Applications of Toxicogenomic Technologies to Predictive Toxicology and Risk Assessment (2007)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

5 Application to Hazard Screening This chapter addresses the application of toxicogenomics to screening chemical compounds for hazard, or the ability to cause harm. A screening test can be defined as one designed to detect a state or property more quickly and cheaply than more elaborate tests for that state or property. In predictive toxi- cology, the property being detected by screening tests is generally hazard. Screening tests may not give complete information on toxicity, such as the time course, chronic effects, or dose-response characteristics. Therefore, in the con- text of this chapter, screening data provide an input to the hazard identification step in risk assessment but do not allow full determination of risk.1 The chemical and pharmaceutical industries also use screening tests to de- tect desirable properties, such as the ability to bind to specific target receptors. The use of genomic techniques to screen for desirable, pharmacologic properties may be analogous to hazard screening (Lum et al. 2004) but is not the principal focus of this chapter. Toxicogenomic technologies may be incorporated directly into existing, more traditional hazard screening tests; they may be the basis of new tests that substitute for more traditional tests; or they may generate mecha- nistic insights that enable more basic tests to be conducted for screening com- pounds, such as receptor binding or other physicochemical assays. This chapter primarily discusses the first two applications of toxicogenomic technologies. DESCRIPTION AND PROPERTIES OF CURRENT HAZARD SCREENING METHODS Hazard screening can be comprehensive, intended to detect all potential hazards, or it can be more limited, detecting only a specific type of hazard. A 1 Hazard identification is one of four elements of a full risk assessment as described in the National Research Council report (NRC 1983). The other three elements are exposure assessment, dose-response assessment, and risk characterization. 73 

74 Applications of Toxicogenomic Technologies comprehensive hazard assessment for a chemical substance generally requires a variety of in vitro and in vivo toxicologic assays as well as evaluations of physi- cal properties. The selection of individual screening tests depends greatly on the setting and specific regulatory requirements. For example, the current practice of the U.S. Environmental Protection Agency (EPA) under the Toxic Substances Con- trol Act (TSCA), in the absence of more extensive preexisting data, is to screen new chemicals based solely on physicochemical data using quantitative struc- ture-activity relationship models. In this setting, chemical tests may be limited to properties such as boiling point, octanol-water partition coefficient, vapor pres- sure, and solubility. If environmental fate and transport of substances are not primary concerns, short-term in vivo rodent assays may be used, such as a 28- day feeding study, which examines histopathology in most critical target organs. More comprehensive screening programs have adopted batteries of tests that provide information on different types of toxicity but remain insufficient to fully assess chemical risks. As one example, the Organization of Economic Coopera- tion and Development (OECD) has developed the Screening Information Data Set (SIDS), which consists of the 21 data elements shown in Table 5-1. Each toxicity test involves administering a measured amount of a compound to whole organisms or to cells in culture and then measuring indicators of toxic outcomes. Compared with more extensive tests, screening tests tend to use higher and fewer doses of the compound being studied, fewer test subjects, a shorter time period of observation, and less extensive evaluation of the toxic outcomes. To reduce the use of mammals for laboratory testing, there is a strong impetus to develop and validate screening tests that use cultured cells or lower order ani- mals, such as worms. The incorporation of toxicogenomics into screening tests involves measur- ing gene, protein, or metabolite changes in response to specific doses of an ad- ministered test compound at specific time points, with or without the parallel measurement of more traditional markers of toxicity. The critical question about new toxicogenomic techniques is whether they can improve hazard screening by making tests faster, more comprehensive, less reliant on higher order animals, and more predictive and accurate without being prohibitively expensive. For a screening test to be useful, it must be capable of detecting the prop- erty or state being tested when it truly exists. This is the definition of the “sensi- tivity” of a screening test. In many cases, screening tests are designed to be highly sensitive, sometimes at the expense of the specificity of the test or the ability of the test to return a negative result when the property or state of con- cern does not exist. Another way to describe this quality is that hazard screening tests often accept a higher rate of false-positive results to avoid not detecting a hazard because of a high rate of false-negative results. When the data generated by screening tests are continuous, as is the case with gene and protein expression and metabolite assays, the selection of thresh- olds for positive and negative results plays a dominant role in determining the 

