Gene Expression Patterns Distinguish Colonoscopically Isolated Human Aberrant Crypt Foci from Normal Colonic Mucosa

Aberrant crypt foci (ACF) are considered the earliest identifiable preneoplastic colonic lesions; thus, a greater understanding of the nature of genetic changes underlying the transformation of normal colonic mucosa (NM) into ACF may provide insight into the mechanisms of carcinogenesis. ACF were identified by indigo carmine spraying onto colonic mucosa during colonoscopy and isolated as standard pinch biopsies of the mucosal areas containing the ACF. RNAs isolated from ACF and matched NM biopsies from the ascending and descending colons of 13 patients were analyzed on arrays containing 9128 cDNAs. Thirty-four differentially expressed (P < 0.001) genes were found in a paired comparison of the ACF and NM samples, and 25 of 26 matched pairs of ACF and NM could be correctly classified in leave-one-out cross-validation. Differential expression for seven of eight genes was confirmed by real-time reverse transcription-PCR. Furthermore, ACF and NM samples, including six pairs of ACF and NM samples that had not previously been analyzed by array hybridization, can be correctly classified on the basis of the overexpression in ACF of three selected genes (REG4, SRPN-B5, and TRIM29) evaluated by real-time reverse transcription-PCR. In a separate analysis of 13 biopsy pairs from either ascending or descending colon, ACF and NM samples could also be correctly classified by the gene expression patterns. Analysis of gene expression differences in ACF from the ascending and descending colon versus NM samples indicates that ACF from these distinct colonic locations are converging toward similar gene expression profiles and losing differences in gene expression characteristic of NM from the ascending versus descending colon. (Cancer Epidemiol Biomarkers Prev 2006;15(11):2253–62)

An aberrant crypt focus consists of aggregates of abnormal crypts that are larger in size than normal colonic crypts and have a thickened layer of epithelial cells, increased pericryptal space, and irregular (frequently oval or slit-like) lumens. ACF are elevated above adjacent normal colonic crypts and are routinely identified as more densely or less densely stained outcroppings on colonoscopy using methylene blue or indigo carmine staining, respectively.

ACF were originally described in methylene blue–stained whole-mount preparations of colons from carcinogen-treated mice and rats (1). Mutations in K-ras and β-catenin genes, changes in expression of several genes implicated in colorectal tumorigenesis (such as APC, p53, c-fos, and c-Myc; ref. 2), and other features characteristic of colon cancer, including altered mucin secretion (3) and microsatellite instability (4), have been found in rodent ACF. It has been shown that the frequency of ACF (and the number of crypts per focus) was increased by carcinogens in a dose- and time-dependent manner and decreased on treatment with agents known to inhibit dimethylhydrazine- or azoxymethane-induced colon cancer in rodents (reviewed in refs. 5, 6). Similarities in the characteristics of ACF and colon cancers allow consideration of ACF as the earliest preneoplastic colonic lesions, and rodent ACF are increasingly used as biomarkers for evaluation of carcinogenic and chemopreventive agents (7-11).

Human colonic ACF are similar to rodent ACF (12-14) and characterized by increased cell proliferation with some expansion of cell proliferative activity to the upper part of aberrant crypts (15-17). Most, if not all, ACF are monoclonal (18, 19) and frequently have mutations in K-ras (reviewed in refs. 20-22). Mutations in the APC gene have also been found in ACF but with low frequencies (23, 24). Other genetic and epigenetic alterations that are present in colon tumors have also been found in ACF: mutations in the gene BRAF (25); altered DNA fingerprints indicative of gross DNA instability (26); methylation of CpG islands in several genes implicated in colon cancer, particularly the p16 gene (27); silencing by methylation of the tumor suppressor gene SLC5A8 (28); and overexpression of carcinoembryonic antigen (29), P-cadherin (30), and E-cadherin (31). Interpretation of data on gene expression changes in human ACF is, however, complicated because gene expression was assessed mostly by immunohistochemical methods and therefore is rather qualitative. In addition, the pattern of altered gene expression usually varies throughout the aberrant crypts (from the base to the apex) and depends on ACF size (see, e.g., refs. 29-31).

