Gene Expression Profiling of Malignant Mesothelioma¹

Malignant pleural mesothelioma is a uniformly fatal cancer of the pleura, affecting 2,500 Americans annually. The incidence is rising sharply in Europe where 5,000–10,000 cases per year are expected (1). No intervention has proven to be curative, despite aggressive chemotherapeutic regimens and prolonged radiotherapy. The median survival in most cases is only 12–18 months after diagnosis.

Several clinical problems regarding the diagnosis, pathophysiology, and treatment of malignant mesothelioma remain unsolved. Making a diagnosis of mesothelioma from pleural fluid is notoriously difficult and often requires a thoracoscopic or open pleural biopsy. Relatively few mesothelioma-specific markers have been identified often necessitating a “diagnosis by exclusion.” Although it is well established that asbestos exposure is a major risk factor in the development of mesothelioma (2), the molecular steps in carcinogenesis remain unknown. A recent suggestion that SV40 virus may be a cocarcinogen raises additional pathophysiologic questions. Finally, given the poor response to current therapies, a better understanding of the molecular pathways active in this disease could potentially provide new targets for therapy.

One potentially useful approach to solve these issues would be to identify specific gene expression changes in cancerous mesothelial cells. Progress in this area has been limited and, to date, has been performed mostly in cell lines (3, 4). Recent reports from the Brigham and Women’s Hospital using actual tumor tissues identified 20 genes that were overexpressed in mesothelioma tissues (5, 6).

Expression profiling offers the opportunity to analyze gene expression changes in mesothelioma in an unbiased manner. Therefore, we performed microarray analysis of mesothelioma tissues from 16 patients and compared these results with 4 control pleural tissue samples. To our knowledge, this approach has not yet been used to describe mesothelioma tissues in a comprehensive manner. Our first goal was to characterize major pathways altered in malignant mesothelioma. Our second goal was to identify genes that could be involved in the pathophysiology of mesothelioma or that might serve as diagnostic markers to improve the accuracy of tumor classification.

Patients in this study presented to the Hospital of the University of Pennsylvania between October 1999 and December 2000, and underwent a standard thoracotomy for diagnostic or therapeutic reasons. Mesothelioma tissue was obtained from those patients with a confirmed pathological diagnosis and who had not received prior therapy. Intraoperative malignant mesothelial samples and nodules were dissected from associated fat and connective tissue, but no microdissection was performed. H&E staining was performed to verify the presence of tumor cells. Samples were immediately frozen in liquid nitrogen before RNA analysis. Control pleural tissue was obtained from patients undergoing resection for lung cancer at an area distant from the tumor site. Tissue acquisition was approved by the Institutional Review Board at the University of Pennsylvania.

Three-hundred mg of each tumor specimen was subjected to RNA extraction using standard techniques (7). Briefly, tissues were homogenized in guanidinium isothiocyanate buffer at room temperature, extracted with phenol-chloroform-isoamyl alcohol, and precipitated with isopropanol in the presence of sodium acetate. After initial recovery and resuspension of the RNA pellet, a DNase step was performed for 3 h at 37°C using 80 μg of RNase inhibitor (Roche, Alameda, CA), 60 μg of RNAsin (Promega, Madison, WI), and 10 units of RNase free DNase (Roche) in 1 m Tris (pH 7.4) buffer solution. Total RNA was then re-extracted, precipitated, and dissolved in water.

Hybridization was performed on GF211 GeneFilters Microarrays (Research Genetics, Inc., Carlsbad, CA) that contain 4132 named human genes based on the protocol supplied by the manufacturer. Gene names are listed according to the UniGene human-sequence collection (available at UniGene Web Site).⁴

Hybridizations, washes, and scanning were performed as described previously (8). Briefly, the gene filter membranes were prewetted in 0.5% SDS and prehybridized for 2 h at 42°C in 5 ml of Microhyb solution (Research Genetics, Inc.) containing 1.0 μg/ml Cot1 DNA (Life Technologies, Inc., Gaithersburg, MD) and 1.0 μg/ml poly-deoxyadenylate (Research Genetics).

Ten μg of DNase-treated total RNA and 2 μg oligodeoxythymidylic acid (Promega) were incubated at 70°C for 10 min, and rapidly chilled on ice. Using SuperScript II RT (Life Technologies, Inc.), RNA was next reverse transcribed according to the manufacturer’s instructions but in the presence of radioactive [³³P]dCTP. Labeled probes were purified using chromatography columns to remove any unincorporated nucleotides.

