Distinctive Patterns of Gene Expression in Premalignant Gastric Mucosa and Gastric Cancer¹

GC⁴ is one of the most frequent malignant tumors in Asia, Eastern Europe, and South America, with an incidence in many regions ∼10-fold that seen in Western countries. It is presently unclear whether the molecular pathology of GC is similar in Western, Asian, and Eastern European countries (1, 2, 3). GC has a poor prognosis in Western countries, including Australia and United States (4), although better results are obtained in Japan, possibly attributable to aggressive approaches to early detection and surgery (5).

The molecular understanding of GC subtypes is limited, and unlike some other solid tumors, molecular features of GC are not used to individualize treatment. All existing GC classification systems are based on histology (Borrman, Mings, Goseki, WHO, and Lauren classification). The Lauren system uses cell morphology to distinguish three broad GC subtypes: IGC and DGC, which roughly correspond to well and poorly differentiated cancers, and MGC, representing a combination of the two (6, 7). IGC and DGC may result from the transformation of different epithelial cells (8) or distinct molecular changes in a common cell type. β-catenin mutation is more common in IGC (9), whereas mutation in CDH1 is more often seen in DGC, including familial GC (10, 11, 12); however, neither mutation is exclusive to either subtype. Chronic infection by Helicobacter pylori is commonly associated with all subtypes of GC but particularly to IGC. Infection results in ChG and is linked to GC pathogenesis (13, 14) through an intermediate stage involving IM (15, 16).

High throughput methods, such as cDNA and oligonucleotide microarrays, are increasingly being used to systematically compare molecular features of individual cancers to key clinical parameters (17, 18, 19, 20). Previous expression profiling studies of GC highlighted differences in gene expression between tumor and adjacent mucosa (21, 22, 23) but failed to identify gene expression patterns discriminating histological subtypes. There is increasing interest in correlating clinical outcomes with gene expression signatures (24, 25). Hasegawa et al. (26) studied IGC specifically and correlated findings with lymph node status of tumors, whereas Leung et al. (27) describes PLA2G2A (Phospholipase A2 group IIA) expression in relation to patient survival. An additional study has correlated gains in chromosome 17q and gene expression, prompting study into possible mechanisms of GC progression (28). Our study uses a larger number of GC tumor and adjacent mucosa samples to focus on the relationship of gene expression pattern to clinical, pathological, and ethnic parameters.

Tumor and normal specimens were collected from patients with GC requiring curative or palliative resection in Melbourne or in China (Shenyang or Dalian). Written informed consent was obtained before collection. Table 1 summarizes the characteristics of patient samples. Tumors were resected and divided (fresh frozen in liquid nitrogen and 70% ethanol), within 25 min ex vivo. All non-neoplastic samples were isolated from patients undergoing gastrectomy for GC. Mucosa ≥2 cm from the macroscopic tumor margin but from the same anatomical region was stripped of underlying muscle and serosa. In addition, cardia, body, fundus, and antral mucosa at maximal distance from the tumor were collected from 11 patients, pooled, and used to create a single batch of reference RNA.

Ethanol fixed tissue sections were stained with H&E and classified independently by two pathologists (S. L. and P. W.). These results were compared with the pathology record from the contributing institution. Final pathology was determined by consensus and review if necessary. Each specimen was attributed a diagnosis and scored for Lauren classification, differentiation, and percentage of tumor cells and inflammatory cells. Pathological Tumor-Node-Metastasis stage and demographic data were obtained from clinical records with patient permission. H. pylori serological testing was performed by PanBio using the PANBIO H. pylori IgG ELISA on preoperative serum collected on patients recruited by the Australian centers.

Total RNA was isolated by phenol-chloroform extraction (TRIzol; Invitrogen) and column chromatography (RNeasy; Qiagen). Total RNA (50–70 μg) was labeled using reverse transcription with Superscript II (Invitrogen) and dCTP-Cy5 (Amersham) for test samples and with dCTP-Cy3 (Amersham) for reference, similar to previous methods (29). Labeled cDNA was cohybridized in 3.125 × SSC and 50% formamide for 14–16 h at 42°C to spotted cDNA arrays printed at the PMCI Microarray Core facility.⁵ These comprise ∼10,500 elements representing 9,381 unique cDNAs (Unigene Build 144) and were printed onto superamine slides (Telechem) using a robotic arrayer (Virtek). Slides were washed in 0.5 × SSC with 0.01% SDS, then 0.5 × SSC, and finally 0.06 × SSC at room temperature and scanned using a ScanArray 5000 (Perkin-Elmer). Data were extracted using Quantarray (GSI Lumonics).

