Array-based Comparative Genomic Hybridization for Genome-Wide Screening of DNA Copy Number in Bladder Tumors¹

Bladder cancer is a major cause of morbidity and mortality in the United States and Europe. Approximately 75% of cases are superficial (pT_a, pT₁, pT_IS), 20% muscle-infiltrating (pT₂–T₄), and 5% metastatic at the time of diagnosis. The ability to identify carcinomas more likely to recur and/or progress would allow more aggressive treatment of these cases, which might lead to reduced morbidity and mortality from bladder cancer. A better understanding of the underlying molecular mechanisms leading to tumor formation and progression could result in the ability to identify more aggressive tumors, leading to improved survival and identification of potential therapeutic targets.

The development and progression of bladder cancer is a multistep process, the result of a series of genetic alterations occurring over the lifetime of a tumor. The acquisition of chromosomal abnormalities by tumor cells is a central event in carcinogenesis and one that frequently decides the future malignant potential of a cancer. The search for specific alterations associated with the development and progression of solid tumors involves an intensive analysis of known genes and a search for genes the roles of which were previously unappreciated. Multiple studies have identified the prevalence and clinical significance of a limited number of genetic markers in bladder cancer. The recently developed array CGH³ technique allows high throughput analysis of copy number changes at high resolution throughout the genome (1, 2). This quantitative measurement of DNA copy number across the genome may facilitate oncogene identification (3) and can also be used for tumor classification (4).

In this study, we have used array-based CGH for high resolution mapping of copy number changes in different stages of bladder carcinogenesis in 41 primary human tumors. Although a substantial body of work has suggested that low-stage tumors differ from muscle invasive tumors in their genetic alterations (5, 6, 7, 8, 9, 10), our array CGH data did not show a significant association of genomic copy number alterations with tumor stage or grade. However, the high resolution of the array CGH technology allowed a precise identification of amplicons and regions containing homozygous deletions throughout the bladder cancer genome. Analysis of the patterns of alterations among pairs of known oncogenes and tumor suppressors revealed significant correlations between loci and concordant or complementary categorical behavior. These relationships agree with what is known about the relevant pathway biology and suggest that the ability to define genomic alterations at high resolution, genome-wide, in larger sets of primary tumors may be an effective means for further elucidating the structure of pathways important in the progression of bladder cancer.

A series of 9 T_a, 7 T₁, and 25 T₂-T₄ freshly frozen bladder tumors were obtained from the tissue banks of the UCSF Cancer Center and the Memorial Sloan-Kettering Cancer Center. An initial H&E-stained frozen section was examined to allow trimming of the block for exclusion of normal or necrotic tissue. A tumor sample was considered suitable for study if the proportion of tumor cells was higher than 70%. Ten 10 μm sections were cut for DNA extraction, followed by a final 5-μm section for validation of tumor tissue remaining in the block. Genomic DNA was obtained according to standard procedures using proteinase K digestion and phenol-chloroform extraction. Normal DNA was isolated from lymphocytes of healthy persons and was used as reference for two-color hybridizations.

Two arrays were used in this study.⁴ The first (Array1) consisted of 1777 clones covering the human genome at roughly a 1.5-Mb resolution [HumArray 1.11 (2)]. The second array (Array2) consisted of 380 clones specifically selected to contain important tumor suppressor and oncogene loci. The clones on Array1 were prepared by ligation-mediated PCR as recently described by Snijders et al. (2). DNA clones were robotically spotted in triplicate onto chromium-coated glass slides (PTI or Nanofilm), followed by UV cross-linking. For Array2, degenerate oligonucleotide-primed PCR products from 380 large-insert clones were robotically spotted in quadruplicate onto three-dimensional link-activated slides [Surmodics, Inc., Eden Prairie, MN; according to Hodgson et al. (11)]. These slides underwent several pretreatment steps to block nonspecific binding, and the DNA was denatured before use.