The prevalence of ACF is correlated with colorectal cancer risk: the number of ACF is higher in the colons of patients with colorectal adenomas and carcinomas as compared with patients without cancer (17, 32-37), in colons of patients from a population with higher colorectal cancer incidence rate (38), in older patients (17, 39, 40), and in tobacco users (40). It has been found that the number of ACF is increased from proximal to distal colon (33, 34, 38, 39, 41, 42), corresponding to the known distribution of sporadic colorectal cancer. Thus, ACF can be used as a potential biomarker for evaluation of colorectal cancer risk. It has been shown that treatment of patients with the nonsteroidal anti-inflammatory drugs sulindac (17) and aspirin (10) significantly reduces the number of ACF in the colon, even after short (2-10 months) treatments (43), indicating that ACF can be used as biomarkers for the identification of cancer preventive agents (44).

The monoclonal origin of ACF, the altered control of cell proliferation reflected in dysplasia, and the presence of gene mutations characteristic of colon neoplasia in human and rodent ACF support the role of ACF as the earliest precursors to colorectal cancer. The study of changes in gene expression in ACF may contribute to an understanding of the nature of ACF and the identification of genes of which the alteration in expression is related to cancer risk and cancer-preventive drug effects.

In this study, we wanted to explore genome-wide changes in gene expression that are characteristic of ACF. As a first step, we asked if ACF identified during routine optical colonoscopy and isolated as pinch biopsies can be distinguished from normal colonic mucosa (NM) by their pattern of gene expression.

Methods of RNA extraction and amplification, cDNA labeling and microarray hybridization, data acquisition, and statistical analysis were described in detail earlier (45) and will only be outlined here.

Protocols used in the study were approved by Institutional Review Boards of the National Cancer Institute and National Naval Medical Center. Before enrollment, informed consent was obtained from all patients who were undergoing medically indicated screening or surveillance colonoscopy. ACF were identified after spraying indigo carmine (0.5%) onto the colonic mucosa during routine colonoscopy. Standard pinch biopsies of colonic mucosal areas containing ACF (with 10 to 50 crypts per focus) and, in addition, NM biopsies were taken, flash frozen in liquid nitrogen, and stored at −80°C. More than 80% of the cells in the biopsies were colonic crypt cells as determined by H&E staining. Matched ACF and NM samples from ascending and descending colons of 13 patients were selected for RNA isolation and array hybridization experiments (Table 1). In addition, matched pairs of ACF and NM samples from ascending or descending colons of six patients were used for validation purposes. All patients were free of malignancy at the time of biopsy collection. Five of 13 patients in the main set and three of six patients in the validation set had previous colorectal cancers. Cancer-related colon surgery was done more than 2 years before the taking of biopsies for all but one of the patients who had undergone colorectal surgery 1 year before the collection of biopsies.

Total RNA was isolated from flash-frozen specimens (usually two standard ACF or NM pinch biopsies per one RNA sample), homogenized with a Disposable Generator and Micro-H Omni Homogenizer in lysis buffer RLT (Qiagen), and purified using Qiagen RNeasy Mini Kit columns (Qiagen) according to the instructions of the manufacturer. mRNA was amplified according to a modified Eberwine's protocol (46).

Fluorescently labeled cDNA was synthesized by reverse transcription of amplified colon antisense RNA and human testis antisense RNA [prepared from total testis RNA (Clontech, Inc., Mountain View, CA) as described for colon antisense RNA] with random oligonucleotide primers in the presence of Cy3-dUTP or Cy5-dUTP (Amersham Pharmacia Biotech, Piscataway, NJ), respectively. For each hybridization experiment, 5 μg of colon antisense RNA and 6 μg of testis antisense RNA (common reference) were used to prepare a mixture of labeled cDNAs. Microarrays containing cDNA clones (from Research Genetics, Inc., Huntsville, AL) spotted on lysine-coated glass slides were obtained from the Advanced Technology Center. Detailed information on printed cDNA can be found on the mAdb web site.³

Microarrays contain 9,128 sequence-verified cDNAs, among which 7,102 represent named genes and 1,179 EST clusters, including 8,556 cDNAs with UniGene cluster ID (mapping to 7,777 unique UniGene clusters).