Application to Hazard Screening 75 TABLE 5-1 Elements of the OECD SIDS Data Elements Comments Physical-chemical properties Melting point Boiling point Relative density Required for inorganic chemicals and should be provided if readily available for organic chemicals. Vapor pressure Partition coefficient: n- octanol/water Water solubility Dissociation constant For substances normally capable of dissociation. Oxidation-reduction potential Required for inorganic chemicals; may be required for certain organic chemicals. Environmental fate Photodegradation Stability in water Not required for classes of chemicals whose molecular structure does not possess functional groups subject to hydrolysis or that are generally recognized to be resistant to hydrolysis. In these cases, a qualitative statement can be provided. Transport and distribution Including Henry’s law constant, aerosolization, between environmental volatilization, soil adsorption, and desorption, compartments including based on experimental data or, if not available or distribution pathways appropriate, calculated using structure-activity relationships. Aerobic biodegradability Environmental toxicology Acute toxicity to fish Acute toxicity to daphnia Toxicity to algae Chronic toxicity Necessity determined based on physical- chemical properties of the chemical. Any new data required should be collected using the most sensitive species (fish, daphnia, or algae) within limitations of the chemical properties. Terrestrial toxicity The need for testing will normally be addressed at the post-SIDS stage. However, if significant exposure is expected or identified in the terrestrial environment (soil), appropriate terrestrial toxicity tests should be considered at the SIDS level. Taking into account animal (Continued on next page) 

76 Applications of Toxicogenomic Technologies TABLE 5-1 Continued Data Elements Comments welfare considerations, the need for avian toxicity testing should be considered only at the post-SIDS stage. Mammalian toxicology Acute toxicity By oral route, dermal route, or inhalation; required only on the most relevant route of exposure Repeated dose toxicity The protocol for new studies should specify the use of the most relevant route of exposure. Genetic toxicity Two end points required, generally point mutation and chromosomal aberrations. Reproductive toxicity Requires data to assess fertility and developmental toxicity. Experience with human exposure If available. Source: Adapted from OECD 2006. Reprinted with permission; 2006, Organisation for Economic Co-operation and Development. sensitivity and specificity of the test. When larger values of the test are more likely to indicate the presence of a particular hazard, selection of a relatively low value as the threshold for a positive result will lead to greater sensitivity and lower specificity (that is, fewer false-negatives and more false-positives). Con- versely, a high threshold for a positive result will lead to lower sensitivity and higher specificity (that is, more false-negatives and fewer false-positives). A critical challenge in designing and validating toxicogenomic screening tests is to identify and define a “gold standard” for hazard screening—the indication of the true state of toxicity against which the sensitivity and specificity of the screening test can be measured. IN VITRO VERSUS IN VIVO TESTING Current screening tests are done with whole animals in vivo and with tis- sue slices, cultured cells, or artificial biologic systems, including individual re- ceptor molecules, in vitro. Because toxicogenomic technologies measure gene, protein, and metabolite responses, they require whole animals, intact tissues, or cell cultures. Other in vitro tests with receptor binding or other molecular end points may be used in combination with toxicogenomic-based tests but are not discussed in detail in this chapter. In vitro tests provide many benefits in a screening setting but also pose serious challenges. They are generally less ex- pensive, more readily automated, and therefore faster than in vivo testing. How- ever, questions about their relevance to in vivo toxicity limit their use in deci- sion-making processes without additional supporting data (Boess et al. 2003; Jessen et al. 2003). Toxicogenomic technologies may offer in vitro methods for 

