Toxicogenomics in drug discovery: from preclinical studies to clinical trials

https://doi.org/10.1016/j.cbi.2004.09.013Get rights and content

Abstract

Gene expression analysis applied to toxicology studies, also referred to as toxicogenomics, is rapidly being embraced by the pharmaceutical industry as a useful tool to identify safer drugs in a quicker, more cost-effective manner. Studies have already demonstrated the benefits of applying gene expression profiling towards drug safety evaluation, both for identifying mechanisms underlying toxicity, as well as for providing a means to identify safety liabilities early in the drug discovery process. Furthermore, toxicogenomics has the potential to better identify and assess adverse drug reactions of new drug candidates or marketed products in humans. While much still remains to be learned about the relevance and the application of gene expression changes in human toxicology, the next few years should see gene expression technologies applied to more stages and more programs of the drug discovery and development process. This review will focus on how toxicogenomics can or has been applied in drug discovery and development, and will discuss some of the challenges that still remain.

Introduction

The application of microarray analysis towards toxicology, also referred to as toxicogenomics, is rapidly being embraced by the pharmaceutical industry as a useful tool to identify safer drugs in a faster, more cost-effective manner. Almost all major pharmaceutical companies have devoted significant resources to develop and apply microarray analysis towards toxicology. Further evidence of the interest in this field is exemplified by the vast numbers of workshops and consortia dedicated to establishing guidelines for determining how and when genomic data should be submitted to regulatory agencies. A number of examples of the application of toxicogenomics towards drug discovery and development have been reported in the literature. These have included gene expression studies that have been used towards identifying the mechanism of toxicity of pharmaceutical agents or standard toxins [1], [2], [3]. In other examples, gene expression analysis has been applied in a predictive mode, to identify potential safety liabilities before the development of other outward manifestations of toxicity such as clinical chemistry or histopathology [4], [5], [6]. In addition, there are some indications that toxicogenomics may be a valuable tool for bridging animal to human toxicity studies [7]. Yet, despite all of these encouraging examples, the true value that toxicogenomics can bring to the drug discovery and development process has yet to be demonstrated.

Gene expression profiling is still an emerging discipline for which efforts have so far essentially focused on validating approaches through proof-of-concept studies with reference or tool compounds. Such compounds usually induce general or specific, pronounced changes, and will not develop into human pharmaceuticals or have been on the market for many years. However, in the last few years, several pharmaceutical companies have clearly moved beyond this proof-of-concept stage and have applied this technology to drug development programs to address critical, development-limiting toxicologic issues, as indicated by recent submissions of gene expression profiling datasets to various regulatory agencies. This review will focus on how toxicogenomics has been or can be applied in drug discovery and development, and will address the following topics:

  • 1.

    The generation and use of reference databases.

  • 2.

    The application of toxicogenomics towards understanding the mechanism of toxicologic changes.

  • 3.

    The identification of candidates with potential toxic liabilities early in drug discovery using expression profiling.

  • 4.

    The application of toxicogenomics as a high-throughput assay using in vitro systems.

  • 5.

    The use of toxicogenomics to understand the relevance to humans of toxicologic or pathologic changes observed in preclinical species.

  • 6.

    The identification of selective gene expression signatures that could be used as sensitive biomarkers.

  • 7.

    The submission of pharmacogenomic data to regulatory agencies.

Gene expression changes, when viewed in isolation, can often lead to more questions than answers [8]. Abundant historical data are not yet available for the meaningful interpretation of thousands of gene expression changes detected with microarrays. In addition, gene expression changes associated with toxicologic changes typically reflect a large number of complex pharmacological, physiological and biochemical processes, most of them interacting with each other and related to a multitude of toxicological endpoints [9]. Consequently, numerous gene expression changes are not necessarily involved in the toxicologic process being investigated, but are simply secondary or indirect consequences of this toxicologic change. For instance, a general, non-specific sign of toxicity is anorexia or decreased food consumption, which can affect the homeostasis of several tissues. This has, for instance, been elegantly studied in rats using a variety of pathologic endpoints [10].

It is evident that to generate a plausible mechanistic hypothesis for the pathogenesis of a toxicologic change, the gene expression changes related to the toxicity need to be identified and separated from those that are adaptive, beneficial or unrelated to the development of the toxicologic change. This requires an appropriate study design, including the evaluation of multiple time points and the access to large reference databases, or at the very least, the use of reference or positive control compounds. The use of reference compounds may clarify which gene expression changes are related to a specific lesion or how the lesion develops. However, contextual information from large, established reference databases are optimal to properly interpret gene expression data by relating unique gene changes with compound classes or specific toxicologic mechanisms [8]. The large number of compounds, tissues, corroborative toxicologic and pathologic changes and gene expression data in these reference databases allows one to strengthen statistical inferences. The concept of databases in toxicology is not novel, and to better illustrate the requirement for toxicogenomics databases, one has only to relate to the use of databases for serum chemistry, hematology, pathology or carcinogenesis. Without proper historical databases and experience, interpretation of these data reveals little about the mechanism of action of toxicologic changes.

