Glucose and Insulin Treatment of Insulinoma Cells Results in Transcriptional Regulation of a Common Set of Genes

The MIN6 insulinoma cell line was a gift from Dr. Jun-ichi Miyazaki (Osaka University, Japan). MIN6 cells were maintained in Dulbecco’s modified Eagle’s medium (DMEM) containing 25 mmol/l glucose, supplemented with 15% heat-inactivated fetal bovine serum (FBS), 100 units/ml penicillin, 100 μg/ml streptomycin, 100 μg/ml l-glutamine, and 5 μl/l β-mercaptoethanol in humidified 5% CO₂, 95% air at 37°C (15). MIN6 cells were used between passages 27 and 33. Insulin secretion of the cells was checked by a standard static incubation method, and those maintaining glucose responsiveness with normal glucose threshold for islet β-cells (between 5 and 10 mmol/l glucose) as well as maximum secretion capacity (maximum secretion at least five times higher than the basal secretion) were used for experiments. Cells were grown in 150-mm dishes, and once they reached 60% confluence, medium was switched to DMEM containing 5 mmol/l glucose and 5% FBS for 18 h before stimulation (Fig. 1A). The stimulating agent (25 mmol/l glucose, 50 mmol/l KCl, or 100 nmol/l insulin) was added directly to the medium, and cells were incubated for 45 min. When indicated, LY294002 (50 μmol/l; Cell Signaling, Beverly, MA) was added to the medium 15 min before treatment with stimulating agents. After stimulation, cells were harvested and total RNA was extracted from cells using TRIzol reagent or RNeasy columns according to the manufacturer’s protocol (Invitrogen and Qiagen, respectively).

First-strand cDNA was generated by oligo-dT primed reverse transcription (Superscript II; Invitrogen) using the 3DNA Array 50 kit (Genisphere) (16). Modified oligo-dT primers were used in which a fluorophore/dendrimer specific oligo sequence was attached to the 5′ end of the dT primer. For RNA expression level comparison, samples were paired and then purified and concentrated using Microcon microconcentrators (Millipore) according to the manufacturer’s protocol.

After the sequencing of cDNA libraries derived from various mouse pancreas tissues gathered by the EPCon (12,13), expressed sequence tags (ESTs) were evaluated for similarity to the existing nonredundant GeneBank entries (http://www.ncbi.nlm.nih.gov/Genbank/index.html) and for redundancy within the EST clone set using standard BLAST analyses (17).

A nonredundant set of clones was selected and PCR amplified in preparation for the microarray slide production. PCR products were purified using Millipore 96-well plates, DNA quantity was determined using Picogreen (Molecular Probes) reagents, and, finally, each of the PCR products was normalized with Millipore purified water to a standard concentration of 400 ng/μl. Before microarray spotting, all PCR products were adjusted to a final DNA concentration of 200 ng/μl in a spotting buffer of 3× sodium chloride-sodium citrate (SSC) and 0.75 mol/l betaine. PCR products were spotted onto epoxy-coated slides (MWG Biotech) using an ArrayMaker2 arrayer (designed by P. Brown’s laboratory; Stanford University). Spotted slides were incubated 12–14 h at 40% humidity in a 42°C oven. After incubation, the slides were cross-linked using the UV Stratalinker 2400 (Stratagene) at 700 μJ ×100 and processed by the following steps: 1) gently shaking the slides in a 0.2% SDS bath for 2 min; 2) three room temperature water bath washes, each at 1 min; 3) a 50°C water bath for 20 min; 4) a 95°C water bath for 2 min; and 5) spinning the slides dry. The slides were stored in a desiccant cabinet for future hybridization experiments.

The different clones on the array were identified through BLASTn similarity matches using the NCBI BLAST (http://www.ncbi.nlm.nih.gov/blast) and WU-BLAST version 2.0 (17) against the nonredundant subsets of the public Mouse databases and RefSeq (http://www.ncbi.nlm.nih.gov/RefSeq), and only matches with an E value below at least 1 × 10⁻⁵⁰ and a score superior to 200 were considered as a match. Annotation was further confirmed comparing the results to the annotation provided on the EPCon website (http://www.cbil.upenn.edu/EPConDB) for the PancChip 5.0 and manually curating discrepancies. Further annotation (official names and symbols, gene ontology [GO] functions) was gathered from Source (http://source.stanford.edu) and in some cases imported directly from LocusLink (http://www.ncbi.nlm.nih.gov/LocusLink).

