Gene Expression in Colorectal Cancer^{1, 2}

CRC⁴ is one of the most frequent cancers in the Western world. The prognosis in advanced cases is poor, and more than one-third of the patients will die from progressive disease because the overall survival is about 40% (15–65%) after 5 years (1). Current treatment is only curing a fraction of the patients and has the best effect on the early-stage disease. CRC was one of the first major epithelial cancers in which molecular alterations were described to occur in a systematic fashion during disease progression (2). It was hoped that this knowledge, e.g., of p53 inactivation, could form the basis for new treatments, and several of these, such as gene therapy, have been tried without any improvement in survival. Therefore, new studies aimed at identifying further molecular alterations in CRC and thereby new diagnostic biomarkers and new treatment targets are needed.

Normal cell growth depends on a balanced expression of growth-promoting and growth-suppressing genes. If growth-promoting genes are activated to a state of hyperfunction, either by mutation or quantity, they are termed “oncogenes,” which exert a positive effect on cell growth. One common example is K-ras, which is activated in CRC and many other tumors (3). Growth-suppressing genes are defined as “tumor suppressors” (4) and are commonly lost following Knudson’s “two-hit” hypothesis (loss of one allele and inactivation of the other allele by mutation or promoter methylation; Ref. 5). The best-characterized tumor suppressors are p53 on chromosome 17p13.1, which shows a LOH frequency of 75% in CRC, Smad4/DPC4 (deleted in pancreatic cancer locus 4) on chromosome 18q21, and APC (adenomatous polyposis coli) on chromosome 5q21–22 (6, 7, 8). Most frequent genomic losses in CRC are found at chromosome 18q21 (9), and most frequent gains are found at 20q13 as determined by comparative genomic hybridization (10). Identification of new oncogenes and tumor suppressors would be of benefit to our understanding of the biology of CRC, and they might constitute new targets for therapy.

Several techniques have been used to monitor gene expression, but most of the methods used are time- and labor-intensive. The recently developed microarray technique permits us to investigate the expression of thousands of genes within a single patient sample. Recently, Kitahara et al. (11) and Notterman et al. (12) published expression array-based studies analyzing CRC tumors with corresponding noncancerous colonic epithelia. They identified important new clusters of genes that showed alterations in cancer tissue. In the present study, we have used a similar methodological approach based on Affymetrix GeneChip microarrays with immobilized oligonucleotide probes, and we monitored the expression of 6,500 known genes and 35,000 ESTs, representing unknown genes or genes with unknown function. Our aim was to identify candidate tumor suppressor genes and oncogenes relevant for CRC development. Furthermore, genes with a specific behavior at certain Dukes’ stages were identified as potential classifying genes for those stages.

We found that 908 known genes and 4155 ESTs changed vastly from normal to tumor tissue or in one of the Dukes’ stages. Based on filtering of these data, we ended up with identification of 226 known genes and 157 ESTs of very high relevance for CRC, covering a spectrum of candidate oncogenes and tumor suppressors as well as classifiers. The data were validated by microarray analyses of an independent set of 25 CRC samples and by real-time PCR of the samples used for pool analysis.

Hierarchical cluster analyses clustered normal tissue with Dukes’ A and clustered Dukes’ B with Dukes’ C. Dukes’ D was separated from the other tissues but was connected with the Dukes’ B/C cluster, clearly pointing out the difference between normal tissue and invasive tumor tissue.

The genes discovered were assigned to functional classes and to chromosomal location to gain insight into the functional status of CRC cells and to identify chromosomal “hot spots”. Some of the hot spots for down-regulation were shown by microsatellite-based LOH analysis to be frequently lost and were compared with a large microsatellite-based study on 55 cases of sporadic cancer from Finnish patients that showed common loss of regions possibly harboring tumor suppressors.

The genes presented in this study not only represent new biomarkers but also represent potential targets for application of chemical genetics such as, for example, target-based screenings (13). According to the functional categories presented in this study, the genes selected could be screened rapidly with a limited number of ligands of interest. Application of such target-based screens will probably identify potential novel therapeutic targets usable for prevention or therapy of CRC.

A detailed description of materials and methods is presented as online supplementary data.