Labeled double-stranded cDNA was added to the prehybridization buffer and the filters hybridized for 18 h at 42°C. Posthybridization washes were performed twice at 50°C in 1× SSC (2× SSC, 15 mm sodium citrate, and 150 mm NaCl) and 1% SDS for 20 min, and once at room temp in 0.5× SSC and 1% SDS for 15 min. After drying, the membranes were placed in cassettes and scanned using the phosphorimager (Hewlett Packard). After each hybridization the filters were stripped by boiling in 0.25% SDS solution and reused for up to three times.

The images resulting from the phosphorimager were imported directly into the image analysis “Pathways” software (Research Genetics, Inc.). The background radiointensity for each array was simultaneously recorded. A full description of the data analysis is beyond the scope of this paper, however, is available at our website.⁵ Software used for data analysis included Microsoft Access, Microsoft Excel, Visual Perl, and Visual Basic. Briefly, our initial step was to remove the background intensity from every hybridization experiment. Normalization was performed in three separate ways. Global normalization was used to calculate gene expression levels based on the average total intensity on each filter. A second method normalized using the average intensity of the 250 genes with the least amount of variability across hybridization experiments. The third method used genes that had the most similar expression levels in each array. In this approach, a Gaussian curve was fitted to a plot of number of genes versus intensity level for each array. The genes that fell within one SD of the mean in all of the arrays were chosen. In our data set, there were 200 genes that fit these criteria.

Three gene prediction techniques to identify significantly changed gene expression levels were used from the data from each normalization process: Student’s t test (9), significant analysis of microarrays⁶ (10), and patterns of gene expression⁷ (11).

Using combinations of three normalization tools and three gene prediction techniques, this process generated nine separate lists of genes with differential expression. To be considered a “significantly changed gene,” a gene had to satisfy the following criteria: (a) the gene must appear on at least four separate lists of significant genes; (b) the P using the Student’s t test after at least one normalization method must be <0.001; and (c) a gene must have at least a 2-fold change in gene expression level between sample groups.

Genes were categorized using the vocabulary defined by the GO⁸ Consortium.⁹ The complete vocabulary is structured into three broad categories, reflecting the biological roles of genes: (a) molecular function: tasks performed by individual gene products; (b) biological process: broad biological goals accomplished by ordered assemblies of molecular functions; and (c) cellular component: subcellular structures, locations, and macromolecular complexes.

Significant genes were grouped under the appropriate GO function and then ranked in groups by the most up-regulated functions (see Table 2). In addition, we examined the list of significant genes and pursued verification studies on a subset that might be useful diagnostically or those that had potentially interesting functions.

Several genes were noted to be consistent with infiltration with WBCs in our tumor samples. Immunostaining with an anti-CD45 antibody (see below) confirmed the presence of infiltrating leukocytes in all of the tumors and pleural samples that were examined. Accordingly, a “virtual microdissection” was performed to remove genes that we suspected to be present because of contaminating WBCs. Genes with the following GO vocabulary words were removed: class II MHC antigen, antimicrobial humoral response, antipathogen response, defense response, response to bacteria, immune response, and response to pathogenic bacteria.

To validate a subset of genes with significant changes, real-time, reverse transcription-PCR was performed. Four pools of RNA (300 μg each) were created. The first pool consisted of RNA (60 μg each) from 5 normal pleural tissues. The second and third pools consisted of RNA from 5 patients (60 μg each) each, randomly selected, who had been analyzed previously on microarrays. The fourth pool consisted of RNA from 5 patients (60 μg each) who had not been studied on the microarrays and was designated our “prospective” pool.

Three-hundred μg of RNA from each pool of total RNA was reverse-transcribed using 0.5 μg oligodeoxythymidylic acid (Promega), 10 mm deoxynucleoside triphosphates (Clontech, Palo Alto, CA), 1 unit of Powerscript Reverse Transcriptase in 5× First-Strand Buffer and 100 mm DTT (Clontech) for 80 min at 42°C. Gene sequences available at the National Center for Biotechnology Information GenBank and Unigene databases were selected to design primers. Optimum primer sequences were selected after verification for gene-specific complementation using the National Center for Biotechnology Information Blast program.¹⁰ Semiquantitative analysis of gene expression was performed using a Cepheid Smart Cycler using the manufacturer’s protocol for the Sybr-Green kit supplied by Roche (Cepheid, Sunnyvale, CA). cDNA concentrations from each pool were normalized using two control genes that showed no change in expression on the arrays: ubiquitin and cytochrome p450 reductase. Standard curves were generated by preparing serial dilutions, and the relative level of expression of each of the verified genes was determined.

