Compound Characterization Using Gene Microarrays

Abstract

Ligands for peroxisome proliferator activated receptors (PPARs) are used as hypolipidemic agents and for the treatment of type II diabetes, PPARa and PPARg, respectively. PPARs are members of the nuclear receptor family of ligand activated transcription factors which exert their effects primarily by regulating gene transcription. We have used gene microarrays to characterize PPAR ligands (profiling) and to investigate their mechanism of action. Compound profiles based on gene expression allow for novel ligand groupings when compared with groupings based on compound structure or on the results from single variable assays.

INTRODUCTION

Compounds and drugs are typically characterized by their chemical structures. However, when used as treatments, compounds and drugs within a single structural group seldom exhibit common physiological effects. For example, thiazolidinediones (TZDs) are ligands for the nuclear receptor PPARγ (reviewed in Elbrecht et al., 2000). Not all TZDs are ligands for PPARγ, but, nevertheless, they are a class of compounds, which, through their interaction with PPARγ, exert certain physiological effects seen as typical for the class. Specifically, they function as insulin sensitizers, and treatment of patients suffering from non-insulin dependent diabetes melitus (NIDDM) with TZDs improves glucose tolerance and reduces circulating glucose, insulin and triglyceride levels. There are many TZDs that do not bind to PPARγ and consequently don't work as insulin sensitizers, yet they still fall into the same chemical structural class as those that do. Once a lead chemical structure has been identified, the key to drug discovery lies in the ability to find physiologically active and specific members of the group. We would like to suggest that grouping based on physiological effects, as measured by transcription profiles, would be advantageous in the identification of drugs from a group of compounds.

Although their interaction with PPARγ is well documented (Lehmann et al., 1995, Elbrecht et al., 1996, Elbrecht, 1998), the mechanism of action of TZDs remains unclear. PPARγ belongs to the nuclear receptor family, and like other members of this family, it probably exerts its effects largely by regulating gene transcription (reviewed in Mangelsdorf et al., 1995). Some of the genes regulated by PPARγ have been identified, and models have been developed to describe their possible roles in the process of insulin sensitization, still, the picture is far from complete.

Similarly, the fibrates are compounds which bind and activate PPARα and function as hypolipidemic agents in humans (reviewed in Elbrecht et al., 2000). As the name for the receptor implies, ligands for PPARα cause peroxisome proliferation in rodents; which serves to increase fatty acid oxidation in these organelles. Interestingly, while fibrates also improve lipid profiles in humans, they do not appear to cause the peroxisome proliferation with ensuing hepatomegaly and hepatic cancer that has been observed in rodents. This fortunate difference has permitted the use of a variety of fibrates for the treatment of human disease. As with TZDs, the fibrates tend to be viewed as a class of compounds with a similar set of physiological profiles.

Together, PPARγ and PPARα, mediate numerous beneficial effects for obese individuals who might be predisposed to developing NIDDM, or who have already been affected by the disease. Even though specific ligands have been identified for these receptors, there are crossover physiological effects; fibrates have modest insulin sensitizing effects and thiazolidinediones lower triglycerides and free fatty acids. Since these ligands are believed to exert their effects primarily through their interactions with their cognate receptors, which in turn regulate gene expression, and since the use of gene arrays has dramatically increased our capacity to measure gene expression, we have used gene arrays to better characterize the ligands and to identify their modes of action. For further characterization using quantitative real-time polymerase chain reaction (RT-PCR), we selected a subset of approximately two dozen genes identified from the array experiments, as well as from the literature. Commercially available software applications, Gene Spring (Silicon Genetics, Redwood City, CA), and applications developed in-house, were used for the analysis of gene expression data.

MATERIALS AND METHODS

Animal Dosing: C57/BL6J mice were dosed with compound, by gavage, using 0.5% methylcellulose as the vehicle. After treatment, mice were anesthetized and killed using carbon dioxide. The tissues of interest were removed, snap-frozen in liquid nitrogen and stored at −80°C. Fenofibric Acid, Bezafibrate and Gemfibrozil were obtained from Sigma, and BRL-49653 (Rosiglitazone), Wy-14,643, Mγ-PhAc, Mγ-part 2, Mγ-part 4, Mα-1 were synthesized in house (Jones, 2001, Berger et al., 2001).

