Gene Expression Profiling in Clinically Localized Prostate Cancer: A Four-Gene Expression Model Predicts Clinical Behavior

CaP⁷ is one of the most common malignancies and the second cause of cancer death among Western men (1). Its morbidity and mortality in the growing elderly population makes CaP a major public health problem. Treatment is ineffective in advanced stages, and the disease must, therefore, be diagnosed early, when the tumor is still confined to the prostate and can be eradicated by radical prostatectomy. Early diagnosis of CaP is currently based on a combination of digital rectal examination and PSA assay and is confirmed by biopsy. However, these tests lack sensitivity, and radical prostatectomy carries a substantial risk of incontinence and impotence. Moreover, about one-third of men who undergo radical prostatectomy for localized tumors relapse, and this risk cannot be accurately predicted by using available markers (clinical stage, serum PSA level, Gleason score, pathological stage, and the number of positive biopsies). New diagnostic and prognostic markers and new therapeutic targets are urgently needed.

The initiation and progression of CaP involves multiple changes in gene expression. cDNA microarray technology offers the possibility of comparing the expression of large numbers of genes between normal and malignant tissues and was recently used to identify disease-related gene expression patterns in prostate samples (2, 3, 4). Discrepancies among reported CaP gene expression signatures could be attributable to the fact that prostate tumors contain many cell types, in addition to carcinoma cells, such as epithelial cells, stroma cells, endothelial cells, adipose cells, and infiltrating lymphocytes. cDNA microarray studies require abundant starting material, and neoplastic epithelial cells cannot be precisely microdissected from their normal counterparts. In addition, although cDNA microarrays can be used to compare the mRNA expression of thousands of genes analyzed in parallel, they cannot discriminate between gene clusters with a high degree of homology (e.g., p14, p15, and p16 on chromosome arm 9p). Finally, cDNA microarrays cannot reveal small variations in gene expression or changes in important but weakly expressed genes (e.g., CYP19 and hTERT in prostate tumors).

To circumvent the technical problems inherent in the cDNA microarray technique, we used a quantitative real-time RT-PCR assay that is a reference in terms of its performance, accuracy, sensitivity, wide dynamic range, and high throughput capacity for nucleic acid quantification.

In this study, we applied RT-PCR to 291 selected genes involved in pathways (e.g., cell cycling, signal transduction, apoptosis, angiogenesis, and metastasis) that are known to be dysregulated in solid tumors, including CaP (Fig. 1). We studied 14 clinically localized prostate tumors from well-documented sources and 7 normal prostate tissues. Our aim was to identify new candidate diagnostic and prognostic markers and therapeutic targets.

Primary prostate tumor samples were obtained from patients undergoing prostate surgery at St. Louis Hospital (Paris), La Cavale Blanche Hospital (Brest), and Nancy University Hospital, France.

Clinically localized tumors were removed by radical prostatectomy. The surgical specimens were first sliced thickly, and samples were then cut from suspect areas.

Suspect areas were stained with H&E for histopathological examination in the surgery suite, and a thick shave of an adjacent section was immediately frozen at −80°C for RNA extraction and additional examination. This preselected tumor specimen was then sliced in the laboratory and again examined histologically. Samples were considered suitable for molecular studies when all examined epithelial cells were neoplastic. Malignant areas were carefully dissected with a scalpel, yielding a homogeneous cell population and avoiding dilution of tumor-specific genetic changes by nucleic acids from normal and reactive cells present in the same specimen. The histological diagnosis and clinical stage (based on the Tumor-Node-Metastasis system and Gleason score) were determined after surgery. Eight of the 14 selected tumors were stage pT2, and six were stage pT3. The Gleason score of the tumor area selected for RNA extraction was confirmed histologically by an experienced pathologist in the laboratory. The Gleason scores were 5–6 in five cases, 7 in eight cases, and 8 in one case.

With 2 years of follow-up, 7 of the 14 patients (three pT2 and four pT3) had progressed (as defined by two successive PSA values exceeding 0.2 ng/ml), whereas the other 7 patients (5 pT2 and 2 pT3) had not progressed.

