Transcriptome Profile of Human Colorectal Adenomas

In developed countries, sporadic adenomatous colorectal polyps are found in roughly one third of asymptomatic adults below the age of 50 who undergo colonoscopy. Depending on their characteristics (multiplicity, size, histologic features, and degree of dysplasia), these lesions can be associated with a substantial risk of recurrence (up to 60% at 3 years) and the development of advanced neoplastic disease (reviewed in ref. 1 and references therein). It has been estimated that 15% of all adenomas measuring ≥1 cm will progress to carcinomas within 10 years of their detection (2).

Although adenomatous polyps are not the only precancerous lesions in the colorectum, they are the most common, and they are the precursors of most of the cancers in this organ. In these neoplasms, the transformation process begins in the epithelial crypts and seems to result from qualitative, quantitative, and spatial subversion of the Wnt signaling pathway, the physiologic regulator of epithelial homeostasis (3-5). This adenoma-carcinoma pathway of tumorigenesis is characterized by mutations involving various components of this pathway (e.g., APC, whose germ-line mutations are responsible for familial adenomatous polyposis; CTNNB1, which encodes a subunit of the cadherin protein complex known as β-catenin; and Axin, the gene encoding a multidomain scaffold protein that is essential for β-catenin degradation). The result of these mutations is an accumulation of β-catenin, first in the cytoplasm and then in the nucleus, where it associates with DNA-binding proteins of the T-cell factor (TCF)/lymphoid enhancer factor family, transforming them from transcriptional repressors into transcriptional activators that affect the expression of numerous genes involved in epithelial homeostasis.

Although the key role played by adenomatous polyps in colorectal tumorigenesis is widely acknowledged, the gene expression changes that trigger or accompany their development have never been comprehensively studied. We therefore conducted a transcriptomic analysis of prospectively collected colorectal adenomas using a standardized oligonucleotide microarray covering the entire human genome. This study not only provided new information that is fundamental for future molecular characterization of these precancerous lesions but also allowed us to identify a putative marker of colorectal tumorigenesis.

The focus of our study was the adenoma-adenocarcinoma pathway of colorectal carcinogenesis, which is closely linked to deregulation of the Wnt signaling pathway. To gain insight into the early steps of this process, we confined our investigation exclusively to sporadic, pedunculated colorectal adenomas (type 0-Ip of the Paris classification; ref. 6). Nonpolypoid and sessile polypoid lesions were not included because in some cases their transformation is believed to proceed along nonadenomatous pathways (7). Details on our case selection criteria are provided in Materials and Methods.

Thirty-two pedunculated adenomatous polyps, each with matched samples of normal mucosa, were prospectively collected from 28 patients (Table 1). The total number of synchronous and previously excised adenomas was <3 in 18 of 28 patients and 3 to 15 in the remaining 10. In this latter subgroup, the absence of APC- or MYH-associated multiple adenomatosis had been confirmed by genetic testing of lymphocyte DNA. Histologic analysis of one polyp (case NM) revealed superficial infiltration of the submucosa, but this case was not excluded because the region sampled for microarray analysis was clearly adenomatous. (As noted below, this finding was consistent with the results of hierarchical cluster analysis shown in Supplementary Fig. S1.).

Analysis of microarray data for the 32 adenoma/normal mucosa tissue pairs revealed that 31,033 of the probes were expressed in one or both of the tissue groups. The normal tissues were effectively segregated from the adenomas in four unsupervised analyses of the expression levels of these genes [hierarchical clustering, principal component analysis (PCA), correlation analysis, and correspondence analysis (CA); see Materials and Methods for details; Fig. 1]. In a separate analysis, these two tissue groups were also unequivocally distinguished from a previously described set of 25 colon cancers (8), which we reanalyzed for this study with the same microarray used to characterize the adenomas and normal mucosa (Supplementary Fig. S1).