ISOLATION OF RNA AND REAL-TIME QUANTITATIVE PCR:

Total RNA was isolated using the Ultraspec reagent as described by the supplier (Biotecx, Houston, TX). In some cases, an equivalent volume of bromo chloro propane was substituted for chloroform to extract the RNA. This substitution was found to improve separation of DNA and RNA and thereby provide a purer RNA preparation. Some RNA preparations were treated with RQ1 RNase-free DNase (Promega, Madison, WI). The RT-PCR reaction was carried out in a 96-well plate using TaqMan Reverse Transcription Reagents (PE Biosystems, San Jose, CA). RT conditions were 25°C for 10 min, 48°C for 30 min, 95°C for 5 min, for 1 cycle on a DNA Thermal Cycler 9600 (PE Biosystems). The PCR reaction was performed using TaqMan Universal PCR Master Mix (PE Biosystems) and the TaqMan primers and probe. PCR conditions were 50°C for 2 min, 95°C for 10 min for 1 cycle, then 95°C for 15 sec and 60°C for 1 min for 40 cycle on an ABI PRISM 7700 Sequence Detector (PE Biosystems) according to the manufacturer's protocol (PE Biosystems TaqMan Gold RT-PCR kit protocol, 1997). The TaqMan primers and probes were designed using the Primer Express program from PE Applied Biosystems. The genes used in this study and sequences for their respective TaqMan primers and probes are shown in Table 1.

Table 1.

TaqMan Primer and Probe Sequences. Forward and Reverse primers are labeled with the suffix “_F” and “_R”, respectively, while the TaqMan probes are labeled with the suffix “_TqM”.

QUANTITATION OF GENE EXPRESSION USING AFFYMETRIX ARRAYS.

Hybridization samples were prepared according to Affymetrix instruction as described by Lockhart, et al., (1996). Briefly, 50 μg of total RNA was treated with an equal volume of RNAmate (Biochain, Hayward, CA) to remove polysacchrides. Five μM primer encoding the T7 RNA polymerase promoter linked to olgo-dT₂₄ primer was used to prime double-strand cDNA synthesis from each total RNA sample using the Superscript Choice System kit (Life Technologies, Rockville, MD). Each double-stranded cDNA sample was purified by sequential phenol/chloroform extraction (Ambion, Austin, TX) and QIAquick kit (Qiagen, Hilden, Germany) according to the manufacturer's instructions. Half of each cDNA sample was transcribed in vitro into the copy RNA (cRNA) using MEGAscript T7 kit (Ambion, Austin, TX) with incorporation of Biotin-11-CTP and Biotin-16-UTP (Enzo Biochemicals, New York, NY). These cRNA transcripts were purified by using mini RNeasy Kit (Qiagen) and quantitated by measuring absorption at 260nm/280nm. Ten μg of each cRNA sample was fragmented at 95 °C for 35 min in 40 mM Tris-acetate, pH8.0, 100 mM KOAc, and 30 mM MgOAc to a mean size of 50–100 nucleotides, added to hybridization buffer (0.1 M MES, pH6.7, 1M NaCl, 0.01% Triton, 0.5 mg/ml BSA, 0.1 mg/ml H. Sperm DNA, 50 pM Bio948 stock) and hybridized to Mu74 A microarrays (Affymetrix, Santa Clara, CA) at 45 °C for 16 h. The chips were washed and stained with streptavidin phycoerythrin (Molecular Probes) on the Fluidic Station (Affymetrix) and scanned with the Hewlett-Packard Gene Array Scanner to capture a fluorescence image.

Data from each microarray were normalized to data from a single vehicle microarray using global scaling based on overall hybridization intensities. Normalization, assessments of replicates, and calculations of gene expression levels as average difference values were performed using GeneChip v3.1 and Data Mining v1.2 software (Affymetrix). Each treatment was represent by two replicate samples using two microarrays.