Microarrays were scanned with an Axon 4000 laser scanner and image analysis was done with GenePix Pro 3.0 Software (Axon Instruments, Inc., Sunnyvale, CA). Data analysis was done with the BRB-ArrayTools (version 3.2) software package developed by the Biometric Research Branch of the Division of Cancer Treatment and Diagnosis of the National Cancer Institute and The EMMES Corporation (Rockville, MD; ref. 47). All arrays used were printed on the same day (as a one printing batch); before statistical analysis, background intensities were subtracted and data were filtered for minimal spot intensity (100 units) in one of the two channels (if in the other channel, background-subtracted intensity was <100, it was adjusted to 100 units) and for missing values (not in >20% of arrays). Fluorescence intensity ratio data were log transformed and normalized by loess smoother.

The Class comparison and Class Prediction modules of BRB-ArrayTools were used to determine if the pattern of gene expression allowed discrimination of ACF and NM samples.

First, a paired t test was done on average differences in normalized gene expression log-ratios in pairs of samples (ACF and NM) from the same patients to select genes that showed univariately statistically significant differences (P < 0.001) in expression between ACF and NM samples. The multivariate class label permutation test was used to compute the number of false-positives among selected genes (48, 49).

Several multivariate classification methods available in BRB ArrayTools 3.2 (compound covariate predictor, diagonal linear discriminant analysis, nearest neighbor and nearest centroid predictors, and support vector machine) were used for classification of ACF and NM samples and gave comparable results, which will be illustrated with the Compound Covariate Predictor (CCP) and Support Vector Machine Predictor (SVMP) classifiers. CCP was calculated as a linear combination of log-ratio differences weighted by univariate t values (50). A positive sign was assigned to t-test values for genes that show higher log-ratios in NM and a negative sign for genes that show higher log-ratios in ACF. The CCP value was calculated for each pair of samples and, as a classification threshold, the sign (positive or negative) of the CCP was used for classifying. SVMP was calculated as a linear function of the log-ratio differences for genes selected in paired t test with weights estimated by linear kernel support vector machine algorithm to minimize the number of misclassified samples.

The misclassification rate was estimated by leave-one-out cross-validation. Specifically, a pair of ACF and NM samples (from one patient) was omitted and a CCP and a SVMP were developed from scratch using the remaining samples. The cross-validation procedure includes performing t tests for a selection of genes on arrays for which differences in expression are significant at a stated P value for remaining samples, recalculating the CCP, SVMP, and classification threshold for the remaining pairs of samples, and applying the new CCP, SVMP, and threshold values to classify the omitted pair of samples. This was done independently for each omitted pair of samples. The ratio of pairs of samples correctly classified in cross-validation to the total number of sample pairs yields the misclassification rate. Permutation P value for a classifier was calculated by performing 2,000 random permutations of class labels and repeating the cross-validation procedure for each permutation. The proportion of random permutations that gave the same or smaller misclassification rate as was obtained with the true class labels is presented as a (permutation) P value for the classifier, and a value of P < 0.0005 was reported when no random permutation of the class label was found out of 2,000 permutations with the same or smaller misclassification rate as for the true class labeling. The effects of multiple factors on gene expression were estimated by ANOVA for fixed and mixed effect linear models, which was done as implemented in R-Plugins of BRB-ArrayTools.