Left-sided colorectal tumor samples from the upper rectum, sigmorectum, or sigmoideum from Dukes’ stages A−D were obtained fresh from surgery, taken from the luminal aspect of the tumors in the surgical specimens, transferred immediately to a solution containing SDS and guanidinium isothiocyanate, snap-frozen in liquid nitrogen, and stored at −80°C. The Dukes’ classification of the clinical stage of disease was applied according to the following criteria: (a) Dukes’ A, tumor confined to the bowel wall without penetration of the muscularis propria (equivalent to stage I or modified As-Co stage A and B₁); (b) Dukes’ B, tumor has penetrated the muscle wall and possibly infiltrated the pericolic or colorectal fat, with no detectable metastatic lymph nodes (equivalent to stage II or As-Co stage B₂); (c) Dukes’ C, lymph node metastasis is detectable (equivalent to stage III or As-Co stages C₁-C₂); and (d) Dukes’ D, metastases detected in distant organs (e.g., liver; equivalent to stage IV or As-Co stage D). Paired “normal” control samples were obtained from the oral resection margins of the operative specimens by taking mucosal biopsies from the luminal aspect of the bowel wall. Informed consent was obtained from patients to use their specimens and clinicopathological data for research purposes. All tumors were sporadic because none of the patients belonged to families with hereditable CRC or other cancers. The project was approved by the local scientific ethical committee.

Total RNA was isolated from about 100 mg of single tissue samples using a Polytron homogenizer or FastPrep FP120 (Bio101 Savant) followed by treatment with RNAzol (WAK-Chemie Medical) according to the manufacturer’s instructions. Five pools were made from equal amounts of total RNA from the following tissues: (a) normal (n = 6; median patient age, 69 years; tissue from oral resection margin of three Dukes’ stage A and three Dukes’ stage B tumors,); (b) Dukes’ stage A tumors (n = 5; median patient age, 72 years); (c) Dukes’ stage B (n = 6; median patient age, 74 years); (d) Dukes’ stage C (n = 6; median patient age, 67 years); and (e) Dukes’ stage D (n = 4; median patient age, 61 years). An independent set of samples that had not been included in the pools was used for single GeneChip analysis. This included five samples each from normal tissue [tissue from oral resection margin of three Dukes’ stage B tumors, one Dukes’ stage A tumor, and one Dukes’ stage C tumor (median patient age, 68 years)], Dukes’ stage A tumors (median patient age, 68 years), Dukes’ stage B tumors (median patient age, 79 years), Dukes’ stage C tumors (median patient age, 68 years), and Dukes’ stage D tumors (median patient age, 55 years). Detailed clinical information on the samples as well as the approximate percentages of the volume fractions of tumor cells and stromal cells, estimated semiquantitatively by an experienced pathologist using paraffin-embedded diagnostic tissue sections (4–12 sections/tumor), are shown in Supplementary Table 7. All samples had a volume fraction showing at least >53% malignant tumor cells, and two-thirds of the samples had a volume fraction showing >80% malignant tumor cells. The diagnostic samples from paraffin-embedded sections contained both normal and tumor tissue as well as transmural tissue from the colon wall (submucosa and muscularis). The estimated percentage of tumor cells is a conservative estimate because tissue used for RNA extraction was from the most superficial tumor-rich areas, avoiding most of the deeper stroma-containing layers, and the percentage of tumor cells is probably higher in the arrayed samples as in the screened, paraffin-embedded, diagnostic histological tissue sections.

Reverse transcription was performed on 12 μg of pooled total RNA for 1 h at 42°C using a T7-oligo(dT)₂₄ primer and Superscript II reverse transcriptase (Life Technologies, Inc.). Second-strand cDNA synthesis was performed for 2 h at 16°C using Escherichia coli DNA polymerase I, DNA ligase, and RNase H (Life Technologies, Inc.), followed by incubation in 50 mm NaOH/0.1 mm EDTA for 10 min at 65°C, leading to degradation of rRNA and tRNA. After phenol-chloroform extraction in vitro, transcription was performed for 6 h at 37°C using Bio-16-UTP (Enzo), Bio-11-CTP (Enzo), and T7-Megascript Kit (Ambion). The cRNA was purified on RNeasy spin columns (Qiagen) followed by fragmentation for 35 min at 95°C.

Samples were analyzed on GeneChips (Affymetrix Inc., Santa Clara, CA) HuGeneFL (Hum6.8k all) with 6,800 genes and on EST chips 35K (subA−D) with 35,000 ESTs. To check the quality of each sample with regard to GAPDH and β-actin, 15 μg of labeled cRNA were run on Test2 arrays (Affymetrix). Expression CHIPs were hybridized with 15 μg of labeled cRNA for 16 h at 45°C under rotation. CHIPs were stained in an Affymetrix Fluidics station with streptavidin/phycoerythrin, followed by staining with an antistreptavidin antibody and streptavidin/phycoerythrin. The CHIPs were scanned with a HP-Laserscanner, and the data were analyzed with GeneCHIP software and Microarray Suite Software 4.0 (Affymetrix). Each microarray was scaled to “150” as described by Thykjaer et al. (14). Expression patterns of normal tissue were compared with carcinoma tissue derived from Dukes’ stage A−D.