RESULTS AND DISCUSSION

Members of the nuclear receptor family function as ligand regulated transcription factors (reviewed in Mangelsdorf et al. 1995). Their effects on the regulation of gene transcription are relatively rapid so we chose seven hours post treatment for our initial analyses. The Affymetrix software application, Gene Chip, was used to convert fluorescence to the value referred to as “Average Difference”. The Average Difference value for a particular gene on the Affymetrix array is a measure of the expression of the gene in the RNA sample; for the current analysis we used only Average Difference values which were marked “P” for “present”. Note that this approach will underestimate the number of genes affected by the various treatments, but it also provides for a higher quality dataset. The Average Difference values for the genes on each chip were imported into the GeneSpring Datamining software program. Values from duplicate chips were averaged and then values from treated samples were normalized by division by values obtained from untreated samples; hence, the results were expressed as ratios. The graphical view of this data along with utilization of the “pseudogene” ¹ feature of GeneSpring permitted us to rapidly identify lists of genes with specific attributes, e.g., genes which are up-regulated ² by activation of PPARα (treatment with WY-14,643 or fenofibric acid), or up-regulated by activation of PPARγ (treatment with BRL-49653). These lists (A. Elbrecht et al, manuscript in preparation) contain genes that were previously shown to be regulated by PPARα and PPARγ ligands, as well as newly identified genes. For example, although the up-regulation of carnitine acyltransferase by fenofibric acid and WY-14,643 (results not shown) has not been described previously, it is known that PPARα agonists increase fatty acid metabolism. The increased carnitine acyltransferase activity could result from peroxisomal proliferation and the need to shuttle acetyl-CoA out of the peroxisomes and into mitochondria. The enzyme carnitine palmitoyl transferase I is rate-limiting for transfer of fatty acids into mitochondria, and its regulation by PPARα ligands has been described previously. Here we report that the expression of carnitine palmitoyl transferase II mRNA is also increased by PPARα ligands (results not shown).

Hierarchical representations of gene expression results have proven useful for clustering and analysis of the results (Figure 1a). This representation clusters gene expression patterns on the basis of how they responded to the various treatments. Clusters can be selected at various nodes along the dendrogram, and members of each cluster should share some attribute(s), however, it is often difficult to identify the attribute(s). Generally, this is achieved by analysis of the genes comprising the cluster; usually involving a literature search and could extend to an analysis of the promoter regions. Clustering in this dimension organizes the gene profiles into islands such that common features, or distinctive features, can be identified for the treatments. It is important to note that the clusters result from the information provided by the three treatments used here, and that information from more treatments could change the clustering pattern. Nevertheless, this process has proven to be very useful in identifying clusters of interest.

Figure 1.

Two dimensional clustering of genes and compounds.

As described above, we used the “pseudogene” feature of GeneSpring to identify lists of genes regulated in a particular manner. There are two lists of genes mapped onto the dendrogram in Figure 1b. There are nine genes which are down-regulated twofold, or more, by the PPARα ligands fenofibric acid and WY-14,643, and there are 18 genes that are up regulated two-fold, or more, by these ligands. As shown in Figure 1b, each set of genes is not represented by a single cluster, although it is certainly possible that given a larger set of treatments, the fragmentation seen in each list could be resolved. Alternatively, for example, the fragmentation seen in PPARα down-regulated genes could be representative of genes belonging to different PPARα regulated pathways. In this case, the identification and analysis of genes adjacent to those currently identified by the lists could provide additional details on the nature of these pathways.

Clustering in the second dimension, i.e., according to treatment, allows for comparison of the gene expression profiles. As expected, the two PPARα ligands, fenofibric acid and Wy-14,643, are paired closely, with the PPARγ ligand, Rosiglitazone, occupying a separate node. This organization is based on a comparison of gene expression results for 2,421 genes and thus might reflect more accurately the physiological effects of these treatments than using single markers like peroxisomal proliferation.

This array-based approach has helped us to identify several genes not previously known to be regulated by PPARs. We have selected genes from these lists and combined them with genes identified from the literature to establish a set of genes for follow-up studies (Table 1). The goal was to use real-time quantitative PCR to obtain more accurate measurements of gene expression for use in compound profiling. Since nuclear receptors exert their effects predominantly by regulating gene transcription, it should be possible to group the compounds based on their gene expression profiles so that two agonists, with similar physiological effects, should pair in a dendrogram from a hierarchical clustering protocol, even if they belong to distinct chemical classes. Perhaps the most important feature of this approach is that it avoids the classical definitions of drugs as agonists, partial agonists or antagonists. These definitions are entirely dependent upon the assay used to measure their activity.