Immunostaining was performed on a subset of genes. Five μm, frozen tissue sections were mounted on slides, permeabilized with acetone, and fixed in 5% blocking serum (PBS/BSA/Azide) for 20 min. The slides were incubated with 5–10 μg/ml of the following antibodies: cytokeratin (5/18; NovoCastra, Newcastle upon Tyne, United Kingdom), anti-Grp94 (StressGen, Victoria, Canada), the M_r 67,000 laminin receptor (LabVision, Fremont, CA), and CD-45 (Sigma, St. Louis, MO). Visualization was achieved by the use of Vectastain kit or by Alkaline Phosphatase kit (Vector Laboratories, Burlingame, CA) using the manufacturer’s protocols.

Sixteen patients with malignant mesothelioma underwent resection and debulking for their pleural disease. Their clinical characteristics are presented in Table 1. The average age at time of operation was 63.0 years, with 1 female. The average smoking history was 33.3 pack-years. Fifteen of the patients had some history of asbestos exposure. Thirteen of the patients underwent a pleurectomy. Pathology was either epithelioid (75%), sarcomatous (6%), or biphasic (19%).

Tumors from the 16 cancerous patients was compared with 5 samples of normal pleura. No microdissection was performed. The normal pleura was a thin layer of tissue “stripped” from the chest wall that consisted of mesothelial cells and a small amount of adherent connective tissue.

RNA from each tissue was extracted, reverse transcribed, and labeled with [³³P]dCTP, and arrayed on “unichannel” nylon microarrays containing 4132 genes. Our entire data set is available on the web.⁵

In validation studies, the average radiointensity of the same gene varied by 8%. Because 1 of the 5 normal pleural arrays demonstrated a variability of >15%, we removed this sample from additional data analysis. Another test of experimental validity was performed by taking the tumor from a single patient and repeating hybridizations on 3 separate days on three new arrays from the same manufacturing lot. Regression analysis demonstrated 12% variability of the same tumor run at three different times.

Of the 4132 genes analyzed, 166 (4.0%) genes were classified as “significantly” up-regulated and 26 (0.6%) genes “significantly” down-regulated based on the criteria described in the “Materials and Methods.” The degree of gene expression up-regulation varied from 2-fold to 11-fold, and the degree of down-regulation ranged from 2-fold to 5-fold decrease in gene expression. A complete list of these genes is available at our website.

We used the GO Consortium vocabulary to categorize significantly up-regulated genes by important molecular functions (Table 2). The categories with the most number of changed genes included those involved with cytoskeletal reorganization (GO categories: extracellular matrix genes, epidermal development, cell shape and cell size control, and cell migration and motility), protein synthesis (GO categories: protein synthesis, RNA processing and modification, and translation factors), and metabolic pathways (GO categories: oxioreductase and energy generation). We additionally analyzed a group of significantly up-regulated genes that might have potential use as surface receptors for diagnostic markers, as well as other genes with possible therapeutic and prognostic implications. Table 3 lists the significantly down-regulated genes. We did not identify any pathways with consistent changes. Of interest was the down-regulation of the retinoblastoma gene, a finding consistent with cell cycle dysregulation.

Hierarchical cluster analysis was also used to interpret patterns of gene expression (12). We used Cluster and TreeView¹¹ to create the clustering and dendograms to visualize all of the genes in selected gene categories (Fig. 1). The various clusters examined involved in cellular metabolism were mRNA translation genes, oncogenes, cytoskeletal reorganization genes, and genes responsible for apoptosis. As shown in Fig. 1 (also available at our website),⁵ mesotheliomas demonstrated consistent up-regulation of gene expression in three major pathways: glucose metabolism (Fig. 1,A), mRNA translation (Fig. 1,B), and cytoskeletal reorganization (Fig. 1,D). Interestingly, a number of other pathways that might have been expected to be up-regulated in mesothelioma [genes classified in oncogenesis (Fig. 1,C) and in apoptosis (Fig. 1 E)] showed no consistent changes.