A total of 42,843 data sets (7,129 data sets from HU6800 arrays and 35,714 datasets from 35K subA−D EST arrays) were sorted according to stringent criteria (see overview in Fig. 1). For filter 1, data sets were excluded if (a) the Abs Call was absent (A) in all five pools; (b) the Diff Call was not changed (NC) in all four comparisons or the Diff Call was not changed in three of four comparisons and one comparison was called marginal decreased or marginal increased (MD or MI); or (c) it was marked as AFFX internal control. In the remaining data sets, the absolute value of the sort score had to be ≥0.5 in at least one of the four comparisons. This resulted in 908 known genes (12.7%) and 4,155 ESTs (11.6%; data not shown). For filter 2, candidates were selected by a combination of absolute analysis (Abs Call: P, present; M, medium present) and comparison analysis (Diff Call: NC, not changed; I, increased; D, decreased; MI medium increased; and MD, medium decreased). As a convention, losses of expression from normal to all Dukes’ stages [represented by PAAAA, present (P) in normal and absent (A) in Dukes’ A−D] always had to be accompanied by Diff Call D in all Dukes’ stages and gains of expression by Diff Call I in all Dukes’ stages’ (APPPP). Losses of expression at one or more Dukes’ stages had to be accompanied by Diff Call D at that stage (e.g., PPAPP). For the known genes, the exclusion limit of the Avg Diff was arbitrarily set to ≥50, and for the ESTs, it was set to be ≥300, referring to Avg Diff of (a) normal data if decreased or (b) tumor data if increased from normal to tumor. Fold change criteria of genes present in normal and tumor (PPPPP) were set to be ≥3- or ≤−3-fold (≥ 5- or ≤−5-fold for ESTs), respectively in at least one of the comparisons. Avg Diff and fold change criteria for the selection of ESTs were arbitrarily set to a higher value to exclude false positives. Also, ESTs below these criteria are regarded as candidates of interest and will be subjected to a closer investigation in the future.

Twenty-nine genes with fold changes between <3 and >−3 and 21 ESTs with fold changes between <5 and >−5, respectively, were nevertheless regarded as important candidates because they were accompanied by Diff Calls decrease or increase in all four comparisons. This resulted in the most remarkable 226 known genes (3.1%) and 157 ESTs (0.4%; Supplementary Table 4, a and b, respectively). Purported or hypothetical proteins derived from ESTs were identified using the NCBI UniGene database.⁵ Functional categories of known genes or EST-derived homologous proteins were attributed using databases OMIM, SWISS-PROT and TrEMBL.⁵

RNA from 25 single samples (normal colon tissue and Dukes’ stages A−D, five samples each, not included in the pools) was analyzed on HuGeneFL (Hum6.8k) GeneChips. One hundred comparison analyses corresponding to 25 comparisons for each Dukes’ stage were run comparing each normal sample versus each tumor sample using Affymetrix Microarray suite 4.0, Microdatabase 2.0, and Datamining Tool 2.0 (DMT 2.0). t tests and Mann-Whitney tests with an exclusion limit of P < 0.05 and based on Avg Diff analyses were made using Datamining Tool 2.0 software from Affymetrix. Genes showing an increase or decrease in >52% of all comparisons within at least one Dukes’ stage (>13 of 25 comparisons) were selected. (Detailed results on this array set will be published elsewhere).⁶

Using the log Avg Diff of selected genes and EST obtained from pool analysis, clustering was performed by Cluster 2.11 and Treeview 1.5 [Life Sciences Division, Lawrence Berkeley National Laboratory, Department of Molecular and Cellular Biology, University of California at Berkeley, Berkeley, CA (15)].

cDNA was synthesized from single samples analyzed previously as pools on GeneChips. Reverse transcription was performed on 1 μg of total RNA for 1 h at 42°C using a T7-oligo(dT)₂₄ primer and Superscript II reverse transcriptase (Life Technologies, Inc.). Identical amounts of cDNA from each normal sample or each tumor were pooled, and real-time PCR analysis was performed on selected genes using the primers shown in Supplementary Table 1. Triple determinations were performed on a Lightcycler using the FastStart DNA Master SYBR Green I kit (Roche), and GAPDH levels were determined using Lightcycler Primer Set Human GAPDH (Search LC GmbH). All samples were normalized to GAPDH as described by Mensink et al. (16). Avg Diffs from GeneChip analyses were compared with the normalized real-time PCR data.

