Progression of Barrett's Metaplasia to Adenocarcinoma Is Associated with the Suppression of the Transcriptional Programs of Epidermal Differentiation

Barrett's esophagus is a specialized intestinal metaplasia of normal squamous to columnar epithelium. Barrett's esophagus is thought to be a premalignant transformation and has been identified in 80% to 100% of esophageal adenocarcinoma of the distal esophagus (1). The etiology of Barrett's esophagus is poorly understood, but chronic gastroesophageal reflux is considered a major contributing factor (2). The presence of Barrett's esophagus increases the risk of developing adenocarcinoma 40- to 125-fold (3). The incidence of adenocarcinoma has increased 3.5-fold over the past three decades, which exceeds that of all other types of cancer (4, 5). Patients with adenocarcinomas of the esophagus present with advanced disease, and 5-year survival is ∼25% (6). Currently, endoscopic surveillance is the only method of identifying patients with early-stage esophageal cancers arising in Barrett's esophagus. In spite of endoscopic intervention, the infrequent association of cancers in Barrett's esophagus makes endoscopy a relatively low-yield intervention, especially considering the costs (7). However, identification of biological markers of Barrett's esophagus progression might identify high-risk patients in whom endoscopy might have a higher yield (8). Expressional profiling represents one such method (9–12). Molecular markers, which identify patients at the higher risk for subsequent transformation of Barrett's esophagus to adenocarcinoma, have yet to be identified. We examined differential gene expression patterns in resected esophageal specimens composed of normal esophageal epithelium, Barrett's esophagus, and adenocarcinoma obtained from the same individual patients. We selected and analyzed gene expression obtained from DNA microarrays and reported that transition from normal esophageal epithelium to Barrett's esophagus and esophageal adenocarcinoma is associated with the suppression of epidermal differentiation programs, including genes localized in the epidermal differentiation complex (EDC). The EDC represents the chromosomal region in 1q21 containing >40 genes responsible for epidermal differentiation. Major families of these genes are S100 proteins, small proline-rich proteins, precursors of the cornified envelope, and intermediate filament-associated proteins (see refs. 13, 14 for reviews). In our investigations, we focused on two members of the EDC, S100A2 and SPRR3. In quantitative reverse transcription-PCR (QRT-PCR) experiments, we found that SPRR3 combined with two transcriptional factors (GATA6 and HOXB7) can successfully discriminate among normal epithelium, Barrett's dysplasia, and Barrett's esophagus–associated adenocarcinomas. These data show that selection of a limited number of markers based on DNA array experiments and their translation into PCR format allows discrimination of normal esophageal epithelium, Barrett's esophagus, and adenocarcinomas with relatively simple techniques.

Clinical samples. Samples of normal, Barrett's, and adenocarcinoma were obtained from fresh pathologic specimens of patients with known Barrett's esophagus and esophageal adenocarcinoma who had undergone esophagectomy. These specimens were processed by pathology within 15 minutes of resection. Samples representative of the various gross histologic types were obtained from experienced gastrointestinal pathologists. These samples were labeled and snap frozen in liquid nitrogen and stored at −80°C for future RNA extraction.

Preparation of RNA and hybridizations. RNAs were purified by combination of column chromatography and TRIzol (Life Technologies, Carlsbad, CA) purification as described in ref. (15). Preparation of labeled cRNA and hybridization with U133A chips was done according to the manufacturer's instructions (Affymetrix, Santa Clara, CA). Data were acquired using MAS 5.0 software (Affymetrix) and exported to MS Excel.

Submission of DNA array data. Data were submitted to the Microarray Analysis and Data Management System database of the University of Chicago, constructed according to the Minimum Information About a Microarray Experiment recommendations and can be accessed through the Web address of the Microarray Analysis and Data Management System database (http://madam.bsd.uchicago.edu:8080/) using public log on. Data were also submitted to the GEO database (National Center for Biotechnology Information), with the accession no. GSE1420.

Data analysis. Throughout this section, patients are denoted by K = 1, …, 8 and genes by letterj.

For data normalization, the expression levels of each array were multiplied by M̄/M, where M is the median expression of the array and M̄ is the overall median expression level. This rescaling makes median expression levels equal across all arrays. For data filtration, we first excluded genes based on present (P) or absent (A) calls as defined by MAS 5.0. Genes were excluded if they were absent for all three tissue types in three or more patients. The genes were further filtered based on signal intensities using receiver operating characteristic analysis as described previously (16, 17). The total number of remaining genes was 8,636.