Application to Hazard Screening 77 answering questions that currently require in vivo tests, when results can be ex- trapolated from one species to another. In the following discussion, current prac- tices and published studies describe the state of the art with respect to the devel- opment of toxicogenomic applications for both in vitro and in vivo screening tests. DEVELOPMENT OF PREDICTIVE ALGORITHMS FOR SCREENING PURPOSES Development of screening tools involves both identifying signatures that can distinguish between various toxic and nontoxic end points and developing predictive algorithms that can use these data effectively to make predictions. Both represent significant challenges, particularly given that most published studies involve far fewer samples than the number of measurements on each sample (the “n < p” problem discussed in Chapter 3) and the well-known prob- lem of multiple testing. Given the broad interest in developing toxicogenomic predictors, these remain areas of active research in statistics, bioinformatics, and computer science. Although several studies have compared various combina- tions of predictive algorithms (Natsoulis et al. 2005; Van Delft et al. 2005), no single method has emerged as being superior for all applications. This result is not surprising given the work of Wolpert and McReady, who present a series of “no-free-lunch” theorems showing that, although a particular algorithm might be optimal for one class of problems, it might be greatly surpassed by another algo- rithm when applied to a slightly different class (Wolpert and MacReady 1997). As most toxicogenomic studies presented to date have looked at small numbers of samples and have not arrived at optimal solutions for predictive classification, these results suggest that significant additional work is necessary to fully ad- dress these issues. STATE OF THE ART Current Practices for Pharmaceutical Hazard Screening With the advent of combinatorial chemistry, pharmaceutical companies are now able to consider thousands of drug candidates at once in the drug devel- opment process. Drug companies are seeking a variety of in vitro and in silico methods to screen drug candidates for effectiveness and toxicity before more expensive, and slower, in vivo testing (Johnson and Wolfgang 2000; Ulrich and Friend 2002; van de Waterbeemd and Gifford 2003; Suter et al. 2004; Butcher 2005). Toxicogenomic technologies are being explored in combination with more traditional in vitro cytotoxicity assays to improve the information obtained from this more rapid screening. The goals are to reduce attrition of compounds during more costly phases of drug development (that is, Phase 2 and Phase 3 

78 Applications of Toxicogenomic Technologies clinical trials) by identifying potential adverse drug reactions and other possible hazards at much earlier stages. One study has reported that current use of toxicogenomic technologies for screening drug candidates varied widely among leading pharmaceutical compa- nies (Branca and Peck 2004). Of the 12 companies surveyed, 7 had made or were planning to soon make extensive use of toxicogenomics for drug candidate screening, and the other 5 limited the use of toxicogenomics to analysis of failed candidates or reference compounds. Most companies applied toxicogenomic technologies in the discovery and preclinical phases, and, whereas most pre- ferred using in vivo assays, several said they were developing in vitro screening tests. Among the principal barriers to broader use and acceptance of toxicoge- nomic technologies in drug screening were uncertainty about biologic relevance, resource limitations, and regulatory consequences. In addition to major pharmaceutical companies, smaller biotechnology companies specializing in toxicogenomic technologies provide screening ser- vices on a contractual basis. Two such companies, Gene Logic and Iconix, pur- port to identify compounds that are most likely to demonstrate unacceptable toxicity based on a molecular profile evaluated by using their proprietary toxi- cology signature models. These companies derived their signature models from analyzing databases containing hundreds or thousands of profiles from reference compounds tested in both in vivo and in vitro systems. Gene Logic uses short-term studies of investigative compounds and their proprietary models to provide predictive profiles of liver, kidney, and heart tox- icity (Gene Logic Inc. 2005). Iconix conducts short-term studies on investigative compounds using up to five time points and at least two doses to sample 12 dif- ferent tissues in laboratory animals. Gene expression data are combined with an assortment of in vitro receptor binding and enzyme tests on the investigative compounds and the results are compared