Ideally, toxicogenomic databases consist of many known pharmaceutical agents, toxicants and control compounds, at multiple doses and time points, with biological replicates for each condition [3], [4], [11], [12], [13], [14]. The reference compounds profiled in the database should reflect a variety of toxic mechanisms, and represent different structure–activity relationships. In some situations, time-course data can be very useful to identify gene expression changes linked to a time-dependent toxic response and increase the chances of observing a true toxic response over that of a single time point. The number of animals required for each time point or dose is also an important consideration for a given study. Typically, at least three animals are used per treatment. Such biological replicates are useful to establish both the biological and technical variability.

A number of publicly available, commercial or proprietary databases exist for the analysis of gene expression datasets. For instance, the National Institute of Environmental Health Sciences has recently created the National Center for Toxicogenomics to create a reference knowledge database that would ultimately allow scientists to understand mechanisms of toxicity through the use of gene expression analysis as well as proteomics and metabolite profiling [15]. Likewise, companies such as GeneLogic (Gaithersburg, MD) or Iconix Pharmaceuticals (Mountain View, CA) are leading toxicogenomics reference database providers. Their databases contain extensive gene expression profiles of a large number of prototypical reference compounds with corroborating toxicologic and pathologic endpoints.

At the present time, the impact of gene expression profiling in drug discovery and development has been most evident when this technology has been used for mechanistic purposes. The decision regarding whether a specific toxicologic change needs to be mechanistically understood is based on multiple factors, including the nature of the toxicologic change (development-limiting or not), the exposures at which the change occurs, the availability of good backup compounds, the stage of the program, and to some extent the culture of the company. For instance, if several backup compounds with very similar physicochemical and pharmacological properties are available early in a program, it may not be worth investigating the mechanism of a specific toxicologic change occurring with a single compound in an exploratory study. However, if backup compounds are not yet available, understanding the molecular basis of toxicologic changes may be useful when it comes time to properly select backup compounds without this liability through early structure–toxicity relationship studies during lead optimization.

Changes in expression of small gene sets have been shown to accurately discriminate among mechanistically distinct drug classes. These gene sets can be used to mechanistically classify compounds and match a compound with an unknown toxicologic mechanism into predefined classes of similar mechanism [12], [16]. For instance, our laboratory has investigated with oligonucleotide microarrays the hepatic effects (hepatomegaly and increased serum transaminases and alkaline phosphatase) of A-277249, a thienopyridine inhibitor of NF-κB-mediated expression of cellular adhesion molecules, in a 3-day repeat-dose toxicity study [6]. Using a proprietary gene expression database of known hepatotoxins, agglomerative hierarchical cluster analysis demonstrated that A-277249 had a gene expression profile quite similar to Aroclor 1254 and 3-methylcholanthrene (3MC), two well-characterized activators of the aryl hydrocarbon nuclear receptor (AhR), indicating that A-277249 hepatic changes were, at least in part, mediated by the AhR either directly or through effects on NF-κB [6].

In drug discovery and development, testicular toxicity is of particular interest, since testicular changes are typically subtle in early stages without well-established correlating biomarkers (such as changes in testicular weight) or striking morphologic changes [17]. Several elegant studies have demonstrated how gene expression profiling can elucidate the molecular basis of testicular toxicity [18], [19], [20]. For instance, gene expression changes in the testis were evaluated following exposure of mice to bromochloroacetic acid, a known testicular toxicant. Using a custom nylon DNA array, numerous changes in gene expression were detected in genes with known functions in fertility, such as Hsp70-2 and SP22, as well as genes encoding proteins involved in cell communication, adhesion and signaling, supporting that the toxicologic effect was the result of disruption of cellular interactions between Sertoli cells and spermatids [18], [21]. Likewise, Lee and co-workers, using several testicular toxicants, demonstrated that upregulation of Fas is a common and critical step for initiating germ cell death; thus, if Sertoli cells were the target of the toxicant, a concomitant upregulation of Fas ligand would also occur to eliminate Fas-positive germ cells [22], [23]. Such gene expression studies provide rapid methods to not only understand the mechanism of testicular toxicity, but also to develop counter-screens to analyze backup compounds.