Two hybridizations were carried out in a sequential manner. The primary hybridization was performed by adding 38 μl of sample to the microarray under a supported glass coverslip (Erie Scientific) at 50°C for 16–20 h at high humidity in the dark. Before the secondary hybridization, slides were gently submerged into 2× SSC, 0.2% SDS for 12 min; transferred to 2× SSC for 12 min; transferred to 0.2× SSC for 12 min; transferred to 95% ethanol for 2 min; and then spun dry by centrifugation. Secondary hybridizations were carried out using the complimentary capture reagents provided in the 3DNA Array 50 kit (Genisphere).

Slides were scanned on a Perkin Elmer ScanArray Express HT scanner to detect Cy3 and Cy5 fluorescence. Laser power was kept constant for Cy3/Cy5 scans, and photo-multiplier tube was varied for each experiment based on background fluorescence. Gridding and analysis of images was performed using QuantArray (Perkin Elmer). Each spot was defined on a pixel-by-pixel basis using a modified Mann-Whitney statistical test.

To compare each of the conditions with every other condition, samples were paired as indicated in Fig. 1B. For each pair, transcripts from one condition were labeled with either Cy3 or Cy5 and hybridized with the Cy5- or Cy3-labeled transcripts from the partner condition. Another pair of RNAs was also labeled inverting dyes (“dye-flip” hybridization). This sample pairing and dye flip required a total of 12 microarrays per experiment (each arrow in Fig. 1B represents one slide of microarray with a pair of RNAs). As a result, each sample for an experiment was hybridized six times: against three other conditions twice, once with Cy3, and the other with Cy5. RNA from the first experiment was hybridized to two sets of microarray, a 5,700 clone set and then to a 9,700 clone microarray consisting of a superset of the 5,700 clone set; therefore, in each case, 12 slides were used. Therefore, a total of 24 microarray slides was used for these analyses.

In the microarray studies performed herein, the probes that gave a signal intensity ≥2.0-fold above the corresponding background intensity in both Cy5 and Cy3 channels were chosen for calculation of the fold change. The intensity of each spot was first adjusted by subtracting background intensity. Then, the log ratio of Cy5 and Cy3 channel intensity of each spot was calculated, and the median of the log ratio from all probes was subtracted from each log ratio (the global normalization procedure [18]). This is based on the assumption that the expression levels of most genes are unchanged, and thus log fold change should center at zero. Indeed, data from all hybridizations indicated that the expression of the majority of genes (∼95%, data not shown) were unchanged.

The log ratios from six pairs of dye-flip hybridizations were used to estimate probe-specific measurement variation, usually derived from self-vs.-self experiments (19). Probe-specific measurement variation was subtracted from log fold change to calculate the final normalized log fold change (probe-specific normalization).

From our pairing scheme, two direct and two indirect ratios were obtained when two conditions were compared. When glucose-stimulated and -unstimulated samples were compared, for example, two ratios were acquired from dye-flip hybridizations directly comparing those two samples. Two additional ratios were calculated indirectly from eight hybridizations involving all glucose, KCl, insulin, and unstimulated samples. These four ratios were used to calculate average fold change as well as standard error. The 95% CI was calculated with attention to the change of distribution when converting log fold change to fold change (20). Criteria were set to assess changes in gene expression that included a fold change of 1.3 or more compared with “unstimulated.” This criterion was selected after observing that the mean variance of all probes derived from a single self-to-self hybridization from the first 5,700 and 9,700 probe sets was 0.12 (Fig. 1C). Because we had four measurements, that mean variance was further reduced to 0.06 (= 0.12/4). With this small variance, the false-positive rate is 0.0032%. With similar multiple hybridization and cDNA microarray, a fold change of 1.3 was used to identify significant gene expression change in other studies (21–23). Additionally, significant fold changes required that the 95% CI of the fold change, two standard errors away from the mean fold change, excluded 1. Thus, any fold changes meeting these criteria could be considered significant with P < 0.05.

Hierarchical clusters were performed using the Genesis software version 1.3 (Institute for Biomedical Engineering, Graz University of Technology, Graz, Austria) (24). Heat maps generated by hierarchical clustering were created using the average linkage clustering (Euclidian distances) directly from the logarithmic ratios (base e), and all color ranges for the “heat maps” were adjusted to a maximum ratio of 1.