Almost half of the expressed probes (15,059 of 31,033) displayed significant expression changes in adenomas. Those with fold changes ≥2 (1,190 probes up-regulated and 2,469 down-regulated in adenomas) were subjected to gene ontology analysis to identify the biological processes involved in the transition of normal mucosa to adenoma. The most significant results of this analysis are listed in Supplementary Table S1. The processes that were most markedly overrepresented among genes that were up-regulated in adenomas included mitosis, DNA replication, and spindle organization. Down-regulated genes were predominantly involved in host immune defense, inorganic anion transport, organ development, and inflammatory response, although a small group of genes involved in the latter process was up-regulated in adenomas (Supplementary Fig. S2).

We then analyzed the transcript levels of 319 genes believed to be components of the complex Wnt signaling pathway (Supplementary Table S2). Sixty-six of these genes (21%) were not expressed in either the normal or adenomatous tissue, and 34% were expressed similarly in both tissue groups. The remaining 144 genes displayed significantly altered expression in adenomas, and 78 of 144 displayed fold changes of ≥2.

A supervised extension of CA (9), canonical CA (CCA), was then used to identify possible correlations between gene expression patterns and clinical or pathologic variables. Four of the variables considered (adenoma diameter, colon segment of origin, degree of dysplasia, and adenoma recurrence; see Table 1) were clearly associated with distinct clusters of expression profiles (Fig. 2, variables in A and clusters for adenoma diameter in B; more details in the legend to this figure). The profile of adenomas measuring >20 mm could be easily distinguished from those of smaller (≤20 mm) adenomas. As shown by CCA and visualized on the corresponding CCA score plot (Fig. 2B), the centers of the three adenoma size clusters are distributed along the principal CCA axis (the vertical axis in Fig. 2B, the most important axis of separation of the expression profiles) in a definite order, with increasing diameters corresponding to progressively higher CCA scores. The variable large adenoma diameter was closely correlated with the vertical CCA axis (i.e., its vector “d>20mm” in Fig. 2A is almost parallel to this axis). It is interesting to note that the same correlation can be observed for the variable high-degree dysplasia (i.e., represented in Fig. 2A by vector “Hd”). This finding confirms the expected correlation between larger diameters and higher-degree dysplasia.

The CCA plot of the 11,709 modeled probes (loading plot, not shown) suggested that the distinction between the three size groups of adenomas is due to a complex network of relatively small changes in the expression of numerous genes (as opposed to marked changes involving a limited number of genes). Nevertheless, to maximize the use of the extensive data sets, we selected the 500 probes with the highest loading scores along the CCA axis 1 and isolated a set of genes whose expression changes displayed significant positive or negative correlation with adenoma size (Supplementary Table S3). Although their association with adenomas must be validated in a larger series, these are the expression changes most likely to play causal roles in the progression of these tumors.

It should be mentioned that normal mucosa from the sigmoid colon had an expression profile that differed significantly from that of tissues from other colon segments (Fig. 2A). This finding will be explored in a future study conducted on a large series of normal mucosa samples from different colorectal segments.

The transcriptional profile of the 32 adenomas was thoroughly analyzed to identify genes likely to be involved in the development and evolution of these lesions. One of the first features that attracted our attention was the marked up-regulation of KIAA1199 (Supplementary Table S4), a gene encoding a protein with unknown function. Its overexpression was striking in all colorectal adenomas we examined (average increases of 54.8-fold compared with normal mucosa) and in carcinomas (8). These findings were fully confirmed by real-time reverse transcription-PCR analysis of RNA extracted from samples used for the microarray study and from additional samples collected after the present study was completed (Supplementary Fig. S3).

In light of these findings, it was natural to wonder whether KIAA1199 might be a novel positively regulated target of Wnt signaling, which is characteristically deregulated in colorectal tumors. Previous microarray studies indicated that genes coregulated at the transcriptional level under different conditions tend to be involved in the same processes and pathways, and the analysis of transcriptional coexpression has been used to predict the function of novel genes (10-12). Therefore, we conducted a search for known Wnt targets (listed in Supplementary Table S5) among the genes whose expression patterns in all the tissue samples significantly correlated with those of KIAA1199. (The procedure used in this analysis is summarized in Materials and Methods and Supplementary Fig. S4.) Forty-nine percent of the known Wnt targets that were overexpressed in our adenoma samples had expression patterns that were positively correlated with that of KIAA1199 (Fig. 3A and B) as opposed to only 7.9% of the overexpressed genes that are not considered Wnt targets (P < 0.0001).