One-step TaqMan real-time reverse transcription-PCR (RT-PCR) was done to study expression of several cDNAs (genes) using an ABI Prism 7700 Sequence Detection System. Primers and hybridization FAM-labeled probes were selected with PrimerExpress software (Applied Biosystems, Foster City, CA) by using complete cDNA sequences that have the same UniGene cluster ID as cDNAs printed on array (Supplementary Table S1). TaqMan Gold RT-PCR kit (Applied Biosystems) and the protocol of the manufacturer (30 minutes at 48°C for RT reaction, 10 minutes at 95°C for activation of TaqGold Polymerase, and 40 cycles consisting of 15 seconds at 95°C and 1 minute at 60°C) were used, and 20 ng (for gene-specific PCR) or 5 ng (for ribosomal 18S RNA–specific PCR) of total RNA were assayed in 25 μL of one-step RT-PCR reaction mixture with gene-specific or ribosomal 18S RNA–specific primers and probes, in triplicates for each sample and each gene. Serial dilutions of a mixture of colon total RNAs from several other colon biopsies were used to calculate PCR efficiency for a gene in the range of 5 to 40 ng of total RNA per reaction. Gene expression was normalized to the amount of ribosomal 18S RNA in a sample, for which RT-PCR reactions were done on the same 96-well plate in separate wells, together with serial dilutions of colon total RNA (from 1.25 to 10 ng RNA per reaction) to calculate PCR efficiency for ribosomal 18S RNA. Data were analyzed, and normalized (by ribosomal 18S RNA quantity) gene expression was calculated with Q-Gene software (51).

Among patients undergoing screening or surveillance colonoscopy, we have selected 13 patients who have ACF in ascending as well as descending colon (Table 1). As might be expected (33, 34, 39, 41), the number of ACF found in the descending colon (11.23 ± 1.653, N = 13) of these patients is higher than the number of ACF in the ascending colon (1.54 ± 0.1831, N = 13; t = 5.827, P < 0.0001). However, the number of crypts per focus in the ascending colon (26.05 ± 2.370, N = 20) is higher than in the descending colon (17.60 ± 0.8479, N = 146; t = 3.447, P = 0.007). The tendency of ACF from ascending colon to be larger than the ACF from the descending colon was noted before (38). Addition of patients selected for validation into analysis does not change the conclusion that there is prevalence of ACF in descending compared with ascending colon, although the size of ACF (number of crypts per ACF) in ascending colon is bigger than in the descending colon in studied patients.

In a combined set of 19 patients, eight patients had previous colorectal cancer. There are no statistically significant differences in number or size of ACF in ascending and descending colon of patients with previous colorectal cancer compared with patients without cancer. Matched pairs of ACF and NM samples from ascending and descending colons of 13 patients were analyzed on a total of 52 arrays, with each sample analyzed on one array. The effect on gene expression of various factors, including the tissue type (ACF or NM), biopsy location (ascending or descending colon), history of previous colon cancer, gender, and age of patient, was studied by ANOVA using linear models with only fixed effects (blocked by array ID or patient ID) or mixed effect models where array ID or patient ID was treated as a random factor. All of the listed factors affect gene expression, with biopsy location being the most dominant factor followed by effect of individual between patient variations. However, in this study, only two factors are balanced: tissue type and biopsy location, whereas other factors are unbalanced, and comparisons of gene expression can be biased by confounding factors. For example, in a main group of 13 patients, there is a higher ratio of male patients with previous colon cancer than of female patients, and patients with previous colon cancer are older (average age, 70 years) than patients without previous cancer (51 years). Given the small number of patients (13) and possible confounding factor effects, we restricted our analysis to comparison of gene expression in ACF versus NM.

ACF and NM samples can be successfully classified by their gene expression profiles without regard to their origin from the ascending or descending colon (data available on request). However, the patterns of gene expression in NM from ascending (right) and descending (left) colon are different (ref. 45, and see below). Therefore, we believe that it is more appropriate to compare gene expression in ACF and NM samples by paired tests that take into consideration correlation in gene expression in ACF and NM from the same side of the colon [e.g., ACF and NM from the right colon (ACFR versus NMR) and ACF and NM from the left colon (ACFL versus NML), resulting in two pairs of samples per patient].

Thirty-four genes show differential expression at a univariate P < 0.001 in ACF and NM on analysis of 26 pairs of ACF and NM according to a paired t test (Table 2). The probability of getting by chance 34 differentially expressed genes out of 8,983 genes (which passed the filtering criteria) is 0.015 as calculated by a multivariate class-label permutation test. Twenty-five of 26 and 24 of 26 sample pairs were correctly classified in leave-one-out cross-validation using the CCP or SVMP. Random permutations of class labels showed that the probability of getting by chance as small a cross-validated misclassification rate as that obtained with the true class labels (4% for CCP or 8% for SVMP) is less than 5 × 10⁻⁴ and 0.001 for CCP and SVMP, respectively.