Because gene expression changes occur before hispathologic or clinical pathologic changes, gene expression analysis has the potential to identify compounds with toxic liabilities early in the drug discovery process. By identifying gene expression changes that strongly correlate with toxicity, it may be possible to estimate safety margins and eliminate compounds with undesirable toxicologic profiles from 1- or 3-day rat studies. This would result in a tremendous savings in both time and resources for pharmaceutical companies, because it may be possible to discard compounds with unacceptable safety margins early in the drug discovery process [24]. However, this task has proven to be substantially more difficult than previously anticipated. In most cases, it is highly unlikely that a change in expression in any one gene can be strongly correlated with a particular toxicity. Rather, changes in expression of a number of genes, some of which may be increased, some decreased, are more likely to correlate with toxicity. In addition, the degree to which these gene expression changes occur may be crucial for determining if treatment with the compound at a certain dose level is likely to result in a toxic reaction. To address this, complex statistical tools for identifying and classifying patterns of gene expression changes have been employed for correlating gene expression profiles to toxicity. Some of these statistical methods are reviewed further.

Because a typical profile can contain thousands of gene expression changes, the first step in developing models for predictive toxicogenomics is to reduce the number of parameters in the classification model. Two major classes of approaches have been used for feature selection. The most commonly used approach is to rank genes with respect to differences in expression between experimental groups using parametric or non-parametric statistical comparisons, such as standard or permutation t- or F-test; ad hoc signal-to-noise statistics; Wilcoxon statistics; or significance analysis of microarrays [4], [11], [25], [26]. The genes that are differentially expressed at a specified significance level are then selected for inclusion in the prediction model. For example, in an effort to distinguish phenobarbital, an enzyme inducer, from peroxisome proliferators, gene expression profiles from livers of rats treated for 24 h with phenobarbital were compared to the pooled expression data from samples treated with Wyeth 14,643, gemfibrozil and clofibrate [11]. Application of a single-gene ANOVA method was able to identify signature genes that have the most consistent and dramatic expression difference between the two classes of compounds.

The second approach of feature selection is to use noise reduction methods, such as principal component analysis (PCA) and wavelet transformation [27]. PCA reduces a large set of genes into several components, where each new component is a weighted linear combination of all the genes. By selecting the first n components as predictors, one is able to vastly reduce the dimension of gene expression data and, at the same time, retain the major information. Wavelet transformation is another noise reduction method extensively used in areas such as medical imaging, signal processing, and spectral analysis in chemometrics [28], [29]. In our laboratory, both signature selection approaches have been incorporated into gene expression analysis. For example, over 200 gene expression profiles were collected from rat livers treated with a number of hepatotoxicants at two dose levels for 3 days. Initially, about 200 of the most relevant genes for hepatotoxicity were selected using ANOVA analysis at a p-value of 0.00001. Using wavelet transformation, the expression changes from the set of signature genes were then reduced to seven parameters. Fig. 1 shows the signature gene expression plot using the original value or the transformed value from four individual rats treated with 40 mg/kg/day 1-naphthylisothiocyanate (ANIT) or 150 mg/kg/day aspirin for 3 days. The similarities between the signals for the two rats treated with ANIT, for instance, are much more evident after wavelet transformation is applied.

Once the parameters have been reduced and signature genes have been selected, a predictive model can be established to classify a drug at a given concentration as toxic and to determine its potential toxicological liability. Classification modeling usually starts with known samples called a training set. By applying some computational algorithms, a set of rules or formulas can be established that correlates the signature gene expression changes with known toxicological endpoints; these rules or formulas can then be applied to classify unknown samples. An array of computational algorithms has been used in gene expression data from the fields of yeast biology, oncology, and certainly, toxicology. Examples of these methods include logistic regression, linear discriminant analysis (LDA), naïve Bayesian classifiers, artificial neural networks (ANN), and support vector machines (SVMs).

Both logistics regression and LDA use statistical inference to weigh the contributions of each signature gene expression value in sample prediction. Therefore, these methods require enough samples to adequately estimate a distribution. Tens to hundreds of samples might be needed depending on the variability of the gene expression data. In situations where there is a clear difference between groups, these methods can be quite robust [11].

Naïve Bayesian classification is another statistical approach that estimates the classification probabilities based on the prior knowledge of where the sample should belong and the new information provided by the current study. The application of naïve Bayesian classification is illustrated in a recent study by Thomas et al. [13]. In this study, liver gene expression profiles were collected from mice treated with 12 compounds at a given dose and several time points. The 12 compounds represented five well-characterized toxicological classes including peroxisome proliferators, aryl hydrocarbon receptor agonists, non-coplanar polychlorinated biphenyls, inflammatory agents, and hypoxia-inducing agents. A forward selection method was also used to rank order genes according to their discriminative power based on the Bayesian model. Using this approach, a classification model consisting of 12 genes was established to accurately classify all of the 24 samples into their associated chemical groups.