Evidence of the potential involvement of KIAA1199 in the Wnt signaling pathway had also emerged from another study by our group (13). A combined analysis of microarray data of tissues and cell lines placed KIAA1199 at the top of a list of genes [Supplementary Table S1 of ref. 13] that were up-regulated in colorectal adenomas and down-regulated in colon cancer cell lines that had undergone stable transfection with doxycycline-inducible forms of dominant-negative TCF1 or TCF4 to suppress Wnt signaling (14, 15). In the present study, KIAA1199 was also found to be markedly down-regulated in LS174T colon cancer cells in which Wnt signaling had been blocked by the induction of β-catenin small interfering RNA or NH₂-terminal–deleted TCF4 (15, 16). The dramatic decrease in KIAA1199 mRNA levels associated with this inhibition of the Wnt pathway was confirmed by Northern blotting (Fig. 3C).

In general, Wnt target genes are expressed predominantly in the proliferating compartment of normal intestinal crypts (lower portion), and their expression is appreciably increased in adenomatous glands (15). Our analysis of human tissues with preserved architecture indicated that these are also attributes of KIAA1199. In in situ hybridization studies, KIAA1199 mRNA was detectable only in the lower portion of normal colonic epithelial crypts (Fig. 4A and B), and its expression levels were much higher in dysplastic glands (Fig. 4C). These patterns were confirmed at the protein level by immunohistochemistry done with an antibody raised in our laboratory (Fig. 4D-J). This analysis also revealed that the KIAA1199 is a cytoplasmic protein whose expression is most intense near the cell membrane, particularly on the luminal side of the dysplastic cell multilayer (Fig. 4F-J).

Adenomatous colorectal polyps are one of the most common human tumors and the most frequent precancerous lesions in the colorectum, but their transcriptome has been only partially analyzed, and the data are generally based on a limited number of cases (17-20). We attempted to fill this gap by doing a comprehensive whole-genome microarray analysis of a large, highly homogenous set of adenomas that was collected prospectively.

A comparison of the transcriptomes of adenomatous polyps and segment-matched samples of normal colorectal mucosa revealed evidence of broad-scale remodeling. As a starting point for future verification studies, we have drawn up a list of 478 genes that were significantly up-regulated (n = 153) or down-regulated (n = 325) in the adenomatous tissues (fold changes of ≥4; Supplementary Table S4). Space constraints preclude more than a cursory examination of this list, but we have highlighted in Table 2 certain aspects that we feel are particularly interesting in terms of their relevance to the process of adenoma formation. For instance, transcription regulation seems to be extensively modified. Twenty-nine molecules involved in this process were expressed in adenomas at levels >4 higher or lower than those observed in the normal mucosa, but there were also several smaller changes in this category (Supplementary Table S6) that might also have dramatic effects on gene expression. Several other alterations reported in Table 2 are noteworthy in terms of their potential effect on cell proliferation, differentiation, apoptosis, and cell adhesion: (a) up-regulation of four members of the REG (regenerating) family of genes (21, 22), which would lead to increased tissue mitogen expression; (b) up-regulation of LCN2 (23) and down-regulation of ZFHX1B/SIP-1 (24) in the absence of significant changes in the expression of the epithelial cadherin CDH1 (E-cadherin), which would prevent or delay the epithelial-mesenchymal transition [changes were also noted in the expression of other cell adhesion genes of the cadherin and claudin families, including the striking overexpression of the placental cadherin gene CDH3, which is associated with early events in the transformation process (25, 26)]; (c) down-regulation of ZFHX1B/SIP-1 and Max dimerization protein 1 (MXD1/MAD1; decreased only 3.3-fold and therefore not listed in Table 2; refs. 27, 28) and overexpression of the RTEL1 helicase, which should facilitate telomere elongation (29); (d) alterations that would diminish apoptosis [e.g., overexpression of the decoy receptor for Fas ligand, TNFRSF6B, which is reportedly coregulated with RTEL1 on chromosome 20q13.3 (30-32); decreased expression of the netrin-1 receptor, UNC5C (33); and expression changes involving three Fas apoptosis inhibitory molecules (FAIM), including FAIM1, which was increased 2.3-fold and is thus not listed in Table 2]; and (e) marked down-regulation of several genes that would result in reduced tumor suppression activity [e.g., those encoding the antiangiogenic factor ANGPTL1 (34), the cyclin-dependent kinase inhibitor CDKN2B/p15, and the prostaglandin catabolism enzyme HPGD (35)].

