The use of animal models in expression pharmacogenomic ... - Nature

The Pharmacogenomics Journal (2001) 1, 48–58

 2001 Nature Publishing Group All rights reserved 1470-269X/01 $15.00 www.nature.com/tpj

REVIEW

The use of animal models in expression pharmacogenomic analyses BE Gould Rothberg CuraGen Corporation, New Haven, CT, USA Correspondence: BE Gould Rothberg, CuraGen Corporation, 555 Long Wharf Drive, New Haven, CT, USA Telephone: +1 203 481 1104 ext 4219 Fax: +1 203 481 1102 E-mail: bgould얀curagen.com

ABSTRACT

Expression pharmacogenomics applies genome/proteome scale differential expression technologies to both in vivo and in vitro models of drug response to identify candidate markers correlative with and predictive of drug toxicity and efficacy. It is anticipated to streamline drug development by triaging towards lead compounds and clinical candidates that maximize efficacy while minimizing safety risks. As the majority of expression pharmacogenomics will be performed on preclinical therapeutic candidates, compatibility with favored preclinical animal model systems will be essential. This review will address expression pharmacogenomics in the context of those animal model systems commonly used for pharmacokinetic, pharmacodynamic and toxicologic analyses. Specific discussions will cover: (A) relative robustness of genomic and proteomic technology platforms used to generate drug response data in critical model systems; (B) animal handling, treatment and other experimental design optimizations; (C) data analysis strategies for extracting and validating candidate pharmacogenomic markers; and (D) overarching limitations in applying expression pharmacogenomics to animal model systems. The Pharmacogenomics Journal (2001) 1, 48–58. Keywords: expression pharmacogenomics; toxicogenomics; animal models; differential expression profiling

Received: 18 December 2000 Accepted: 26 February 2001

Pharmacogenomics, the application of high-throughput genomic technologies and data analysis strategies to unravel the molecular basis of drug response, is rapidly becoming standard practice in both preclinical and clinical drug development. Pharmacogenomics can be further subdivided into two discrete disciplines: ‘expression pharmacogenomics’ and ‘pharmacogenetics’. Each addresses a unique aspect of drug response. Pharmacogenetics is the classical discipline that identifies gene-based genetic variations (eg single nucleotide polymorphisms) contributing to the individual variability in drug response across a population. The earliest pharmacogenetic studies date back to the 1950s where correlations between enzymatic polymorphisms and variable response to drugs including succinylcholine, paraaminosalicylic acid and various antimalarials were made.1 In contrast, expression pharmacogenomics is a relatively young discipline, having evolved and matured over the past 4 years. Expression pharmacogenomics refers to the application of differential gene/protein expression technologies to both in vivo and in vitro model systems of drug response, bridging the fields of genomics and medicinal chemistry.2 These data not only provide a comprehensive molecular composite of toxic or efficacious endpoints but also serve as diagnostic markers useful in the high throughput triage of lead compounds and other chemical/biological entities. The advent of pharmacogenomics to the drug development industry is particularly timely. Not only are the overall costs of developing new drugs reaching historic highs of $400–500 million but also, as the average timeline from patent

Animal models in expression pharmacogenomics BE Gould Rothberg

49

to market approaches 15 years, pharmaceutical companies are challenged to recuperate exorbitant development costs in 4 years of sales.3 Moreover, 30 of the most recognized blockbuster drugs, including Prozac (fluoxetine), Prilosec (omeprazole), and Claritin (loratidine), are expected to lose patent protection by 2002. Generic formulations of these drugs will be highly attractive to managed care and healthmaintenance organizations adding additional costing pressures to the next generation of therapeutic breakthroughs. The pharmaceutical industry anticipates that, in its maturity, pharmacogenomics will significantly streamline the time and cost of drug development by guiding the triage of lead compounds, preclinical and clinical candidates through all stages of drug development. The focus of this review will be to summarize the application of expression pharmacogenomics to the most relevant preclinical animal models. In addition to reviewing experimental design and analysis strategies, genomic and proteomic technology platforms will be discussed within the context of expression pharmacogenomics. Specifically, two sample datasets will be presented: (1) rat cardiac differential gene response to selective serotonin reuptake inhibitors (SSRIs) and other serotonin modulators; and (2) the comparison of wild-type and insulin-resistant rat liver response to thiazolidenedione oral hypoglycemics useful for type II diabetes mellitus. While these examples demonstrate the triage of small molecule compounds, expression pharmacogenomics is equally useful in assessing efficacy and toxicity of biologics including protein therapeutics and humanized monoclonal antibodies. ANIMAL MODEL SYSTEMS MOST USEFUL IN EXPRESSION PHARMACOGENOMICS Expression pharmacogenomics will enjoy a prominent role in preclinical drug development to both sort among the thousands of ultra-high-throughput screening (UHTS) chemicals and select leads that maximize efficacy while minimizing the risk of toxicity. In fact, toxicogenomics has evolved to specifically triage potentially toxic compounds based upon their differential expression profiles in selected model systems. As most expression pharmacogenomic experiments will be performed on compounds prior to human administration, these analyses must be highly compatible with the animal model systems commonly used in preclinical development. The most important of these preclinical models are those used for toxicogenomic analyses. Expression toxicogenomics will drive molecular toxicology by predicting deleterious outcomes from a drug’s differential gene expression (DGE) profile on target tissues prior to its detection by more laborious and time-consuming traditional toxicological assays.4 The goal of toxicogenomics is not to supplant standard, routine toxicology including 7-day and 14-day acute and 24month chronic toxicity screens. These and other traditional screens are not only considered the gold-standards of toxicological metrics but are also mandated by the US Food and Drug Administration and other international regulatory agencies prior to the initiation of human clinical trials on