DNA was extracted from microdissected tumor and normal colon tissue. Microdissection was performed using ×100 magnification in a microscope and serial sections of 4-μm-thick tissue-TEK or 10-μm-thick paraffin-embedded tissues. Sections were rinsed in Xylol, and DNA was extracted using a Puregene DNA extraction kit (Gentra Systems, Minneapolis, MN). DNA from the microdissected tissue was analyzed for allelic deletions (LOH) using microsatellite markers. Sequences of the fluorescence-labeled primers for nine different chromosomal locations are listed in Supplementary Table 2. PCR products were analyzed using an ABI Prism 377 DNA sequencer and Genescan software. LOH was defined as a loss of one allele in case of heterozygosity. In case of homozygosity, scoring was done by peak height, and a decrease of >50% was defined as LOH. MSI was defined as the presence of new bands after PCR amplification of tumor DNA that were not present in corresponding normal DNA. If MSI was present, data could not be used for LOH detection.

The 5′ CpG island methylation of selected down-regulated genes (i.e., H1F2, H2BFB, sigma 3B, and hsp70) was examined using a PCR-based methylation assay (17, 18). One μg of genomic DNA from nonmicrodissected tumor and corresponding normal tissue was digested with one of the methylation-sensitive restriction enzymes HapII (identical to HpaII), CfoI, or FnuDII or with the methylation-insensitive restriction enzyme MspI for 2 h using 10 units of restriction enzyme. Subsequently, 50 ng of digested or undigested DNA were amplified using the primers listed in Supplementary Table 3. Exon 2 of the p16 gene has been shown to be methylated in bladder cell lines (17). A methylation analysis of this exon in the bladder cell line T24 was carried out as a positive control for detection of methylation by the PCR-based methylation assay.

Approximately 500 bp upstream from the start codon of the promoter region of Acyl-CoA dehydrogenase (Z80345 SCAD) were sequenced as described previously by Gregersen et al. (19) to find an explanation for down-regulation of expression of this gene. Apart from a known polymorphism at position −171, no mutation was detected in DNA from normal and microdissected tumor tissue of 16 different patients (data not shown).

By using oligonucleotide arrays, we analyzed the expression level of approximately 6,500 known genes and 35,000 ESTs in biopsies from the left colon (sigmoid and upper rectum). We compared pools of RNA from tumors to a pool of RNA from normal biopsies taken from the oral resection margin. Each pool consisted of equal amounts of four to six samples with identical histology and staging.

In an effort to reduce the very large data set, we applied different filtering methods (Fig. 1). As the first step, we eliminated genes and ESTs that were absent in all samples and genes that did not show a variation from normal to tumors (see “Materials and Methods”). That left us with approximately 12% of the initial genes and ESTs. We then applied a set of more stringent criteria, resulting in a reduced number of informative alterations (Fig. 1). In the supplementary data, the 226 known genes (Supplementary Table 4a) and 157 ESTs (Supplementary Table 4b) are listed that fulfilled all criteria. The most interesting genes and ESTs from the selected candidates that have been validated with single sample analysis are shown in Table 1 (53 genes) and Table 2 (46 ESTs).

The group of 226 known genes that alter their expression from normal to cancer was subdivided into candidate tumor suppressors that are lost or much reduced in cancer tissue (70 genes) and candidate oncogenes that are gained de novo or much increased in cancer tissue (88 genes). A third set of genes showed a behavior that was related to only one or two of the Dukes’ stages, peaking at this stage or being severely reduced at this stage. Those genes were named candidate Dukes’ classifiers (68 genes; 7 for Dukes’ A, 18 for Dukes’ B, 26 for Dukes’ C, and 17 for Dukes’ D) because they may be used to identify a specific stage of the disease.

To validate the gene alterations on an independent microarray-based data set, we analyzed 25 single samples from normal mucosa and Dukes’ A, B, C, and D (5 samples each), all from the same location in the left colon. The data from the pools were cross-validated against these 25 arrays and showed matching alterations in 72%, corresponding to 161 genes. The criterion used was conformity in >52% of samples (>13 of 25 comparisons within at least one Dukes’ stage; t test and Mann-Whitney test, P < 0.05). The genes fulfilling this criterion are marked with an asterisk in Supplementary Table 4, a and b. Of these, 73 genes should be regarded as general tumor markers because they showed conformity in >52% of comparisons of consecutive Dukes’ stages (28 genes in Dukes’ A, B, C, and D; 40 genes in Dukes’ A, B, and C; and 5 genes in Dukes’ B, C, and D). The correlation was surprisingly good in the case of 90 genes because their alteration could be reproduced in >80% of comparisons in at least one Dukes’ stage (>20 of the 25 comparisons made). In all of the nonconfirmed cases, an alteration was detected, but it was either only valid for <52% of comparisons or not significant upon Mann-Whitney analysis.