RTQ-PCR was used to confirm expression levels of 18 genes by comparing three pools of cDNA derived from RNA of mesothelioma patients to one pool of RNA from normal pleural samples. (Fig. 2). Two of the mesothelioma pools (5 patients in each pool), combined in Fig. 2, contained RNA obtained from the same patients that had been used on the original array. The third mesothelioma pool contained RNA from 5 new patients not used to generate the array data. Two genes (ubiquitin and cytochrome p450) that had equivalent expression on the arrays were used to normalize the RTQ-PCR data across samples, because the expression of genes that have traditionally considered to be housekeeping genes and used for normalization, i.e., β-actin or GAPDH, were actually highly elevated in tumor patients. As shown in Fig. 2, there was a very high level of agreement between the array and the real-time PCR data supporting the validity of the array data.

One of the most striking changes we observed in the mesotheliomas was up-regulation of many genes in the pathway involving glycolysis and the Krebs cycle (Fig. 1,A; Fig. 3): Seven of the 10 enzymes present were up-regulated with an average fold increase of 2.8. Especially prominent increases were seen in GAPDH (6.4-fold increase; P < 0.00077), LDH (5.5-fold increase; P < 0.00001), and phosphoglycerate kinase 1 (3.9-fold increase; P < 0.0011; Table 2; Fig. 3). These changes were confirmed by RTQ-PCR (Fig. 2).

These findings are consistent with observations that cancer cells maintain high aerobic glycolytic rates, and produce high levels of lactate and pyruvate despite the presence of oxygen, a phenomenon known historically as the Warburg effect (13, 14). Preferential reliance of glycolysis is correlated with disease progression in several cancers such as breast cancer, non-small cell lung cancer, uterine cancer, and hematological malignancies (14, 15), and the activities of several of the glycolytic enzymes such as LDH, hemokinase, pyruvate kinase, and phosphofructokinase have been reported to be significantly increased in cancer cells.

It is interesting to consider the mechanisms of this metabolic activation. A ChoRE, a 5′-CACGTG-3′ motif, controls transcription of several of these metabolic enzymes including hexokinase, GAPDH, pyruvate kinase, enolase, and LDH. The binding site for the ChoRE contains an E-box sequence, CACGGG; however, the transcription factors binding this site still remain poorly understood (16, 17, 18). Although glucose is a major regulator of the ChoRE promoter, two other potential participants include HIF-1 and c-myc (19, 20, 21). DNA sequence and functional analyses have revealed that the ChoRE promoter has an active HIF-1 and c-myc binding site, and that it stimulates expression of ChoRE-dependent glycolytic enzymes.

To study a possible relationship between glycolytic enzyme expression with HIF-1 and c-myc expression, we used RTQ-PCR to correlate expression of these two transcription factors with two of the up-regulated glycolytic enzymes that have the ChoRE promoter: GAPDH and LDH. We made cDNA from 4 patients who had relatively low glycolytic enzyme expression levels and 4 patients who had high glycolytic enzyme expression levels on the microarray hybridization experiments. We then used RTQ-PCR to measured their GAPDH and LDH levels, as well as the gene expression levels of HIF-1α and c-myc. There was a very strong and significant (P < 0.05) correlation between HIF-1 and LDH (r² = 0.98, Spearman coefficient) and HIF-1 and GAPDH (r² = 0.85, Spearman coefficient). The correlation with c-myc was much lower pronounced (GAPDH r² = 0.59 and LDH r² = 0.41).

Although these data are only correlative, and do not prove cause and effect, the significant association between HIF-1 levels and up-regulation of glycolytic enzymes in mesothelioma is intriguing. Hypoxia is a usual feature of many solid cancers, and has been linked to malignant transformation, metastasis, and treatment resistance (22, 23). Thus, the up-regulation of HIF-1 in tumors is common where it functions as a key transcription factor that potentially regulates 9 of the 11 glycolytic enzymes (21, 24, 25) in addition to other genes such as VEGF and erythropoietin. It is thought that HIF-1 then orchestrates the adaptation of cancer cells to hypoxia by inducing glycolysis, angiogenesis, and erythropoiesis.