novel pharmacologic entities. Expression toxicogenomics must be completely integrated and complementary with those toxicological assays required by the regulatory agencies. To comply, expression pharmacogenomics and toxicogenomics must be accurate and robust in those strains and species commonly used for required Investigational New Drug (IND) filings. Included among these are Sprague–Dawley and Wistar outbred rats, dogs and cynomolgous monkeys. Other species used in secondary, follow-up toxicological assays as well as drug efficacy assays include mouse, guinea pig, rabbit, sheep, pig and human ex vivo or in vitro models. Differential gene and protein expression profiling technologies selected for expression pharmacogenomics must be sufficiently robust for each of these species, as well as all others routinely used to assess pharmacodynamics and drug efficacy parameters. Currently available genomic and proteomic technologies and their relative usefulness in expression pharmacogenomics will be discussed in a following section. To facilitate DGE data integration with traditional pharmacologic and toxicologic parameters, it is essential that histopathology, clinical laboratory and, where possible, pharmacokinetic data be collected for the same animals whose tissues are processed for expression pharmacogenomic analyses. We routinely assay hematology, coagulation and chemistry parameters on all treated and control animals (Table 1) which can serve as parameters for multivariate analyses to identify genes whose quantitative levels of expression correlate with these treatment-related changes. It is also standard to prepare treated tissue for routine histopathology. In liver and other unpaired organ studies, we recommend preserving a discrete, quantifiable section (eg left lobe of the liver); for kidney and other paired organ studies, we recommend preserving the contralateral organ. Hematoxylin and eosin stains are routinely collected on all tissues. Additional stains including Oil Red O and Masson’s Trichrome are obtained as appropriate. Tissues are collected and fixed in RNAse-free medium for subsequent use in RNA in situ hybridization or immunohistochemistry to validate tissue sublocalization of certain compelling gene/protein markers where necessary. Administered drug pharmacokinetics is a similarly important parameter to measure. Ideally, the route of drug administration should match clinical formulations. While in the majority of dosings, the animals are correctly handled and the drug is appropriately introduced, failed administrations will occur. Furthermore, the outbred nature of most toxicologically relevant animal models introduces variability of both drug absorption and drug metabolism across the individual animals used. The ability to correlate each animal’s peak, trough and average drug levels as well as rate of drug metabolism with the resulting differential expression profiles can add significant value. However, many of the small animals routinely used for expression pharmacogenomics have insufficient blood volumes to permit simultaneous pharmacokinetic and clinical chemistry studies while leaving a sufficient residual circulating blood volume to maintain adequate target organ perfusion. In

www.nature.com/tpj


50

Table 1 Summary of reference clinical laboratory tests obtained for expression pharmacogenomics studies Hematology: Hematocrit Platelet count Coagulation: Prothrombin time Clinical Chemistry: Sodium Blood urea nitrogen Magnesium Globulin Free fatty acids Lactate dehydrogenase Aspartate aminotransferase

Hemoglobin Reticulocyte count

Leukocyte differential Mean corpuscule volume

Chloride Glucose Total protein Triglycerides Direct bilirubin Alanine aminotransferase

Bicarbonate Calcium Albumin High density lipoproteins Alkaline phosphatase

Activated partial thromboplastin time Potassium Creatinine Inorganic phosphates Cholesterol Total bilirubin Creatinine phosphokinase Gamma glutamyltransferase

these small animal studies, some utility can be gained by incorporating a parallel cohort of animals in a separate ‘pharmacokinetic arm’ but this is not viewed as essential. Large animal studies, however, should always be accompanied by pharmacokinetic evaluation. EVALUATION OF TECHNOLOGY PLATFORMS APPLIED IN EXPRESSION PHARMACOGENOMICS Expression pharmacogenomics is driven by the differential expression technologies available to monitor levels of gene and/or protein expression. Open architecture gene expression, open architecture protein expression and closed architecture gene expression are the three classes of genomics/proteomics technologies most commonly applied for expression pharmacogenomics. Protein chip technologies (eg SELDI) are still investigational5 and will not be addressed here. The relative merits and weaknesses of each will be presented here and are summarized in Table 2 (reviewed in Rininger et al6 and Green et al7). DNA gene chip microarrays and redundant real-time quantitative polymerase chain reaction (RTQ-PCR) are representative of closed architecture differential gene expression. RTQ-PCR technology is most useful when a limited number of candidate genes need to be quantitated across several hundred samples. RTQ-PCR is particularly advantageous when the amount of starting RNA is limited; successful experiments have been executed with as little as 500 fg of starting total RNA.8,9 The need to precalibrate and normalize each experiment against a non-differentially expressed housekeeping gene prior to data acquisition suggests a diminishing time benefit unless a large sample set will be assayed for each gene. DNA microarrays are designed to simultaneously measure the expression levels of several thousand genes in a single assay.10,11 Specific nucleotide probes representing each gene to be assayed are spotted onto a solid support and levels of gene expression are subsequently measured by detecting the fluorescence of the drug-treated sample floruophore (Cy3) compared to the control sample fluorophore (Cy5) at each unique spot on the microarray.12,13 Differences in labeling

The Pharmacogenomics Journal

Leukocyte count Blood cell morphology

intensity are converted into a quantitative output of relative gene expression. The standard output from microarray experiments includes the list of differentially expressed genes, their n-fold modulation and the relative confidence level that each data point is a true positive. There are two principal types of microarrays: (1) spotted arrays that consist of presynthesized DNA affixed to supporting (eg glass) slides; and (2) high-density oligonucleotide arrays in which short oligonucleotide sequences (苲25 base pairs in length) are synthesized in situ on chip wafers using a photolithographic manufacturing process.10,11 Spotted microarrays have the advantage in that individual investigators can not only customize these to include only those genes most relevant to each experiment but that they also can be produced and utilized in a regular laboratory setting. Specific technological advantages afforded by microarray technology include speed of experimental completion and sensitivity; mRNA transcripts occurring as sparsely as 1:300 000 have been detected on microarrays.11 DNA microarrays have been applied towards the characterization of hepatotoxicity in rodent models (D Mendrick, International Life Sciences Institute (ILSI) Toxicogenomics Consortium, unpublished data4) and are rapidly becoming the technology of choice for rodent toxicogenomics programs including that of the National Institute of Environmental Health Sciences.14 However, generalized use of microarrays across all relevant animal model systems is restricted to those species where genome sequencing efforts have already characterized ⬎75% of the expressed genome. Closed architecture systems are defined by their requirement of a priori sequence knowledge for each gene or clone to be assayed in an experiment. Microarray assay robustness is directly correlated with the level of source organism genome sequence characterization and annotation. The mouse genome effort is expected to produce a finished version by 2003 and the working draft of the rat genome has been undertaken.15 As expected, mouse-derived gene chips offer reasonable coverage of its genome. Rat chips are more erratic. Less than 30% of the genes spotted on genomeencompassing commercially available microarrays corre-