with Iconix’s proprietary database of gene expression, molecular pharmacology, literature review, and histopathology to predict potential toxicities (Iconix Pharmaceuticals 2005). Both companies highlight the ability of gene expression arrays to provide mechanistic insight into any potential toxicities that are identified. Neither company claims the abil- ity to predict outcomes of chronic toxicity, such as cancer or neurodegenerative diseases. The Japanese government, through its Ministry of Health, Labour and Welfare and National Institute of Health Sciences (NIHS), has teamed up with a consortium of 17 pharmaceutical companies on the project “Construction of a Forecasting System for Drug Safety Based on the Toxicogenomics Technique and Related Basic Studies.” The long-term goals of this 5 billion yen (approxi- mately U.S. $43 million) project are to facilitate drug development and improve toxicity prediction by linking this database with others, including one being de- veloped at NIHS for environmental chemicals (Urushidani and Nagao 2005). For each of the initial 150 compounds investigated, the following tests are con- ducted: toxicogenomic, biochemical, and histopathologic studies of acutely and 

Application to Hazard Screening 79 repeatedly dosed rats; in vitro tests on rat hepatocytes; and in vitro tests on hu- man hepatocytes. Current Practices for Environmental Chemicals Whereas pharmaceutical companies are clearly conducting toxicoge- nomic-based screening of compounds, there is little documentation of such screening going on in the chemical industry. Publicly available information on the application of toxicogenomic technologies for screening environmental chemicals has come from government institutions rather than from industry. EPA and NIHS are conducting two projects of interest. Through a contract with Iconix, EPA is assessing the usefulness of gene expression signatures for categorizing environmental chemicals according to hazard. The Iconix contract involves five chemicals and to date, because of funding constraints, has assessed only hepatic toxicity (R. Kavlock, EPA, personal communication, 2006). EPA scientists are investigating how to adjust the statistical thresholds to optimize sensitivity and specificity for environmental chemicals as opposed to drug can- didates, upon which the Iconix system was originally designed (R. Kavlock, EPA, personal communication, 2006). The Japanese NIHS is creating a database of mouse gene expression array data corresponding to four doses and four time points for roughly 100 environmental and industrial chemicals. The platform for this data system, which they have dubbed “millefeuille” (French for thousand leaves), is based on Affymetrix microarrays. Data obtained are normalized on a per cell basis (Kanno et al. 2005; Urushudani and Nagao 2005). Proof-of-Concept Studies from the Published Literature Although some of the efforts to incorporate toxicogenomic technologies into hazard screens suggest the development of robust predictive models, the published literature does not yet provide details on these systems. Early proof- of-concept studies established that gene arrays can be used to classify toxicants with different mechanisms (Thomas et al. 2001; Waring et al. 2001; Hamadeh et al. 2002a,b). More recent studies, discussed below, apply data classification al- gorithms to gene expression data to predict the toxicity of unknown compounds or to demonstrate the use of proteomic methods and in vitro models to classify toxicants. Steiner at al. (2004) at Hoffman La Roche Ltd. reported an early attempt at predicting the hazard of unknown compounds. Their study used a class pre- diction approach with support vector machines (SVMs), a class of supervised learning algorithms that recognize informative patterns within an input dataset and then classify previously unseen samples. The training set consisted of mi- croarray gene expression data from 28 known hepatotoxicants and 3 compounds with no hepatotoxicity. The data were correlated with histopathology findings and a variety of clinical chemistry markers associated with liver toxicity. To 