Total RNA was isolated, and 1 μg was used to prepare cDNA, primed with random hexamers, and reverse transcribed with Superscript II (Invitrogen) according to manufacture’s protocol. Quantitative RT-PCR (qRT-PCR) was performed by monitoring in real time the increase in fluorescence of the SYBR Green dye (ABI) as described (25,26) using the ABI 7000 sequence detection system (Applied Biosystems). For comparison of transcript levels between samples, a standard curve of cycle thresholds for serial dilutions of a cDNA sample was established and then used to calculate the relative abundance of each gene. Values were then normalized to the relative amounts of 18S ribosomal RNA, which were obtained from a similar standard curve. All PCRs were performed at least in replicates of four. Standard error of the quantity of transcript normalized to the amount of 18S ribosomal RNA was calculated from a formula with consideration of error propagation. When gene expression levels of two conditions were compared, the ratio was expressed with standard error calculated from the same formula. Specificity of each primer pair was confirmed by melting curve analysis and agarose gel electrophoresis of PCR products. Sequences of primers used in this study are included in an online appendix at http://diabetes.diabetesjournals.org.

The MIN6 insulinoma cell line provides an excellent experimental model, because these cells retain gene expression and insulin secretory responses to physiological changes in glucose concentrations (3,4,6,15,27). Cells were starved for glucose and serum overnight and then either harvested at the 0 time point (unstimulated) or further incubated with 25 mmol/l glucose for 45 min. Another set of cells was treated with 50 mmol/l KCl to ascertain the effects of depolarization on gene expression. A third set of cells was treated with insulin (100 nmol/l) to examine potential autocrine/paracrine effects on gene expression (Fig. 1A). Concentrations of those stimuli were chosen to elicit maximal short-term responses, after observing that the media of insulinoma cells contains 17.4–39.8 nmol/l insulin (M.O., M.A.P., unpublished data). At the end of the 45-min incubation, RNA was extracted and pairs of RNA samples (e.g., unstimulated vs. stimulated) labeled with either Cy3 or Cy5 were hybridized together (in duplicate with dye flip) onto cDNA microarrays, as illustrated in Fig. 1B. For this experiment, the same RNA samples were hybridized to two different sets of microarrays (5,700 or 9,700 cDNA probe sets) providing four to eight observations for each pair. Criteria for significant alteration of expression were defined as 1) fold stimulation or inhibition of ≥1.3 and 2) 95% CI of the fold change excluding 1.0. Based on the small variances we observed (Fig. 1C), a 1.3-fold change would yield false-positive findings <0.01% (see research design and methods). Independent confirmation of these fold changes was obtained on a subset of the genes with altered expression as described later.

Stimulation by glucose resulted in a significant increase in mRNA levels for 72 of the ∼9,700 cDNAs on the arrays. The 50 genes most increased by glucose treatment are shown in Table 1. The complete list of regulated genes is shown in Table 1S in the online appendix. Glucose-stimulated genes of diverse GO functions, described more fully below, included two immediate early genes previously identified as glucose responsive in islet β-cells, early growth response 1 (Egr1) (10,28), and an inhibitor of DNA binding 1 (Idb1) (29). Several other immediate early genes found to respond to mitogenic stimuli in other tissues (immediate early response 2 [Ier2] and 3 [Ier3] and inhibitor of DNA binding 2 [Idb2]) were also shown to respond to glucose treatment. Interestingly, among the glucose-regulated genes, the overwhelming majority was not previously known to be glucose responsive. Glucose treatment also resulted in significant downregulation of 21 genes (Table 1), of which only one, Chop (Ddit3), was previously described as glucose regulated (3). Chop has been incriminated in cell cycle arrest and apoptosis (30). At least 13 of the glucose-responsive genes do not yet have a match in LocusLink or UniGene.

Because islet β-cell glucose metabolism is accompanied by closure of ATP-sensitive potassium (K_ATP) channels followed by depolarization, we next sought to study the genes regulated by KCl-induced depolarization and the degree of overlap of these genes with those activated by glucose. A total of 202 genes were activated by KCl treatment (139 upregulated and 63 downregulated [Table 1S in the online appendix]). Notably, examination of Table 1 revealed that 47 out of the 50 glucose regulated genes listed here were also activated by KCl. KCl treatment also resulted in a decrease in mRNA levels for at least half of the glucose-repressed genes (Table 1). Comparison of all the glucose- or KCl-regulated genes revealed that >80% of genes activated by glucose were also activated by KCl (Table 1S). This is graphically represented in a hierarchical cluster analysis, as shown in Fig. 2A. The results of these experiments suggested that depolarization is a major mechanism of transcriptional regulation induced by glucose.