51

Table 2 Comparison of differential expression technologies (adapted from Rininger et al 6 and Green et al 7) Technology

Sensitivitya

Template required

Coverageb (%)

DNA microarray RTQ-PCR Differential Display SAGE RDA GeneCalling TOGA 2DE/MS Proteomics

1:300 000 1:300 000 1:100 000 ⬍1:10 000 ⬍1:300 000 1:125 000 ⬍1:100 000 苲1:10 000d

5–10 ␮g total RNA 1 ng mRNA 0.5 ␮g mRNA 5 ␮g total RNA 0.02 ␮g mRNA 0.02 ␮g mRNA

Variable 100%c 96% (240 random primer reactions) 92% (15 000 sequencing reactions) N/A 95% (96 restriction enzyme pairs) 98% (4 enzymes ⫻ 256 reactions each) 50–66%e

Variable

a

Sensitivity refers to the number of copies a single gene must possess as a fraction of all gene transcripts present in the cell/tissue in order for that gene to be correctly assayable as differentially expressed at the n-fold limit of the specified technology, assuming that the minimum starting material required by the individual technology has been introduced. b Coverage refers to the % of a cell’s/tissue’s expressed transcriptome that is detectable (differently regulated or unchanged) using any technology with its minimum starting material. For those technologies that require parallel iterations of key chemistry steps (ie sequencing reactions, restriction endonuclease digestions) to achieve maximal % coverage, the number of these iterations required is listed in parentheses. c RTQ-PCR is unique in that its detection limit is low enough to detect, with proper PCR primer design, any assayed gene as long as it is expressed in the tissue source. This includes rare genes expressed at low levels in complex tissues. d While exact sensitivity values have not been determined for large-scale proteomic analyses, it has been noted that low-abundance proteins including protein kinases and transcription factors are not readily detectable.26 e 2DE PAGE allows for the maximal detection of 4000–5000 protein units.31 This allows for 50% coverage of tissues/cells expressing 10 000 proteins or 66% of those expressing 7500 proteins.

spond to GenBank annotated mRNA sequences; the remainder derives from expressed sequence tag (EST) sources and offers minimal functional data.12 Two competing commercially available rat genome chips are available. The Rat U34 Set covers 24 000 elements on a 3-chip set. These include 7000 known and functionally annotated rat genes (U34A) and 17 000 unannotated EST clusters (U34B and U34C) (www.affymetrix.com). The Rat GEM3 chip (www.incyte.com) offers coverage of 5931 unique annotated genes or EST clusters represented across 8951 elements. The lack of any sequence data on many as-yet-uncharacterized rat orthologs of known human and mouse genes precludes their inclusion on any chip. Yet, for rat liver expression pharmacogenomics, DNA microarrays can be useful. Many rat genes involved in basic metabolism, xenobiotic metabolism, stress response and other molecular pharmacologic and toxicologic processes have already been fully characterized and accessible to microarrays.4 The Rat Toxicology U34 array (www.affymetrix.com) specifically covers over 850 genes and EST clusters previously associated with toxicological processes and derived from literature searches in collaboration with pharmaceutical toxicologists. Sequence databases for dogs and monkeys are sparse to nonexistent.15 This has been a limiting factor in successful application of closed architecure systems for their analysis. Open architecture DGE platforms do not assume prior sequence knowledge of target genes and can, therefore, survey all transcripts in any tissue source regardless of organism to identify all drug-responsive genes, both known and novel. Examples of open architecture DGE technologies include: differential display,16 serial analysis of gene expression (SAGE),17 representational differential analysis (RDA),18 subtractive hybridization,19 total gene expression analysis (TOGA)20 and GeneCalling.21 Common to all technologies are: the preparation of mRNA or cDNA from both

drug-treated and control animals, chemical parsing of transcripts and electrophoretic separation of the resulting fragments. Fragment patterns on gels are used to infer gene expression. Fragments showing levels of differential expression between drug treatment and control are then selected for subsequent identification. As fragment pattern drives candidate prioritization, resulting genes are as likely to be known or novel. The relative strengths of each open architecture technology are not only linked to sensitivity (ranging from transcript detection at 1:10 000 to 1:300 000) and actual measured genome coverage (60–98%) (reviewed in Rininger et al6 and Green et al7) but also correlated with the speed confirmed gene identities can be assigned to interesting differentially expressed fragments. Differential display (DD) and related technologies PCRamplify poly-A primed cDNA using a combination of random primers and up to 16 dinucleotide-followed-by-oligodT anchored primers. The resulting PCR products from treated and control samples are subsequently run on a gel and fragments whose intensities differ between experimental states are identified. For subsequent gene identification, each targeted fragment must be physically excised from the source gel and included clones to be isolated and sequenced.16 Typically, this exercise yields several genes per fragment. Identifying the true positives requires additional effort and false positives continue to remain a significant failure mode for these technologies. SAGE sorts genes following limited sequencing reactions to identify unique 10–14 bp sequence identifiers at the 3⬘ end of each cDNA in the sample.17 The coverage and sensitivity of SAGE is dependent upon the number of clones sequenced. The likelihood of detecting a transcript present at three copies per cell, assuming 300 000 cellular transcripts is 92% if 10 000–15 000 clones are sequenced.22 SAGE can confound gene families with significant sequence overlap