In a separate analysis of ACF and NM samples from the ascending colon (13 pairs), 46 genes show differential expression at P < 0.001 (Table 3). As calculated by multivariate class-label permutation test, the probability of getting by chance 46 differentially expressed genes out of 8,983 genes is <0.001. All 13 pairs of ACFR and NMR were correctly classified in leave-one-out cross-validation using CCP and SVMP, with probabilities of getting by chance a 0% misclassification rate 0.005 and 0.004, respectively. Similar results were obtained in an analysis of the 13 pairs of ACF and NM samples from the descending colon. Forty-two genes show differential expression at P < 0.001 (Table 4; probability of getting by chance 42 differentially expressed genes out of 8,983 genes is 0.006), and 12 of 13 and 13 of 13 sample pairs were correctly classified in leave-one-out cross-validation using CCP and SVMP, respectively. Random permutations of class labels showed that the probability of getting by chance 8% and 0% misclassification rate is 0.036 and 0.007 for CCP and SVMP, respectively.

To check if the differences in gene expression defined by array hybridization can be used for predictive classification of ACF and NM samples, we applied classifiers developed on the basis of the 26 pairs of ACF and NM, as well as on the basis of the 13 pairs of ACFL and NML, to array data for newly obtained ACF and NM samples from descending colons of three patients that were not included in the main group of 13 patients. All six samples (three ACFL and three NML) were correctly classified in both cases.

Thus, ACF can be distinguished from NM regardless of the site of origin based on common differences in gene expression. In addition, ACF from the ascending colon can be distinguished from the normal ascending colonic mucosa, and ACF from the descending colon can be distinguished from the normal descending colonic mucosa.

To validate data on differential gene expression in ACF and NM, we selected eight genes for real-time RT-PCR analysis. Twelve pairs of RNAs from ACF and NM samples from the ascending and descending colons of six patients (that were analyzed by array hybridization) were analyzed by real-time RT-PCR (data for three genes are presented in Supplementary Figs. S1 and S2). In general, there is good agreement between array and real-time RT-PCR data on differences in expression of the eight selected genes in ACF and NM: Spearman rank correlation coefficient (r) for ratios of mean normalized gene expression values obtained by real-time RT-PCR and geometric mean ratios from array data are 0.833, 0.762, and 0.857 for ACF/NM, ACFR/NMR, and ACFL/NML log 2 ratios. Perhaps due to the high individual variability in gene expression and the low number of samples, only some of the eight genes show statistically significant differences (paired Wilcoxon signed rank test, P < 0.05) in expression between ACF and NM when their expression is analyzed individually, although all but one display a trend in differential expression expected from the array data (Table 5).

In addition, real-time RT-PCR data can also be used to identify ACF and NM: all 12 pairs of ACF and NM samples can be correctly classified on the basis of overexpression of at least two of three genes (REG4, SRPN-B5, and TRIM29) in ACF when compared with NM from the same location and the same patient (Supplementary Figs. S1 and S2). To confirm this observation, we selected six pairs of ACF and NM samples (that had not been analyzed previously by array hybridization) from the ascending (two pairs) and the descending (four pairs) colons of six patients. As shown in Fig. 1, all samples can be correctly classified by the pattern of expression of REG4, SRPN-B5, and TRIM29 genes: in all six pairs of samples, expression of each of the three genes is at least twice higher in ACF than in the corresponding matched NM.

Thus, there is a good correlation for data on differences in gene expression in ACF and NM obtained in array hybridization and in real-time RT-PCR experiments, and real-time RT-PCR data can be used to correctly classify ACF and NM.