To further examine the role of depolarization and glucose-mediated gene expression, diazoxide, a K_ATP channel activator, was used to assess the glucose-activated gene expression in the absence of depolarization. A subset of genes expressed within all ranges of fold induction was analyzed with qRT-PCR, as shown in Table 2. For all five genes tested here, glucose- and KCl-induced gene activation was markedly reduced by diazoxide. These data therefore further substantiate the role of depolarization in gene regulation by glucose.

There were 129 genes whose expression was significantly increased by insulin treatment after a 45-min incubation and 25 genes whose expression was significantly suppressed. Comparisons to the top 50 genes activated by glucose stimulation are shown in Table 1 (complete list in Table 1S). Whereas some of these genes had been previously reported to be regulated by insulin treatment in other tissues (31), none had been noted to be altered by insulin treatment in insulinoma cells.

Examination of the genes whose expression was significantly affected by insulin treatment revealed a surprising result, i.e., there appeared to be considerable overlap between the genes regulated by insulin and by glucose, as can be readily seen in Table 1, where glucose- and insulin-activated genes are compared (bold). The expression profiles after the two treatments are also graphically assessed in Fig. 2B by a hierarchical cluster analysis. Immediately apparent is that 8 of the top 10 genes activated by glucose were not at all stimulated by insulin treatment. On the other hand, most of the remainder of the glucose-regulated genes was expressed to the same extent by insulin treatment.

The observation by microarray analysis that genes appeared to be activated by both glucose/depolarization and insulin required validation by independent means. This was performed by qRT-PCR on RNA samples from a subset of genes that were shown to be activated on the microarrays (experiment 1). As can be seen in Table 3, in general, the fold changes noted with qRT-PCR were larger than those observed with microarray, yet there was a highly significant correlation (R² = 0.90, data not shown) between the levels of gene expression measured by the two methods for experiment 1.

Additionally, biological replication was sought by qRT-PCR analysis of RNA obtained from two additional independent experiments (experiments 2 and 3). As shown in Table 3, a total of 13 genes was chosen for this analysis from genes with various expression profiles that appeared to be activated on the microarrays by at least one stimulus. These results clearly demonstrated that small fold changes (e.g., <2) can be observed by microarray and validated with qRT-PCR and that the changes in gene expression after stimulation by all three agents for this representative set of genes were reproducible in three independent experiments. Importantly, these results confirmed that some genes could rapidly be activated by both glucose and insulin under these conditions.

The observation of a common set of genes activated by glucose and insulin was consistent with glucose/depolarization and insulin inducing gene transcription via completely separate pathways that target the same genes or that the two stimuli activate pathways that converge on a common signaling pathway that activates the same genes. To differentiate between the two possibilities, gene expression was evaluated in another set of experiments using a pharmacological inhibitor (LY294002) for PI 3-kinase, known to be involved in insulin signaling (32). KCl and insulin treatment were compared in this experiment, because KCl induced depolarization and glucose stimulated considerably common sets of genes, but the fold changes by KCl were more robust (Table 1). After 45 min of KCl or insulin treatment, the cells were harvested and analysis was performed in identical fashion to the previous experiments. As shown in Fig. 3, the genes were represented in log fold changes to allow a representation of the transcripts on a same scale for both up- and downregulation. The addition of LY294002 resulted in almost complete elimination of gene regulation by each stimulus. These results suggest that glucose, acting through depolarization, and insulin share common signaling pathways leading to gene expression and that this PI 3-kinase-dependent pathway seems to play a major role.

From earlier studies, genes acutely regulated by glucose and depolarization were mostly transcription factors or transcriptional regulators, such as Egr1, Egr2, c-myc, c-fos, Idb1, and Idb2. In the current experiments, the expression of a common set of genes was activated or inhibited by glucose and insulin. More than half of these genes can be characterized according to GO functions (Fig. 4). Two major categories comprised over 70% of the functions, with catalytic (i.e., enzymatic) activity comprising the largest. Binding activity, the second largest category observed, includes ligands, receptors, and DNA binding proteins. The next most abundant categories included genes classified as involved in translational or transcriptional regulator activity. Notably, genes suppressed by glucose/depolarization and by insulin included several genes classified as pro-apoptotic, e.g., Chop and Scotin. The major conclusion from this analysis is that glucose/depolarization and insulin treatment resulted in alteration of genes not limited to transcriptional regulators but rather genes with pleiotropic effects in terms of GO functions.