www.nature.com/tpj


52

between individual members including cytochromes and other drug metabolism and disposition genes. RDA, technically a form of subtractive hybridization, is one of the earliest well-established DGE technologies. Original RDA methods detect only those genes in the treated samples that do not have counterparts in the control but subsequent modifications have allowed resolution of genes present in varying amounts between case and control.18 RDA is limited in that it does not quantitate levels of DGE (although Northern verification has demonstrated the detection limit to be ⫾2-fold) and is prone to high false positive and negative rates.23 Relative strengths of RDA include detection of transcripts as rare as 1% and its success in characterizing levels of DGE in non-eukaryotic species.24 GeneCalling and TOGA both parse the source cDNA using restriction endonucleases, PCR amplify representative fragments using directed primers and resolve the fragments using either slab gel or capillary electrophoresis.20,21 In both cases, fragment detection is achieved through the addition of a fluorescent label to one end of each PCR product. The typical readout is a gel chromatogram with the x-axis measuring length in basepairs of each gene fragment and the yaxis measuring relative fragment fluorescence intensity. The direct overlay of chromatograms generated by drug treatment and its vehicle control will yield the subset of fragments differentially expressed following the selected drug treatment. Coupling the length of a differentially expressed gene fragment with the a priori sequence knowledge of the terminal base pairs required to have allowed restriction enzyme digestion, can prompt in silico identification of the differentially expressed gene through a sequence database query in the relevant species. Simultaneous parsing of a single cDNA pool with multiple restriction enzyme chemistry reactions generates redundant fragment coverage of the genome and raises confidence in predicted gene assignment. When database lookup yields no reasonable candidate genes, gene identification is achieved by gel elution and clone sequencing that can reduce the efficiency of data collection. This latter scenario is common for assays in dog, monkey and other poorly sequenced species. Construction of proprietary sequence databases in advance of expression profiling greatly enhances database lookup to increase process efficiency (J Irzyk, unpublished data). GeneCalling possesses the additional advantage of eliminating false-positive gene assignment to differentially expressed fragments. This is accomplished by an independent confirmation step where unlabelled sequence-specific oligonucleotides are constructed for each candidate gene determined by directed clone sequencing or in silico database query and added to the original chemistry reaction in 100-fold excess prior to PCR-reamplification. Following electrophoresis, the resulting traces from the competitive reaction are superimposed upon the original GeneCalling traces. The true positive only will show quenching of the differentially expressed peak relative to the original reaction: oligonucleotide poisoning.21 Proteomic assays characterize the full complement of proteins in a specific tissue sample. The technical components


of all proteomic analyses include: recovery of all sample protein constituents, fractionation of the mixture, structural characterization of each fraction’s components and database lookup to identify isolated proteins.25 At present, proteomic platforms that integrate either two-dimensional gel electrophoresis (2DE) or liquid chromatography fractionation with mass spectroscopy identification offer an effective method by which several thousand proteins can be separated, detected and quantified in a single operation.26 Proteomics offers the advantage over mRNA expression in that it not only quantitates levels of protein expression but also surveys post-translational modification events. Both equally impact protein activity. In this manner, proteomics evaluates the dynamics of metabolic, signaling and regulatory networks and how these become altered following drug treatment.27 Proteomics has already generated intriguing toxicogenomic analyses in mouse,28,29 hamster29 and rabbit30 but several technological hurdles are still limiting. 2DE fractionation methods bias against low abundance proteins and other proteins with low codon bias values such that the number of spots on 2D gels is not representative of the gene set expressed in the analyzed sample.26 Many proteins are not soluble in the available detergents used for 2DE sample preparation.31 Similarly, for liquid chromatography, the dynamic range is insufficient to be effective in resolving highly complex mixtures and protein quantitation is not readily achievable.26 Each class of differential expression technology has both strengths and weaknesses. Yet, each technology has matured such that high throughput generation of quality raw data on drug response models is no longer rate limiting. It is trivial to generate large volumes of data in short periods of time. In moving forward, the greatest challenge facing expression pharmacogenomics is the development of robust, statistically valid analysis strategies that can distill large datasets to extract the subset of differentially expressed genes/proteins that are true predictive markers of toxicologic and efficacious endpoints. Two components are essential here: robust experimental design and appropriate data analysis tools capable of distilling large datasets to represent the key findings in a simple and visually appealing fashion. The latter is the goal of many academic and commercial bioinformatics efforts and their products are essential tools for all motivated in expression pharmacogenomics. Michael Eisen’s laboratory at the Lawrence Berkeley Laboratories in Berkeley, CA has produced a suite of gene expression analysis software programs that are available free to academic groups and non-profit organizations and licensable to all others (http://rana.lbl.gov/EisenSoftware.htm). The package includes: ScanAlyze, an image analysis program to process fluorescent microarray images and features semi-automatic definition of grids and complex pixel and spot analyses, Cluster, an analysis package that performs a choice of hierarchical clustering, self-organizing maps, k-means clustering or principle components analysis on complex DGE datasets, and TreeView, a graphical output program that promotes easy browsing of clustered datasets. Spotfire, a Swedish bioinformatics company (http://www.ivee.com), also offers DGE data vis-


53

Figure 1 Individual drug triage experimental design. An investigational drug is flagged following the phenotypic manifestation of unanticipated toxicity or efficacy. Comprehensive molecular characterization of the unanticipated phenotype requires the selection of positive controls, related compounds with literature documentation of a similar drug response, and negative controls, related compounds where literature reference indicates absence of this phenotype. Marker genes/proteins can then be represented as those genes similarly differentially expressed in both the flagged compound and the positive controls but not altered in the negative controls (green shaded area). These markers can then be prospectively applied to test drugs, next generation compounds to be triaged for presence/absence of the flagged phenotype.

ualization tools. In addition to hierarchical and k-means clustering, the package offers additional functions including comparison of expression data across multiple array experiments, superimposing metabolic pathway data onto genes of interest and connections to GATC and Incyte LifeArray databases. CuraGen Corporation’s GeneScape web-based bioinformatics portal (http://www.curagen.com) allows clients the ability to integrate their DGE data (both open and closed architecture platforms are supported) with public and proprietary sequence databases, yeast-two-hybrid protein– protein interaction maps in multiple species and a comprehensive human SNP database and genotyping resource. Other companies offering bioinformatics solutions for gene and protein expression visualization include Lion Biosciences and Gene Logic (http://www.genelogic.com). The third essential expression pharmacogenomics component is a series of robust experimental designs. As stated, the field of expression pharmacogenomics is still in its infancy. This is most apparent during critical evaluation of the experimental paradigms and analysis strategies used by both industry and academia to discover biological markers correlative with drug efficacy and toxicity. At present, there exist very few peer-reviewed studies that either substantiate or refute individual experimental or analysis strategies as effective; this emerging field is still highly speculative. In this light, I put forth here three experimental paradigms that show promise towards obtaining relevant pharmacogenomic markers from animal model drug response data.