It is also important to recall the size-related differences noted in the adenoma gene expression profiles (Fig. 2; Supplementary Table S3). When validated in a larger series of tumors, these differences should provide important clues to the molecular basis of the well-known link between the dimensions and malignant potential of colorectal adenomas (1).

Our study also furnishes a complete picture of expression changes involving gene components of the Wnt pathway across the transition from normal to adenomatous epithelium (Supplementary Table S2) as well as evidence for the existence of a novel Wnt target: KIAA1199. This gene, which encodes a protein of unknown function, was strikingly overexpressed in all the adenomas included in this study and in 25 adenocarcinomas of the colon described in a previous report (8). Even more intriguingly, its expression was significantly correlated with that of several genes that are well-established targets of Wnt signaling. Our hypothesis that KIAA1199 is up-regulated by the TCF(s)/β-catenin transcription complex was considerably strengthened by the marked decreases in KIAA1199 expression observed in cultured colorectal cancer cells when the Wnt pathway was inhibited by overexpression of dominant-negative TCF4 proteins or by β-catenin knockdown. It is not yet clear whether this is a direct effect, but this possibility is supported by the results of a recent genome-wide TCF4 ChIP-on-chip analysis, which indicates that the KIAA1199 locus is surrounded by four TCF4-bound regions.¹⁰

Other features of KIAA1199 expression are also compatible with its putative role as a Wnt target gene. KIAA1199 mRNA and protein are both confined to the proliferative compartment of normal intestinal crypts, where Wnt signaling is normally active, and they are highly overexpressed in colorectal adenomas and carcinomas, where this pathway is almost always aberrantly activated.

In normal and tumor tissues, KIAA1199 is expressed in the cytoplasm of epithelial cells. In glands with low-degree dysplasia, higher concentrations are observed in the mucin vacuoles of goblet cells, but cytoplasmic expression of the protein in tumor cells remains elevated even after goblet cell differentiation has been lost (Fig. 4). These features, together with the localization of KIAA1199 in the luminal portion of the cytoplasm, are suggestive of a secreted and/or membrane protein. This conclusion is consistent with our in silico analysis of KIAA1199 (see Supplementary Data and Supplementary Fig. S5), which strongly predicts the presence of a signal peptide at its NH₂-terminal end. In addition, the central region of KIAA1199 contains a TMEM2 homology domain, which is present in several eukaryotic proteins, including TMEM2, polyductin (PKHD1), and fibrocystin L (PKHD1L1; Fig. 5), all large receptor proteins characterized by an NH₂-terminal signal peptide or a single transmembrane helix and a short cytoplasmic tail (36).

A study based on yeast two-hybrid screens suggested that KIAA1199 may interact with plexin A2 (KIAA0463; ref. 37). The transmembrane plexins interact with transmembrane semaphorins on nearby cells, providing “stop” and “go” signals that are crucial for cell motility and invasive growth (38, 39). KIAA1199/plexin A2 interaction could thus play important roles in colorectal tumorigenesis not only in the invasive stages but also earlier during the formation of abnormal glands in benign adenomas.