The clinical characteristics of the 14 patients are shown in Table 1. Specimens of normal prostate tissue were obtained from 7 of these 14 patients who underwent radical prostatectomy and were used to determine basal target-gene mRNA expression. Normal-looking areas of each surgical specimen were examined histologically to confirm the absence of cancer cells and benign hyperplasia.

Total RNA was extracted from tissue specimens by using the acid-phenol guanidinium method (5). The quality of RNA extracts was determined by electrophoresis through agarose gels, staining with ethidium bromide, and visualization of the 18S and 28S RNA bands under UV light. cDNA was synthesized as described previously (6).

Quantitative values are obtained from the Ct number at which the increase in signal associated with exponential growth of PCR products starts to be detected (using PE Biosystems analysis software, according to the manufacturer’s manual).

The precise amount of total RNA added to each reaction mix (based on absorbance) and its quality (i.e., lack of extensive degradation) are both difficult to assess. We, therefore, also quantified PPIA (the peptidylprolyl isomerase A gene encoding cyclophilin A) and RPLP0 (encoding the human acidic ribosomal phosphoprotein P0) transcripts as endogenous controls. RPLP0 is also known as 36B4 and is widely used as an endogenous control for Northern blot analysis.

Results, expressed as N-fold differences in target gene expression relative to PPIA (or RPLP0) expression and termed “Ntarget,” were determined by the formula: Ntarget = 2^ΔCtsample, where the ΔCt value of the sample was determined by subtracting the average Ct value of the target gene from the average Ct value of the PPIA (or RPLP0) gene.

The Ntarget values of the prostate samples were subsequently normalized such that the mean of the seven normal prostate Ntarget values was 1.

Primers for PPIA, RPLP0, and the 291 target genes were chosen with the assistance of the Oligo 4.0 program (National Biosciences, Plymouth, MN). We performed Basic Local Alignment Search Tool (7) searches against dbEST, htgs, and nr (the nonredundant set of the GenBank, European Molecular Biology Laboratory, and DNA Data Bank of Japan database sequences) to confirm the total gene specificity of the nucleotide sequences chosen as primers and the absence of DNA polymorphisms. In particular, the primer pairs were selected to be unique relative to the sequences of closely related family member genes or corresponding retropseudogenes. Primer sets were also tested by PCR to ensure they yielded a single band on agarose gel, and PCR products were purified and sequenced to confirm primer specificity. To avoid amplification of contaminating genomic DNA, one of the two primers was placed at the junction between two exons or in a different exon (except for intronless genes). A primer pair lying in intron 12 of the albumin gene was used to check the DNA-free status of RNA samples (data not shown).

Moreover, genomic DNA was used as a template to confirm the RNA specificity of the primer set in each experiment.

For each set of primers, a no-template control and a no-reverse transcriptase control (reverse transcriptase-negative) assay, which produced negligible signals (usually Ct > 35), were used to confirm the absence of primer-dimer formation and genomic DNA contamination.

A relative kinetic method was applied using standard curve. The latter was constructed with 4-fold serial dilutions of cDNA from a pool of normal prostate tissues. Standard curves were produced for the 291 target genes, PPIA, and RPLP0.

All PCR reactions were performed using an ABI Prism 7900 Sequence Detection System (Perkin-Elmer Applied Biosystems) and the SYBR Green PCR Core Reagents kit (Perkin-Elmer Applied Biosystems). The thermal cycling conditions comprised an initial denaturation step at 95°C for 10 min and 45 cycles at 95°C for 15 s and 65°C for 1 min.

In a first step, prostate tumor mRNAs were mixed to obtain two prostate tumor pools [one representative of patients without relapse 2 years after radical surgery (n = 7) and one with relapse (n = 7)] and a pool of paired normal prostate specimens.

Tissue samples were considered suitable when PPIA Ct values were between 17 and 19, reflecting an appropriate starting amount and quality of total RNA.

The mean Ct values of the prostate samples included in the pools were 18.17 ± 0.29 (patients without relapse), 18.00 ± 0.47 (patients with relapse), and 18.22 ± 1.45 (normal prostate samples).

The distributions of mRNA levels were characterized on the basis of their medians and ranges. Relationships between the target genes and clinical and histological parameters were tested with the nonparametric Mann-Whitney U test. Differences were considered to be significant at confidence levels greater than 95% (P < 0.05).