In this study, the glucose-induced early transcriptional responses of mouse insulinoma cells were observed as a model for defining initial signaling events that ultimately link the β-cell’s environment with long-term effects such as proliferation and function. The results of the current experiments have potentially important biological implications, although the conclusions are to some extent based on small fold changes in gene expression resulting from each stimulus. Through technical replicates and dye-flip hybridizations, we ensured that the technical variations on our experiments could be minimal (33), allowing us to reliably detect fold changes of ≥1.3 for >99.9% of the probes. The significance of these somewhat modest changes is further supported by the consistency of results from duplicate probes, the low false-positive rate with self-vs.-self analysis, the qRT-PCR data, and the biological replicates.

A number of biologically relevant observations were made by the results of these experiments. 1) Over 90 potential candidates either up- or downregulated by glucose were uncovered in this study. Many of these genes are currently unknown ESTs and are not yet listed in UniGene or AllGenes (DoTS assemblies). 2) By comparing gene expression profiles after glucose- and KCl-induced depolarization, there was considerable overlap of genes activated by both agents, consistent with the hypothesis that glucose activation of rapidly responding genes is mediated predominantly via depolarization. These results are in accord with those of earlier studies showing that expression of several early response genes depended on depolarization (10,29). This premise was further confirmed by the results of the study of the measurements of gene expression in the presence of the depolarization inhibitor diazoxide in a subset of these genes. 3) The current results showed that exogenous insulin treatment at 100 nmol/l activated transcription of a number of newly described early response genes in insulinoma cells. The most notable finding, however, was that glucose and insulin treatments each activated a common subset of genes among several thousands on the arrays.

The validity of the observed microarray results ideally requires both technical validation and biological replication. Technical validation can be seen in that several clones corresponding to the same gene were present on our microarray set, for example, splicing factor arginine/serine rich (Sfrs5) and inhibitor of DNA binding 2 (Idb2) (Table 1). We also used qRT-PCR for validation of the microarray results and observed excellent correlation between microarray and qRT-PCR results testing the same samples (Table 3). Further, biological replication was achieved by two additional independent experiments in which expression of a subset of genes was measured by qRT-PCR (Table 3).

Although a short 45-min treatment is suitable for evaluating initial transcriptional responses, the long-term changes that may result from these early changes remain to be determined. When MIN6 cells were incubated in high glucose for 4 h, preliminary results indicated that the majority of genes regulated at 45 min returned to baseline. This finding was consistent with characteristics of early response genes (8). On the other hand, new sets of genes not previously regulated at 45 min were now found to be regulated at 4 h (M.O., C.C.-M., Y.Z., E.B.-M., M.A.P., unpublished data). This may suggest cascades of gene expression in which genes responding early to stimuli may initiate or suppress expression of other genes.

The maximum concentrations of stimuli for glucose, KCl, and insulin were used. One may raise an issue regarding potential effect of osmotic stress as a glucose stimulus. In previous publications (10,29) and unpublished observations (E.B.-M. and M.A.P.), we have shown failure of nonmetabolizable glucose analogs to activate transcription, that the effects of glucose were blocked by diazoxide, and that KCl in the presence of a PI 3-kinase inhibitor showed no activation of genes (Fig. 3).

Another limitation of the current study was that the levels of expression of genes activated by glucose and insulin were for the most part in the 1.3- to 1.5-fold range. However, multiple determinations of fold change by microarray resulted in small variances of measurements, as seen in Fig. 1C, and several of the glucose and insulin response genes were validated by other means (Table 3). Although we believe that in general these 30–50% changes in gene transcriptional rates are significant, rather than speculating on the importance of these modest fold changes in gene expression, we found that the main conclusion of the study is that common signaling pathways are similarly activated by two distinctive important β-cell stimuli: glucose and insulin.

Not all β-cell early response genes such as c-myc and c-fos were observed in the present study (7,34). These genes were not included in the original EPCon arrays, presumably because of the low abundance of those transcripts. Despite those omissions, our current studies are confirming and extending the previous findings, such as Egr1, and highlighting the use of microarray to assess similarities of treatments (e.g., glucose and insulin) and to identify signaling pathways activated by those treatments.