80 Applications of Toxicogenomic Technologies reduce the number of genes analyzed by the SVM algorithms, the authors used a recursive feature elimination approach that excluded genes that did not contrib- ute significantly to the classification of compounds as hepatotoxicants. This re- sulted in a much smaller set of informative genes being identified. This ap- proach correctly predicted 63/63 vehicle-treated animals as belonging to the untreated control group. Although this may reflect overfitting of the parameters in the classifier, it is worth noting that approximately 90% of test compounds with known hepatotoxicity, but not previously tested by the algorithm, were correctly predicted to have toxic effects on the liver, suggesting that the ap- proach has identified a hepatotoxicity signature. Although there were no false- positive predictions of toxicity, there was clearly a significant (10%) false- negative classification of toxicants. These may be compounds with mechanisms of action not represented in the original training data, but it does suggest that additional work is necessary before these results can be broadly applied. Steiner et al. (2004) next asked if the SVM approach could be used to pre- dict the subtype of liver histopathology that a test compound would induce. They first classified compounds in the training set based on their mechanisms of toxicity: direct-acting hepatotoxicants, steatosis inducers, cholestasis inducers, or PPAR-α agonists (peroxisome proliferators). They based classification on the most prevalent type of histopathology induced at specified doses. They then generated five different SVM algorithms to distinguish among the controls and treatment classes. Not surprisingly, the number of genes required, as well as the accuracy of prediction, differed for each type of toxicity; in some cases, the same genes were included in classifying the different types of toxicity. The au- thors evaluated data from each new microarray experiment with each algorithm and predicted the mechanism of toxicity with a calculated discriminant value. This approach did not misclassify any of the controls. Although only three test compounds were analyzed, the results attested to the ability of the algorithm to accurately classify hepatotoxicants on the basis of their mechanism of action. Moreover, this approach seemed to be capable of discerning compounds with mixed modes of action. Scientists from Hoffman-LaRoche (Suter et al. 2003; Ruepp et al. 2005) recently described another example of an early proof-of-concept study. The study provides a rare insight into the value of toxicogenomic approaches when investigators have access to a densely populated, high-quality database—in this case, the Roche toxicogenomic database, which classifies compounds into sub- categories based on their histopathologic manifestations. This study used a retrospective approach to determine whether gene ex- pression profiling could have predicted the toxicity of a failed candidate com- pound before it was evaluated by more time-consuming methodologies of clas- sic toxicology. The compound in question was Ro-Cmp-A, an antagonist of the brain-specific, 5-hydroxytryptamine (5-HT6) receptor with predicted value in improving memory deficit in patients with Alzheimer’s disease. Toxicologic studies in rats and dogs found significant hepatotoxicity, characterized by steatosis, despite the absence of the 5-HT6 receptor in liver 

Application to Hazard Screening 81 cells. The gene expression profiles induced in rat liver in vivo by Ro-Comp-A and a nonhepatotoxic homolog, Ro-Cmp-B, were determined as a function of dose and time after exposure. The authors then compared the profiles with those in the reference gene expression database. Gene expression profiles induced at subchronic doses ad- ministered over 7 days made it possible to identify Ro-Cmp-A as a steatotic hepatotoxicant. Remarkably, gene expression profiles induced by acutely toxic doses also identified Ro-Cmp-A as a steatotic hepatotoxicant as early as 24 hours after exposure, despite a lack of characteristic histopathology. If specific marker genes with mechanistic links to toxicity were selected, steatotic activity was predicted by expression profiling as soon as 6 hours after an acutely toxic dose. In contrast, the nonhepatotoxic analog Ro-Comp-B induced gene expres- sion patterns in rat liver that more closely resembled those of the untreated con- trol animals. In addition to the in vivo studies, exposure of rat hepatocytes to Ro-Cmp-A in vitro at doses that also did not induce significant toxicity induced a subset of the predictive genes. This example speaks to several of the issues related to the initial use of toxicogenomic data for screening or hazard identification, at least for application to pharmaceuticals. First and foremost is the necessity for access to a high- quality database that is densely populated with high-quality toxicogenomic data that represent many compounds and that are phenotypically anchored (associ- ated with more traditional indicators of clinical pathology). This was a key fac- tor in classification of the candidate compound as a steatotic agent across doses and exposure times. In fact, the authors indicated that, when using this database, essentially the same set of differentially expressed genes could be used for clas- sification regardless of the type of statistical or computational approaches used to mine the data. Another significant finding of this study was that a subset of the expression profile that predicted steatosis could be replicated in a short term