The same compound is capable of inducing expression of one gene and at the same time repressing another. Contradictory activities for compounds can also extend to organs. There is ample evidence to show that Tamoxifen works as an antagonist in breast tissue, and indeed was developed as a treatment for breast cancer, but also has agonist activity in other estrogen-responsive tissues like the uterus (reviewed in Mourits, 2001). Thus, there is nothing inherent to Tamoxifen that makes it either an agonist or an antagonist. It would be much more informative to group drugs by their physiological effects. Especially, but not exclusively, in cases where the drugs have profound effects on gene expression, it would be logical to group them using gene expression measurements.

Hierarchical clustering of PPAR ligands is shown in Figure 2. Expression profiles from 17 genes were used to cluster the PPAR ligands shown in Panel A. It is clear that the 3 PPARα ligands cluster together even though gemfibrozil (Elbrecht et al., 2000) is in a distinct chemical class from fenofibric acid and bezafibrate. Further experimentation with four PPARγ ligands, Mγ-PhAc, Rosiglitazone, Mγ-Part4 and Mγ-Part2, and three PPARα ligands, fenofibric acid, WY-14,643 and Mα-1, results in two major clusters in Panel B. Mγ-PhAc and Rosiglitazone are considered full agonists in transactivation assays (results not shown). Although the clustering results were obtained with expression values for a limited set of genes, it is interesting to note that the physiological effects of Rosiglitazone are similar to those of Mγ-PhAc, and they are paired in the dendrogram. In db/db mouse models of diabetes, both compounds decreased serum triglycerides, glucose and insulin levels during a similar time course and with a similar dosing regimen (results not shown). Mγ-PhAc is not a thiazolidenedione so, as with the PPARα ligands, this clustering is not based on chemical structure. Mγ-Part2 is also structurally distinct from Mγ-PhAc, and has been characterized as an antagonist in that it can block the activity of Rosiglitazone in transactivation assays (results not shown). Clustering based on gene expression indicates that this compound belongs to the PPARγ cluster, yet is distinct from the full agonists Mγ-PhAc and Rosiglitazone. Nevertheless, it is clear from the profile that some genes are up-regulated by all three compounds (Figure 2b). Had any of these genes been selected as the reporter for gene expression and PPARγ activity, all three compounds would have been characterized as agonists. This also applies to Mγ-Part4, another structurally distinct PPARγ partial agonist.

Figure 2a and b.

Compound profiling using gene expression. Db/db mice were treated with 30 mpk per day for 11 days as described in “Materials and Methods”. RNA was isolated from the livers of six animals for each treatment group. After averaging the results, the expression profiles of the genes listed in Table 1 were used to cluster the compounds. The numbering and order of the genes in the figure (shown by the vertical arrow) matches the numbering and order in the table. Pearson correlation was used as the metric for hierarchical clustering.

Thus, based on the results shown in Figure 2, the dendrograms do not represent clustering based on structure. These compounds also have different potencies as measured by binding and transactivation studies. For example, fibrates are generally considered as weak binders to PPARα with K_ds in the sub mM range while WY-14,643 exhibits a K_d of 0.6 μM (Elbrecht et al., 200 and references therein). Clearly, clustering is not based on potency. Since the grouping of these compounds seems to reflect their physiological effects, it suggests that grouping compounds based on gene expression profiles should be useful in their classification. We are not suggesting that potency and chemical structure are not important, but rather that the information gained from gene expression profiling of compounds should be used in conjunction with structural information and the information gained from other assays to obtain a better classification of compounds and drugs.

Acknowledgements

The authors would like to thank Dr. Tom Doebber and his lab for maintaining and dosing animals, Dr. Derek Von Langen for providing Mg-PhAc, John Acton for providing Mg-Part4, and Tony Lialin of Silicon Genetics for help with GeneSpring.