Oral cancer in vivo gene expression profiling assisted by laser capture microdissection and microarray analysis

Abstract

Large scale gene expression profiling was carried out on laser capture microdissected (LCM) tumor and normal oral epithelial cells and analysed on high-density oligonucleotide microarrays. About 600 genes were found to be oral cancer associated. These oral cancer associated genes include oncogenes, tumor suppressors, transcription factors, xenobiotic enzymes, metastatic proteins, differentiation markers, and genes that have not been implicated in oral cancer. The database created provides a verifiable global profile of gene expression during oral carcinogenesis, revealing the potential role of known genes as well as genes that have not been previously implicated in oral cancer.

Introduction

High-throughput technologies, such as DNA microarrays, comprehensively profile and monitor gene expression in biological processes, including cancer. While this approach has been applied to hematopoietic tumors (Alizadeh et al., 2000; Golub et al., 1999), its application to the major type of human cancer, solid tumors, has only been reported for tumor homogenates and cell lines (Alon et al., 1999; Perou et al., 2000; Sgroi et al., 1999). The cell-specific profiling of solid tumor gene expression has been hampered by the inability to procure specific pure cell populations. This obstacle was recently overcome through the development of laser capture microdissection (LCM), which allows the harvesting of specific cells from complex tissues such as solid tumors.

In this study we demonstrate the application of LCM to the harvesting of normal and tumor cells from a solid tumor site, oral cavity cancer. The mRNAs were linearly amplified by three rounds of T7 RNA polymerase reaction, biotinylated, and then hybridized to HuGeneFL^® microarrays. Analysis of the hybridization outcome revealed that 26 to 40% of the genes queried were present in all samples analysed. Data analysis revealed intriguing clues to the biological pathways involved in oral malignancy. These include genes associated with metastatic/invasion, transcription factors, oncogenes/tumor suppressor genes, differentiation markers and members in the xenobiotic pathway. To validate the GeneChip^® data, we selected three genes that are consistently altered (5/5) in the metastatic/invasion pathway (collagenase, urokinase plasminogen activator (UPA) and cathepsin L) and validated their differential expression by real time quantitative PCR. Our approach should be applicable to any solid tumor gene expression profiling study using currently available microarrays to reveal relevant mechanistic pathways and events in normal and pathological processes.

Results and discussion

Gene expression profiling of normal and tumor oral epithelia

The use of LCM to harvest cells from their native tissue environment and the use of high-density oligonucleotide probe arrays to identify gene expression differences between normal and malignant oral epithelial cells provide powerful means to decode the molecular events involved in the genesis and progression of head and neck/oral cancer. We isolated paired normal and malignant epithelium from five snap frozen biopsies (10 samples total). Most of these patients have a history of smoking and alcohol consumption, which are the major etiological causes of oral cancer (see Materials and methods). The quality and quantity of isolated RNA was examined by reverse transcription polymerase chain reaction (RT–PCR) of five cellular maintenance gene transcripts of high to low abundance (glyceraldehyde-3-phosphate dehydrogenase; tubulin-α; β-actin; ribosomal protein S9; uniquitin C) (Ohyama et al., 2000). The quantity of isolated RNA was also assessed with RiboGreen RNA Quantitation Reagent and kit (Molecular Probes, Eugene, OR, USA) using spectrofluorometry (Bio-Rad, Hercules, CA, USA). Only those samples exhibiting PCR products for all five cellular maintenance genes were used for subsequent analysis. The biotinylated cRNA from the 10 samples (normal and cancer) were further used to hybridize the Affymetrix Test-1 probe arrays to determine cRNA quality and integrity. The arrays contain probes representing a handful of maintenance genes and a number of controls (Ohyama et al., 2000). Analysis of the arrays confirmed the RT–PCR findings. cRNA linearly amplified from human oral cancer tissue produced no nonspecific or unusual hybridization patterns and the transcripts for the maintenance genes were detected. The 5′ region of the RNA was degraded but enough 3′ transcript was intact to proceed for hybridization using the HuGeneFL probe arrays. In addition, probes synthesized on the arrays are biased to the last 600 bp in the 3′ region of the transcripts. Yields from the LCM and amplification steps are shown in Table 1a. Linear amplification of the total RNA began with ∼100 ng of total RNA.