To visualize the ability of a given molecular marker to discriminate two populations (in the absence of an arbitrary cutoff value), we summarized the data in a ROC curve (8). This curve plots sensitivity (true positives) on the Y axis against one specificity (false positives) on the X axis, considering each value as a possible cutoff. The AUC were calculated as a single measure for the discriminatory capacity of each molecular marker. When a molecular marker has no discriminatory value, the ROC curve lies close to the diagonal and the AUC is close to 0.5. When a molecular marker has strong discriminatory value, the ROC curve moves to the top left or bottom right corner and the AUC approaches 1 or 0.

Hierarchical clustering was performed using GeneANOVA software (9). The data sets were rank-transformed before analysis by the WARD method.

mRNA levels of 11 (3.8%) of the 291 target genes were very weak in both prostate tumor pools and in the normal prostate pool, being detectable but not reliably quantifiable (Ct > 30) by means of real-time quantitative RT-PCR with fluorescence SYBR Green methodology.

mRNA levels of each of the 291 genes were determined in the two prostate tumor pools and were compared with the mRNA levels of each gene in the normal prostate pool. The 88 genes displaying markedly altered expression (at least a 2-fold difference between at least one prostate tumor pool and the pool of normal prostate tissues) were selected for additional analysis in the 14 individual prostate tumors and seven paired normal prostate specimens.

Mean expression values obtained with the individual prostate tumors were very similar to those obtained with the corresponding pooled samples (data not shown). Median mRNA levels of 46 (52.3%) of these 88 genes differed significantly between the 14 prostate tumors and the seven normal prostate samples (P < 0.005, Mann-Whitney U test; Table 2). Five genes showed significantly higher expression in the tumor specimens, and 41 showed significantly reduced expression.

With regard to the five genes with significantly increased expression, the ratio of the median value in the 14 prostate tumors and the median value in the seven normal prostate tissues ranged between 1.8 (MMP9) and 5.7 (CDKN2A). With regard to the 41 genes with significantly reduced expression, the ratio of the median value in the 14 prostate tumors and the median value in the seven normal prostate tissues ranged between 0.07 (NRG1) and 0.63 (TYMS and UNC5C).

The capacity of each of these 46 genes to discriminate between tumoral and normal prostate tissue was assessed by using the AUC-ROC method. The AUC values shown in Table 2 represent the capacity of each gene to discriminate between normal and tumoral prostate tissue. The Ntarget values indicated in Table 2 (calculated as described in “Materials and Methods”) were based on the amount of the target message relative to the PPIA endogenous control, to normalize the amount and quality of total RNA; similar results were obtained by using a second endogenous RNA control, the RPLP0 gene (also known as 36B4; data not shown).

The only associations between clinical and pathological features (e.g., age, serum PSA level at diagnosis, stage, and Gleason score) and the expression of individual target genes were between ILK and F3 expression and the Gleason score (P = 0.05 and P = 0.009, respectively); increased ILK expression was associated with moderately or poorly differentiated tumors, and reduced F3 expression was associated with poorly differentiated tumors.

With 2 years of follow-up, 7 of the 14 patients (three pT2 and four pT3) had progressed (as defined by two successive PSA values exceeding 0.2 ng/ml), whereas the other 7 patients (five pT2 and two pT3) had not progressed (Table 1).

The capacity of each gene to discriminate patients who relapsed from patients who did not relapse was estimated by the AUC-ROC method. By combining the two genes with the AUC-ROC values closest to 0 (XLKD1/LYVE1 and CGA; 0.061 and 0.143, respectively) and the two genes with the AUC-ROC values closest to 1.0 (F2R/PAR1 and BCL-G; 0.878 and 0.714, respectively), we obtained a four-gene model that discriminated perfectly between the seven patients with and the seven patients without relapse, independently of stage and grade. XLKD1/LYVE1 and CGA expression was significantly lower (P < 0.05, Mann-Whitney U test) in patients who relapsed than in patients who did not relapse (median values: 0.06 versus 0.24 and 0.36 versus 3.55, respectively), whereas BCL-G and F2R/PAR1 expression was significantly higher (1.04 versus 0.48 and 1.27 versus 0.58, respectively; data not shown).