EXPERIMENTAL PARADIGM NO. 1: INDIVIDUAL DRUG TRIAGE Figure 1 depicts the experimental outline for individual drug triage. This strategy is applied in later preclinical stages once a certain quantity of toxicologic or efficacy data has been collected and one or more assayed compounds display unexpected clinical parameters. These outliers stimulate molecular characterization of the observed phenotype. A minimum of nine treatments is required to execute this experiment. In addition to the flagged drugs, investigational compounds displaying the unanticipated physiologic response, three positive controls (literature-referenced compounds capable of producing the same clinical outcome) and three negative controls (related, literature-referenced compounds devoid of the cited drug response) are required for comparison. Vehicle controls and untreated controls complete the treatment set. Dose and time course selection is arbitrary and driven by the investigator but all compounds should be dosed at either similar bioequivalent doses or clinical efficacy doses. Covering a spectrum of dose/time combinations allows the investigator the opportunity to determine, from among the markers, which are early response genes, late response genes or mid-phase (U-shaped curve) genes and make additional scientific inferences about potential mechanisms of therapeutic/toxic activity. However, for the core purpose of novel lead compound triage assay, a single dose/time point that clearly differentiates the positive and negative controls is the most direct critical path towards

www.nature.com/tpj


54

achieving a desirable marker set. Following animal treatment, target organs are procured at necropsy and mRNA/protein extracts are prepared according to standard procedures. Expression profiles are then generated on the sample sets and binary comparisons between each drug treatment and the vehicle control are performed. The untreated controls are used as a reference point to establish the level of background variability in certain gene classes (eg stress response) where levels of differential expression can emanate from sources other than drug administration. Candidate markers are the set of genes/proteins that are coordinately differentially expressed in both the flagged drug and the positive controls but are unchanged or oppositely modulated in the negative controls. These genes/proteins can be subsequently applied as markers in a high throughput assay of animal cohorts treated with backup lead series, second generation compounds or other test compounds for prospective triage against the presence/absence of the flagged phenotype. We applied Analysis Strategy No. 1 to comprehensively characterize the rat cardiac response to a series of serotonin receptor agonists. Selective serotonin reuptake inhibitors (SSRIs) dexfenfluramine and fenfluramine were withdrawn from the market in 1997 following clinical evidence of cardiac valuvulopathy among healthy obese women taking these medications alone or in combination with the indirect sympathomimetic amine phentermine.32 This withdrawal has all but eliminated the further development of novel SSRIs as anti-obesity agents due to the lack of adequate preclinical markers capable of eliminating valvulopathy-prone lead compounds. The cost of a late-stage clinical failure would outweigh potential benefits of a successful market launch. Because SSRIs are effective anorexic agents, identifying gene-based markers capable of predicting cardiac valvulopathy could reinvigorate development of novel SSRIs for obesity. A total of six serotonin agonists were administered to male 10-week-old Sprague–Dawley rats at doses equivalent to 10fold the accepted human ED100 dose for 3 days. Both dexfenfluramine and fenfluramine were designated as flagged compounds. Fluoxetine and sibutramine, SSRI compounds currently prescribed for obesity management that do not produce valvular fibrosis33,34 were selected as negative controls. Positive controls included the ergot alkaloid dihydroergotamine35 and the 5-HT1B/D receptor agonist sumatriptan. The 5-HT1B/D receptor subtype has been documented to be expressed on cardiac valves,36 but no correlation has been established with the incidence of valvular fibrosis. Nineteen thousand unique rat cardiac gene fragments were assayed for differential expression by GeneCalling. Figure 2a depicts the distribution of candidate marker genes. Of the 1911 genes responding to at least one of the administered compounds, only 78 demonstrate coordinate DGE in the flagged compounds ⫾ positive controls. Similarly, 45 genes exhibited coordinated dysregulation across all six serotonin modulators, indicating the presence of baseline serotonin effects on whole cardiac tissue. Figure 2bi and 2bii list 26 representative dexfenfluramine/fenfluramine-selective gene


Figure 2 Evaluation of potential serotonin cardiotoxicity in a rat model. Fenfluramine and dexfenfluramine, drugs flagged for incidence of cardiac valvulopathy were administered to male Sprague– Dawley rats for 3 days. Levels of DGE in rat heart tissue were compared to those seen with positive control compounds dihydroergotamine and sumaptriptan and negative controls fluoxetine and sibutramine. (a) Of the 1911 gene fragments differentially expressed by GeneCalling across all treatments, a total of 78 were coordinately differentially expressed in the flagged drugs ⫾ positive controls. Thirty-four of these fragments were common to the flagged drugs to the exclusion of all other therapies. These 78 genes are candidate markers for the subsequent triage of novel SSRI-derived antiobesity agents for potential to induce cardiac valvulopathy. (bi) Partial gene list of representative candidate dexfenfluramine/ fenfluramine-triaging genes. Three classes of genes: cardiac stress response genes (red), cell proliferation regulators and mediators (green) and contractility regulators (blue) are most compelling in offering physiologically tractable mechanisms of cardiac dysfunction. (bii) Partial gene list presenting representative genes coordinately modulated across all serotonin modulators assayed.

markers and eight coordinate serotonin-responsive genes, respectively. Three classes of genes stand out among the putative valvulopathy marker set: genes previously associated with cardiac stress response, genes involved in growth regulation and genes functioning in cardiac muscle contractility. As shown, the putative cardiotoxicity marker set includes abundant (eg ribosomal proteins) as well as less common (eg rho-associated kinase beta) transcripts. A similar distribution of genes is found among the general serotonin response set (see Figure 2bii). Finally, more than 50% each of the putative cardiotoxicity and serotonin marker genes are previously unidentified rat orthologs of well-annotated human or mouse genes. This latter result is the direct benefit of having selected an open architecture technology for data generation. Independent physiologic validation and tissue localization of candidate markers is ongoing.