To validate the expression alterations using an independent method, we used real-time PCR on pools of 4–6 RNA samples/Dukes’ stage analyzed previously on microarrays (Fig. 2). We used triple determinations and normalization based on GAPDH level. A remarkably good correlation for the 10 known genes and 5 ESTs analyzed was found between the two methods as shown in Fig. 2, indicating that the array-based determinations were highly reproducible.

In case of the ESTs, we detected 45 candidate tumor suppressors and 32 candidate oncogenes as well as 81 ESTs (Supplementary Table 4b) that could be used as classifiers of Dukes’ stages (15 Dukes’ A, 18 Dukes’ B, 16 Dukes’ C, 24 Dukes’ D, 3 Dukes’ A and B, and 5 Dukes’ B and C). In the case of the ESTs, no cross-validation was done with single arrays; however, five of the ESTs were analyzed by the independent real-time PCR method (Fig. 2). As with the known genes, a remarkably good correlation was found between the alterations detected by the two methods on the same RNA sample.

The identity of 93 of the 157 ESTs was established using the NCBI Unigene database as either already known genes or genes encoding hypothetical proteins. The function of the proteins encoded by these 93 genes as well as by the 226 known genes was estimated by the OMIM, SWISS-PROT, and TrEMBL databases. Based on the classification of Lander et al. using 12 functional categories (20), these functions were divided into 15 main functional categories and 15 subcategories, resulting in 29 functional groups (miscellaneous group excluded) used as subheadings in Supplementary Table 4, a and b, and in Fig. 3. Multifunctional proteins were categorized according to their most important function (e.g., an ion transporter is found under transporters, although it is also a membrane protein), and proteins with unknown function were listed in a separate group. For the calculations that follow, we only used genes with a known function because the purported function ascribed to some ESTs may still be error-prone. The numerically most prominent group of genes that change expression during cancer progression of the colon encode proteins related to metabolism (22%), in particular, mitochondrial metabolism, followed by genes encoding proteins related to transcription and translation (11%), cellular processes (9%) including cell cycle proteins and proteins involved in growth and differentiation, cell adhesion (8%), protein folding and degradation (7%), transport (6%), immune system (6%), and nucleic acid interaction (6%). Remarkably, proteins related to apoptosis or signaling and signal transduction were only rarely altered. For some groups, it was remarkable that the alterations detected were mainly either up-regulation or down-regulation (Fig. 3). Most of the genes encoding proteins related to cell cycle, methylation, DNA and RNA metabolism, translation, cell adhesion, or proteases were up-regulated. Genes encoding proteins that were mainly down-regulated belonged to the groups membrane and protein trafficking, lipid metabolism, and membrane proteins and kinases/phosphorylases. Nuclear-encoded mitochondrial proteins showed a distinct behavior because the 15 genes encoding these were all down-regulated in at least one of the tumor stages (Table 3). Fourteen of these 15 genes were cross-validated on the 25 single samples and confirmed to be decreased significantly in all samples (P ≤ 0.05). One of these, the SCAD gene, was also shown to be decreased by real-time PCR validation (Fig. 2).

Using the log Avg Diff of pool analysis, a hierarchical cluster analysis clustered normal tissue together with Dukes’ A and clustered Dukes’ B together with Dukes’ C. Dukes’ D was separated from the other tissues but was connected with the Dukes’ B/C cluster, clearly pointing out the difference between normal tissue and invasive tumor tissue (Fig. 4). Six selected clusters are shown in Fig. 4, and the complete cluster analysis is shown in Supplementary Fig. 1. Gene clustering did not clearly identify larger gene families.

We used the NCBI UniSTS database to precisely map as many of the genes and ESTs as possible. Interestingly, many of the candidate oncogenes and candidate tumor suppressors clustered to well-defined chromosomal regions (Table 4). A detailed mapping of each of these hot spots for expression alteration is shown in Supplementary Table 5. Apart from clusters of genes (marked with a black line) showing a similar alteration in expression, some genes encoding candidate oncogenes and candidate tumor suppressors were located next to each other on the same chromosomal arm, as seen, for example, at chromosomes 4q and 6p (Supplementary Table 5). For several locations, it was found that genes within a cluster were very close or arranged next to each other, as shown in an exemplary fashion in Table 5 for chromosomes 5, 10, 12, 13, 15, and 19, respectively, and in Supplementary Table 5. One possible meaning of this could be a common regulation of genes close to each other.