One interesting clinical implication of these observations is that the marked up-regulation of glycolysis-related genes may represent the molecular explanation for the increased metabolic activity seen in mesotheliomas when imaged using PET scans using radiolabeled ¹⁸FDG, a compound that correlates directly with glucose metabolism (26). ¹⁸FDG is transported into cells and phosphorylated; however, ¹⁸FDG-phosphate is an unsuitable substrate for the next enzyme (phosphoglucose isomerase) in the glycolytic pathway. Increased hexokinase activity, together with increased glucose transporter expression in tumor cells compared with that in surrounding tissue, results in selective ¹⁸FDG accumulation in the tumor (27). Because ¹⁸FDG is labeled with the positron-emitting nuclide ¹⁸F, an image of the tumor can be seen using PET. Up-regulation of glycolytic enzymes could potentially be useful therapeutically if, or when, new treatments are developed that target enhanced glucose metabolism in tumors.

A second pathway in which many genes were markedly up-regulated in mesothelioma was the mRNA translation pathway (Fig. 1,B; Fig. 4). Protein synthesis occurs on the ribosome; however, the ribosome does not bind to mRNA directly, but must be recruited to mRNA by the concerted action of many eIFs (28, 29). Among our top up-regulated genes, 4 were ribosomal proteins and 6 were elongation factors (Table 2). As shown in Fig. 4, significant overexpression of genes was observed in almost all parts of the translation initiation pathway including eIF1γ, eIF4A1, eIF3, eIF2β, eIF3, and eIF4G1.

This up-regulation may be highly significant because there is both experimental and observational evidence linking overexpression of eIFs to oncogenesis. Experimentally, overexpression of eIF-4E (30) or eIF4G (31) in NIH 3T3 cells causes tumorigenic transformation. Cells overexpressing eIF-4E showed a 130-fold increase in VEGF protein production (32). Conversely, down-regulation of eIF4E levels in transformed fibroblasts (33) or human head and neck squamous carcinoma cell lines (34) using antisense technology leads to a loss of tumorigenicity. In observational studies, eIF4E was shown to be overexpressed in a broad range of rat tumors and cells lines compared with normal tissues (35), and in human breast and prostate carcinomas, as well as non-Hodgkin’s lymphomas (36, 37). Several additional translational initiation factors have been discovered subsequently to be up-regulated in human tumors, including: eIF2G in 30% squamous lung carcinomas (38), eIF-4AI in human melanoma cells, eIF-2B in human breast cancer cell lines, and eIF-3 in breast cancer (39), squamous cell esophageal carcinomas, and prostate cancer (40).

Many genes in the cytoskeletal reorganization pathway are also up-regulated (Table 2; Fig. 1,D). Given the epithelial differentiation of mesothelioma cells compared with normal mesothelial cells, it was not surprising to observe marked up-regulation of a number of cytokeratin genes: cytokeratin 8 (9.3-fold increase; P = 0.06), cytokeratin 18 (8.9-fold increase; P = 0.058), cytokeratin 5 (5.32-fold increase; P = 0.029), and cytokeratin 7 (2.66-fold increase; P = 0.042). The relatively high Ps suggest wide variation among tumors. Keratins constitute the major intermediate filaments in several simple epithelial tissues such as liver, intestine, and pancreas, and their presence has been used extensively in the diagnosis of tumors from epithelial and nonepithelial origin. For example, keratin 8 and keratin 18 are commonly associated with both well- and poorly differentiated carcinoma cells (41). Staining of the mesotheliomas with an anticytokeratin 18 antibody showed very strong staining of tumor cells with no staining of normal mesothelium (Fig. 5). The expression of the intermediate filament vimentin was also increased 2.3-fold (P = 0.01).

Given the known tendency for malignant mesothelioma cells to generate abundant stromal tissue (42), it was also not surprising to see laminin (3.62-fold increase; P = 0.003) and several collagen genes among the most strongly up-regulated: collagen III, α1 (6-fold increase; P = 0.004), collagen I, α2 (5.55-fold increase; P = 0.0048), collagen IV, α1 (4.48-fold increase; P = 0.003), and collagen VI, α2 (2.58-fold increase; P = 0.011). A number of these increases were confirmed with semiquantitative PCR (Fig. 2).

Actin filaments are key molecules in the regulation of cell adhesion, cell spreading, and cell configuration, and are critical for the regulation of various cell functions, including proliferation (43). Many studies have demonstrated genes that regulate the actin-based cytoskeleton are important in the oncogenic process and have been reported to correlate with prognosis in patients with various cancers (44). For example, small GTPases of the rho family (such as rho, rac, cdc42, ralA, ral-GDS, and ral-BP1) induce particular surface protrusions generated by actin-remodeling reactions that change cell shapes, and influence cell adhesion and locomotion (43, 45). In addition, the rho family of genes have roles in cytoskeletal transformation, regulation of expression of growth-promoting genes, and progression of cell cycles through the G₁ phase of the cell cycle (43, 46). Rho family proteins induce tumorigenic transformation of rodent fibroblasts (47). Our experiments demonstrate increases in expression of rho family genes including rho G (3.4-fold increase; P = 0.003), ral-A (2.4-fold increase; P = 0.00015), and rac (2.3-fold increase; P = 0.00012).