55

Figure 2 Continued.

EXPERIMENTAL PARADIGM NO. 2: CONSTRUCTION OF DRUG RESPONSE DATABASES FOR USE IN HIGH THROUGHPUT SCREENING LEAD COMPOUND TRIAGE The major limitation of individual compound triage strategies is that the results are only validated for the surveyed compounds. There is a growing trend among toxicologists to create a robust, universal database describing the molecular correlates of quantifiable toxicologic endpoints. Expression profiles generated independently for novel chemical entities would be superimposed upon this outcomes database to indicate which novel leads portend toxic outcomes and which are relatively benign.37 Figure 3 depicts an optimized experimental design for a robust single organ toxicogenomics database. Following organ selection, a careful literature review is required to define all unique, measurable subtypes of toxicity incurred. For example, hepatotoxicity can be subdivided by histopathologic subtype (eg zone 3 hepatic necrosis, steatosis, cholestasis) or by mechanistic subtype (eg apoptosis, stress response, cell proliferation). Drugs representative of each subtype are then selected. Ideal drugs

include those that produce high incidence and high severity for the desired toxic phenotype to the exclusion of all others. It is recommended that each drug be dosed using a dose/time matrix that will include datapoints manifesting overt toxicity upon histopathology as well as dose-time combinations that are clearly benign to the animal model. A minimum of thirteen ‘high incidence/high severity’ drugs per subtype are required to attain statistical significance in marker selection (JS Bader, personal communication). Of these thirteen, eight are arbitrarily designated as the ‘training set’ (Figure 3a(i)) and are processed for differential expression profiling along with the vehicle and untreated control. The remaining five drugs are set aside as a ‘test set’ (Figure 3a (ii)) and exclusively used to validate the fidelity of the markers assigned by the training set. Twenty additional drugs, 16 with varying levels of incidence and/or severity (Figure 3a (iii–v)) and four that do not manifest the toxicity (Figure 3a (vi)) are suggested. Following differential expression profiling of the ‘training set’ compounds, several nested analyses should be perfor-

www.nature.com/tpj


56

med on the raw data. First, dataset normalization to control for run-to-run variability38 is required. The normalized data are then hierarchically clustered39 to determine overall relatedness of each drug response. Subsequent analyses include identification of all genes/proteins differentially expressed following treatment with any of the compounds and selection of candidate marker genes from among those coordinately altered across all training set compounds. These analyses should be conducted using statistically robust strategies and parameters. Our experiences with data modeling support using a nested analysis of variance (ANOVA) on the normalized dataset to assess levels of differential expression across all treatments and to extract the relevant markers through either a univariate ANOVA or a multiple analysis of variance (MANOVA)/principle components analysis (PCA) (R Gerwien, unpublished data). Histopathology and clinical laboratory values serve as surrogate endpoints to rank training set drug response severity. As the univariate ANOVA and MANOVA datasets tend to be complementary, we perform both. The final two steps involve biological validation of the marker gene set for each toxicologic subtype. First, each candidate marker gene must be validated against each ‘test set’ treatment. Only those genes that show reproducible differential expression across both training and test samples should be considered markers for the selected toxicity sub-

type. Robustness can be enhanced through RTQPCR/Western blot analyses on the samples collected from the remaining high incidence/high severity drug dose/time points and the graded incidence/graded severity drug set. Finally, all marker genes should be functionally annotated and placed within biologic context based upon known and presumed mechanisms of toxicity. The Gene Ontology (GO) hierarchy system40 is particularly useful for functional categorization of both known and newly characterized genes. If marker genes suggest novel mechanisms of toxicity, these should be independently validated through standard physiologic, biochemical or cell biologic assays as appropriate. Following analysis of all toxicity subtypes, the individual marker sets determined for each must be integrated into a single high-throughput assay consistent for the whole organ. At this step, marker classification is further refined to identify those genes/proteins unique and specific to a single subtype, severity and incidence and to distinguish these from genes that overlap across several subtypes, levels of severity or percent incidence. Candidate markers that hamper the ability to distinguish among different endpoints may be discarded at this stage. The final gene set, representative of the organ composite, is then spotted as a custom ‘toxicity’ microarray or serial RTQ-PCR/Western Blot assay that can be iteratively used to evaluate novel therapeutic agents.

Figure 3 Organ composite toxicogenomics marker database experimental design. (a) Construction of a marker set for a single toxicity subtype. A minimum of 24 drugs capable of selectively producing the targeted toxicity is required (3ai–3av). Eight compounds showing both high levels of incidence and high degree of symptom severity in the animal population for the designated subtype will be apportioned as the initial training set to generate a putative candidate marker list. Five more high incidence/high severity drugs are selected as a test set to validate the candidate markers (3aii). Drug sets (3aiii), (3aiv), (3av) represent series of drugs with decreasing relative incidence and/or severity. Assessing marker gene prevalence across these can reduce the triage error rate by anticipating marker behavior among less severely toxic chemical entities. (3avi) represents a series of negative controls. (b) Organ composite creation through integration of marker sets from all toxicity subtypes. Marker classification, at this step, is further refined to identify those markers unique and specific to a single subtype and to distinguish these from genes that overlap across several subtypes.



57

21 of 468 differentially expressed gene fragments responded in similar fashion.