Several mechanisms can lead to altered expression of a gene, and we decided to look at two of these, methylation of proximal promoter CpG islands and loss of an allele. Both mechanisms can lead to a reduced gene expression; methylation can do so because of reduced binding of transcription factors, and loss of an allele can do so because of a gene dose effect. We searched databases for promoter CpG islands for four genes, H1F2, H2BFB, sigma 3B, and hsp70, that all were vastly reduced in tumor tissue compared with normal tissues. All four promoter sites were methylated at specific sites as detected by a PCR-based methylation assay and validated by sequencing of PCR products (data not shown). However, we did not detect any difference in this pattern between tumor tissue and normal tissue as shown in Supplementary Table 6. Because we did not use microdissected tissue for analysis, methylation involvement in down-regulation cannot be excluded here.

Allelic loss in Danish samples was determined by the use of microsatellites on microdissected tumor tissue followed by comparison with normal tissue in 24 sets of tissues, 16 of which had been analyzed previously on single GeneChips. Analysis of a locus close to TN and trans-Golgi p230 on chromosome 3p22-p21.3 showed 25% LOH, and analysis of a locus close to Cdx1 homeobox transcription factor on chromosome 5q32-q33 showed 17% LOH. Analysis of a locus close to ESTs AA171913 and AA151674 (carbonic anhydrase XII) on chromosome 15q22 showed 13% LOH, and analysis of a locus close to sigma 3B and aminopeptidase N/CD13 on chromosome 15q25 showed 28% LOH. No LOH was found on chromosome 6p21.3, where tenascin-XB1 is located. We concluded that loss of an allele could be one of several mechanisms involved in the down-regulation in these chromosomal areas. Through collaboration, we got access to a Finnish CRC patient material in which LOH had been scored throughout the genome by the use of 372 microsatellites in 55 patients with sporadic CRC (21). Ten chromosomal areas (1p36, 4q21, 5q31, 6p19, 12q13, 14q, 15q11, 17p13, 18q11, and 22q13) showed LOH in >25% of the patients and had adjacent candidate genes detected on microarrays. A comparison with the present expression analysis data showed that in all 10 cases, some or all adjacent candidate genes showed down-regulation. As an example, the microsatellites D15S153 and D15S127 showed 38% and 33% LOH, respectively, in the Finnish material and surrounded the candidate tumor suppressors sigma 3B and aminopeptidase N/CD13. This correlated well with our own findings described above. Microsatellites D17S849 and D17S938 were found to show 52% and 56% LOH, respectively, in the Finnish samples and surrounded the candidate suppressors very-long-chain acyl-CoA dehydrogenase (VLCAD) and EST AA447145 (KIAA0399 protein) as shown in Supplementary Table 5. These four genes are located relatively close to their corresponding marker, which increases the likelihood that these genes are lost in carcinogenesis. We hypothesize that matching regions of LOH and decreased or lost gene expression could represent hot spots for new tumor suppressors.

Oligonucleotide microarray analyses of pools of approximately 6,500 known genes and 35,000 ESTs led to the identification of 226 known genes (3.1%) and 157 ESTs (0.4%) that seemed to be of interest for CRC. For the known genes, 72% could be reproduced by microarray analysis of 25 different single tumor samples not included in the previously analyzed pools, and for both known genes and ESTs, it was possible to find a remarkable correlation with real-time PCR. The number of varying genes detected agrees with recent findings, in which cluster analysis and mathematical selection of expression profiles identified about 2.5% and 1.8% of the genes detected on the microarray to differ between adenocarcinoma and normal colon tissue (12). Our samples for array analysis were not microdissected as in a previous publication (11) and did contain RNA from stromal components such as histiocytes, endothelial cells, muscle cells, and inflammatory cells. However, because of the recent literature on the important signaling between tumor cells and surrounding stromal and inflammatory cells (22), we believe that it would lead to a substantial loss of information if only tumor cells were analyzed. Furthermore, it was recently shown that working with tissue samples that were not microdissected did not influence the ability to cluster adenomas, carcinomas, and normal tissue into distinct trunks (10). In that study, 18 adenocarcinomas and paired normal samples were analyzed (12) using the same microarray system as in the present study. Although analyses were applied to CRC tumors in general without referring to the grading and location of the samples, and candidate genes were selected according to cluster analyses based on fold changes, those data are highly consistent with our data. Examples of up-regulated genes in common are melanoma growth-stimulatory activity (MGSA), transcription factor IIIa (TFIIIA), and S-adenosylhomocysteine hydrolase (AHCY). Examples of down-regulated genes in common are guanylin, TN, adipsin/complement factor D, or carbonic anhydrase IV. Some of those genes have also previously been confirmed by serial analysis of gene expression. Beyond the candidates in common, we present several additional candidates validated by single sample analyses to change significantly, such as MDP4/MDP7 microsomal dipeptidase, aminopeptidase N/CD13, or NGAL/LCN2 lipocalin 2 (neutrophil gelatinase-associated lipocalin). Analyses of EST chips resulted in known genes and hypothetical proteins with and without predicted functions. Those hypothetical proteins could be of use as new molecular markers for CRC and deserve further investigation.