Artificial neural networks, on the other hand, do not use statistical inference to weight genes. Instead they are analogous to a biological nervous system. A typical neural network is composed of a number of highly interconnected processing elements called neurons and is tied together with weighted connections (Fig. 2). Because an ANN has the advantage of learning complex patterns, it has gained increasing popularity for classification of gene expression profiles. For example, Khan et al. reported accurate class prediction among small, round blue cell tumors of childhood using a two-layer neural network [27]. The inputs to the ANN were the first 10 principal components of all genes and the four outputs corresponded to four distinct diagnostic categories. After the model was calibrated, the contribution of each gene to the classification was measured by the sensitivity of the classification to a change in the expression level of each gene. The prediction accuracy was further improved by including only the 96 highest ranked genes into the ANN model. Although this specific example came from cancer classifications, the similarity of gene expression data from toxicological studies suggests the potential application of an ANN approach in the area of predictive toxicology.

Support vector machines are a relatively new type of learning algorithm [30]. When used for classification, SVMs operate by finding an optimal hyper-surface that separates the positive samples from the negative samples as far as possible. Unlike other classification algorithms, SVMs can perform very well even with a large number of classifiers. This feature makes it especially attractive to the classification of gene expression data, which usually have a large number of gene expression endpoints and a limited number of samples [31].

One can easily be confused with the choice of various classification algorithms. An important criterion to select the optimal prediction model is the prediction accuracy, which is estimated using a testing set that is different from the training set. Depending on the size of the gene expression database, a testing set can be selected by split-sample method, n-fold cross-validation, or the leave-one-out validation method [32]. Although most studies evaluate the robustness of the prediction using one of the validation approaches, no exhaustive survey of the various prediction methods has been reported in the literature. In the field of predictive toxicogenomics, the choice of methods could be very likely depend on the experimental design. For instance, a LDA approach could perform as well as other methods if there is a clear difference between classes and/or the sample size is relatively large. If one wants to differentiate a more complicated system, such as the general hepatotoxicity, from a limited number of gene expression profiles, a machine learning approach like SVMs or ANN would be recommended.

The vast majority of toxicogenomic studies to date have been carried out using tissue from animals dosed in vivo. These studies require a large amount of compound, such that they usually cannot be effectively applied early in the drug discovery process. Moreover, applying toxicogenomics towards samples from animal studies limits the number of compounds that can be analyzed, in part because of the cost and practicalities of these studies, but also because of the cost of the microarray experiments themselves. Thus, application of toxicogenomics using in vitro systems may significantly improve the throughput and value of this technology in drug discovery. An in vitro system would allow for the screening of compounds for potential safety liabilities very early in the drug discovery process, before large amounts of time or resources have been devoted to any one compound. In addition, the combination of toxicogenomics and in vitro systems may identify gene expression changes that could serve as biomarkers of toxicity in clinical studies to monitor possible adverse drug reactions. Finally, the combination of toxicogenomics and in vitro systems, such as isolated human hepatocytes may, in some cases, be more predictive of toxic reactions in humans than preclinical in vivo studies.

There is no clear consensus on what type of in vitro systems would be optimal for conducting in vitro toxicogenomic studies. For instance, an in vitro system for hepatotoxicity could consist of liver slices, isolated hepatocytes or liver cell lines. All of these systems offer advantages and disadvantages. Liver slices have the advantage of maintaining intact cellular interactions and spatial arrangements [33]. However, this system is not well suited for high-throughput analysis. Cell lines offer the advantage of being readily available; they generally give reproducible results over time, and are very cost effective. However, a number of studies have demonstrated that liver cell lines can be strikingly different from intact organs or even from primary cells in terms of function. A recent study compared gene expression patterns from rat livers to rat liver slices, primary rat hepatocytes cultured on collagen monolayer or collagen sandwich, and two rat liver cell lines, BRL3A and NRL clone 9 cells. The results showed that, based on gene expression analysis, liver slices were the most similar to intact rat livers, followed by primary hepatocytes in culture. The two rat liver cell lines showed little correlation, based on gene expression analysis, to intact rat livers. Further analysis revealed that the two cell lines expressed very low or undetectable levels of phase I metabolizing enzymes, both at the RNA and protein levels [34]. Thus, it is very likely that analyzing compounds for toxicity in cell lines may be inaccurate if the toxicity of the compound is due to the formation of a metabolite, such as the case for methapyrilene and acetaminophen.

In contrast, isolated hepatocytes are reasonably similar to intact livers in terms of gene expression analysis [34], [35]. In addition, they do retain their metabolizing capabilities for short-term cultures [36], [37]. However, isolated hepatocytes, especially of human origin, can be very difficult and expensive to generate or to obtain. Furthermore, lifestyle differences of the human donors, such as smoking or drinking habits, medications or general health may cause substantial differences in gene expression responses from one hepatocyte isolation to another. Fig. 3 shows that approximately 2500 significant gene expression changes can be observed between isolated hepatocytes in culture from one donor to another.

Ultimately, the choice of what cell system to choose for toxicogenomic analysis may depend on what questions are being addressed. If one desires to identify the mechanism of toxicity of a compound, more than likely using a cell system that most closely mimics the target organ, such as a primary cell system, would be preferable. However, if one wishes to identify general markers of toxicity, then the choice of cell type may not be as important. In fact, it may be possible to identify markers for hepatotoxicity using a non-liver cell line, such as HeLa or Jurkat cells. Further studies will help address these issues.