Each tumor sample was hybridized to both arrays, as described previously (1, 4), with modifications. One μg of tumor DNA was labeled by random priming with fluorolink cy3-dUTP, and normal reference DNA was labeled in the same fashion with cy5-dUTP (Amersham Pharmacia, Piscataway, NJ). Unincorporated fluorescent nucleotides were removed using Sephadex G-50 spin columns. One-half of the labeled tumor sample was hybridized to Array1, and the remainder was hybridized to Array2. Test and reference DNAs were mixed with 100 μg of Cot-1 DNA (Life Technologies, Inc., Gaithersburg, MD), were precipitated and were resuspended in 30–50 μl of a hybridization solution containing 50% formamide, 10% dextran sulfate, 2× SCC, 4% SDS, and 100 μg tRNA. The hybridization solution was heated to 72°C for 10 min to denature the DNA and then was incubated for 1 h at 37°C to allow blocking of the repetitive sequences. Hybridization was performed for 48 h in a moist chamber on a slowly rocking table, followed by a 15-min posthybridization wash in 50% formamide/2× SSC at 45°C, and for 10 min in phosphate buffer at room temperature. Slides were mounted in 90% glycerol in phosphate buffer containing 4′,6-diamidino-2-phenylindole (DAPI; 0.3 μg/ml).

Sixteen-bit fluorescence intensity images were obtained using a charged coupled device camera (Sensys, Photometrics, equipped with a Kodak KAF 1400 chip) coupled to a 1× magnification optical system. The acquired microarray images were analyzed using Genepix Pro 3.0 (Axon Instruments, Inc., Foster City, CA). DNA spots were automatically segmented, local background was subtracted, and the total intensity and the intensity ratio of the two dyes for each spot were calculated. Spots composed of less than nine pixels, showing bad correlations of the two fluorescent dyes (Genepix 3.0, R² < 0.5), or showing autofluorescent particles over the target were discarded.

Log2 intensity ratios obtained for each array for each case were individually centered by subtracting the median of log2 intensity ratios for that case over all clones that met the quality control parameters described below. Data on the two arrays was then merged into one data set using the genomic mapping information from all of the clones. There were 19 clones in common on the two arrays. A matched-pair t test on each of the 19 revealed no clones with significantly different ratios at the 5% level.

A series of eight normal versus normal hybridizations was used to define the set of clones having consistently good hybridization quality (data not shown). For each analysis, clones were excluded for which none or only one spot remained after the Genepix analysis. For all analyses, the 5% of clones with the most extreme average test over reference ratio deviations from 1.0, and the 1% of clones with the largest SD in this set of normal controls was excluded. This procedure resulted in the exclusion of 174 clones. In addition, all X-chromosome clones were excluded from data analysis (sex-mismatched reference samples were used for quality control). The final set, on which all of the analyses were performed, contained 1747 clones.⁵

We considered three types of questions: (a) whether there were associations between copy number alterations and tumor stage or grade; (b) whether gene pairs exhibited significant correlations; and (c) whether gene pairs exhibited complementary or concordant behavior based on a categorical analysis. The association analyses consisted of statistical correlation with permutation-based assessment of significance, visualization by hierarchical clustering, and automatic pattern classification with cross-validation to assess predictive power, all as in Wilhelm et al. (4) and Olshen and Jain (12). For the gene pair correlations, we selected 24 clones containing known bladder cancer oncogenes or tumor suppressors and 22 clones that were most frequently aberrant. In this analysis, the values of clones spanning the same gene were averaged. We computed the pair-wise correlations of copy number for these clones. Permutation analysis under the null hypothesis of no association between clones was performed to establish the appropriate significance threshold for the correlation coefficient (4, 13). This form of permutation testing corrects for the multiple comparisons present in array analysis. We use a conservative method that computes a null distribution of the maximum magnitude correlation statistic across all genomic loci. We select our significance threshold to be at the 95th percentile of the permutation distribution [details can be found in Olshen and Jain (12)]. Because the foregoing analysis captures only linear relationships between values on a continuous scale, we also performed a categorical analysis to define associations among loci. We selected 10 gene pairs based on knowledge of signaling pathways (but independent of the array CGH data). These genes were also known a priori to be frequently gained or lost in bladder cancer. For each pair, we constructed 2 by 2 contingency tables with categories being “change” and “no change.” Change is defined as the most frequent aberration in a given gene (gain or loss using a conservative threshold of ±0.25 log₂ value). Change may represent gain in the first gene of a pair and loss in the second. In other words, we addressed whether change in one gene makes the change in the other less (complementarity) or more (concordance) likely. The hypothesis was tested using an empirical null population of pairs constructed from the clones that were frequently changed in the data set. Here, “frequent” was defined to be 25%. The standard statistic for testing the difference in binomial proportions was used in the analysis. Null hypothesis of independence is not suitable here because there is bias toward concordance among frequently changed clones.