The order of sample processing was randomized to avoid systematic bias. All tumor and normal RNA samples were competitively hybridized with the common reference RNA described above.

GC and premalignant tissues were ethanol fixed, paraffin embedded, and used to create tissue microarrays. Ki-67 (DAKO Corp., Carpinteria, CA; Ref. 30), E-cadherin (E-Cad; Zymed, San Francisco, CA; Ref. 31), and TFF1 (a gift from Dr A. Giraud) immunostaining was performed either using tissue arrays which were derived from the samples used for microarray analysis (Ki67) as described previously (32), or on an independent data set of samples (E-Cadherin and TFF1). Staining was revealed with a horseradish peroxidase linked secondary antibody using the DAKO LSAB+ kit (DAKO), following the manufacturer’s instructions.

Nine representative genes (CDX1, BENE, MYO1A, CDH11, SFRP4, PDGFRB, DAB2, SPARCL1, and IGFBP7) were selected that were differentially expressed, based on hierarchical clustering of gene expression signatures. Quantitative expression analysis was performed with real-time PCR (Applied Biosystems, Foster City, CA), using oligonucleotide primers to known sequences. SYBR green intercalation to PCR products was estimated using an ABI PRISM 7000 (Applied Biosystems), according to the manufacturer’s specifications. TP63 and BIKE served as control genes, because these showed little variation in expression across the samples used in microarray experiments. Test genes were normalized to the average values obtained for the two controls. Sequences of primers were: TP63 forward primer, 5′-CCAAAGCGAGGCACCCTTA-3′ and reverse, 5′-GGAGAGTAGGCTGCCATGAGG-3′; BIKE forward primer, 5′-CTAAGACCTGGA-AATGGCCCT-3′ and reverse, 5′-CTGAGGAGGTCCCTGACCC-3′, CDX1 forward primer 5′-CCCTGACCTTCT-GGGACATG-3′ and reverse 5′-GGATGCAGAGGGTGGATAGG-3′; BENE forward primer 5′-GCATTGGAAGAAAAGGCTGC-3′ and reverse, 5′-TGGGTCCTT-CACTCCTCGC-3′; MYO1A forward primer, 5′-CTGGTCAGCGAGCATGTGAT-3′ and reverse, 5′-CACAGCCCGGTACATTTTGG-3′; CDH11, forward primer, 5′-AACAGCG-TGGATGTCGATGA-3′ and reverse, 5′-GTCTGCCTCC-TGTATTCTCGTGT-3′; SFRP4 forward primer, 5′-CCTTACAGGATGAGGCTGGG-3′, and reverse, 5′-CATGGCCTTACATAGGCTGTCC-3′; PDGFRB forward primer, 5′-TGTCCAGATGAAGCAAGGCC-3′ and reverse, 5′-CTGACCCCCAGGATGGAAGT-3′; DAB2 forward primer, 5′-ACAAAGCTGATAGCCAGACACG-3′ and reverse, 5′-CAAAGCTGGA-ACAAGGGCAG-3′; SPARCL1 forward primer, 5′-ATCATGTGCACTTCAAGAA-AATGG-3′ and reverse, 5′-ACACGTAAACCACAAAAGAGTAGCAT-3′; and IGFBP7 forward primer, 5′-ATGCTGGAGAATATGAGTGCCA-3′ and reverse, 5′-CTGAAGCCTGTCCTTGGGAA-3′.

Data were imported into GeneSpring (Silicon Genetics) and intensity-dependent normalization carried out using LOWESS (33). A group of 7383 genes satisfied filtering criteria based on the presence of significant signal in ≥80% of hybridizations. Hierarchical clustering was performed using Pearson correlation as a measure of similarity, after average linkage and median centering of values using the program Cluster, and viewed with Treeview (34). Discriminant genes and differences between groups were analyzed using two-tailed ANOVA with the Benjamini Hochberg multiple testing correction factor at P ≤ 0.01 (35), unless otherwise specified.

Non-neoplastic samples were classified histologically as ChG or IM, and tumor samples were classified as IGC, DGC, or MGC. To systematically profile the spectrum of gene expression, we compared each sample to a common reference RNA. Independent hybridizations (192) were conducted, including 37 replicate experiments performed with separate slide batches and 31 direct comparisons of tumor to adjacent mucosa from the same patient. Hierarchical clustering revealed that all repeat samples clustered together (Supporting Information). Similarly, data from multiple clones for individual genes almost invariably coclustered, indicating a high degree of experimental reproducibility.