A continuing challenge associated with in vitro systems in drug safety assessment is the underlying question of their predictive value for toxicity in animals or people. Clearly, in vitro toxicogenomics falls into the same predicament as other in vitro assays used routinely for safety assessment. A number of published studies have begun to address the relationship of in vitro toxicogenomic results compared to in vivo safety liabilities. Several studies have shown that it is possible to distinguish compounds with different mechanisms of toxicity using in vitro systems. In our laboratory, we compared the gene expression profiles of 15 well-characterized hepatotoxins, including allyl alcohol, carbon tetrachloride, Aroclor 1254, and 3MC, in isolated primary rat hepatocytes [38]. Some of the compounds with similar mechanisms of toxicity, such as the AhR ligands Aroclor 1254 and 3MC, resulted in similar expression profiles, and, using unsupervised hierarchical clustering, were clearly distinguished from other agents such as carbon tetrachloride and allyl alcohol that cause toxicity by other mechanisms. In addition, in some cases there is significant correlation between the genes regulated in vivo and in vitro, as demonstrated with the AhR ligands (Table 1).

A similar study was performed by de Longueville et al. where primary rat hepatocytes were exposed to 11 different hepatotoxins, including acetaminophen, amiodarone, clofibrate, and 4-pentenoic acid. Using a low-density array containing only 59 genes, the authors were able to correctly group the compounds into different mechanistic hepatotoxin classes based on gene expression analysis [39]. Harries et al. treated a human hepatoma cell line, HepG2 cells, with carbon tetrachloride and ethanol. Microarray analysis of the treated cells showed that it was possible to distinguish the compounds based on gene expression, and that the gene changes were reflective of the mechanism of toxicity of the compounds [40]. Finally, a study by Hong et al. showed that gene expression analysis was able to distinguish ethanol from two quinone-containing anticancer drugs, mitomycin C and doxorubicin in HepG2 cells [41].

These studies demonstrate that gene expression profiling using in vitro systems can distinguish compounds that have different mechanisms of toxicity. However, this by itself is not sufficient for screening compounds for toxicity, since it would be neither practical nor amenable to high-throughput screening to identify gene expression changes for thousands of genes. For an in vitro toxicogenomics assay to have practical applications in drug discovery, gene expression patterns or markers need to be identified that are highly correlative with the potential to cause toxicity. In addition, in order to increase the throughput of gene expression profiling (for instance by adapting it to a 96- or 384-well format), it is necessary to reduce the number of genes that are being monitored. While much work remains to be done in this area, several studies have demonstrated that it is possible, using in vitro systems, to identify small subsets of genes, or signatures, that could be used to screen compounds for toxicity.

Newton et al. applied toxicogenomic analysis in two different cell lines for characterizing genotoxic compounds. From these studies, the authors were able to identify a small subset of genes that could be used to identify compounds that cause genotoxicity by directly interacting with DNA versus those compounds that are genotoxic via a secondary mechanism [42]. Morgan et al. used microarray analysis in an in vitro system to identify a subset of seven genes that correlated with oxidative stress. In this study, HepG2 cells were treated with various agents known to induce oxidative stress, such as sodium selenite and potassium bromate [43]. In a similar study, Burczynski et al., using microarray analysis in HepG2 cells, identified a small subset of genes that could be used to identify DNA damaging agents and cytotoxic anti-inflammatory agents [44]. Finally, Orr et al. was able to identify a set of genes in primary rat hepatocytes that were highly correlative for steatosis and hepatitis [45]. While a tremendous amount of work remains to be done in this area, extensive efforts are being applied to identify signatures of toxicity in an in vitro system, from academics, large pharmaceutical companies, and companies that specialize in toxicogenomics, such as GeneLogic and Iconix Pharmaceuticals. It is therefore very likely that in the very near future, the application of in vitro toxicogenomics assays will become a routine method for screening compounds early in the drug discovery process.