Figure 4 Drug efficacy screening in animal models of human disease. An animal model of human disease (black mouse) is treated with either test drug (brown pill) or vehicle control (blue pill). Wildtype parental strain animals (brown mouse) are treated in parallel with the vehicle control. Two binary comparisons are created: one to assess the baseline genetic differences between vehicle-treated diseased animal and the wild-type control (upper circle) and the other to assess drug response in the disease model (lower circle). The genes where drug treatment corrects the baseline expression of disease model animals to levels comparable to wild-type levels (red area) can define principle molecular components of drug efficacy.

EXPERIMENTAL PARADIGM NO. 3: DRUG EFFICACY SCREENING IN ANIMAL MODELS OF HUMAN DISEASE The third experimental strategy defines levels of drug efficacy in animal models of human disease (Figure 4). It is highly complementary to routine pharmacology studies and should be considered as a supplement to these standard assays. This assay determines the set of genes differentially expressed between the disease model and its parental wildtype strain that can be restrored to wild-type levels following disease model drug treatment. As an example, we compared the effects of troglitazone, an oral hypoglycemic,41 on the spontaneously hypertensive (SHR) rat model of noninsulindependent diabetes mellitus. SHR rats were treated with either 617 mg kg⫺1 day⫺1 troglitazone or 0.02% DMSO in daily oral doses for 3 days. Parental Wistar–Kyoto (WKY) rats were also treated with 0.02% DMSO in an identical fashion. GeneCalling was performed on the liver, adipose and skeletal muscle tissues of all three animal cohorts. In liver, 1776 gene fragments (5.2%) were differentially expressed at baseline between SHR and WKY rats. Of these, 427 (24%) were restored to wild-type levels in SHR rats following troglitazone treatment. In adipose, 78 of 712 SHR/WKY baseline differences were restored by troglitazone. In skeletal muscle,

LIMITATIONS IN ANIMAL MODEL USE Animal models provide valuable pharmacologic and toxicologic data on novel and existing therapeutics. However, there are certain limitations inherent in the animal models themselves that must be accounted for when assessing pharmacogenomic data. The first consideration is the extent to which the model’s drug response resembles that observed, or expected, in the human. In cases where there are discrepancies between human and animal model response physiology, expression pharmacogenomic data must be interpreted with caution. For example, when hyperlipidemic people are treated with fibric acid derivatives, plasma triglycerides and cholesterol are lowered. In rats and mice, administration of the same compounds not only causes a similar reduction in plasma lipid levels but also causes peroxisome proliferation, most notably in hepatic tissue. Direct consequences of this peroxisome proliferation include: hypertrophic and hyperplastic driven liver enlargement, increase in reactive oxygen species and nongenotoxic hepatocarcinogenesis.42 These effects have not been documented in man. The second limitation relates to the outbred nature of most commonly used animal strains. Traditional pharmacology and toxicology studies address this issue by performing their experiments on large cohorts of animals where outliers can be appropriately identified. Pharmacogenomics, however, has typically been applied on much smaller cohorts of animals. n ⫽ 3 is not uncommon and costs of differential expression profiling often preclude n ⬎ 10. We, and others (D Mendrick, unpublished data) have shown that outbred animals show a significant level of intratreatment gene expression variability, especially following oral drug administration. To discriminate the majority trend from the outliers, a minimum of 6–8 animals must be treated per dose/time point. Moreover, adequate computational support to address and accommodate intratreatment variability must be included. In particular, tracking individual rat response/nonresponse across genes belonging to discrete metabolic, signaling or regulatory pathways may provide additional insight into the molecular basis for variable levels of drug response among identically treated individuals. As animal model expression pharmacogenomics continues to play an increasingly important role in preclinical drug development, improvements to all facets will be implemented. Technology improvements will continue to increase the sensitivity, speed, and accuracy with which differentially expressed genes are identified. Genome sequencing efforts will incrementally enhance the quality of annotated DNA and protein databases and facilitate functional characterization of differentially expressed genes. As mentioned, however, data generation is no longer rate limiting. We anticipate that the most substantial improvements will occur following reevaluation and refinement of data analysis strategies, including those presented here. As large sets of high quality data are processed through each analysis

www.nature.com/tpj


58

strategy, candidate marker genes will be evaluated within the context of their ability to both correctly predict toxicity/efficacy of other known chemical entities and determine if the biological implications of candidate marker perturbations are consistent with known biochemical and physiological toxicity/efficacy mechanisms. This iterative process of marker evaluation and analysis strategy refinement, central to the construction of toxicogenomic and other drug response databases, will ultimately yield useful tools that will be broadly applied by both pharmaceutical companies and regulatory agencies to not only enhance the efficiency of the drug development process but also improve the overall quality of products approved for human disease management. DUALITY OF INTEREST None declared.

ABBREVIATIONS DGE – differential gene expression UTHS – ultra-high-throughput screening IND – investigational new drug RTQ-PCR – real-time quantitative polymerase chain reaction EST – expressed sequence tag DD – differential display SAGE – serial analysis of gene expression RDA – representational difference analysis TOGA – total gene expression analysis 2DE – two-dimensional gel electrophoresis SSRI – selective serotonin reuptake inhibitor ANOVA – analysis of variance PCA – principle components analysis GO – gene ontology SHR – spontaneously hypertensive WKY – Wistar–Kyoto