Unsupervised clustering of the data demonstrated a major division of samples into malignant and nonmalignant groups. Moreover, there was a highly structured partitioning of samples into recognized histological groups of ChG, IM, IGC, DGC, or MGC (Fig. 1). MGC samples, which display both IGC and DGC elements histologically, fell predominantly at the interface of IGC and DGC clusters. Although expression profiling allowed good segregation of samples, several samples clustered into groups inconsistent with their histology, e.g., despite having expression signatures that overlapped with their correct group, samples ab077IM, ab156IM, ab135CG, and ab149N cosegregated in the DGC cluster. With the exception of ab156IM, the samples were derived from patients with DGC. These samples had no histologically detectable tumor cells but appear to have been misclassified because they share a strong stromal signature with DGC samples (Fig. 1, Region K, discussed below). To validate our RNA expression results, we selected nine genes and a representative cohort of samples to perform real-time PCR as a semiquantitative assessment of relative RNA abundance. Table 2 shows that there was an excellent concordance of the results of real-time PCR analysis and microarray findings.

A clear distinction of tumor and adjacent mucosa has been observed in other GC expression profiling studies (22), but we believe this is the first to distinguish between DGC and IGC. The number of genes and specimens in our data set has allowed a high-resolution analysis of GC and its reported premalignant conditions.

GC is highly endemic in Asia and Eastern Europe, presumably because of differences in the balance of predisposing genes and/or environmental risk factors relative to Western countries. We were interested to determine whether the gene expression pattern of GCs arising in Chinese nationals was similar to those in Australians. We did not observe clear segregation of GC samples based on ethnicity using unsupervised clustering (Fig. 1) or ANOVA (data not shown), indicating that the molecular features of GC are similar in Chinese and Australian samples. Although we were able to find statistically significant differences in gene expression in non-neoplastic samples based on country of origin (Supporting Information), this analysis was confounded by unequal distribution of IM and ChG samples from each country.

ChG was defined by expression patterns of groups of genes sharing similar biological function, including metallothionein or elongation factor gene clusters (Figs. 1 and 2, Region B). In particular, there was a preponderance of nuclear genes encoding mitochondrial proteins, especially those involved in the oxidative production of ATP, including enzymes that degrade fatty acids and other oxidative substrates (Fig. 2, and Supporting Information). These included a large number of genes associated with the respiratory complex, such as subunits of cytochrome c oxidase (COX), NADH dehydrogenase (NDUF), ATP synthase H+ transporting, mitochondrial F0 complex (ATP), and mitochondrial solute carrier family 25 members (SLC25).

Infection with H. pylori is almost an invariant feature of ChG in patients with GC (36), suggesting that the ChG signature may be linked to this microbe. H. pylori infection was determined by histology of pathological samples and serological testing of preoperative blood samples collected from Australian centers. We found that 80% of all Australian patients were found to be positive for H. pylori. Strains of H. pylori are diverse, and several genetic loci, including the cag pathogenicity island, babA2, and VacA genes, have been implicated in the pathogenesis of GC (37, 38, 39, 40). These genes produce a number of important pathogenic factors, including CagA, which undergoes tyrosine phosphorylation after type IV secretion into the epithelial cell (41) and deregulates SHP-2 (42) and VacA, which induces mitochondrial damage when applied to gastric epithelial cells (43). The p34 fragment of VacA localizes specifically to mitochondria, causing release of cytochrome c and apoptosis (44). Ultrastructural studies have also demonstrated that mitochondrial damage and loss are notable features of ChG (45). Our findings are consistent with loss or damage of mitochondria in ChG, leading to a high level of mitochondrial biogenesis. H. pylori is a neutralophile (46), and there may be an adaptive advantage in attenuating ATP production required by the proton pump for acid secretion or in triggering apoptosis of the acid-producing parietal cells through the release of cytochrome c. Mitochondrial damage may be important for H. pylori establishment and growth, because secretion of urease by H. pylori appears to be insufficient to adequately neutralize the local environment (46). At least two recent studies have investigated acute H. pylori infection of AGS cells in vitro. Although interesting from the point of view of gene expression changes in immortalized cultured cells, these studies acknowledge the need for in vivo experiments (47, 48). Our findings reflect chronic infection with H. pylori in an in vivo setting, rather than acute infection over hours as described by these studies. Although our findings highlight an impact on mitochondria, among the many postulated effects of H. pylori infection (49), expression analysis of normal and gastritic mucosa from infected and uninfected individuals is needed to determine this conclusively.