in vitro cell-based assay, at least at carefully selected doses. Another example of a toxicogenomic screen for a specific mode of action also used a short-term in vitro cell-based assay (Sawada et al. 2005). The au- thors correlated electron microscopic findings of phospholipidosis at the cellular level (development of lamellar myelin-like bodies in lysosomes) with changes in gene expression measured with microarrays to define a subset of genes that were predictive of phospholipidosis. They developed an mRNA assay based on poly- merase chain reaction (PCR) with this subset of genes and demonstrated its pre- dictive ability with a different set of compounds with a known tendency to cause phospholipidosis. Whether this screening algorithm is generalizable to other model systems remains to be demonstrated. In addition, the use of a PCR-based assay for screening new compounds (after using microarrays to select the genes) makes it more expensive than other testing methods to detect phospholipidosis. The authors noted that converting the PCR-based assay to a gene-chip-based assay would be likely to lower costs and improve throughput for use as a screen. Another example of toxicogenomic hazard screening involving mechanis- tic information is the effort to develop profiles to identify specific carcinogenic 

82 Applications of Toxicogenomic Technologies modes of action. Toxicogenomic profiles for cancer mode of action would also have profound implications for risk assessment methodology. In an in vivo short-term study, Ellinger-Zeiglebauer et al. (2004) tested the hypothesis that genotoxic liver carcinogens deregulate a common set of genes and that those deregulated genes represent defined biologic pathways implicated in early events in tumorigenesis. Although neither a single gene nor pathway sufficiently discriminated genotoxic from non-genotoxic carcinogens, their findings sug- gested that, with further understanding of mechanisms of carcinogenesis as well as further development of data and analytic tools, combinations of pathway- associated gene expression profiles may ultimately be able to predict genotoxic or nongenotoxic carcinogenic potential of compounds in short-term studies (El- linger-Ziegelbauer et al. 2004). Among nongenotoxic carcinogens, however, it may be difficult to generalize responses to derive simple profiles indicative of nongenotoxic mechanisms. Iida and coworkers demonstrated that mouse liver tumor transcriptional response to nongenotoxic carcinogens involved more dif- ferences than similarities in changes in early gene expression (Iida et al. 2003). Further analyses revealed that early gene expression changes appeared to be carcinogen specific and involved apoptosis and cell-cycle-related genes. Thus, the ability of early changes in gene expression to predict carcinogenesis requires further evaluation. Much of the published literature has focused on hepatotoxicity as an end point, because it is one of the best characterized adverse effects and is frequently responsible for candidate drug failure. However, one recent study from Iconix (Fielden et al. 2005) describes the development of a screening algorithm for nephrotoxicity. The authors used a training set of 64 compounds (15 known re- nal tubular toxicants and 49 non-nephrotoxicants, with gene expression meas- ured after 5 days of treatment) and a sparse linear programming algorithm (SPLP) to derive a set of 35 gene signatures for detecting preclinical nephrotox- icity. The SPLP algorithm is similar to the support vector machine approach described earlier, but it produces short lists of discriminant genes to allow more intuitive understanding of the mechanistic roles of individual genes in the signa- ture. These 35 gene signatures were then tested on 21 compounds whose nephrotoxicity was already characterized but that were not in the original train- ing set. Seven of the 9 (78%) known tubular toxicants and 9 of the 12 (75%) known non-nephrotoxicants were correctly identified. Although this moderate degree of sensitivity and specificity is not ideal for a screening test, the authors suggest that the test provides a promising alternative to 28-day in vivo studies with histopathology, which are the only tests now available to screen candidate drugs for nephrotoxicity. Another nonhepatotoxicity example assessed the utility of transcriptome profiling in screening for endocrine disruptors (Daston and Naciff 2005). The investigators examined gene expression profiles in the uterus and ovaries of fetal and prepubertal female rats. They administered three prototypical estrogenic compounds—17-α-ethynylestradiol, genistein, and bisphenol A—at several doses. Transcriptome profiles indicated a similar pattern common to all three 