Toxicologic changes occurring in preclinical species are not necessarily relevant to humans because of species differences in cell biology, physiology or responses to changes induced by compounds [9]. A better understanding of the molecular mechanisms of toxicologic changes can help establishing the relevance to humans and better identify species-specific responses. This offers the opportunity to improve the predictive accuracy of extrapolating from preclinical species to humans, and consequently reduces this source of uncertainty in safety assessment. A classic example includes compounds that act via the peroxisome proliferator-activated receptor α (PPARα), such as the fibrate class of cholesterol-lowering drugs, and that upon chronic administration, cause hepatomegaly and eventually hepatic neoplasms in rats [46]. There are marked species differences in the response to peroxisome proliferators, with mice and rats being highly responsive and humans being poor to non-responders. This differential species response correlates directly with the number of hepatic PPARα; human liver only expresses 5–10% of the number of PPARα in rodent liver. Consequently, humans are at minimal or no risk to the development of hepatic tumors following chronic exposure to peroxisome proliferators, and compounds, such as the fibrates, have been extensively used as therapeutic molecules in humans without significant hepatic side-effects. A number of studies have used microarrays to further our understanding of the molecular mechanisms associated with the various effects of several peroxisome proliferators [12], [16], [47]. Some of these studies have also used other compounds, such as various hepatic enzyme inducers or phenobarbital, that also are associated with increased liver weight or hepatic neoplasia in rats. Because the effects of most of these compounds in humans are relatively well characterized, these studies provide a basic gene expression dataset with which to understand the relevance of changes in the rodent liver occurring with compounds in development. In addition to other correlating changes observed with other endpoints, such as electron microscopy or in vitro metabolism studies, gene expression profiling offers the opportunity to specifically demonstrate the mechanisms of action by which certain compounds lead to rodent hepatomegaly and hepatic carcinogenesis, thereby properly assessing the risk to humans.

The value of gene expression changes to identify or understand species-specific responses was also demonstrated in a study of cyclosporine-induced nephrotoxicity [48], [49]. In the kidneys of cyclosporine A-treated rats, a profound downregulation of calbindin-D 28 kDa, a calcium binding protein, was found to correlate with the accumulation of calcium in tubules, and to be the cause of this renal tubular calcification. Subsequently, treatment of rats with various cyclosporine derivatives and other immunosuppressive drugs, such as FK506 and rapamycyn, was also shown to downregulate calbindin-D 28 kDa expression, while cyclosporine did not regulate calbindin-D 28 kDa expression in the kidneys of dogs and monkeys, two species resistant to cyclosporine-mediated renal toxicity.

Our laboratory has performed transcriptional profiling on different antibacterial compounds from the quinolone class. One of the quinolones, trovafloxacin, caused a severe liver toxicity in a relatively small number of patients. Preclinical studies failed to identify any serious liver toxicity resulting from this compound. Isolated human hepatocytes from four separate donors were treated with five different quinolone compounds including trovafloxacin, and microarray analysis was performed. Fig. 4 shows the results from one of the donors. Using gene expression profiling, trovafloxacin can clearly be distinguished from the other quinolones; overall treatment with trovafloxacin resulted in far more gene expression changes than treatment with the other compounds. In addition, trovafloxacin regulated a number of mitochondrial genes that were not regulated by the other quinolone compounds. Mitochondrial toxicity has been purported to be a possible mechanism for idiosyncratic toxicity [50], [51], [52].

Another example of applying toxicogenomics to identify species-specific toxicity was demonstrated by Kier et al. In this study, isolated primary human hepatocytes were treated with three compounds from the thiazolidinediones (TZD) class of compounds used for the treatment of type II non-insulin-dependent diabetes. One of the TZD compounds, troglitazone, resulted in idiosyncratic hepatotoxicity in a small percentage of patients; it has since been removed from the market [53]. Microarray analysis of the treated hepatocytes showed that treatment with troglitazone resulted in a large number of gene expression changes that were not observed with two other TZD compounds, rosiglitazone and pioglitazone [5]. While studies such as these are admittedly preliminary, they do suggest that the combination of microarray analysis in isolated primary human cells may, in some cases, identify potential safety liabilities that are not evident from preclinical studies.

When toxicologic changes occur in preclinical studies, biomarkers specific to these changes are useful to monitor their possible occurrence in clinical trials and longer-term preclinical studies or simply to establish a No-Effect-Level (NOEL). Various biomarkers are already available for this purpose, such as serum chemistry or hematology parameters or cardiovascular parameters generated through electrocardiograms. While these current biomarkers cover a wide variety of potential toxicologic endpoints, there are a lot of changes for which specific and sensitive biomarkers are missing. Ideally, to be of practical use, a biomarker should be sensitive, specific for the toxicologic change and detectable or quantifiable in easily accessible tissue samples (such as urine or blood).

In some cases, microarray analysis can identify new gene expression changes that highly correlate with toxicity. These specific transcripts can then be evaluated at the protein level and ultimately be used to validate new biomarkers. This approach has been successfully used to identify adipsin as a potential non-invasive biomarker of gastrointestinal toxicity associated with perturbed Notch signaling. In the last few years, functional γ-secretase inhibitors (FGSIs) have been developed as potential therapeutic agents for Alzheimer's disease [54]. FGSIs can block the cleavage of several transmembrane proteins, including the cell fate regulator Notch-1. In rats exposed to FGSIs, gastrointestinal toxicity has been described, characterized by an increase in gastrointestinal weight, distended small and large intestines, and a mucoid enteropathy related to goblet cell hyperplasia in the ileum [54]. Microarray analysis of ileum of FGSIs-treated rats revealed changes in the expression of several genes consistent with Notch signaling perturbation. In particular, the transcript for the serine protease adipsin was found to be significantly increased (up to 12-fold) following treatment. The investigators also demonstrated elevated levels of the adipsin protein in gastrointestinal contents and feces of FGSIs-treated rats. These data collectively provided a potential biomarker of FGSIs-induced gastrointestinal toxicity.