The microarray results with insulinoma cells are limited to β-cells adapted to culture, yet isolated islets are a mixture of several cell types and are further complicated by this issue. Isolation of pure β-cells from isolated islets raises additional concerns regarding the conditions of preparation, so every method has its limitations. The results of our experiments in insulinoma cells can now be used to measure responses to primary islets in culture, both rodent and human. Importantly, our microarray data now identify dozens of new early response genes, some known and some only described in the databases as ESTs.

Remarkably, >90% of the effects of glucose on gene transcription were mimicked by KCl-induced depolarization (Fig. 2B). Although Ca²⁺ influx after glucose-induced depolarization is a critical mechanism resulting in insulin secretion, the role of depolarization in glucose-mediated gene expression has not been fully evaluated. Only few genes were shown to be regulated via depolarization in previous studies of early response genes, but the results of this study demonstrated that the majority of the glucose-regulated genes are activated through depolarization. Considering the results of previous studies performed on limited numbers of genes (10,29), this is likely to result in activation of Ca²⁺-regulated kinases leading to phosphorylation and activation of transcription factors. Although signaling events leading to transcriptional activation or suppression after depolarization remain to be further defined, the current studies indicated that the PI 3-kinase pathway seems to be involved (Fig. 3). It remains to be determined whether PI 3-kinase activation occurs via an autocrine/paracrine effect of insulin.

The observation that glucose and insulin induce an overlapping set of genes is interesting in light of recent data that provide evidence for an important role of insulin in the regulation of gene transcription in β-cells. Insulin stimulation of β-cells in vitro results in transcriptional induction of the preproinsulin, liver-type pyruvate kinase, and acetyl-CoA carboxylase I genes by activation of the insulin receptor substrate/PI 3-kinase pathway (35–37). Mouse models (38–40) clearly indicated the importance of the roles of the insulin receptor (βIRKO mouse) and the downstream targets of the insulin-signaling pathway (IRS2) in islet β-cell survival and function, but its precise mechanisms are still elusive.

Regarding the GO functions of the genes commonly regulated by glucose and insulin, there was a high proportion of regulated genes that belonged to catalytic and binding activity classes, and there were lower proportions in the other categories. The differences in the number of genes in these GO function categories differed slightly from the distribution of genes on the entire array (data not shown), although the array as a whole is not representative of the MIN-6 cells, because it comprises genes expressed in all cells of the pancreas. Further, it would be speculative to suggest that these categories of biological functions represent true changes at the cellular level considering the small numbers of genes in each GO category.

The broadly overlapping set of genes common to both glucose and insulin treatment in the current studies suggested at least three possible explanations. One is that these two islet β-cell stimuli act through independent pathways that finally converge on the same gene promoters (Fig. 5A). Another is that the two stimuli act on two different pathways that converge on a common pathway that ends by activating the same gene promoters (Fig. 5B). This latter alternative seems the most likely because the PI 3-kinase inhibitor almost completely eliminated gene expression induced by both stimuli (Fig. 3). A third possibility is that the major effect of glucose/depolarization on β-cell gene transcription is through an autocrine/paracrine effect of glucose-stimulated insulin secretion (Fig. 5C) (41–43). It has been difficult to test the autocrine/paracrine effects of insulin, because MIN6 cells are immersed in a large amount of insulin and anti-insulin receptor antibody studies have been inconclusive (M.O., E.B.-M., M.A.P., unpublished data). To test the third hypothesis, experiments are currently underway to silence insulin receptor gene expression with small interfering RNA (44) in MIN6 cells.

This work was supported in part by National Institutes of Health Grants DK16746, DK56954, and DK99007 (to M.A.P.) and the Washington University Diabetes Research and Training Center.

We gratefully acknowledge the D. Melton lab (Harvard University) and K. Kaestner and C. Stoeckert labs (University of Pennsylvania), as well as Ellen Ostlund, Jessica Murray, Sandy Clifton, Hiroshi Inoue, Chris Sawyer, Mike Heinz, Elaine Mardis, and other members of the Genome Sequencing Center for their work with EPCon and microarrays. We would also like to thank Gary Stormo for helpful suggestions and Gary Skolnick for preparation of the manuscript.