We tested the quality of the CGH arrays and intrinsic variability of the method by performing eight sex-mismatched normal versus normal hybridizations. These control hybridizations were performed simultaneously with the tumor analyses, using the same batch of arrays, with identical labeling and hybridization conditions. The average replicate SD of these clones was 4%, and the average of the SE per clone, representing the average clone variability in the 8 normals, was 6% in this set of controls. Fig. 1 shows a representative profile and a histogram of the copy numbers across all clones in the 8 control experiments. On the basis of these controls, we used thresholds of 0.2 and −0.2 (log2ratio) for calculating the frequencies of genomic copy number gains and losses, respectively, in the bladder tumor cases.

As a further test of quality within the tumor data, we assessed the concordance among clones in which more than one clone contained the same gene. In a hierarchical clustering constructed without averaging clones spanning the same gene (figure not shown), all of the gene pairs that spanned the same gene clustered together. The average correlation for the 27 pairs of overlapping clones was 0.87, validating the reproducibility of the array analysis within the tumor set.

Fig. 2 shows representative examples of the high-resolution analysis of the 41 transitional cell carcinomas of the bladder (9 T_a, 7 T₁, and 25 T_2–4). Copy number gains and losses can easily be detected for small chromosome regions, chromosome arms, and whole chromosomes. Small genomic regions showing high-level amplifications (defined as log₂ratio >1), as well as regions indicating homozygous deletions (defined as log₂ratio <-1), can also be identified.

Copy number alterations are involved in a large fraction of most tumor genomes. To quantitate that fraction, each clone was assigned a genomic distance equal to the sum of one-half of the distance between its own center and that of its two neighboring clones. The average genomic distance between clones resulting from this calculation was ∼1.6 Mb (Table 1). The entire tumor set involved an average of 281 Mb or 9.9% of the genome, and copy number loss affected 339 Mb or 11.9% of the genome. Only 8 of the 41 tumors showed any copy number alterations in less than 10% of their genome (4 T_a, 0 T₁, and 4 T_2–4 cases). Although the number of tumors in each group is small, there were differences in the extent of the genome affected among the tumor stages. The group of T₁ tumors contained on average the highest fraction of genomic alterations, both copy number gains and copy number losses (P < 0.01 for total changes). On average, superficial bladder tumors showed greater copy number losses (11.1% of the genome) than gains (6.8%). High-level amplifications were seen most frequently in the invasive tumors (0.3%).

Fig. 3 shows a frequency plot of gains and losses for all of the clones. Although a large part of the genome is affected, the distribution of copy number alterations is not uniform throughout the genome. Tables 2,3,4,5 list the most frequently gained or lost clones in this series and those with high-level amplifications or homozygous deletions. High-level amplifications were identified in 22 of the 41 patients (2 T_a, 3 T₁, and 17 T_2–4) and were present in 191 clones, of which 52 showed amplification in more than one case. Oncogenes contained in clones with high-level amplifications included E2F3 (6p22.3), EGFR (7p11.2), FGFR1 (8p12), CMYC (8q24.12-q24.22), CCND1/EMS/INT2 (11q13), MDM2 (12q14.3-q15), ERBB2 (17q12), JUNB (19p13.2), CCNE (19q13.11), and CYP24 (20q13.2). Homozygous deletions were less common, with 74 clones identified in a total of 16 cases (5 T_a, 5 T₁, and 6 T_2–4). Only four clones showed homozygous deletion in more than one case. Two were located within 1 Mb of each other at the CDKN2A/p16 locus on 9p21.3, with nine cases having homozygous deletion for one of these clones. The others were located at 8p23.1 and 11p13. On average, each tumor contained seven clones with high-level amplification and two clones with homozygous deletion.