Some of the selected candidate genes can be regarded as progression markers and molecular predictors (changes in at least two consecutive Dukes’ stages) or Dukes’ classifiers (major changes in one Dukes’ stage only). Examples of progression markers are phosphoenolpyruvate carboxykinase (PCK1) and monocyte-derived neutrophil chemotactic factor (MDNCF; IL-8). The latter, which is an angiogenic cytokine, was shown in a CRC cell line-based study to be produced by the tumor cells (23).

Furthermore, our results on the dramatic reduction in expression of molecules like CgA and TN correspond to previous findings. CgA is a neuroendocrine differentiation marker, and only a minority of stage III and stage IV CRC patients (11% and 22%, respectively) showed a positive staining for CgA (24). TN, a stromal component of tumors and a participant in proteolytic processes through its binding to plasminogen, is regarded as a tumor suppressor. A low plasma TN level is related to a shortened survival (25, 26). The good correspondence between our data and those from previous studies indicates the validity of the large number of new alterations that we detect in gene expression. We only analyzed samples from the left colon and upper rectum because of the large difference often found between the left and the right side of colon (27).⁷

The human genome contains about 30,000–70,000 protein-encoding genes. Lander et al. (20) classified about 40% of these proteins into 12 functional categories. In the present study, the distribution on functional categories was, in some cases, far from what could be expected. As an example, genes involved in cellular processes constitute approximately 2.2% of the genome but 9% of the genes that show changes during CRC progression. This indicates a cancer-specific change of certain functional groups rather than a selection due to the distribution in the genome.

Remarkably, many genes that showed a decrease or loss of expression in at least one Dukes’ stage were mitochondrial proteins. Microarray analysis of 25 CRC single samples confirmed that 14 of 15 selected genes found by pool analyses showed a statistically significant decrease from normal tissue to tumor in at least one Dukes’ stage (t test, P < 0.05). Three of those genes were found to be decreased at all Dukes’ stages. Rhodanese (thiosulfate sulfurtransferase) is involved in forming iron-sulfur complexes and cyanide detoxification and catalyzes the transfer of the sulfane atom of thiosulfate to cyanide to form sulfite and thiocyanate (PROSITE PDOC00322). 3-Hydroxy-3-methylglutaryl CoA synthase (HMG-CoA synthase) catalyzes the condensation of acetyl-CoA with acetoacetyl-CoA to produce HMG-CoA and CoA (PROSITE PDOC00942). The mitochondrial form is responsible for ketone body biosynthesis. SCAD (acyl-CoA dehydrogenase) was found to be significantly decreased in all Dukes’ stages [Dukes’ A, 7-fold (P = 0.018); Dukes’ B, 5-fold (P = 0.008); Dukes’ C, 17-fold (P = 0.002); and Dukes’ D, 4-fold (P = 0.024)]. SCAD is a FAD flavoprotein and catalyzes the β-oxidation of Butyryl-CoA to Acetyl-CoA (PROSITE PDOC00070), and SCAD deficiency results in an increase of butyric acid. Butyrate is the primary source of energy for colonocytes and was shown to act in a contradictory fashion, depending on the availability of other energy sources: at low concentrations, butyrate stimulates growth under glucose- and pyruvate-depleted conditions; whereas it causes apoptosis at the same concentrations in the presence of glucose and pyruvate (28). We hypothesize that one possible cause of CRC carcinogenesis might be located in the mitochondria. Mitochondrial DNA accumulates more damage due to less efficient repair systems in the mitochondria compared with those in the nucleus. Although all of the selected mitochondrial proteins are nuclear encoded, the mitochondrial function is altered directly by expression changes of nuclear-encoded proteins involved in electron transport and oxidative phosphorylation and altered indirectly because oxidative phosphorylation is linked to many pathways of intermediary metabolism, as discussed by Augenlicht and Heerdt (29).

Decreased RNA transcription could be due to hypermethylation of the proximal promoter as shown previously, e.g., the mismatch repair enzyme MLH1 (30). A similar mechanism could cause the down-regulation of the candidate tumor suppressors described in the present study. Consequently, we investigated the 5′ CpG island methylation of four selected down-regulated genes, two ESTs (i.e., H1F2 and H2BFB) and two known genes (i.e., sigma 3B and hsp70). The results of the methylation analyses demonstrated that the 5′ region of all four genes is methylated at specific sites in both normal and tumor tissue. However, we did not detect any difference in this pattern between tumor tissue and normal tissue.

