, genes identified as being highly expressed in the prostatic stroma as per Stuart et al [15] were also removed. In all 1506 stromal genes and 7678 invariant genes were removed from the Affymetrix dataset. SAM generated gene lists with the lowest false discovery rates (FDR) were further analyzed for gene ontology (GO) and pathway annotations using NIH's DAVID annotation tool [18].

For CodeLink arrays, image files were analyzed with the CodeLink Expression Analysis Software version 4.1 (GE HealthCare) with use of the normalized intensity values in downstream analysis. For cross-platform comparison, Affymetrix probe sets and Codelink identifiers were mapped to Unigene ids using the DAVID annotation tool (see above). Expression data from both platforms was compared using z-transformation. Hierarchical clustering was performed using Eisen's Cluster and Treeview [19]. Data from Affymetrix experiments has been submitted to NCBI's Gene Expression Omnibus (GEO) as series GSE6919, with the following accession numbers GSE6604 (normal donor prostate), GSE6605 (metastatic prostate tumors), GSE6606 (primary prostate tumors) and GSE6608 (normal prostate tissue adjacent to tumor). Data from the CodeLink platform have been submitted to GEO with the accession number GSE6752 (primary and metastatic prostate tumors).

Differential expression of ten genes in primary and metastatic prostate cancer samples was verified with quantitative real-time PCR (QPCR) with the ABI PRISM^® 7000 sequence detection system (Applied Biosystems, Foster City, CA). Three selected RNA samples from each patient were pooled together (except for patient FB666 n = 1) and therefore four RNA samples, each representing one patient, were tested. RNA samples were first heat-denatured at 70°C for 10 minutes in an Eppendorf master cycler (Eppendorf, Westbury, NY) and then chilled immediately on ice. cDNAs were reversely transcribed from one microgram of RNA using the M-MLV Reverse Transcriptase kit (Invitrogen, Carlsbad, CA), as recommended by the manufacturer. QPCR was performed based on the manufacturer's instructions with TaqMan Gene Expression Assays (Applied Biosystems) for the following genes: EGR3, SYNPO2, ANGPT2, SPP1, FOXM1, ADM, RDX, TGFBRAP1, MAK and EGR1(assay IDs: Hs00231780_m1, Hs00326493_m1, Hs01048047_mH, Hs00959010_m1, Hs00153543_m1, Hs00969450_g1, Hs00988414_g1; Hs01093285_m1; Hs01048300_m1, Hs00152928_m1). When multiple TaqMan assays for one gene were available, the assay that interrogated the sequence closest to the target sequence in the Affymetrix arrays was chosen. PCR cycles were performed according to the assay instructions in an ABI PRISM^® 7000 Sequence Detection System (Applied Biosystems). Relative quantification of the expression level of each transcript in each sample was calculated using the Delta-Delta CT method in the ABI PRISM 7000 Sequence Detection System Software (Applied Biosystems) [20]. Human reference RNA from Stratagene (Stratagene Corp., La Jolla, CA) was used as the calibrator (untreated control) and human glucuronidase beta (GUSB) gene was used as the endogenous reference gene (Forward primer: GGA ATT TTG CCG ATT TCA TGA; Reverse primer: CCG AGT GAA GAT CCC CTT TTT; Probe: 6FAM-AAC AGT CAC CGA CGA GAG TGC TGG G-TAMRA).

Differential expression analysis of the metastatic and primary tumor samples shows that a large number of the most highly downregulated genes such as TAGLN, ACTG2, TPM1, MYH111 and DES have been previously identified as expressed mostly in the prostatic stromal cells [15]. Since only the epithelial component of prostate cancer is present in metastatic tumors, this result most likely reflects the lack of stroma in metastases, and not a true down-regulation of these genes in the metastatic epithelial cells. Therefore, based on a recent report characterizing cell type specific gene expression in the prostate [15], we removed the set of genes expressed mainly by the stromal cells of the primary tumors. In all 1506 transcripts associated with a stromal signature were deleted prior to further analysis. Since the stromal genes were characterized using the U95Av2 chip and our analysis includes u95Av2, B and C chips, only stromal genes represented by probe sets on U95Av2 were removed in this modified analysis. SAM analysis shows that 1277 genes are up and 977 genes are downregulated at least 2 fold at the lowest FDR (0.01), in metastatic prostate samples (see Additional file 1). A list of the top 50 up and top 50 down regulated genes at the lowest fdr, after removing ESTs and uncharacterized clones is shown in Table 2. This list includes signal transducers, cell cycle regulators, metabolic enzymes and cell adhesion molecules. Some of the most upregulated genes in our list are EIF1AX, AR, HSPD1 and HSPCA, K-ALPHA1, MLL5, UGT2B15, and some of the most downregulated genes include WNT5B5, ANXA11, FOS and SFRP1.

Using immunohistochemistry (IHC) Shah et al. have shown that metastatic samples are highly heterogenous in expression of prostate specific markers leading to the hypothesis that at the molecular level, metastatic prostate cancer may represent multiple diseases even within the same patient [21]. We examined the expression of several transcripts markers including some studied by Shah et al. and confirmed the heterogeneity of expression levels in metastatic prostate cancer tissues. Expression values in donor samples, primary and metastatic samples were compared. Prostate specific antigen (PSA/KLK3) remains high in some metastatic samples and is low or absent in others, even within the same patient (Figure 1). Interestingly, AMACR, another biomarker for prostate cancer [22] expresses a heterogenous expression pattern similar to PSA. HPN, which is overexpressed in primary cancer maintains high expression in the metastatic samples in our study. AR, while overexpressed in 23 out of the 24 metastatic samples, shows highly variable expression values in individual samples. The proto-oncogenes FOS and JUNB, which are both overexpressed in primary tumors, are consistently downregulated in all metastatic samples.