Copy number gain at 6p22 was observed in 11 of the 41 cases analyzed. High-level amplifications were observed in three tumors at 6p22.3. In two of these cases, four clones spanning ∼2.1 Mb were involved; in one case, only two of these clones were involved; and clone RP11-159C8 with the highest ratio (log2ratio = 3.7) contained E2F3.

Almost 50% of all of the bladder cancers showed a loss of the 11p13 region. Two tumors showed intensity ratios indicating homozygous deletions in this region, one case showed 17 deleted clones spanning ∼14 Mb at 11p13 with clone RP11-187A8 having the lowest ratio decrease. This clone was also the only one showing a log2ratio below −1 in the second tumor, again suggesting homozygous deletion. Further support for the involvement of this specific clone came from three tumor cases that showed only a deletion of this clone at the 11p13 region. This clone contains the TNF-associated factor 6 (TRAF6) and the human recombination-activating gene RAG1.

Array CGH also allowed the analysis of frequent break point regions mapped to 8p12. Seventeen cases showed a pattern of transition between distal loss and proximal gain in a small region composed of 17 clones spanning ∼9.2 Mb. In six of these cases, the break point was flanked by clones RP11-258M15 and RP11-274F14, a region of 2.3 mb containing the candidate gene neuregulin 1 (NRG1). A second break point, located 8 Mb proximal to the first one, was present in another 8 cases, spanning a distance of only 1 Mb. The most promising candidate gene mapping to this break point region was FGFR1, a member of the fibroblast growth factor receptor family.

Quantitative real-time PCR analyses were performed on the same set of bladder cancers for the detection of copy number changes for seven genes mapped to chromosome 9.⁶ Six of these genes were contained in clones present on the arrays used. There was a strong concurrence between the log₂ratios for these clones obtained by array CGH and quantitative PCR for the associated genes, with an overall correlation of 0.78. Fourteen homozygous deletions of the CDKN2A/p16 gene were detected by quantitative PCR (mean log₂ratio = −2.16). All 14 of these cases had deletions of the clone containing this gene in the array CGH analyses, with 11 cases having log₂ratio at or below the −1 threshold for homozygous deletion and 3 cases with ratios indicating single copy loss (log₂ratios = −0.63, −0.68, −0.74). In contrast, quantitative PCR analyses showed a normal copy number for CDKN2A/p16 in 20 cases (mean log₂ratio = 0.01). In 18 of these cases, normal copy number was also detected by array CGH; 1 case showed copy number gain (log₂ratio = 0.26); and 1 case showed a copy number loss (log₂ratio = −0.21; mean log₂ratio for 20 cases = 0.02). The copy number correlation between array CGH and quantitative PCR for this specific region was 0.98.

We explored three types of statistical relationships within the data set to determine whether: (a) there were associations between copy number alterations and tumor stage or grade; (b) gene pairs exhibited significant pair wise correlations; (c) gene pairs exhibited concordant or complementary categorical behavior (details in “Materials and Methods”). We found no significant relationship between genomic copy number alterations and tumor stage or grade using our methods. Given previous studies suggesting such a relationship (5, 6, 7, 8, 9, 10) and given the large number of genomic loci being measured, our results likely indicate an insufficient sample size to reveal such associations (we have used statistical methods that correct for multiple testing).