Next, the significance analysis of microarrays approach (18) was used to identify genes significantly overexpressed and underexpressed in the three pairwise comparisons of Barrett's/normal, adenocarcinoma/normal, and Barrett's/adenocarcinoma. Significance analysis of microarrays identifies genes with statistically significant differences between groups by assigning each gene a score based on the difference in gene expression between two groups (e.g., normal and Barrett's) relative to the adjusted pooled SD of the multiple measurements from both groups. Permutations of the measurements are then used to estimate the false discovery ratio (FDR), the percentage of genes identified by chance. As the cutoff point, we used a Δ value, such that the estimated median number of falsely discovered (called) genes was <1, and required at least a 2-fold expression ratio. In contrast to using a cutoff point of a fixed FDR level, this approach resulted in different cutoff Δs and FDR levels for the three comparisons: Barrett's/normal (Δ = 1.270, FDR = 0.33%), adenocarcinoma/normal (Δ = 1.555, FDR = 0.121%), and adenocarcinoma/Barrett's (Δ = 0.892, FDR = 0.876%). Based on these criteria, we selected 447 genes significantly expressed in adenocarcinomas compared with normal epithelium and 200 genes significantly expressed in Barrett's esophagus compared with normal epithelium. We also found 85 genes significantly different between adenocarcinoma and Barrett's esophagus, among which 45 overlapped with genes significantly different in adenocarcinoma versus normal epithelium. Next, expression ratios of all genes between two tissue groups were compared with the reference “same-to-same” distribution to identify genes for which the ratios are larger than expected. In a simple case with two normal samples, the “same-to-same” distribution is the distribution of L_j = log₂N_j1/N_j2 over all genes j (17). We extend this concept to a situation with more than two arrays by considering

where K is an even number of normal samples and N_jk is the expression level of gene j. For every gene j, we consider the C=(KK/2) possible ways the samples can be separated into two equal size groups, obtaining C = 70 possible combinations for each gene based on the eight arrays, hybridized with RNA from normal tissues. For each of the 70 distributions, we compute quantiles q_0.005, q_0.025, q_0.975, and q_0.995, corresponding to nonparametric 95% and 99% confidence limits. Averaging these over the 70 combinations provides cutoff points for where the bulk of the same-to-same log ratios occur. For each gene j Barrett's/normal and adenocarcinoma/normal (“different-to-same”) log ratio

is then compared with the reference “same-to-same” distributions, and genes with expression ratios outside the cutoff limits are considered to be differentially expressed. Using the geometric mean rather than the nonstandardized ratio allows us to compare directly the distributions of the “same-to-same” and “different-to-same” hybridizations, adjusting for the fact that the “different-to-same” ratios are based on K pairs of tumor and normal expression levels and that the “same-to-same” ratios are based only on K normal expression levels. Thus, L_j can be naturally interpreted as average per-patient log ratio.

Functional selection and prognosticators analysis. For the selection of functionally significant groups of genes, we used OntoExpress software described by Draghici et al. (19) and available at http://vortex.cs.wayne.edu:8080/ontoexpress/servlet/UserInfo. We selected functional groups containing at least three genes and used a binomial distribution with a significance level of ≤0.05. Combining results of functional and expression-based selections, we finally selected 214 genes. Two-dimensional hierarchical clustering of these genes was done based on the estimation of the Euclidean distances by Ward's method using log₂X_ijk/X̄_j^(N), the log-transformed expression levels normalized to the average expression level in the normal tissues, X̄_j^(N). Samples T5 and N8 were removed as outliers. For clustering and data presentation, we used JMP and TreeView software as described previously (20).

To define genes that correlate with the progression of Barrett's esophagus to adenocarcinoma, we considered the 96 genes expressed in both Barrett's esophagus and adenocarcinoma. These genes were separated into two groups based on average between-patient expression: the first group contained genes that were up-regulated from normal to Barrett's esophagus and further from Barrett's esophagus to adenocarcinoma and the second group was defined similarly for down-regulated genes. All other potential patterns were excluded from this analysis. The significance of the difference in expression from normal to Barrett's esophagus and from Barrett's esophagus to adenocarcinoma in each group was evaluated by a one-sided paired t test using a P < 0.05 cutoff (taking into account that we already preselected genes that are either up-regulated or down-regulated in each group).