Hierarchical clustering analysis reveals that gene expression in metastatic samples is more variable between patients than between different metastatic sites from each patient (Figure 2). Although the 24 metastatic samples represent tissues from 6 metastatic sites (Table 1), no organ specific clusters were detected (Figure 2) whereas samples from the same patient tend to cluster together. Statistical comparison of organ-specific expression profiles was not attempted due to unequal distribution of samples from different metastatic sites.

In order to identify probe sets that are similarly regulated in every patient, and therefore likely to represent a specific metastatic profile, the SAM differentially expressed gene list at a FDR of 2% was further filtered. For each gene on this list, a patient specific median expression value was calculated from the multiple samples from each patient. Patient P4 had only sample and this sample's signal value was considered the median value. The median values were then compared to the median value of the primary samples and those probe sets whose median value showed equal or more than a 2 fold change in every patient were considered part of the metastatic prostate cancer signature. Under this criteria 415 transcripts are upregulated and 364 are downregulated in all patients with metastasis (see Additional file 2). A truncated gene list consisting of genes regulated at least 3 fold in all patients is shown in Table 3. Upregulation of AR in all samples from metastatic cancer patients represents a known "androgen resistant" or AARPC (androgen ablation resistant prostate cancer) phenotype [12]. The transcripts identified as differentially expressed in our study exhibit similarities with a previous study of AARPC tumors [10, 12]. Cytokeratins 5 and 15 (KRT5/KRT15), markers of basal cells in prostate glands, show uniform downregulation in all metastatic tumors, confirming the absence of basal epithelial cells.

The list of differentially expressed transcripts at least 2 fold in all patients was further analyzed for biological themes and gene ontology (GO) using the NIH's DAVID annotation tool. This analysis revealed that metastatic prostate cancer exhibits altered regulation of amino acid, carbohydrate and nucleotide metabolism consistent with the proliferative capacity and altered energy needs of metastatic tumors (data not shown). In the context of prostate cancer biology, genes involved in cell-adhesion, bone remodeling, cell-cycle and transcription are of particular interest (Table 4). Disruption of cell adhesion and altered interaction with the extracellular matrix is a hallmark of metastatic tumors [3]. In agreement with this, the secreted phosphoprotein and cell adhesion molecule osteopontin (SPP1) is one of the most highly upregulated transcripts in our metastatic samples. Elevated expression of SPP1 has been correlated with poor prognosis in prostate tumors and other cancers and it has often been implicated in metastasis to bone and other organs [2, 23‐28]. In all 29 probe sets representing cell adhesion genes are altered in all metastatic samples. This gene list includes FN1, ITGB8, THBS2, HNT and CDH10. Genes involved in bone remodeling such as BMP4 and ANKH are also altered in expression, although none of the samples in our study are bone metastatic samples suggesting that these proteins may also be involved in cancer metastasis to other organs.

Disruption of the cell cycle is highlighted by the presence of a large number of cell-cycle related transcripts in the list of differentially expresed genes in all metastatic samples. The list contains 37 cell cycle genes, and includes SEP4, SEP7, PTN and VEGF. Similarly, a large number of transcription factors (67) including AR, SRY, FOS and EGR3, are differentially expressed. Two members of the winged-helix family of transcription factors, FoxP1 and FoxM1, show upregulation in the metastatic samples. Interestingly, FoxM1b has been shown to promote progression of prostate carcinomas in an experimental model [29].

The MAP kinase signaling pathway was also identified as being important in the metastatic process, with 26 probe sets involved in this pathway being differentially expressed in all metastatic samples. The regulated genes include DUSP1, DUSP2, DUSP8, MAP3K8, MAP4K4, FGF13, FGFR2 and FOS. Involvement of MAP kinase in androgen receptor signaling has been previously described [30].

Gene expression analysis with the CodeLink Uniset 20 K microarray was carried out for 23 of the metastatic samples and compared to an independent set of 10 primary tumors. Similar to the Affymetrix analysis (see above), hierarchical clustering of the CodeLink data set reveals heterogeneity in expression and no organ-specific clustering (data not shown). Comparison of results with the Affymetrix based dataset, based on genes with common Unigene ids on both platforms, show a similar pattern of differentially expressed genes. Of the top 1000 up and down regulated transcripts from each platform, approximately 70% share common unigene ids and of these 22% of the genes are identified as regulated by both platforms (see Additional file 3). This level of correlation is significant, given the well-documented difficulties in cross-platform comparisons of expression data [31, 32]. Examples of z-transformed expression values for selected genes in both platforms are shown in Figure 3.

Additionally, real-time quantitative PCR (QPCR) assays were performed for a selected set of genes in pooled samples for each patient with metastatic disease and 5 of the primary tumors from the CodeLink set. The transcripts for this analysis were chosen to represent diverse biological processes and were chosen from the differentially expressed genes identified as up/down-regulated in the Affymetrix/Codelink data comparison. As shown in Figure 4, qPCR assays confirmed the results from the microarray platforms. SYNPO2 and EGR3, which are downregulated and RDX and FOXM1, which are upregulated in the microarray analysis exhibit a very similar expression pattern in the qPCR analysis. Interestingly, FoxM1 is consistently upregulated in metastases, while RDX was upregulated in only two of the four patients with metastatic disease, confirming the heterogeneity of metastatic prostate cancer.