Table 1a LCM, RNA isolation and amount of cDNA after two rounds of T7 amplification

Table 1b summarizes the hybridization outcome of the five paired cases of oral cancers. The per cent transcript detected ranged from 26 to 40%, indicating satisfactory quality and representation of the harvested RNA. Note that the difference between the normal and cancer samples from each patient is very similar, indicating little variability among each pair, suggesting that the quality of the RNA isolated from the normal and tumor epithelium is similar.

Table 1b Per cent transcripts detected in normal and tumor tissues

Microarray hybridization results

Differential gene expression using GeneChip^® analysis software revealed 404 probe sets changed in the majority of the cases (3/5). Among the 404,211 were increased in tumor and 193 were decreased in tumor, compared to normal. There were 39 probe sets that changed in all five cases (Table 2a). Sixteen of them were increased in tumor and 23 were decreased in tumor, compared to normal. Table 2b lists a subset of the differentially expressed genes grouped into biological pathways known to be relevant in carcinogenesis.

Table 2a Thirty-nine genes whose expression changed in 5/5 cases

Table 2b Representative sample of differentially regulated genes

Our data revealed that many known genes involved in neoplasia are differentially expressed in the five paired cases of oral cancer. Our analysis also revealed members of known biological pathways whose expression are altered during oral carcinogenesis. These include metastatic and invasion pathways, transcription factors, oncogenes and tumor suppressor genes, and differentiation markers (Table 2b). Of particular importance are the differentially expressed genes that are not yet functionally characterized or genes that have not been studied by classic methods in head and neck/oral carcinogenesis. One such example is neuromedin U (Nmu), which is downregulated in 5/5 tumors (Szekeres et al., 2000). Nmu is a poorly understood protein that manifests potent contractile activities on smooth muscle cells. Recently, two G-protein coupled receptors (NMU1 and NMU2) have been identified to interact with Nmu with nanomolar potency (Fujii et al., 2000; Raddatz et al., 2000). Our data provide strong evidence that Nmu is relevant in the development of oral malignancy and suggest the need for further study of the role of Nmu (down regulated expression in tumor) in carcinogenesis.

In order to validate our findings, three metastatic pathway genes whose expression are consistently altered in the five paired cases of oral cancer, were selected. Real-time quantitative PCR (RT–QPCR) in conjunction with the TaqMan specific probe system or SYBR^® Green system was used to validate the expression levels of interstitial collagenase (a member of the MMP's involved in metastasis), urokinase plasminogen activator (UPA, associated with metastasis) and cathepsin L (a member of the serine proteases). Comparison of the microarray and RT–QPCR data revealed that they approximate each other (Table 3). The actual comparative data is for collagenase is graphically shown in Figure 1a while Figure 1b shows the gel electrophoresis results of the collagenase RT–QPCR products. Similar data were obtained for UPA and cathepsin L (Table 3, gel electrophoresis data not shown). We have further validated a number of other high and low genes including Neuromedin U, GST, cytochrome P450, ALDH-9, ALDH-10 and Wilm's tumor-related protein (data not shown). Figure 2 is a schematic highlighting our findings in the proteolysis pathway that may contribute in the development of oral cancer.