Quantitative reverse transcription-PCR. cDNA was synthesized using SuperScript II reverse transcriptase (Invitrogen Life Technologies, Carlsbad, CA) following the manufacturer's instructions. cDNA was diluted 1:10 in sterile nuclease free water (Ambion, Austin, TX). Quantitative PCR was done on an ABI 7700 system (Applied Biosystems, Foster City, CA) using SYBR Green PCR reagents in a 25 μL reaction mixture containing 2.5 μL of 10× SYBR Green PCR buffer, 0.25 μL of 10 mmol/L primers, 2 μL deoxynucleotide triphosphate mix, 3 μL of 25 mmol/L MgCl₂, 0.25 μL AmpErase, 0.125 μL Amplitaq Gold, and 2.5 μL of the 1:10 diluted cDNA.

Primers for selected genes were designed based on Unigene reference sequences using PrimerExpress software (Applied Biosystems). For the internal control, we used glyceraldehyde-3-phosphate dehydrogenase (GAPDH). PCR was done for 40 cycles at 95°C for 15 seconds and 60°C for 1 minute after initial incubations at 50°C for 2 minutes and 95°C for 10 minutes.

All samples were amplified in triplicate reactions. The expression of each individual gene was calculated based on the difference between amplification of the individual mRNA template and the internal control (GAPDH) mRNA template. These differences were measured by delta ct (dct) values as described in the manufacturer's instructions (Applied Biosystems). dct values were calculated as (ct_X − ct_GAPDH), where ct_X is the ct value of the specific gene X and ct_GAPDH is the amplification of the internal control. Fold induction relative to GAPDH was calculated as 2^−dct and therefore was equal to 2^{−(ctX − ctGAPDH)}. Ratios of gene X relative to gene Y in the same samples were calculated as R_X/Y =2^{−(ctX − ctY)}. To transform these ratios to a range above 1, we multiplied them by a factor of 100,000. Finally, we transformed the data to log₁₀ format to present them in linear scale. The final expressional value was calculated as:

Discrimination among normal esophageal epithelium, Barrett's metaplasia, and adenocarcinomas based on expressional profiling. Based on the results of the statistical analysis, we selected genes differentially expressed (either up-regulated or down-regulated) in Barrett's esophagus and adenocarcinoma. Figure 1 represents these results as a Venn diagram. Compared with normal esophageal epithelium, we identified 200 genes differentially expressed in Barrett's esophagus and 447 genes differentially expressed in the Barrett's esophagus–associated adenocarcinoma. The criteria and techniques for the selection of the differentially expressed genes are described in Materials and Methods. The comparison of genes differentially expressed in Barrett's esophagus and adenocarcinoma showed that 96 genes were commonly overexpressed in Barrett's esophagus and adenocarcinoma. Adenocarcinoma differentially expressed 351 genes that were not expressed in Barrett's esophagus, and Barrett's esophagus differentially expressed 104 genes that were not expressed in adenocarcinoma. Nonoverlapping genes (351 in adenocarcinoma and 104 in Barrett's esophagus) were used in the subsequent selection of significant functional groups using OntoExpress software and these functions are presented in Table 1. We also selected genes by comparison of the “same-to-same” and “different-to-same” hybridizations as described in Materials and Methods using 99% confidence intervals based on the nonparametric quantile analysis. Combining both approaches, we selected 214 genes, presented in Fig. 2, which represent the two-dimensional hierarchical clustering of selected genes and show the actual discrimination among normal samples, Barrett's esophagus, and adenocarcinoma. The list of these genes is presented in Supplementary Table S1. In the first dimension (vertical), the data separate into three expressional clusters according to the dendrogram presented in Fig. 2 (right). Cluster 1 (80 genes) contains the genes up-regulated in adenocarcinoma compared with normal epithelium. Cluster 2 (63 genes) contains the genes that are sequentially suppressed in Barrett's esophagus and adenocarcinoma compared with the normal epithelium. Cluster 3 (71 genes) contains the genes most drastically suppressed in adenocarcinoma compared with normal epithelium and Barrett's esophagus.

Expressional patterns of normal epithelium, Barrett's esophagus, and adenocarcinoma include different functional groups of genes. Figure 3 identifies the major functional groups associated with the three major expressional clusters presented in Fig. 2. The first expressional cluster contains functional groups of genes associated with immune response, cell-cell signaling and cell-extracellular matrix interactions, control of cell cycle/growth/proliferation, and regulation of transcription and receptor activity (Fig. 3A).