A recent report linked high levels of KIAA1199 mRNA with cell mortality in human fibroblasts and in a renal cell carcinoma cell line (40). In that study, however, there was no significant increase in KIAA1199 expression during replicative aging of mortal cells, and this finding contrasts with the documented behavior of other genes involved in cell aging (41). Furthermore, the authors reported wide variation in KIAA1199 mRNA expression in breast cancer cell lines, and this finding raises the possibility that expression of this gene in vivo and in cell lines may differ.

We believe that our microarray data will serve as a springboard and reference point for other studies on the molecular basis of colorectal transformation along the adenoma-carcinoma pathway (and subsequently for the study of alternative pathways). Some of the transcriptional changes reported in this study might one day be used as molecular indices of the susceptibility of adenomas to malignant transformation, information that would be helpful in planning appropriate follow-up of the lesions. As for KIAA1199, its invariably high expression in the colorectal tumors we studied raises interesting possibilities for the development of a new molecular marker for the detection of these neoplasms. For example, because KIAA1199 expression in the normal mucosa is limited to cells in the lower portion of the crypts, which are not yet programmed to be shed into the intestinal lumen, the presence of KIAA1199 peptides in fecal water might prove to be a specific marker of adenomatous lesions. In addition, although due consideration must be given to its probable physiologic role(s) in intestinal crypts and possibly in several other human tissues (40, 42, 43), KIAA1199 may be a potential target of antibody-based therapies.

Pedunculated colorectal polyps and normal mucosa were obtained during colonoscopies carried out in the Gastroenterology Unit of the Belcolle City Hospital (Viterbo, Italy). The tissues were collected prospectively with informed patient consent and the approval of the local Human Research Ethics Committee. Patients with documented familial polyposis, with >15 adenomatous polyps (total: synchronous + previously excised; ref. 44), or currently treated with nonsteroidal anti-inflammatory drugs (including aspirin) were excluded from the study.

For each polyp, three biopsies of normal mucosa were collected from the same colon segment (≥2 cm from the site of the polyp). Immediately after removal, a small sample of epithelial tissue (5-15 mg) was cut from the tip of each polyp, leaving the underlying muscularis mucosae intact. We excluded polyps <1 cm to ensure that the sampling procedure would not interfere with the histologic diagnosis. All polyp samples were collected by a single operator (M.d.P.) using the same procedure to minimize artifacts due to sampling differences. The approach used allowed us to obtain specimens with a high percentage of epithelial cells without resorting to microdissection, which can diminish the quantity and quality of the extracted RNA.

The polyp sample and the three normal mucosal biopsies were immersed in RNAlater (Ambion) for subsequent microarray analysis, and the remainder of the polyp was submitted for pathologic analysis. The cut surface at the tip was labeled with India ink so that the sampled area could be easily identified during routine histologic examination. The tissue was then fixed in buffered formalin and embedded in paraffin. DNA extracted from sections of this specimen was also used to rule out microsatellite instability (reflecting defective DNA mismatch repair) at the BAT26 locus, as previously described (45). All of the polyps included in the study met the following criteria: type 0-Ip (6), maximum diameter of 1 to 4 cm, absence of surface ulceration, histologic diagnosis of adenoma, and absence of microsatellite instability at BAT26.

In some analyses, we also included transcriptomic data from a previously described set of 25 colon cancers (mismatch repair proficient and deficient; ref. 8), which we reanalyzed for this study with the same microarray used to characterize the adenomas and normal mucosa.

Total RNA was extracted (RNeasy Mini kit, Qiagen) from homogenized tissue samples (5-15 mg), and its integrity was verified by capillary gel electrophoresis (Bio Analyzer, Agilent Technologies). Complementary RNA (15 μg/sample), synthesized and labeled as previously described (8, 46), was hybridized with the Affymetrix U133 Plus 2.0 array, which contains in situ synthesized oligonucleotides representing the entire human genome (54,675 probes).