Gene expression changes indicative of a toxicologic reaction occur in affected tissues, as well as potentially in other “surrogate” tissues. Hence, it is possible that while toxicologic changes may be located to soft tissues, correlating changes in gene expression may also occur in other readily accessible tissues, such as peripheral blood mononuclear cells (PBMCs) or buccal mucosal epithelial cells. These changes in gene expression could be directly related to target modulation if PBMCs express the drug target or indirectly as a result of toxicologic perturbations of other tissues. Albeit still in its infancy, this approach has already been evaluated [55]. In particular, allergic drug reactions appear optimal to evaluate this approach, since they are immune mediated. For instance, in a retrospective study using PBMCs and RT-PCR to evaluate various cytokines and cytotoxic markers, patients with severe and mild cutaneous allergic drug reaction were compared to subjects taking the same drugs and not experiencing reactions [56]. Gene expression results could clearly differentiate the three groups with patients experiencing severe reactions having higher expression for cytotoxic markers (perforin and granzyme B). Another study using microarray analysis in PBMCs was able to classify patients by treatment arm in a predictive fashion, supporting the concept that peripheral blood could be used as a surrogate system for biomarker identification [57]. As reagents for storage and purification of RNA from clinical blood samples improve, the analysis of whole blood rather than PBMCs should become readily feasible, minimizing laboratory procedures and variations due to sample handling and processing.

Because of the potential of gene expression profiling to improve the safety assessment of new chemical entities, the FDA has recently issued a draft guidance for the regulatory submission of pharmacogenomic data in an attempt to clarify FDA policy on the use of pharmacogenomic data in the drug application review process (http://www.fda.gov/cder/guidance). At the time this review is written, public comments have been collected and a final guidance is expected during the Fall of 2004. While this guidance is currently at the draft state and will most likely be modified to some extent, it is worth briefly reviewing its general content, as it reflects the current thinking of the FDA, but also because it provides some basis for the future of gene expression profiling in drug discovery and development. This guidance provides recommendations to sponsors on pharmacogenomic data submission requirement, the format to be used for data submission and how the data will be used in regulatory decision-making. In particular, this draft guidance demonstrates that the FDA is open to and expects the submission of gene expression profiling data that were generated to support scientific contentions related to toxicity. An interesting aspect of this draft guidance is the concept of “Voluntary Genomic Data Submission” (VGDS). The FDA recognizes that at the current time, most gene expression profiling data are of an exploratory or research nature, and would therefore not be required for submission. However, to be prepared to appropriately evaluate future submissions, FDA scientists need to develop an understanding of a variety of relevant scientific issues. VGDS would provide the material necessary to develop this understanding, and would be reviewed by a cross-center Interdisciplinary Pharmacogenomic Review Group (IPRG) whose composition is still unknown. Finally, the guidance discusses pharmacogenomic tests as biomarkers, indicating that gene expression datasets could ultimately be recognized as biomarkers. It is worth mentioning that this draft guidance distinguishes between “known valid biomarkers” (that have been accepted in the broad scientific community) and “probable valid biomarkers” (that appear to have predictive value, but may not be widely accepted or have been independently replicated). While this distinction led to some confusion at the workshop organized by the Drug Information Association in November 2003 in Washington, DC, to discuss this guidance, it also re-emphasizes the enormous amount of work and improvement that will be needed in the future to make gene expression data suitable for regulatory decision making. This includes obviously an improved scientific framework for data interpretation through the use of larger, more complete reference databases, but also improved quality control of laboratory procedures, a better understanding of the comparability of different platforms, and some better-defined process to validate biomarkers.

Section snippets

Summary

In last few years, toxicogenomics has moved from being an exploratory technology to a technique that has had a significant impact in the discovery and development of new chemical entities. Large-scale gene expression analysis has already been implemented in several pharmaceutical companies to generate earlier toxicologic data points, thereby enabling a faster and more accurate selection of the best candidate molecules early in the drug discovery process. In the next few years, the value that

Reference (57)

  • J.F. Waring et al.

    Isolated human hepatocytes in culture display markedly different gene expression patterns depending on attachment status

    Toxicol. In Vitro

    (2003)
  • A.P. Li et al.

    Primary human hepatocytes as a tool for the evaluation of structure–activity relationship in cytochrome P450 induction potential of xenobiotics: evaluation of rifampin, rifapentine and rifabutin

    Chem. Biol. Interact.

    (1997)
  • R.G. Ulrich et al.

    Cultured hepatocytes as investigational models for hepatic toxicity: practical applications in drug discovery and development

    Toxicol. Lett.