Our analysis of the correlation between gene pairs showed 22 significant correlations (P < 0.05) and 2 that almost met statistical significance (P < 0.1; see Fig. 4). These associations are present even in the absence of T_a tumors, suggesting that they are not driven by differences between tumor groups. Losses or gains of large genomic fragments account for significant correlations between 19 gene pairs on chromosomes 9, 11, 18, and 20. More interestingly, significant correlations were observed between copy number gain of ERBB2 (17q12) and gain of CCNE (19q13.11), between gain of AIB1 (20q12) and loss of PTEN (10q23), between loss of ABL1 (9q34.2) and loss of CDKN2A/p16 (9p21; possibly attributable to frequent loss of all of chromosome 9), and between loss of TP53 (17p13.3) and gain of CCND1/FGF3 (11q13; these last two significant at only P < 0.1).

To determine whether paired genes from the same biological pathway showed complementary or concordant behavior (i.e., whether only one alteration of genes in a pathway is sufficient to modify the entire pathway or whether both of the alterations tend to occur together), a direct categorical complementarity analysis of pairs of genomic loci containing key components of important biological pathways was performed (TP53, MDM2, MYC, CDKN2A/p16, CCND1, CCNE1, BCL2, CDK4, E2F3). Copy number gain of CCND1 behaved complementarily with gain of E2F3 (P < 0.05) and gain of CCNE1 (P < 0.1). E2F3 had a concordant relationship with CCNE1 (P < 0.1). Finally, loss of TP53 showed complementarity with gain of MDM2 (P < 0.1). The results were robust to the threshold chosen for declaring gain/loss. The significant pairs are shown in Fig. 5.

This study represents one of the first applications of genome-wide copy number analysis by array CGH. Snijders et al. (2) recently showed the potential for this technology to provide precise copy number measurements using an overlapping set of clones. The key biological value of high-resolution array CGH lies in its ability to detect small amplicons and deletions that potentially harbor specific oncogenes and tumor suppressor genes. High-level amplifications involving clones containing known oncogenes such as EGFR, CMYC, CCND1, and CCNE1 were easily detected, as well as small homozygous deletions containing the CDKN2A/p16 tumor suppressor gene at 9p21.3.

Our analysis of clones containing high-level amplifications or homozygous deletions revealed two interesting candidate genes: transcription factor E2F3 on 6p22 and TNF-associated factor 6 (TRAF6) on 11p13. High-level amplifications as well as frequent copy number gain at 6p22 have previously been reported by others by low-resolution chromosomal CGH without oncogene identification (14). Array CGH analysis strongly pointed to E2F3 as the target gene on 6p22. E2F3 is thought to be sequestered by unphosphorylated retinoblastoma protein and then released after RB phosphorylation by the CCND1/CDK4 complex in the G₁ phase of the cell cycle (15). Once released, E2F3 is thought to act as a transcriptional regulator at the G₁-S phase transition (16, 17, 18). E2F-1 has been implicated in bladder carcinogenesis, but alterations of this gene appear to occur at the epigenetic level (19).

A frequently deleted clone mapping to 11p13 contained both the TNF-associated factor 6 (TRAF6) and the human recombination activating gene RAG1. TRAF proteins are thought to be important regulators of cell death and responses to stress. The RAG proteins are known to initiate V(D)J recombination by facilitating double-stranded breaks. TRAF6may be a more likely deletion target than is RAG1 in these tumors because it is presumed to function as a tumor suppressor.

Frequent gains and losses could be defined at high resolution using array analysis, allowing a precise mapping of these genomic regions. For example, two frequent break points on chromosome 8 were identified, containing candidate target genes neuregulin 1 (NRG1) and fibroblast growth factor receptor 1 (FGFR1). NRG1 interacts with the NEU/ERB family of receptor tyrosine kinases, known to be frequently overexpressed in bladder tumors. FGFR1 is a member of the fibroblast growth factor receptor family. Specific binding of fibroblast growth factors to these cell surface-expressed receptors activates tyrosine kinase activity. This activation allows coupling to downstream signal transduction pathways that regulate proliferation, migration, and differentiation of endothelial cells, thus enhancing angiogenesis (20). A recent study by Simon et al. (21) using fluorescence in situ hybridization (FISH)-based high-throughput tissue microarray analysis also showed frequent alterations of this gene in bladder cancers.