Table 3 Comparison of per cent increases for three upregulated genes measured by GeneChip and real-time quantitative PCR data

Figure 1

Comparison and validation of microarray data by RT–QPCR. (a) Comparison of gene expression data (from GeneChip^©) and by RT–QPCR for collagenase. (b) Visualization of actual by RT–QPCR products by agarose gel electrophoresis

Figure 2

Proteolysis, remodeling of the extracellular matrix and cell migration pathway in oral cancer based on differential gene expression data. One of the central themes in this schematic is the activation of plasmin from plasminogen through the action of urokinase plasminogen activator (UPA). Plasmin is the main activator of matrix metalloproteinases (MMP) such as MMP-1, 3, 9 and 13. Activation of insulin growth factors (IGF) by plasmin and internalization of IGF-2 receptor by cathepsins are also of significance for continuous and uncontrolled cell division. Inhibin beta alpha is another molecule that has a function in activating MMP-2 which also activates MMP-9 and MMP-13. In our analysis, UPA as well as several metalloproteinases, cathepsin L, inhibin alpha beta and IGF-2 receptor are upregulated in the tumors. MT-MPP=Membrane Type Matrix Metalloproteinase

SOM and hierarchical clustering analyses

The microarray data, though voluminous, can be analysed by pattern recognition (clustering) software to aid in deriving lists of genes that distinguish and characterize disease versus normal biopsies, thus shedding light on molecular genetic profiles and ultimately the mechanism of the disease under study. Techniques used for clustering include self-organizing maps (SOM), Bayesian, hierarchical, and k-means. SOM was selected for our analysis because of advantages in initial exploration of the data allowing the operator to impose partial structure on the clusters (Tamayo et al., 1999). Other advantages of SOM include good computational properties, computational speed and easy implementation. SOM analysis was applied to the microarray data on the five paired cases of oral cancers. The clusters graphically represent gene expression patterns across all 10 samples (normal and tumor), each cluster differing in gene number and grouping. This method provides candidate set of genes whose differing expression activity can be used to distinguish normal and tumor cell behavior.

By SOM analysis, 178 transcripts were found to be differentially expressed between tumor and normal tissues. An important observation is that many of the differentially down-regulated genes are known to be important enzymes in the xenobiotic metabolic pathway (Jourenkova-Mironova et al., 1999; Katoh et al., 1999; Park et al., 1997; Sato et al., 1999). These include cytochrome c oxidase subunit Vb (coxVb), gamma-aminobutyraldehyde dehydrogenase, microsomal glutathione S-transferase (GST-II), aldehyde dehydrogenase 7 (ADH7), COX C VIII, ALDH8, EPH2 cytosolic epoxide hydrolase and ALDH10. Further data analysis revealed that other xenobiotic pathway genes, not included in this cluster, were also down-regulated in all five cases, suggesting perhaps a general downregulation of xenobiotic pathway genes during oral cancer development.

The xenobiotic pathway is of importance in the degradative metabolism of both foreign/native toxic and carcinogenic products. Phase I and II xenobiotic enzymes are two key sequential steps in the metabolism of toxic substances including alcohol and tobacco products. It is interesting to note that most of the five cases of oral cancer were from heavy smokers and drinkers. These data may suggest that key regulatory events were altered in the xenobiotic pathways during oral carcinogenesis that may contribute to the increased susceptibility towards carcinogens such as tobacco and alcohol, the two major etiological factors for oral carcinogenesis.

Using Matlab analysis, 117 transcripts were identified to be differentially expressed between normal and tumor cells. Hierarchical clustering is shown in Figure 3. The distinct clustering of the normal samples, from the tumor samples suggests that LCM procured pure, homogenous samples. It is interesting to note that the tumor samples 1B and 3B clustered more tightly together than the other three tumor samples. The sample size is too small to draw any conclusions for this however. A larger study is in progress to determine if this tight sub-clustering reflects a developmental stage of oral cancer. Based on the outcome of three analytical methods (GeneChip, SOM and Matlab), ∼600 candidate oral cancer genes were identified. Of this comprehensive set, 27 of the differentially expressed genes were identified by all three methods (Table 4). Of the 600 candidate genes, 41% were detected at low levels, 1–5 copies per cell.