The gene clusters are shown as dendograms (Fig. 2), in which line length and branching reflect the relatedness of the samples according to the expression of the four genes.

We studied the mRNA expression of 291 genes involved in pathways (e.g., cell cycling, signal transduction, apoptosis, angiogenesis, and metastasis) that are known to be dysregulated in solid tumors, including CaP.

Eighty-eight genes showed at least a 2-fold difference in expression between pooled prostate tumors and pooled normal prostate tissues. The expression level of these 88 genes was then studied individually in 14 prostate tumors and seven normal prostate specimens. Analysis of the expression level of these 88 genes in each sample further allowed us to check for potential heterogeneity within each pool. Mean values calculated for individual prostate tumors were very close to the value obtained for the corresponding pooled samples (data not shown), indicating that the initial evaluation based on pooled samples was reliable. The expression of 5 of the 88 selected genes was significantly increased, whereas that of 41 genes was significantly decreased (Table 2). It is noteworthy that the proliferative marker MKI67 had no discriminatory value, the ROC curves lying close to the diagonal and the AUC value being close to 0.5 (AUC-ROC, 0.479; data not shown), suggesting that the dysregulation of the 46 target genes in prostate tumors was not caused by the high proliferative activity of cancer cells.

The gene expression pattern in prostate tumors was related to: (a) complex physiological properties (e.g., angiogenesis and invasion); (b) the activity of specific signaling pathways (e.g., DCC, NTN1, and ROBO); and (c) cell specificity (e.g., stroma-dependent expression of growth factors).

First, most of the dysregulated genes belonged to the ECM and ECM-remodeling categories and were involved in angiogenesis (e.g., XLKD1/LYVE1, PROK1, SERPINB5/Maspin, and THBD). The MMP gene MMP9 was overexpressed in the tumors, whereas all of the other genes (ECM, ECM-remodeling, and angiogenesis dysregulated genes: MMP14, ITGA5, ITGB4, ITGB3, XLKD1/LYVE1, PROK1, SERPINB5/Maspin, and THBD) were underexpressed. Except for CDKN2A, no genes involved in cell cycling, DNA synthesis, or gene transcription activities related to the highly proliferative status of cancer cells were dysregulated. Similarly, no genes involved in apoptotic pathways were significantly dysregulated. However, the antiapoptotic gene BIRC5/Survivin was expressed about 1.6 times more strongly in the tumors than in the normal prostate tissues (data not shown).

Second, several key genes in axon guidance and cell migration were underexpressed in prostate tumors, including NTN1/netrin-1, the chemoattractant gene and its receptor DCC (deleted in colorectal cancer), and SLIT3 and its ROBO (Roundabout) receptors. These genes might, therefore, be involved in the chemotactic behavior of CaP cells.

Finally, most of the growth factor genes dysregulated in prostate tumors were down-regulated (AREG, NRG1, BMP7, GRO1, and so forth), possibly because of the stroma-dependent nature of their expression.

Some of the genes found to be dysregulated here have already been forwarded as cancer biomarkers or mediators of prostate carcinogenesis. For example, PCA3 (also known as DD3) has already been shown to be overexpressed in prostate tumors (10) and is a very sensitive and specific marker of this malignancy (11); GRPR (gastrin-releasing peptide receptor), which has high affinity for gastrin-releasing peptide and bombesin, has also been shown to be overexpressed (12), and therapeutic approaches based on bombesin receptor antagonists and cytotoxic bombesin analogues have been considered in CaP (13). We also found a 9.7-fold decrease in the expression of SERPINB5/maspin, in keeping with the role of the corresponding protein as an angiogenesis inhibitor (14) that acts at the cell membrane to inhibit the invasiveness and motility of prostatic cancer cells (15).

We identified some other genes or their products as potential therapeutic targets in CaP.

MMP9 was overexpressed, and synthetic inhibitors of MMPs have therapeutic efficacy in various cancers. Recently, a novel inhibitor, Ro 28–2653, with high selectivity for MMP9, showed promise in patients with CaP (16).