A paired t test identified 1,340 genes that showed statistically significant (at a univariate P < 0.001) differences in expression in ascending versus descending NM (Supplementary Table S2), confirming our previous results obtained on analysis of a different patient group (45). Differences in expression of 312 genes were also found to be significant at P < 0.001 in a paired t test between ACF from ascending versus descending colon (Supplementary Table S3). However, the observed differences in gene expression between ACFR and ACFL likely reflect for the most part the presence of NM cells in pinch biopsies of ACF. Most (270) of the 312 genes are present in the list of genes differentially expressed in NMR and NML samples, being overexpressed or underexpressed in both NMR relative to NML and ACFR relative to ACFL, and the magnitude of differences in expression of the 270 genes is lower in ACFR versus ACFL compared with NMR versus NML. In addition, when genes differentially expressed between NMR and NML are omitted from the analysis, ACFR and ACFL can no longer be successfully distinguished in leave-one-out cross-validation using CCP or SVMP (Supplementary Table S4).

Thus, the comparisons of differential gene expression in ACFR versus ACFL and NMR versus NML show that ACF are losing differences in gene expression characteristic for NM from the ascending and the descending colon. Most of the differential gene expression observed between ACFR and ACFL is not acquired de novo but is related to the differential gene expression between NM from the ascending versus the descending colon, and this can be a reflection of the presence of NM in ACF samples.

We have used the Ingenuity Pathway Analysis (IPA)⁴

program to map genes affected in ACF to categories (“High Level Functions”) and known gene networks available in the IPA database. In this analysis, we have used lists of genes that show differences in expression between ACF and NM samples at P < 0.005 (available on request), as well as at P < 0.001 (Tables 2-4). The Fisher exact probability test is used in IPA to identify categories containing more affected genes than may be expected by chance given the genes present on the arrays. According to IPA global function analysis, there is enrichment in ACF of genes implicated in cancer, lipid metabolism, cell death, tumor and tissue morphology, cell growth and proliferation, and other categories (Supplementary Table S5). Genes affected in ACF are mapped with scores >10 to twelve known gene networks that control cell growth and proliferation, cell death and cell movement, cellular and tissue morphology, and hematologic, skeletal, and muscular system development and function (Supplementary Table S6; the score is the anti-logarithm of the probability of getting by chance a number of genes equal to or greater than the actual number of affected genes in a network, given the genes present on the array; a score of 2 means P = 0.01).

The histopathologic identification and classification of aberrant crypt foci has not achieved a consensus standard and is the subject of a variety of distinct classification schemes reported in the literature (12-14, 52-56). Classification is complicated by the spectrum of foci that include appearances only subtly distinguishable from NM at one end of the histopathologic spectrum and indistinguishable from frank adenomas at the other end. We chose in this study to approach ACF beginning with the modality that is the most likely to have an immediate clinical effect, optical colonoscopy.

We thus applied indigo carmine staining during routine colonoscopic surveillance examinations and isolated ACF containing 10 to 50 aberrant crypts as standard pinch biopsies. We can now conclude that these colonoscopically distinct foci are distinguishable from the NM from which they arose on the basis of their pattern of gene expression. A recent paper (57) studied gene expression profiles within azoxymethane-induced ACF derived from a high-risk versus a low-risk inbred strain of mice. Whereas some of the specific genes that distinguished these two strains overlap with the broadly defined pathways that we found to distinguish NM from ACF, these pathways are quite broad in their definition and we could not pinpoint a specific gene or genes that were of particular relevance in this comparison.

Correct classification of ACF and NM samples can be achieved using gene expression differences between ACF and NM samples combined from the ascending and descending colon. Classification can be also achieved for ACF versus NM from either the ascending or the descending colon. However, the lists of genes emerging in those classifiers are different. Only 11 and 12 of the 34 genes that are found differentially expressed at P < 0.001 in a comparison of the combined set of 26 pairs of ACF and NM samples (i.e., genes that show differences in expression common between all ACF and NM from both ascending and descending colon) show differential expression at P < 0.001 in separate analyses of ACF and NM samples from the ascending or the descending colon, respectively. The reason for this is mainly the decrease in the number of sample pairs from 26 to 13 pairs. The differences in expression for most of these 34 genes are significant at a less stringent P value (P < 0.05) in separate comparisons of ACFR versus NMR (29 genes) and ACFL versus NML (28 genes), with 23 genes being present in both lists of significant genes.