Figure 3

Hierarchical clustering was done on intensity values standardized by dividing by root mean square. Cosine correlation of similarity coefficient and complete linkage clustering classified the samples as shown. Normal ‘A'; Tumor ‘B'

Table 4 Differentially expressed genes identified by all three methods

Shillitoe et al. (2000) and Leethanakul et al. (2000a,b) have created expression libraries of human oral cancer cell lines and LCM-generated oral cancer tissues. Their studies revealed 52 genes to be differentially expressed at more than twofold in at least three of the cancer tissue sets. Of these 52 genes, 26 were present on the Affymetrix GeneChip^®. By our analysis of these 26 overlapping genes, 18 were called absent (not detectable) in both normal and tumor samples (DP-2/U18422; TIMP-4/U76456; VEGF-C/U43142; FGF3/X14445; FGF5/M37825; FGF6/X63454; IGFBP5/M65062; EGF cripto protein CR1 and 2/M96956; APC/M74088; ERK6/X79483; GDI dissociation protein/U82532; MAP kinase p38/L35253; MKK6/U39657; MEKK3/U78876; Frizzled/L37782; FZD3/U82169; Dishevelled homolog/U46461; Patched homolog/U43148;); one gene shows no difference between normal and tumor tissues (cyclin H/U11791); one gene was upregulated in 5/5 tumors (beta1-catenin/X87838); three genes were upregulated in 4/5 tumors (thrombospondin2 precursor/L12350; inhibitor of apoptosis protein/U45878; Caspase 5 precursor/U28014) and one gene was upregulated in 3/5 tumors (MMP-10/X07820); one gene was downregulated in 4/5 tumors (RhoA/L25080). Finally one gene was upregulated in two tumors, downregulated in two tumors and called absent in the fifth oral cancer (TRAF2/U78798).

Of the 52 genes, two genes were detected present only through our LCM/GeneChip^® analysis. They are human SPARC/osteonectin (J03040) and 5T4 oncofetal antigen (Z29083), which are consistently altered in the same manner in all five oral cancers examined.

Of interest is that a number of genes were identified by either our LCM/oligonucleotide microarray approach or the LCM/cDNA library approach (Leethanakul et al., 2000a,b; Shillitoe et al., 2000) to be highly expressed/upregulated in oral cancer tissues. These include: ferritin heavy polypeptide I, urokinase plasminogen activator, ATP-binding cassette transporter, interleukin-1 receptor antagonist and keratin 4.

In addition, there are genes that were differentially expressed and detectable in the cell line study (Shillitoe et al., 2000), not in the Head and Neck CGAP (HNCGAP) libraries (Leethanakul et al., 2000a,b), but were detected present in our dataset. Good examples of these genes are the collagen type 1 alpha 2 genes and the heat shock protein 70 kD gene. An example of a gene that was not identified by either LCM approach (HNCGAP libraries or our method), but detected present in the cell line filtered cDNA microarray analysis is the transforming growth factor alpha gene, suggesting perhaps the elevated expression of this gene maybe associated with in vitro culturing.

The different outcome of the various studies are likely reflective of the experimental approaches and methods of analyses. First, by using LCM-generated RNA, contamination of heterogeneous cellular elements is avoided. Second, sample number and the type of microarray used in the respective studies may be relevant to the discrepancies. Third, the stage of the tumor, source and anatomical site of the oral cancers, and handling methods can further result in different gene expression levels. In our study, the detection of 39 cellular genes consistently altered in 5/5 different paired cases of human oral cancer lends strong support to the experimental approach using LCM-generated RNA, linearly amplified by T7 RNA polymerase and subsequently analysed by high-density oligonucleotide GeneChip^® probe arrays. These oral cancer associated genes will now be tested to determine their usefulness as classifiers to predict the normal/malignant nature of oral epithelial tissues. The biology associated with these genes could also be explored to evaluate their role in oral cancer development. A number of these genes are secretory proteins that are upregulated in cancer tissues and could be evaluated as biomarkers of oral malignancy. These include osteonectin, ferritin, cathepsin L, proteoglycan (secretory granule) and oncofetal trophoblast glycoprotein. Our results also indicate that our approach is applicable to the molecular analysis of solid tumor, providing a means for obtaining information about consistent molecular alterations that advance both the understanding of the basic biology of this tumor as well as the clinically relevant aspects of the molecular epidemiology of oral cancer. Our data supports the use of LCM-derived RNA to be used on microarrays and that array hybridization coupled with hierarchical and non-hierarchical analysis methods provide powerful approaches for identifying candidate genes and molecular profiling. A larger study is underway to validate these findings and improve our understanding of the molecular changes associated with oral cancer.