CDKN2A exon 1α (p16) transcripts were overexpressed in prostate tumors; indeed, 8 (57%) of 14 prostate tumors displayed increased expression of p16CDKN2A, varying from 2- to 10-fold higher than normal prostate tissues. In contrast, a decrease in p16CDKN2A expression was found in three prostate tumors (data not shown). The product of this gene is a cyclin-dependent kinase inhibitor that functions as a cell growth regulator controlling cell cycle G₁ progression through engagement of the Rb-cdk4/6-cyclin D pathway (17). p16CDKN2A overexpression has already been described in solid tumors including CaP (18). An altered RB1 axis could trigger p16CDKN2A overexpression in certain systems. However, we found no association, at the mRNA level, between p16CDKN2A and RB1 or other cell cycle regulator expression.

It is of note that the p14^ARF product encoded by this gene, through an alternate open reading frame, was expressed 1.6 times more strongly in the tumors than in the normal prostate tissues (data not shown).

Other interesting candidate targets are key genes in axon guidance and cell migration in the central nervous system, such as the secreted proteins netrin-1 and slit3.

We then analyzed the prostate tumor gene expression signature according to tumor recurrence. Recurrence is a major problem in this setting, because 30% of men undergoing radical prostatectomy for localized tumors eventually relapse, and none of the current indicators of progression (stage and grade at diagnosis) can predict outcome (Table 1). There is, thus, an urgent need for robust prognostic markers capable of identifying patients at risk of relapse after radical surgery.

In this study, the patients were all assessable with respect to recurrence after surgery; seven patients relapsed (defined as two successive PSA values >0.2 ng/ml), and seven remained relapse free for at least 2 years. We used the AUC-ROC test to assess the prognostic value of each gene in terms of relapse. A four-gene model based on the two genes with the AUC values closest to 0 (XLKD1/LYVE1 and CGA) and the two genes with the AUC values closest to 1 (F2R/PAR1 and BCL-G) perfectly discriminated between the seven patients who relapsed and the seven patients who did not relapse after 2 years of follow-up (Fig. 2). XLKD1/LYVE1 and F2R/PAR1 encode angiogenic factors BCL-G, a proapoptotic factor, and CGA, a growth factor.

CGA is the α-subunit of the human glycoprotein hormone chorionic gonadotropin. CGA was overexpressed in patients who did not relapse, in keeping with our previous work showing that CGA is a specific ERα -responsive gene in CaP and that its overexpression may be associated with a good prognosis in CaP (19) as in breast cancer (20).

XLKD1/LYVE-1 is an endocytic receptor for hyaluronan in lymphatic endothelium (21). The functional role of XLKD1/LYVE-1 in lymphatic vessels, and its use as a marker of tumor lymphangiogenesis, are important areas of investigation. In our study, the association between low XLKD1/LYVE-1 expression and poor outcome is in keeping with the known function of this gene; indeed, XLKD1/LYVE-1-positive structures have been observed in lung tumor margins but not within the mass of such tumors (22). Moreover, negative intratumoral XLKD1/LYVE-1 staining does not rule out metastasis (23), suggesting that functional lymphatics in the tumor margins are sufficient for lymphatic metastasis (22).

We also found that increased F2R/PAR1 expression was associated with tumor recurrence. This coagulation factor II (thrombin) receptor is involved in regulating the thrombotic response. Thrombin promotes angiogenesis by enhancing vascular endothelial growth factor synthesis and inducing its secretion and has already been implicated in prostate tumorigenesis and metastasis (24).

Last, overexpression of BCL-G, encoding a BCL2 family protein, has been shown to induce apoptosis in COS-7 cells (25). However, little is know about the expression of this gene in tumor samples, and BCL-G has not been linked previously to clinical outcome.

Our results indicate that although F2R/PAR1 and BCL-G overexpression are associated with poor outcome, optimal outcome prediction is obtained by their combination with genes, such as CGA, that are associated with good outcome. The resulting four-gene model was perfectly predictive of tumor recurrence, independently of tumor stage and grade. It is noteworthy that F3 and ILK, the expression of which correlates with the Gleason score, a current prognostic marker, were excluded from this model.

In conclusion, our data identify potential new therapeutic targets in CaP and show that gene expression profiling can be used as a predictor of outcome. Routine clinical use of genomics-based outcome predictors must await confirmation in larger, independent data sets.