    (1995)
  • J.F. Waring et al.

    Microarray analysis of hepatotoxins in vitro reveals a correlation between gene expression profiles and mechanisms of toxicity

    Toxicol. Lett.

    (2001)
  • F. de Longueville et al.

    Gene expression profiling of drug metabolism and toxicology markers using a low-density DNA microarray

    Biochem. Pharmacol.

    (2002)
  • H.M. Harries et al.

    The use of genomics technology to investigate gene expression changes in cultured human cells

    Toxicol. In Vitro

    (2001)
  • Y. Hong et al.

    Discriminating two classes of toxicants through expression analysis of HepG2 cells with DNA arrays

    Toxicol. In Vitro

    (2003)
  • O. Grenet et al.

    The cyclosporine A-induced decrease in rat renal calbindin-D 28 kDa protein as a consequence of a decrease in its mRNA

    Biochem. Pharmacol.

    (1998)
  • M.C. Varela et al.

    Cyclosporine A-induced decrease in calbindin-D 28 kDa in rat kidney but not in cerebral cortex and cerebellum

    Biochem. Pharmacol.

    (1998)
  • U.A. Boelsterli

    Diclofenac-induced liver injury: a paradigm of idiosyncratic drug toxicity

    Toxicol. Appl. Pharmacol.

    (2003)
  • Y. Masubuchi et al.

    Role of mitochondrial permeability transition in diclofenac-induced hepatocyte injury in rats

    Hepatology

    (2002)
  • R.G. Ulrich et al.

    Metabolic, idiosyncratic toxicity of drugs: overview of the hepatic toxicity induced by the anxiolytic, panadiplon

    Chem. Biol. Interact.

    (2001)
  • K.G. Tolman et al.

    Hepatotoxicity of the thiazolidinediones

    Clin. Liver Dis.

    (2003)
  • G.H. Searfoss et al.

    Adipsin a biomarker of gastrointestinal toxicity mediated by a functional gamma-secretase inhibitor

    J. Biol. Chem.

    (2003)
  • S.J. Posadas et al.

    Gene expression levels of cytokine profile and cytotoxic markers in non-immediate reactions to drugs

    Blood Cells Mol. Dis.

    (2002)
  • J.F. Waring et al.

    Development of a DNA microarray for toxicology based on hepatotoxin-regulated sequences

    Environ. Health Perspect.

    (2003)
  • H. Ellinger-Ziegelbauer et al.

    Characteristic expression profiles induced by genotoxic carcinogens in rat liver

    Toxicol. Sci.

    (2004)
  • J. Uetrecht

    Prediction of a new drug's potential to cause idiosyncratic reactions

    Curr. Opin. Drug Discov. Dev.

    (2001)
  • Cited by (60)

    • Transcriptomics in toxicology

      2017, Food and Chemical Toxicology
      Citation Excerpt :

      In vitro cell culture models are attractive alternatives to in vivo animal models to evaluate the potential of drug candidates to result in DILI. Cell culture models are cheaper, faster, and more convenient to screen a large number of drug molecules for liver toxicity (Yang et al., 2004). Several investigators have attempted to integrate transcriptome analysis in cell culture models as an alternate approach to the use of expensive in vivo models for early detection of liver toxicity.

    • Potential new targets involved in 1,3-dinitrobenzene induced testicular toxicity

      2012, Toxicology Letters
      Citation Excerpt :

      Toxicogenomics integrates toxicology and gene expression evaluation, where altered transcript levels in tissues or cells following exposure to chemicals are analyzed in association with conventional toxicity endpoints. Anchoring the molecular changes to the resulting phenotypic lesions can help to define the sequence of key molecular events associated with the mechanism of action, and can allow a more effective safety assessment of chemicals or new drug candidates (Cui and Paules, 2010; Daston, 2008; Suter et al., 2004; Yang et al., 2004). This approach has been successfully applied to the study of several types of toxicities, including hepatotoxicity (Blomme et al., 2009; Heinloth et al., 2004), nephrotoxicity (Amin et al., 2004), genotoxicity (Ellinger-Ziegelbauer et al., 2004) and testicular toxicity (Ludwig et al., 2011; Rouquie et al., 2009).

    • Potential of 'Omics' Technologies for Implementation in Research on Phytotherapeutical Toxicology

      2012, Advances in Botanical Research
      Citation Excerpt :

      This assumption is based on the hypothesis that TGx can detect low toxicity of compounds, which is below the detection limit of conventional methods. TGx was rapidly embraced by the pharmaceutical industry as useful tool to identify safer drugs in a quicker, more cost-effective manner (Yang et al., 2004). Studies have already demonstrated the benefits of applying gene expression profiling towards drug safety evaluation, both for identifying mechanisms underlying toxicity, as well as for providing a means to identify safety liabilities early in the drug discovery process.

    View all citing articles on Scopus
    View full text