Application to Hazard Screening 83 estrogenic compounds. Moreover, there was overlap in the responses in hor- mone-sensitive organs, including genes known to be regulated by estrogen. Most efforts to apply toxicogenomic techniques to hazard screening have used gene expression (transcriptome) profiling, but a small number of studies have demonstrated the potential usefulness of proteomic and metabonomic stud- ies to classify putative drugs or other compounds into categories of toxicity or biologic activity. Fella et al. (2005) reported the use of proteomic techniques to discover early markers of liver cancer in a 25-week study of Wistar rats exposed to N-nitrosomorpholine. The authors retrospectively identified several proteins characteristic of active cancers that were also deregulated at 3 weeks of expo- sure, suggesting that they might prove to be efficient indicators of liver carcino- genesis in shorter studies than standard cancer bioassays. Researchers at Pfizer (Kikkawa et al. 2005; Yamamoto et al. 2005) added proteomic analyses to in vitro cell culture assays in two proof-of-concept studies using the hepatotoxicants acetaminophen, amiodarone, and tetracycline. In the first study, they identified 31 candidate marker proteins together with more standard measures of hepatotoxicity, such as lactate dehydrogenase release and altered mitochondrial respiration. They concluded that proteomic methods were superior to lactate dehydrogenase release at detecting toxic changes at earlier time points. In the second study, the authors identified three markers of oxida- tive stress that might be more reliable and earlier indicators of toxicity than measures of secondary cell membrane damage. These studies do not demon- strate the performance of these in vitro proteomic methods for actual screening purposes, but they do suggest that in vitro proteomic methods have potential for classifying responses to toxic drug candidates and chemicals. The Consortium for Metabonomic Toxicology published a proof-of- concept paper demonstrating the use of metabonomic profiles for drug and chemical screening (Lindon et al. 2005b). The authors reported using a set of 147 compounds to develop an expert system for predicting toxicity based on nuclear magnetic resonance (NMR) spectra from urine specimens. The system correctly predicted histopathologic organ changes in approximately 50% of cases; in 5% of cases, the system made incorrect predictions. The remaining samples did not display conclusive histopathologic changes, metabolites ob- scured critical portions of the NMR spectrum, or overlapping patterns of toxicity precluded accurate prediction. POTENTIAL APPLICATIONS Use of Toxicogenomics in Existing Chemical Screening Programs Within larger chemical screening initiatives, toxicogenomic technologies could enhance screening-level determinations of basic mode of action. Two such applications involve the High Production Volume (HPV) Chemical Challenge 

84 Applications of Toxicogenomic Technologies Program and the initial evaluation of premanufacturing notices (PMNs) for new chemicals under the TSCA. Categorizing HPV Chemicals The EPA HPV Chemical Challenge Program is designed to make basic toxicity data publicly available on approximately 2,800 HPV chemicals or chemicals manufactured in quantities greater than 1 million pounds. The pro- gram began in 1998 as a result of a collaborative effort between the EPA Office of Pollution Prevention and Toxics, Environmental Defense, and the American Chemistry Council. To date, chemical manufacturers and importers have com- mitted to providing data for more than 2,200 chemicals to fill in any gaps in the SIDS battery (for further details, see EPA 2006a). More than 400 companies and 100 consortia have sponsored 1,371 chemicals directly in the program. An addi- tional 851 chemicals have been sponsored indirectly in an international counter- part to the HPV Challenge Program, the International Council of Chemical As- sociations HPV Initiative. Since the inception of the HPV Challenge Program, 365 robust summaries have provided data on approximately 1,328 chemicals. To obtain information on as many high production chemicals as possible, the HPV program allows chemicals to be grouped into “categories.” In theory, chemicals within a category share common modes of action, allowing testing results to be extrapolated from a few representative members of the category for various end points to the other members of the category. At present, there are few objective data to justify category classification. In vivo studies of acute tox- icity and 28-day repeat dosing are the most common traditional toxicology tests used to assert the existence of categories in the HPV program, but these tests are done on only a small subset of chemicals in the category. Submitting all mem- bers of a category to a short-term study using transcriptome, proteome, or me- tabonome analysis could allow more rigorous confirmation of categories and