Raw gene expression data generated by GeneChip Operating Software (Affymetrix) were imported into the GeneSpring software program (Agilent Technologies) and normalized per chip (i.e., to the median of all values on a given array) and per gene (i.e., to the median expression level of the given gene across all samples). Analysis was done using the log expression values with GeneSpring's cross-gene error model turned on. Probes were excluded from analysis unless they were listed as “present or marginal calls” and/or had expression values ≥100 in ≥50% (≥16 of 32) of the samples in at least one of the tissue groups (adenomas and normal mucosa).

Expression data were subjected to four different unsupervised analyses: (a) hierarchical clustering using the Pearson correlation coefficient as a similarity measure and the average linkage algorithm for branch merging; (b) PCA, which reduces the dimensionality (number of variables) of a data set while retaining most of its variance (8); (c) correlation analysis, which involved computation of Pearson correlation coefficients for all possible sample pairs and visualization of correlation values as tile plots; and (d) CA, another dimension-reducing method (47), which was used to identify samples associated with particular gene expression levels. In typical CA, a matrix of n gene expression levels from p samples is treated as a two-way contingency table (genes by samples or vice versa) with n and p specifications for the “factors” gene and sample, respectively. Each intensity value thus reflects the abundance of a given transcript in a given sample. Like PCA, CA identifies independent “factorial components” that account for variance within a multidimensional gene data set, but in this case, the components are identified and ranked according to the correlation between gene and sample scores. A supervised or constrained extension of CA (9), CCA, was then used to identify possible correlations between gene expression patterns and clinical or pathologic variables. CA and CCA, as well as the corresponding plots, were computed using R software and the ade4 and made4 packages furnished by Bioconductor.¹¹

The Mann-Whitney test was used to select genes differentially expressed in normal mucosa and adenomas; Benjamini-Hochberg multiple testing correction was applied with a false discovery rate of 0.01. The genes in this set that were differentially expressed with fold differences of ≥2.0 were then analyzed with ErmineJ software (48) to identify any biological processes from the Gene Ontology database (49) that were overrepresented.

Pearson correlation was used to identify correlation between KIAA1199 expression and the expression of other genes in the entire set of tissue samples. Fisher's exact test was used to identify possible overrepresentation of known Wnt targets among genes whose expression was closely correlated with that of KIAA1199 (correlation values ≥0.8).

Reverse transcription-PCR and Northern blotting were done as previously described (46, 50) to verify the expression level of KIAA1199 in tissue samples and in LS174T colon cancer cells in which inducible inhibition of the Wnt pathway had been achieved with previously described methods (14-16).

Digoxigenin-labeled KIAA1199 antisense riboprobes were synthesized from a PCR product amplified from human colon cDNA with KIAA1199-specific primers (sense: 5′-cacatcggggaggagataga-3′; antisense, containing a T7 RNA polymerase-binding site: 5′-taatacgactcactatagggttccagacttgaca-3′). This product was transcribed in vitro using the DIG RNA labeling kit and T7 RNA polymerase (Roche Diagnostics). In situ hybridizations were done on paraffin-embedded sections of human colon fixed with 4% buffered formalin as described elsewhere (51).

Our in silico analysis of KIAA1199 (see Supplementary Data) indicated that residues 202 to 217 (IHSDRFDTYRSKKESE) form a loop between a conserved β-strand and the following helix of the NH₂-terminal GG domain. This charged, surface-exposed peptide was used to raise a rabbit polyclonal antibody, which was purified by affinity chromatography on Thiopropyl Sepharose 6B (Amersham) derivatized with the antigenic peptide. A 1:1,000 dilution of this antibody was used, as previously described (45), to evaluate KIAA1199 expression in formalin-fixed, paraffin-embedded sections of adenoma and normal mucosal tissues.

We thank the patients who participated in this study; R. Haider, E. Pani, T. Lehmann, A. Patrignani, and U. Wagner for technical assistance; A. Mueller, E. Veljkovic, and F. Bannwart for critical reading of the manuscript; and M. Kent for editorial assistance.