Materials and methods

Matched normal and malignant human oral cancer biopsies

See Table 5 which summarizes the key demographic data of the five cases of human oral cancers used in the study.

Table 5 Summarization of the key demographic data of the five cases of human oral cancers used in the study

RNA isolation, linear amplification (aRNA) from laser capture microdissection (LCM)-generated cells and target sample preparations

These procedures were carried out as previously described (Ohyama et al., 2000). Normal mucosa was obtained from the contralateral side of the patient's oral cavity.

Hybridization of biotinylated cRNA to Test 1 and HuGeneFL^® probe arrays

The cRNA was fragmented as described by Wodicka et al. (1997) All array washing, staining and scanning was carried out as described in the Gene Expression Manual (Affymetrix, Inc. 1999) (Tamayo et al., 1999).

GeneChip^® probe arrays

The probe sets consist of oligonucleotides 25 bases in length. Probes are complementary to the published sequences (GeneBank) as previously described (Lockhart et al., 1996). The sensitivity and reproducibility of the GeneChip^® probe arrays is such that RNAs present at a frequency of 1 : 100 000 are unambiguously detected, and detection is quantitative over more than three orders of magnitude (Redfern et al., 2000; Warrington et al., 2000 Warrington et al., 2000). In this set of experiments with oral cancer samples, the bacterial transcript (BioB), spiked before the hybridization at concentration of 1.5 pM which translates to three copies per cell (based on the assumption that there are 300 000 transcripts per cell with an average transcript length of 1 kb), were called present in nine out of 10 experiments (Lockhart et al., 1996). Array controls, and performance with respect to specificity and sensitivity are the same as those previously described (Lockhart et al., 1996; Mahadevappa and Warrington, 1999; Wodicka et al., 1997). Information regarding the genes represented on the arrays used in this study can be found at www.netaffx.com.

GeneCluster/Self-Organizing Maps (SOM)

For GeneCluster analysis, we input gene expression levels and geometry of nodes. Before the computation of the SOM, two preprocessing steps took place. First, a filter was applied to exclude genes that did not change significantly across the pairs. Genes were eliminated if they did not show a relative change of x=2 and an absolute change of y=35, (x,y)=(2,35). Second, normalization of expression levels across experiments was carried out, thus emphasizing the expression pattern rather than the absolute expression values. Data was normalized using GeneChip software. Description of normalization procedure can be found on pp. A5–14, GeneChip Expression Analysis Technical Manual, 1999 (Tamayo et al., 1999).

Real time quantitative PCR (RT–QPCR)

The cDNA product of the reverse transcription was used as the template for the RT–QPCR. For the RT–QPCR reaction we used iCycler IQ^TM Real Time PCR detection system (Bio-Rad, Hercules, CA, USA) with TaqMan specific probes and primers for Cathepsin, and SYBR^® Green buffer and reagents (Perkin Elmer/Applied Biosystems Foster City, CA, USA) for Urokinase Plasminogen Activator and Collagenase I (Heid et al., 1996). For designing the specific primers and probes we used PE/ABD Primer Express software as well as MacVector. Primer sequences used are: Collagenase forward: 5′-ACACGGAACCCCAAGGACA-3′; Collagenase Reverse: 5′-GTTTTGTTGCCGGTGGTTTT-3′; UPA forward: 5′-GCACCATCAAACAAACCCCCTTAC-3′; UPA reverse: 5′-CAGACAGAAAAACCCCTGCCTG-3′; Cathepsin L forward: 5′-CAGTGTGGTTCTTGTTGGGCT-3′; Cathepsin L reverse: 5′-CTTGAGGCCCAGAGCAGTCTA-3′. The final PCR products were run on 2% minigel to ensure single product amplification during the PCR assay.