Associations of alterations in pairs of known oncogenes and tumor suppressors could be characterized using array-based CGH. Genes functioning upstream or downstream of each other in the same biological pathway may exhibit complementary alteration, because an alteration of a single gene in this pathway may suffice for altering the effects of the entire pathway (e.g., enhanced proliferation, apoptosis). The retinoblastoma tumor suppressor pathway plays a critical role in the control of cellular proliferation by regulating the activity of the E2F family of transcription factors (22, 23). Many of the genes involved in this pathway appear to be affected in bladder carcinogenesis, most notably cyclin D1 and cyclin E by amplification, and CDKN2A/p16 and RB1 by inactivation (24, 25, 26, 27). Clear associations were observed in a number of members of the RB proliferation pathway. There was a significant complementary association between copy number gain of the locus containing cyclin D1 and gain of the transcription factor E2F3 locus. Also, there was a trend for complementarity between gain of the CCND1 and gain of CCNE1 loci. This is consistent with the idea that amplification of CCND1 and either CCNE1 or E2F3 may activate the RB pathway, but there is no advantage for more combined gene amplification. Geng et al. (28) showed that cyclin E can functionally replace cyclin D1 and suggested that cyclin E is the key downstream effector of cyclin D1. High-throughput tissue microarray analysis of a large series of bladder cancers (26) showed that cyclin E gene amplification was present in a subset of bladder carcinomas, especially during early invasion. These findings suggest that the RB pathway appears to be deregulated in a majority of the bladder cancers by a gene dosage increase of one of the activators cyclin D1, E2F3, or cyclin E.

Significant correlation was observed in the gain of clones containing ERBB2 (17q12) and CCNE1 (19q13; P < 0.05). Activation of ERBB2 receptor tyrosine kinase pathway is thought to lie upstream from cyclin E activation; therefore, it is unclear why the gain of both genes would be selected for during tumor progression. It is possible, for example, that CCNE activation may play a larger role in the presence of activated receptor pathway. A different relationship may explain the correlation between cyclin D1 gain with p53 loss in individual tumors. Cells with both alterations may have a selective advantage because they play key roles in different cellular pathways controlling proliferation and programmed cell death, respectively. Interestingly the categorical analysis revealed a trend for complementarity between p53loss and MDM2 gain, consistent with the known interaction of these two proteins.

Two distinct pathways are involved in bladder cancer development, with low-grade superficial lesions exhibiting clear clinical and prognostic differences from higher-grade CIS and invasive tumors. Genetic differences have been proposed to identify these tumor patterns, and candidate genes have been proposed to distinguish among them (29, 30). Array CGH has been used to differentiate among histological types of renal cancer (4). However, in this study, no statistically significant association was observed between the pattern of copy number alterations and tumor stage or grade. This may be explained by the presence of small genetic differences between superficial and invasive stages of bladder cancer, or the large heterogeneity within individual groups. Furthermore, the sample size for this study was relatively small, with limited power to discern such differences.

In conclusion, array CGH detected the large chromosomal alterations previously identified by chromosomal CGH and other cytogenetic approaches. Higher-resolution analysis allowed specific identification of alterations in smaller copy number transition regions, which suggest the presence of specific candidate genes. Relationships were observed between alterations of loci containing genes known to be key players in bladder cancer biological pathways. This pilot study must be repeated with much larger numbers of well-characterized primary tumors and must be validated with other molecular, cytogenetic, and immunohistochemical approaches.

We thank Karen Chew and the UCSF Cancer Center Tissue Core for assistance with the accrual and review of bladder tumor specimens at UCSF. We thank Joe Gray for his assistance with production of Array 2, and Dan Sudilovsky for help during the pathology review.