Chromosomal instability and MSI in sporadic CRC are thought to constitute two major pathways for CRC progression. With regard to chromosomal instability, the combination of loss of an allele and inactivation of the other allele by methylation or mutation is considered a general mechanism for inactivation of tumor suppressors. As we identified clusters of down-regulated genes at certain chromosomal locations, we hypothesized that loss of an allele in the same location would further strengthen the likelihood that we had identified candidate tumor suppressors. Thus, we made two approaches: one in which we analyzed microsatellites located in those areas; and one in which we compared our locations with a large study of 55 Finnish sporadic tumor patients scrutinized with 372 microsatellites. Due to the diversity of the genetic background, those comparisons are not universally valid, but the colocation of clusters of candidate suppressors and a raised LOH frequency might strengthen the likelihood of defining new tumor suppressor locations.

Surprisingly, we found several genes of interest located on chromosome 15q, although this is reported to be a region infrequently affected by alterations in CRC. Neogenin at 15q22.3-q23 has been reported to be generally involved in genetic disorders (31, 32), and Park et al. (33) proposed thrombospondin 1 (THBS1) to be a new tumor suppressor gene on chromosome 15q21.1. THBS1 (U12471) was found to be down-regulated (Avg Diff = 83 in normal tissue; pattern, PAPAA; decreased in Dukes A–D). Other genes, e.g., integrin ITA3, a receptor for THBS1, are concomitantly decreasing. Down-regulation of expression of EST AA171913 (carbonic anhydrase XII, CA12) on chromosome 15q22 and aminopeptidase N/CD13 (APN/CD13), sigma 3B protein (adaptor-related protein complex AP-3, sigma 2 subunit) and CIB (calcium and integrin-binding protein) on chromosome 15q25-q26 coincides with a LOH frequency of 38% and 33% in Finnish samples and 13% and 28% in our Danish samples, respectively. Thus far, neither information about the role of chromosome 15q25 nor an involvement of APN/CD13 or sigma 3B protein in CRC carcinogenesis is found in the literature, but according to the data provided here, both genes might resemble novel potential tumor suppressors. APN/CD13 is a member of the peptidase family M1 (zinc-binding metalloproteinase), resembling a type II membrane surface antigen glycoprotein, and catalyzes final protein degradation by removal of single amino acids of small peptides. ANP/CD13, described to be expressed on colon carcinoma Caco-2 cells (34), was found to be substantially decreased in renal cancer tissues (35), and, surprisingly, highly expressed APN/CD13 probably plays a role in the invasion and metastasis of prostate cancer cells (36). Sigma 3B protein, which facilitates the budding of vesicles from the Golgi membrane, is involved in trafficking to lysosomes and might play a role in the recognition of intracellular, tyrosine-based sorting signals (37, 38). Keeping in mind that the basic definition of a tumor suppressor reveals one copy to be lost and the other to be inactivated by methylation or mutation, future research should be directed toward methylation analysis and sequencing of these genes.

It is remarkable that both up-and down-regulated genes occurred in clusters along the chromosomes. In a few cases, this was separated by individual genes behaving in a contradictory fashion, but this does not remove the impression of coregulation of downstream-located genes. Whether this is due to the use of common transcription factors, opening or closing of the double strand, or other events is not known at present. This area deserves additional study because it might reveal important mechanisms for tumor progression.

As can be seen from Supplementary Table 4a, most known genes (70%) that are up-regulated or down-regulated do so during the transition from normal to early-stage Dukes’ A tumors: 37 genes are generated de novo (pattern, APPPP), 30 are lost from normal to tumor (26 PAAAA and 4 PPAAA), 51 are increased in all Dukes’ stages, and 40 are decreased in all Dukes’ stages. Far fewer genes are changing their level of expression during the progression through the different Dukes’ stages. This indicates that the basic properties of tumor cells are acquired in the early tumor stages and that only minor changes, probably those involving the stromal components (22), are needed later on. From a therapeutic point of view, this is important because the same targets seem to be present at most Dukes’ stages.

We thank Bente Devantier, Hanne Steen, Inge-Lis Thorsen, and Susanne Bruun for excellent technical assistance; project nurses Birgitte Gustafsson and Edith Kirkedahl for the collection of CRC tissue samples; and Margrethe Kjeldsen (Research Unit for Molecular Medicine, Skejby, Aarhus, Denmark) for sequencing of the SCAD promoter region.