We also observed up-regulation of thymosin β-4 (3.7-fold change; P = 0.00028), a protein that binds monomeric actin, a component of the cytoskeleton, and may act as an actin buffer, preventing spontaneous polymerization of actin monomers into filaments supplying a pool of actin monomer when the cell needs filament (48). It is not clear how increased expression of thymosin β-4 might promote metastasis, but it is likely related to the need for cells to migrate (49). Thymosin β-4 has already been demonstrated to be highly up-regulated in several tumors including renal, bladder, prostate, colorectal, and thyroid neoplasms (50, 51).

In addition to those genes that were part of defined activated pathways, our analysis identified a number of other significantly up-regulated genes that might be useful in understanding the pathogenesis of mesothelioma and/or assist in diagnosis (Table 2). A subset of these was chosen for additional analysis. Up-regulation of these genes was verified by real-time PCR (Fig. 2) and/or immunohistochemistry (Fig. 5).

Gp96 (also known as adenotin, endoplasmin, tumor rejection antigen 1, gp100, grp 94, or stress-inducible tumor rejection antigen gp96) is a cytoplasmic and cell surface-expressed member of the cellular hsp family, most closely related to hsp90 (52). On the array, gp96 was 4.7 fold up-regulated in mesothelioma tissue compared with normal pleura (P = 0.0004). RTQ-PCR demonstrated a 4–10-fold difference in mesothelioma tissues compared with normal pleural samples (Fig. 2). Immunostaining showed variable gp96 up-regulation in 10 of 19 tumors (Fig. 5). None of the normal pleura demonstrated any significant gp96 staining. Gp96 has been implicated as an important cellular hsp that has been known for its ability to induce tumor-specific immunity in animals that are immunized with it (53, 54).

Our array data show marked overexpression of a chemoresistance gene called LRP. LRP gene expression was up-regulated by 5.5-fold on the array (P = 0.00001) and by 4–6-fold using RTQ-PCR (Fig. 2). LRP is a M_r 110,000 protein that is the major “vault” protein in humans. Vaults are cytoplasmic organelles that are localized to the nuclear membrane and act as a transporter, mediating nucleocytoplasmic exchange and have been shown to remove cytostatic drugs such as doxorubicin, vincristine, VP-16, Taxol, and gramicidine-D (55). Overexpression of LRP has been demonstrated to select for doxorubicin resistance in colon and non-small cell lung carcinoma cells lines (56), and elevated expression levels of LRP have been seen in colorectal and ovarian carcinomas and may serve as prognostic factors (57, 58).

Human malignant mesotheliomas are extremely resistant to chemotherapy with very low response rates to a wide variety of chemotherapeutic agents such as doxorubicin. Our data suggest that overexpression of LRP may be involved. This finding could be important therapeutically, because it has been shown that ribozymes capable of degrading LRP decrease the levels of doxorubicin that accumulate in the nucleus. If clinically useful (i.e., small molecule) inhibitors to LRP are developed, they could potentially be used to great advantage in the treatment of mesothelioma.

Galectin-3 binding protein (also known as Mac-2) is an endogenous β-galactoside-binding protein that has been implicated in cell growth, differentiation, adhesion, and malignant transformation (59). The protein has been shown to bind collagen and fibronectin, to be located in the extracellular matrix, and to promote cell adhesion and spreading by binding to β1-integrins. Galectin-3 binding protein has been demonstrated to have prognostic significance in several tumors. In head and neck squamous cell carcinomas, levels of galectin-3 binding protein contributed additional prognostic value to conventional clinical staging of patients (44). Investigators in Michigan have demonstrated expression has been correlated with advanced tumor stage in colon cancer, although direct evidence for a role in metastasis is lacking (60). Similarly in breast cancer and non-small cell lung cancer, increased expression levels have been proven to be an indicator of poor survival (61). In mesothelioma, galectin-3 binding protein demonstrated a 10-fold up-regulation (P = 0.00026) from hybridization experiments. Its role in mesothelioma tumor progression awaits additional experimentation.