A detailed description of these groups in the terms of the Gene Ontology (GO) database is presented in Table 1. The second expressional cluster also includes control of the cell cycle/proliferation but also contains the specific functions of intracellular transport, bile acid transport, and aldehyde and lipid metabolism (Fig. 3B). The third expressional cluster contains functional groups of genes that may specifically lead to the development of adenocarcinoma. They include ectoderm development/epidermal differentiation, cytoskeleton, control of cell shape and cell-to-cell and cell-to-extracellular matrix interactions, Ca²⁺ binding and metabolism, and a group of proteases and protease inhibitors. Many of these genes are specifically associated with epidermal differentiation and malignant transformation (see Discussion).

Analysis of genes common to Barrett's and adenocarcinoma. Ninety-six genes were commonly expressed in Barrett's esophagus and adenocarcinomas. The list of these genes is presented in Supplementary Table S2. We assumed that these genes might be helpful prognostic and diagnostic markers and/or potential therapeutic targets if they satisfy several conditions. These conditions are as follows: (a) common differential expression (up-regulation or down-regulation) in Barrett's esophagus and adenocarcinoma relative to normal esophageal epithelium, (b) significance of differences between expression in Barrett's esophagus versus normal epithelium and adenocarcinoma versus Barrett's esophagus, and (c) identical expressional changes (e.g., up-regulation in Barrett's esophagus relative to normal and further up-regulation in adenocarcinoma relative to Barrett's esophagus). Based on these assumptions (see Materials and Methods), we selected the 21 genes described in Table 2. Three of these genes (HLA-F, HOXB7, and RRBP1) were presented by two probe sets. These genes are progressively up-regulated or down-regulated in Barrett's esophagus and adenocarcinoma and have significant differences between expressional levels in Barrett's esophagus versus normal epithelium and adenocarcinoma versus Barrett's esophagus.

Quantitative reverse transcription-PCR experiments. For confirmation of our DNA array-based expressional data, we used QRT-PCR as described in Materials and Methods. Individual samples of RNA were profiled using six gene primer pairs for GATA6, HOXB7, TCF3, S100A2, SPRR3, and SCCA1. These genes were selected based on potential prognostic value and levels of differential expression on the arrays. S100A2 and SPRR3 were also selected as components of the EDC (14). Data from these experiments are presented in Figs. 4 and 5. Fig. 4A shows a trend in three transcriptional factors (GATA6, HOXB7, and TCF3) toward up-regulation in the transition from normal esophageal epithelium to adenocarcinomas. In contrast, three genes related to keratinocyte differentiation (S100A2, SPRR3, and SCCA1) trended toward down-regulation during the transition from normal squamous to malignant epithelium (Fig. 4B). These data are consistent with the results from the entire set of tissue types in the microarray analysis as shown in Fig. 4C and D. To select expressional markers most closely correlated with premalignant and malignant changes, we calculated Ps and correlation coefficients for single genes and their combinations, and these are presented in Table 3, where we see that each single marker can significantly discriminate normal esophageal epithelium from adenocarcinoma. However, only HOXB7 can discriminate normal tissues from Barrett's. Neither of these markers alone can discriminate Barrett's from adenocarcinomas and simultaneously normal tissue from Barrett's. We therefore decided to check whether combinations of markers may provide better discrimination than each marker alone. As is shown in Table 3, combinations of GATA6/SPRR3, HOXB7/SPRR3, and GATA6+HOXB7/SPRR3 discriminate better than each of the genes alone. For these gene combinations, the t test provides highly significant values for discrimination between each condition: Barrett's esophagus and normal, adenocarcinoma and normal, and Barrett's esophagus and adenocarcinoma (P = 0.0012, 1.462 × 10⁻⁶, and 0.0013, respectively). Additionally, mixed-effects ANOVA models were used to determine whether there are differences in expression of GATA6/SPRR3, HOXB7/SPRR3, and GATA6+HOXB7/SPRR3 combinations between the three groups, accounting for the presence of intrasubject correlation due to the presence of several subjects with multiple samples. These analyses confirmed that the expression levels of these combination markers are significantly different among normal, Barrett's, and tumor tissue types (data not shown). In addition, for the combinations listed, the correlation between expression and tumor progression is higher than for either gene alone. However, increasing the number of combined markers may not improve correlation with clinical diagnosis. As is shown in Table 3, double markers (GATA6/SPRR3 and HOXB7/SPRR3) did as well as triple markers (GATA6+HOXB7/SPRR3). These differences suggested the use of the GATA6/SPRR3 ratio as a potential molecular marker of transformation. The performance of this ratio as a marker is presented in Fig. 5, which shows that the ratio of GATA6 to SPRR3 progressively increases along the spectrum from normal epithelium to Barrett's dysplasia to adenocarcinomas. In the range of the 95% confidence interval (estimated by normal epithelium values), specificity of the test (measured by percentage of negative cases) is equal to 89% (8 of 9). Sensitivity of the test (measured by percentage of positive cases) is equal to 28.6% (2 of 7) for Barrett's esophagus and 100% (7 of 7) for adenocarcinoma. With a cutoff level equal to 67% confidence interval, specificity of the test is also equal to 89%, sensitivity to Barrett's esophagus is 86% (6 of 7), and sensitivity to adenocarcinoma is 100% (7 of 7). These data show that selection of the markers based on DNA array experiments allows significant separation of pathologic conditions and that the number of these markers can be reduced to several specific genes that can make such an estimation feasible in large populations.