As previously shown (45), the ascending and descending colon can be considered as two separate entities on the basis of their gene expression pattern. We observed the tendency of ACF from the ascending and the descending colon to converge toward similar gene expression profiles. Most of the genes differentially expressed in ascending versus descending NM, of which the expression is changed in ACFR or ACFL, show smaller or no differences in expression in ACFR versus ACFL. Among 46 genes that are differentially expressed in ACFR versus NMR (Table 3), 40 genes are also expressed differentially at P < 0.001 in NM samples from ascending versus descending colon (Supplementary Table S2). Five genes of which the expression is increased in ACFR relative to NMR are underexpressed in NMR relative to NML, and 35 genes that are underexpressed in ACFR relative to NMR are overexpressed in NMR relative to NML. Fourteen of 42 genes differentially expressed in ACFL versus NML (Table 4) also show differential expression at P < 0.001 in NM from ascending and descending colon (Supplementary Table S2). Eleven of the 14 genes follow the above-mentioned trend: seven genes overexpressed in ACFL are underexpressed in NML relative to NMR and four genes underexpressed in ACFL are overexpressed in NML relative to NMR.

Eight of the 12 gene networks, to which genes differentially expressed in ACF (P < 0.005) have been mapped with a score >10, are linked by the MYC, FOS/JUN, and/or TP53 genes (Supplementary Table S6). Although it indicates the importance of this particular subset of gene networks for ACF emergence and/or maintenance, the possible role of MYC, AP1, and TP53 proteins by themselves remains unclear. For example, the MYC gene is overexpressed in ACFR compared with NMR (1.7-fold; P = 0.002), and expression of several genes known to be directly controlled by MYC protein is changed in ACFR in the predicted direction. In Network 5 (Supplementary Table S6), CPT1A, NDRG1, and TAT genes are underexpressed in ACFR (MYC decreases expression of these genes), and RHOB is overexpressed in ACFR (MYC increases expression of this gene). In ACFL, the MYC gene is expressed at the same level compared with ACFR, but its expression is not changed in ACFL compared with NML (MYC is overexpressed in NML compared with NMR 2.12-fold; P = 1.4 × 10⁻⁶).

We did not confirm overexpression in ACF of several genes that had been reported in the literature: those encoding carcinoembryonic antigen (31), E-cadherin (31), protein kinase C βII (58), inducible nitric oxide synthase (59), c-fos (60, 61), and K-ras (61, 62). Some of the genes of which the expression was found altered in ACF, such as glutathione S-transferase π (62) and sodium transporter SLC5A8 (28), were not present on our arrays. The earlier described overexpression of P-cadherin (CDH3; ref. 30) and gastric MUC5B genes (3, 63, 64) was confirmed by our data, although only for ACF from the descending colon if P < 0.001 is used. Most of the published data were obtained by analyzing expression on the protein level, mainly by immunohistochemical methods. It is known that RNA and protein levels only moderately correlate (65, 66), and discordances in RNA and protein levels may affect conclusions about differences in gene expression. In addition, it is possible that differences in expression of some genes do not reach statistical significance due to the presence of NM in our ACF biopsies, which may artificially decrease the magnitude of differences or increase the variance in apparent RNA levels in ACF and NM. For example, the NOS2A gene seems to be overexpressed in ACF (ratios: ACF/NM 1.6, ACFR/NMR 1.76, ACFL/NML 1.46) as may be expected from published data (59), but the differences are only significant at P = 0.0062 for the ratio ACF/NM.

It is known that nonsteroidal anti-inflammatory drugs, and particularly celecoxib, decrease the frequency of ACF in humans as well as in rodent models of carcinogen-induced colon cancer (67-70). We compared the changes in gene expression found in normal descending colonic mucosa of patients treated with celecoxib for 1 year (71) to the differences seen in gene expression between ACF and NM from the descending colon. It is provocative that for 21 of the 22 genes that are differentially expressed in ACFL versus NML and for which expression was found to be affected by celecoxib treatment, the celecoxib-driven changes in expression are in the opposite direction: genes with increased expression in ACF compared with NM show decreased expression in celecoxib-treated NM compared with untreated (data available on request).