One of the most highly up-regulated genes on the array was the M_r 67,000 laminin receptor, which showed an 11.6-fold increase (P = 0.0018). However, unlike the close correlation we observed between the real-time PCR data and the array data for most of our other up-regulated genes (Fig. 2), we observed only a very small (1.3–1.6-fold) increase using PCR. Interestingly, when we performed immunohistochemistry, we observed very little expression of the M_r 67,000 laminin receptor on the tumor cells but did note strong staining of blood vessels within the tumor (Fig. 5 C, arrow). In contrast, blood vessels in normal lung tissue did not show expression.

The M_r 67,000 laminin receptor is a nonintegrin protein of M_r 67,000 that was isolated on a laminin affinity column in the late 1980s (62). It binds to a cysteine-rich domain in the short arm of the laminin β1 chain. This receptor plays a role in tumor development, progression, and metastasis. For example, its up-regulation on tumor cells is associated with the malignant phenotype and prognosis in breast, lung, and ovarian cancer (63, 64, 65, 66). Given a report by Kallianpur et al. (67) that mesothelioma tissues expressed the M_r 67,000 laminin receptor, we were somewhat surprised by the lack of tumor cells staining in our tissue samples. Although we have no clear explanation, it is possible that the protease and acid treatment that Kallianpur et al. (67) used on their formalin-fixed, paraffin-embedded material changed the type of staining that we observed using acetone-fixed frozen sections and a different antibody. Instead of tumor cell staining, we observed strong expression on the vessels within the tumor, those vessels presumably involved in angiogenesis.

This vessel-staining pattern is consistent with a second function attributed recently to the M_r 67,000 laminin receptor, specifically that this molecule is involved in angiogenesis in retinal tissues (68). Little data has yet accumulated on the role of the M_r 67,000 laminin receptor in tumor angiogenesis, although a synthetic laminin peptide that binds to this receptor has been shown to inhibit experimental tumor angiogenesis (69). On the basis of our data, we would propose that up-regulation of the M_r 67,000 laminin receptor may play a role in the development of tumor vessels in mesothelioma.

Two genes highly up-regulated in mesothelioma were the VDAC 1, which was increased 6-fold (P = 0.00025), and VDAC 2, which was increased 6.5-fold (P = 0.00046). VDAC 1 was up-regulated 2–3-fold using real-time PCR. VDAC is the primary pathway for metabolite diffusion across the outer mitochondrial membrane, that in its open configuration is permeable to molecules as large as M_r 5,000 (70). VDAC has been linked recently to cellular apoptosis through its interaction with the Bcl-2 family of proteins, although the effects of these proteins on VDAC are controversial. VDACs appear to interact with Bcl-2 family members and regulate levels of apoptosis. The importance of up-regulation of VDAC in mesothelioma is currently unclear. It has been shown that mesotheliomas express relatively high levels of both bax and Bcl-x_L, and that transfection of mesothelioma cells with additional Bcl-x_L protein enhances resistance to apoptosis (71, 72). Having higher levels of VDAC may amplify any imbalance between pro- and antiapoptotic Bcl-2 family members that may be created by proapoptotic therapies.

Although differential display is not quantitative, it is interesting to compare our results with the study conducted by Gordon et al. (5). In their comparison, 14 known genes (and about an equal number of expressed sequence tags) were identified to be more highly expressed in mesothelioma than in normal pleura, including the α-folate receptor and IAP-1. Of the 5 genes that were on our array (histone acetyltransferase 1, ribosomal proteins S15a and L27a, IAP-1, and palmitoylated erythrocyte membrane protein), all showed up-regulation with a range of 2.7–4.7-fold, although the expression levels of some were quite low. The correlations in gene up-regulation between genes up-regulated in mesothelial and mesothelioma cells in culture was much lower (3, 4), likely as a result of the major changes occurring in cells during the explanation and culture process.

Gordon et al. (6) have reported recently results using transcript profiling to discover 5 genes (MRC OX-2, KIAA097, VAC-β, calretinin, and PTGIS) that were preferentially expressed in malignant mesothelioma versus lung adenocarcinomas. Our array experiments contained VAC-β (also called annexin VIII), and we similarly found it to be one of our most highly up-regulated genes (11.8-fold; P = 0.008; see Table 2). Annexin VIII is a specialized calcium and phospholipid-binding protein normally present on lung endothelium, skin, and liver (73). Annexin VIII is also expressed in acute promyelocytic leukemia cells, although not in any other lymphoid malignancies (74, 75). In acute promyelocytic leukemia, annexin VIII plays a role in signal transduction of cellular proliferation (75, 76). This gene can potentially serve as a diagnostic marker in biopsy samples.