Several gene families are consistently down-regulated in Barrett's and adenocarcinoma compared with normal esophageal epithelium and/or adenocarcinoma compared with Barrett's esophagus (Fig. 3B and C). These genes include aldo-keto reductases, aldehyde dehydrogenases (located predominantly in expressional cluster 2), dual-specificity phosphatases, annexins, chloride channels, keratins and genes involved in the formation of desmosomes, and the cornified envelope of squamous epithelium (see refs. 21–23 for reviews). The genes involved in these processes are S100 proteins, small proline-rich proteins, involucrin, periplacin, desmocollins, and desmoglein 3. Many of these genes are in the specific chromosomal loci called the EDC (13, 14). The EDC is the chromosomal loci containing >40 genes involved in epidermal differentiation and localized in 1q21 (see refs. 13, 14, 24). Excluded from the genes presented in Figs. 2 and 3 were the 96 genes commonly expressed in Barrett's esophagus and adenocarcinoma (see also Supplementary Tables S1 and S2). Among these, two genes are also involved in epidermal differentiation: transglutaminase-3 (25) and NICE-1, with the latter also localized in 1q21 (14). Our data indicate that the formation of Barrett's esophagus–associated adenocarcinoma is accompanied by the suppression of the genetic programs of epidermal differentiation, including genes clustered in the EDC, which raises questions regarding the mechanisms of suppression of the EDC. One explanation may be the deletion or loss of heterozygosity of 1q21, although data from the literature are controversial. Loss of heterozygosity of 1q21 was shown in six gastrinomas (26), but in the majority of other cases, cytogenetics, comparative genomic hybridization, and fluorescence in situ hybridization analysis reveal amplifications of 1q21 or translocations involving this locus (27–31). Recent data indicate that activity of EDC genes may be controlled by highly coordinated changes of local chromatin structure, specific spatial distribution of chromatin regions, and temporally and spatially ordered transcriptional regulation (13, 32, 24). These data show that in interphase nuclei of keratinocytes (but not lymphoblasts) the EDC is positioned external to chromosome 1 (13). The same situation was shown for the transcriptionally active MHC locus on chromosome 6 (33). Interestingly, that suppression of histone deacetylases led to the “open” conformation of the EDC and increased expression of some genes. Silencing of genes in this locus was connected with CpG methylation (24). Such coordinated changes of conformation may be tightly coordinated with transcriptional control. From that point of view, our data provide interesting correlations between transcriptional activity of the EDC and expression of different transcriptional and chromatin remodeling factors, mostly grouped in expressional cluster 1 (see Figs. 2 and 3A; Table 2). For example, the evolutionarily conserved CBX3 protein involved in the condensation of heterochromatin regions (34) increased 1.3-fold in Barrett's esophagus and 2.8-fold in adenocarcinoma relative to normal epithelium, consistent with the possibility of condensation of the EDC region (see Fig. 3A; Supplementary Table S1). HOX genes have been implicated in human skin development (35). HOXA7 has been shown recently to regulate transglutaminase-1 and suppress keratinocyte differentiation (36). HOXB7 was identified as the promoter of ovarian carcinomas (37) and identified recently as an important factor in the tumor-associated angiogenic switch (38). TCF3 (see Table 2) was identified as a suppressor of epidermal differentiation (39) and GATA factors as regulators of epithelial differentiation and skin development (40, 41). In Caenorhabditis elegans, genes encoding GATA factors (elt-5 and elt-3) control differentiation of epidermal cells throughout animal development (42). In addition to the transcriptional factors, several other genes indicated in Table 2 are related to epidermal differentiation and oncogenesis. RRBP1 is involved in the epithelial-mesenchymal transformation (43), and KANK has been reported as a tumor suppressor (44). Its orthologue, VAB-19, is critical for epidermal differentiation in C. elegans (45). Based on the progressive behavior of these genes and significance of differences, we may suggest that they may be not only useful markers but also potential targets for pharmacologic intervention.