Consistent with the acquisition of an intestinal phenotype, the IM group showed prominent expression of genes characteristic of intestine (Figs. 1 and 2, Regions E and F). These include CDX1 (a homeobox gene that regulates gut morphogenesis and differentiation and a recognized marker of IM), MYO1A (encodes a brush border myosin found in intestinal epithelium), MTP or microsomal triglyceride transfer protein (encodes an enzyme involved in lipid metabolism), cholecystokinin (encodes a gut hormone), and Villin1 (a major component of intestinal microvilli; Fig. 2, Region F). Interestingly, many of the markers of mature intestinal cell function are lost or show diminished expression in IGC, possibly reflecting a less differentiated state of tumor compared with metaplastic tissue. Other genes of interest that define the IM group include FAT (Fig. 2, Region F), a putative tumor suppressor, and TFF1 (Fig. 2, Region E). There was a marked attenuation of TFF1 expression in more than half of all DGC and IGC. Somatic mutations in TFF1 are implicated in gastric lesions (50), and TFF1 knockout mice develop GC (51). We stained tissue microarrays with TFF1 antisera to validate our RNA expression findings. Down-regulation of protein in malignant tissue was seen in IGC and particularly in DGC, consistent with our microarray findings (Fig. 5). The down-regulation we describe in IGC was also observed in another microarray study specifically investigating IGC (26).

A tightly coexpressed group of genes that includes several carcinoembryonic antigens and PSG (Fig. 2, Region E) exemplify genes that are up-regulated in both IM and IGC. PSG proteins have been shown to interact with Kruppel-like factor 4 in placental development, and KLF-4 showed similar expression with PSG in our samples. TGFα was persistently up-regulated in IM and IGC (Supporting Information, Fig. 1 Region E). Given its role as a mitogen (52), TGFα may play a role in the progression from IM to IGC. It remains to be addressed whether genes in region E, such as TGFα, represent markers of terminal differentiation of intestinal epithelium that retain expression after transformation or are early participants in the transformation process.

In Fig. 3, we used ANOVA with the Benjamini Hochberg multiple testing correction factor (P < 0.01) as an alternative algorithm to identify genes that are most significantly differentially expressed between ChG and IM. Of 189 genes discovered by ANOVA, almost all were present in the regions B, E, and F in Fig. 2, providing validation of our approach to identifying differentially expressed genes between ChG and IM.

A notable feature of the IGC samples was the relatively strong expression of proliferative markers (Figs. 1, Regions D1–D3 and 4). These included a large number of genes required for G₂-M transition, DNA replication, spindle assembly, and chromosome segregation. The list includes genes shown previously to be overexpressed in GC and other cancers, such as topoisomerase (TOP2A; Ref. 53) and CDC25B (54, 55, 56). Although it is expected that tumor cells should express genes involved in cell division, it was noteworthy that a number of the DGC samples did not show a similar elevation of these markers (Figs. 1 and 4). In principle, the infiltrative nature of DGC tumors and consequently lower proportion of tumor cells in DGC samples could result in dilution of proliferative markers. However, staining of tissue arrays containing the corresponding samples with anti-Ki-67 antibodies showed that the IGC group generally had a higher proliferation score compared with DGC (Figs. 4 and 5).

Rapidly dividing cells might also be expected to up-regulate protein synthetic machinery to meet the demands of cell growth; however, examination of Fig. 1 suggests that is not the case. Region C comprised a tightly coregulated group of ribosomal protein genes, which were more strongly expressed in DGC and ChG than in IGC (Fig. 4). The expression patterns of a largely overlapping set of ribosomal subunit genes have been found to vary substantially between different tumor types (57), indicating that not all tumors strongly express genes involved in protein synthesis. Furthermore, expression profiling of germinal center B cells, which are among the most rapidly proliferating cells in the body, has revealed that high level expression of cell cycle markers was associated with relatively low level expression of many of the same ribosomal subunit genes present in region C (58). It appears there are circumstances where proliferation occurs at the expense of cell growth, such as in germinal center B cells and IGC. Our findings accord with the high nuclear:cytoplasmic ratio characteristic of malignant cells and typical of IGC.