Although microarray technology can provide vast amounts of data, there are a number of limitations and caveats that must be considered in the interpretation of this and any other such experiment. First, selection of genes that have “real” and biologically significant differences between normal and tumor samples is still an imperfect science. There are large amounts of variability at every step of the array process. Clearly using large sample numbers is an advantage to obtaining reliable results. Data analysis is also important. Rather than depending on any one normalization or gene selection technique, we used a multiplicity of approaches and selected genes that were commonly identified in multiple combinations of analysis algorithms. The goal of our gene identification strategy was not necessarily sensitivity (there are likely many gene differences that we did not detect), but specificity. On the basis of our real-time PCR validation, this goal was achieved. Of the 16 genes validated using real-time PCR (those shown in Fig. 2 plus data showing no change in ubiquitin and cytochrome p450 reductase), there was a striking similarity between the array data and the PCR data in all but 1 of the genes (up-regulation of the M_r 67,000 laminin receptor being the one exception). The validity of our results was additionally supported by showing very similar results with PCR when analyzing a pool of RNA from five mesothelioma tumors that had not been included in the original array. Thus, we feel that the normalization and gene selection protocols used accurately identified significantly changed genes and feel confident in the validity of the “non-PCR confirmed” data points.

Another limitation of our microarray hybridization experiments is they failed to predict genes that have been proven to be differentially expressed in malignant mesothelioma by reverse transcription-PCR or other molecular biology techniques. Several potential reasons explain this finding in our experiments. Microarray sensitivity for selection of differentially expressed genes is unknown (77). Many well known genes (i.e., thrombospondin, p21, NCAM, p16, and NF2) were not on our nylon microarrays (4132 genes). Furthermore, to minimize our false-positive rate of prediction of differentially expressed genes, we filtered those genes with low intensities and heterogeneous expression. As a result, we eliminated genes (i.e., platelet-derived growth factor and p53) that could have important clinical implications to maintain a rigorous selection process. As bioinformatic tools improve, investigators can use the data available publicly to reanalyze the data and extract more differentially expressed genes.

A third caveat to be considered in our study is that microdissection of tumor tissue was not performed. Thus, a mixture of cells including tumor, infiltrating WBCs, stromal cells, and tumor vessels were all analyzed. The limitations of this approach were evident from our initial analysis where we noted that a number of WBC-specific genes were up-regulated. Immunostaining with an antibody against the common leukocyte antigen CD45 confirmed that tumors and normal pleural tissues were infiltrated to various degrees by leukocytes. This confounding factor was reduced to some degree by a “virtual microdissection” (i.e., using gene ontology data to “subtract out” leukocyte-specific genes), but this process is by nature incomplete and cannot remove genes common to all cells.

Another example of potentially confusing data are illustrated by our findings with the M_r 67,000 laminin receptor. Although this receptor was up-regulated in the tumor samples, it was not expressed on the tumor cells, but on the infiltrating vessels. On the other hand, using nonmicrodissected tissues has the advantage of revealing what genes will be expressed in a standard clinical biopsy specimen where only macrodissection will be feasible. If one is looking for potentially diagnostic or prognostic genes, it may be more important to sample the entire tissue milieu including white cell, stromal, and endothelial cell genes. Clearly a number of validation steps including quantitative RNA analysis, protein measurements, and immunostaining are important when trying to assign significance to a specific gene identified in microarray data.

Microarray analysis of mesothelioma tumors revealed activation of a number of key pathways including genes involved in glucose metabolism and mRNA translation. These findings are consistent with the need of the tumor to enhance energy and protein production, and provide a molecular basis for the clinical observation that mesotheliomas show strong uptake of ¹⁸FDG on PET scanning. HIF-1 may have an important role in regulation of energy metabolism and will be the subject of additional study. Other up-regulated genes included gp96, LRP, galectin-3 binding protein, the M_r 67,000 laminin receptor (on tumor vessels), and voltage-dependent anion channels. Additional study of these proteins may be important to improving our understanding of the pathogenesis of mesothelioma, and will hopefully improve diagnostic methods and treatments.

We thank Adam Henry for valuable assistance in preparation of this article.