Two recent publications where DNA arrays or subtractive libraries were used for analysis addressed the issue of differential gene expression in Barrett's esophagus, adenocarcinoma, and squamous cell carcinoma formation compared with normal esophageal epithelium and are germane considering our observations. Barrett et al. compared normal esophageal, gastric and duodenal squamous epithelium, and Barrett's epithelia (12). Of the 203 genes included in the squamous epithelium cluster, 37 matched with our selection (Fig. 2; Supplementary Table S1). According to the observations of Barrett et al., genes associated with keratinocyte differentiation were up-regulated in the squamous epithelial cluster compared with duodenum and gastric epithelium. According to our observations, these genes are down-regulated in adenocarcinoma. In another report, human esophageal squamous cell carcinoma and normal esophageal tissues were used for construction of subtractive λ-libraries, which were screened by custom nylon arrays (9). Fifteen genes identified in this study as down-regulated were common or homological to the genes found in our current report. These genes also belong to the small proline-rich proteins, annexins, S100 proteins, keratins, and transglutaminase-3. These data show that results from different laboratories with different platforms for expressional profiling analysis lead to the consistent selection of genes involved in epidermal differentiation. In contrast to the observations with normal squamous epithelium where previous studies show up-regulation of differentiation-associated genes in squamous epithelium (12), our data show marked down-regulation of the same genes in Barrett's esophagus–associated adenocarcinoma. We suggest that such dedifferentiation is important for tumor clones to release from tight junctions with the surrounding cells through desmosomes and cornified envelope. This is a potential mechanism for invasive growth and formation of metastasis. Understanding of the regulatory mechanisms of suppression of epidermal differentiation programs and the EDC are critical for potential control of adenocarcinoma development. Our data show that development of Barrett's esophagus–associated adenocarcinoma is connected with the suppression of the epidermal differentiation program and associated with activation or alteration of expression of specific regulatory genes. Identification of these genes may provide insights into the etiology and development of Barrett's esophagus and esophageal adenocarcinomas. These findings may also provide a set of markers to determine pathologic stage and risk of adenocarcinoma development as well as define new potential therapeutic targets. QRT-PCR experiments provide further specification of such markers. Six selected genes (GATA6, HOXB7, TCF3, S100A2, SCCA1, and SPRR3) showed significant correlation during progression from normal tissue to Barrett's esophagus and adenocarcinoma and were able to be quantified by QRT-PCR in RNA samples extracted directly from surgical specimens. All are directly involved in or associated with epidermal differentiation. Our experiments showed that the ratio of GATA6 to SPRR3 can significantly discriminate normal from malignant esophageal epithelium. Most importantly, our data suggest that this ratio may vary in Barrett's dysplasia, perhaps reflecting risk of Barrett's transformation. With relatively small size of the sample investigated in this report, we relate definition “significant” to discrimination of groups inside this experimental sample. Further investigations with a larger number of patients with clinical follow-up may reveal whether it can be used for estimation of the risk of Barrett's progression to adenocarcinoma.

These investigations show that the selection of markers based on DNA array experiments may provide molecular criteria for discrimination of pathologic conditions of esophageal epithelium. They also show that these data may be used for selection of a limited number of markers that can be screened in large populations by conventional techniques like PCR.

Grant support: NIH grant PO1 CA 71933 (R.R. Weichselbaum) and Claire and Dennis Nardoni Fellowship in Surgical Oncology (E.T. Kimchi).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

We thank Edwardine Labay for active help with the preparation of this article.