REFERENCES 1 Evans WE, Relling MV. Pharmacogenomics: translating functional genomics into rational therapeutics. Science 1999; 286: 487–491. 2 Bailey DS, Bondar A, Furness LM. Pharmacogenomics—it’s not just pharmacogenetics. Curr Opin Biotechnol 1998; 9: 595–601. 3 R & D Directions Staff. Beyond genomics. R & D Directions 1999; 5: 40–44. 4 Farr S. Dunn II RT. Concise review: gene expression applied to toxicology. Toxicol Sci 1999; 50: 1–9. 5 Davies H, Lomas L, Austen B. Profiling of amyloid beta peptide variants using SELDI ProteinChip arrays. Biotechniques 1999; 27: 1258–1261. 6 Rininger JA, DiPippo VA, Gould Rothberg BE. Differential gene expression technologies for identifying surrogate markers of drug efficacy and toxicity. Drug Discov Today 2000; 5, 560–568. 7 Green CD, Simons JF, Taillon BE, Lewin DA. Open systems: panoramic views of gene expression. J Immunol Meth 2001; 250: 67–79. 8 Roberts CA, Dietzgen RG, Heelan LA, Maclean DJ. Real-time RT-PCR fluorecent detection of tomato spotted wilt virus. J Virol Meth 2000; 88: 1–8. 9 Heid CA, Stevens J, Livak KJ, Williams PM. Real time quantitative PCR. Genome Res 1996; 6: 986–994. 10 Schena M, Shalon D, Davis RW, Brown PO. Quantitative monitoring of gene expression patterns with a complimentary DNA microarray. Science 1995; 270: 467–470. 11 Lockhart DJ et al. Expression monitoring by hybridization to high density oligonucleotide arrays. Nat Biotechnol 1996; 14: 1675–1680. 12 Harrington CA, Rosenow C, Retief J. Monitoring gene expression using DNA microarrays. Curr Opin Microbiol 2000; 3: 285–291. 13 Lipshutz RJ, Fodor SP, Gingeras TR, Lockhart DJ. High density synthetic nucleotide arrays. Nat Genet 1999; 21 (1 Suppl): 20–24.


14 Lovett RA. Toxicologists brace for genomics revolution. Science 2000; 289: 536–537. 15 Pennisi E. Rat genome off to an early start. Science 2000; 289: 1267–1269. 16 Liang P, Pardee AB. Differential display of eukaryotic messenger RNA by means of the polymerase chain reaction. Science 1992; 257: 967–971. 17 Velculescu VE, Zhang L, Vogelstein B, Kinzler KW. Serial analysis of gene expression. Science 1995; 270: 484–487. 18 Hubank M, Schatz DG. Identifying differences in mRNA expression by representational difference analysis of cDNA. Nucleic Acids Res 1994; 22: 5640–5648. 19 Lee S, Tomasetto C, Sager R. Positive selection of candidate tumor suppressor genes by subtractive hybridization. Proc Natl Acad Sci USA 1991; 88: 2825–2829. 20 Sutcliffe JG et al. TOGA: an automated parsing technology for analyzing expression of nearly all genes. Proc Natl Acad Sci USA 2000; 97: 1976–1981. 21 Shimkets RA et al. Gene expression analysis by transcript profiling coupled to a gene database query. Nat Biotechnol 1999; 17: 798–803. 22 Bertelsen AH, Velculescu VE. High throughput gene expression analysis using SAGE. Drug Discov Today 1998; 3: 152–159. 23 Hubank M, Schatz DG. cDNA representational difference analysis: a sensitive and flexible method for identification of differentially expressed genes. Meth Enzymol 1999; 303: 325–349. 24 Bowler LD, Hubank M, Spratt BG. Representational difference analysis of cDNA for the detection of differential gene expression in bacteria: development using a model of iron-regulated gene expression in Neisseria meningitidis. Microbiology 1999; 145: 3529–3537. 25 Chalmers MJ, Gaskell SJ. Advances in mass spectrometry for proteome analysis. Curr Opin Biotechnol 2000; 11: 384–390. 26 Gygi SP, Rist B, Aebersold R. Measuring gene expression by quantitative proteome analysis. Curr Opin Biotechnol 2000; 11: 396–401. 27 Anderson NL, Matheson AD, Steiner S. Proteomics: applications in basic and applied biology. Curr Opin Biotechnol. 2000; 11: 408–412. 28 Anderson NL et al. The effects of peroxisome proliferators on protein abundances in mouse liver. Toxicol Appl Pharmacol 1996; 137: 75–89. 29 Giometti CS, Tollaksen SL, Liang X. Cunningham ML. A comparison of liver protein changes in mice and hamsters treated with the peroxisome proliferator WY-14643. Electrophoresis 1998; 19: 2498–2505. 30 Kanitz MH et al. Alterations in rabbit kidney protein expression following lead exposure as analyzed by two-dimensional gel electrophoresis. Electrophoresis 1999; 20: 2977–2985. 31 Celis JE, Gromov P. 2D protein electrophoresis: can it be perfected? Curr Opin Biotechnol 1999; 10: 16–21. 32 Connolly HM et al. Valvular heart disease associated with fenfluraminephentermine. N Engl J Med. 1997; 337: 581–588. 33 Cookson J, Duffett R. Fluoxetine: therapeutic and undesirable effects. Hosp Med 1998; 96: 622–626. 34 McNeely W, Goa KL. Sibutramine: a review of its contribution to the management of obesity. Drugs 1998; 56: 1093–1124. 35 Redfield MM, Nicholson WJ, Edwards WD, Tajik AJ. Valve disease associated with ergot alkaloid use: echocardiographic and pathologic correlations. Ann Intern Med 1992; 117: 50–52. 36 Roy A, Brand NJ, Yacoub MH. Expression of 5-hydroxytryptamine receptor subtype messenger RNA in interstitial cells from human heart valves. J Heart Valve Dis 2000; 9: 256–260. 37 Pollack A. DNA chip may help usher in a new era of product testing. New York Times 28 November 2000. 38 Brazma A, Vilo J. Gene expression data analysis. FEBS Lett 2000; 480: 17–24. 39 Eisen MB, Spellman PT, Brown PO, Botstein D. Cluster analysis and display of genome-wide expression patterns. Proc Natl Acad Sci USA 1998; 95: 14863–14868. 40 Ashburner M et al. Gene ontology: tool for the unification of biology. The gene ontology consortium. Nat Genet 2000; 25: 25–29. 41 Johnson MD, Campbell LK, Campbell RK. Troglitazone: review and assessment of its role in the treatment of patients with impaired glucose tolerance and diabetes mellitus. Ann Pharmacother 1998; 32: 337–348. 42 Schoonjans K, Staels B, Auwerx J. The peroxiosme proliferator activated receptors [PPARs] and their effects on lipid metabolism and adipocyte differentiation. Biochim Biophys Acta 1996; 1302: 93–109.