DGC is an infiltrating cancer characterized by a marked stromal reaction (6). This was reflected in the expression profile, which included genes encoding components of the extracellular matrix (collagens, biglycan, osteoglycin, proteoglycan, matrix metalloproteinases, cadherin 11, Thy-1 SERPINS, and fibrillin), as well as smooth muscle and cell adhesion molecules (Figs. 1 and 4, Region K; also in Supporting Information Regions K1, K2, L, and M). A number of genes encoding extracellular proteins was up-regulated in DGC that are linked to positive and negative regulation of cell proliferation, including secreted frizzled-related protein 4, NMB, DAB2, SPARC, Wnt5a, DLG5, DKK3, FDZ1, and IGFBP7. Increased SPARC expression has been linked to more aggressive tumor types in many tissues (59, 60) but has not been implicated previously in GC. One study shows abrogation of tumorigenicity by inhibiting SPARC function in human melanoma cells (61). Wnt proteins are involved in cell transformation and signal transduction via frizzled and frizzled-related receptors. Consistent with a previous report (62), Wnt5a was up-regulated in both DGC and IGC (Fig. 4). Surprisingly, dickkopf 3 (DKK3), a putative negative regulator of Wnt signaling through the β-catenin pathway (63, 64, 65, 66), and the Wnt-signaling protein frizzled homologue 1 (FDZ1) clustered together. Discs large homologue 5 (DLG5), a putative negative regulator of cell growth (Fig. 4) cosegregated with DKK3 and FDZ1. DAB2 is also implicated as a negative regulator of growth (67). The co-overexpression of both positive and negative regulators of cell growth in DGC samples suggests a complex interplay between their stromal and epithelial elements. There was a remarkable concordance between a cluster of genes expressed in Region K and a gene cluster associated with SPARC in breast cancer (57). It seems likely that this group of genes reflects a characteristic signature associated with the reaction of stroma to infiltrating tumor cells.

Down-regulation of E-cadherin (CDH1), which participates in cell-cell contact through the formation of tight junctions, is a recognized feature of sporadic DGC (10). In addition, germ-line mutations in CDH1 are associated with high penetrance familial DGC (11). Consistent with this, we found a significant attenuation of CDH1 expression in DGC relative to other tissues (Fig. 4) but not in IM. Immunohistochemical staining of tissue microarrays with E-cadherin antibodies confirmed the pattern of expression seen in our microarray experiments (Fig. 5). This finding agrees with previous studies (68). Occludin, a major membrane component of tight junctions, and actin-related protein 2/3, coclustered with CDH1 and were similarly down-regulated in DGC. It is recognized that a loss of Occludin and carcinoembryonic antigen 1 has been associated with loss of cell polarity in prostate cancer and could be involved in tumorigenesis (69). Occludin is expressed in well-differentiated intestinal epithelium (70), consistent with our findings in IM (Fig. 4, E-Cad Region). It will be interesting to test whether these genes or other components of tight junctions are mutated in GC and result in function analogous to CDH1 mutations.

A group of genes that showed pronounced differential expression between samples but independent of histological subtype were genes involved in inflammation (Fig. 1, Regions H–J). These regions included a comprehensive set of B- and T-cell markers, pro-inflammatory cytokines, and mediators of interferon signaling and interferon-induced genes (Fig. 1, and see Fig. 4 for a subregion of H). These expression changes were consistent with infiltration of plasma cells histologically (data not shown).

In summary, we found that expression profiling of a large set of GC and non-neoplastic specimens was able to distinguish premalignant and malignant subtypes of GC, and the structure of the major gene clusters reflect identifiable biological functions. It is noteworthy that the structure of the cluster reflects an expression continuum from ChG to GC. In this regard, the clustered gene expression data molecularly recapitulates the model of GC etiology first proposed by Correa (15). Our work on gene expression changes in the transition from proposed premalignant mucosa to GC provides a valuable database for functional studies that investigate the pathogenesis of GC. The identification of apparent mitochondrial stress in ChG has important implications for the major molecular effects of H. pylori infection on parietal and epithelial cells and possibly the initiation of the premalignant process of GC. IGC is characterized by a proliferative signature in contrast to DGC and suggests that patients with IGC may particularly benefit from antiproliferative chemotherapeutic agents. Finally, it is apparent that tumor cells and a large part of the DGC signature correspond to a dynamic interaction between tumor and stroma that calls for further investigation.

We thank members of the Bowtell Lab in assistance with manuscript preparation and Dr. A. Giraud for providing the anti-TFF1 antibody.