provide a mechanism to identify potential outliers. However, assays limited in dose range and time course could not be expected to exclude the possibility of differential toxicity within a category and toxicogenomic data would not be ex- pected to be the sole source of information for category validation. As databases grow and the quality of the interpretative methods improves, gene and protein assays may ultimately support quantitative interpolation within categories as well as provide mechanistic insight to inform further testing protocols. Enhancing New Chemical Evaluations Under TSCA Under TSCA, EPA requires chemical companies planning to bring a new product to market to file a PMN. The EPA then has 90 days to determine whether to allow the chemical to proceed to commercialization or to require additional testing or limitations on production based on evidence of toxicity (for additional information, see EPA 2006b). Because TSCA does not require any 

Application to Hazard Screening 85 specific toxicity tests be performed before PMN submission, 85% of PMNs con- tain no health data and EPA must make a determination based on consideration of structure-activity relationships in most cases. Categorization based on chemical structures eventually could be combined with transcriptome or proteome profiling. PMNs might then contain toxicoge- nomic data that show different, nontoxic profiles compared with other chemicals of similar structures. Alternatively, EPA scientists may request rapid toxicoge- nomic assays of new chemicals and use the profiles generated to make more rigorous determinations of safety. Additional database development and demon- stration of predictive accuracy would be required before true toxicogenomic screens would be useful in this setting. Use of Toxicogenomics to Screen for Hazard in Future Chemical Programs Current toxicogenomic assays are not sufficiently developed to replace more traditional toxicologic screening tests, but there are important opportuni- ties to build the familiarity and databases needed to inform future toxicogenomic screening regimens by adding toxicogenomic tests to existing chemical pro- grams. For example, The National Toxicology Program uses established meth- odologies to test relatively large numbers of animals, which is both costly and time-consuming. However, the chronic exposure experiments provide a rare source of biologic material that could be used to advance toxicogenomic appli- cations. Strategically adding appropriate toxicogenomic assays to these testing protocols will generate novel, phenotypically correlated data that would be es- pecially useful in building databases to aid in predictive toxicology. Such data will be suited for helping determine whether short-term in vivo studies using toxicogenomic technologies can substitute for chronic bioassays. Similar opportunities exist outside the United States. Within OECD screening programs, toxicogenomic assays could be added to the basic SIDS dataset. In the European Union, a new approach to chemical regulation known by the acronym REACH (Registration Evaluation and Authorization of Chemi- cals) will require hazard data generation as a condition of continued marketing of chemicals, with the amount of data dependent on the volume of the chemical produced. Although the program relies on traditional toxicology assays, the in- crease in data-generation activities expected under REACH presents another opportunity to expand the datasets and acquire more experience in large-scale applications of toxicogenomic assays. Future Use in the Pharmaceutical Industry Whether the pharmaceutical industry will expand the use of toxicoge- nomic technologies to detect potential toxicity in lead candidates is an open question. For in vivo studies, toxicogenomics may allow ongoing noninvasive sampling (especially for proteomic or metabonomic end points) and provide 

86 Applications of Toxicogenomic Technologies useful mechanistic information. For certain end points (as demonstrated cur- rently with different types of hepatotoxicity), reliance on toxicogenomic data to predict potential hazard will be possible. Several companies appear to be working on using in vitro toxicogenomic assays to evaluate multiple cellular effects. Surrogate screens using panels of genes in in vitro systems may be deployed rather cheaply and produce large amounts of data quickly. The critical question for the industry is whether adding toxicogenomic technologies to more standard toxicology testing or other in vitro methods will result in overall cost savings. To date, the economic value of toxi- cogenomic testing in drug development has not been demonstrated through rig- orous studies. In the absence of such data, it is likely that pharmaceutical com- panies will make individual judgments about incorporating toxicogenomic assays. Regardless of the degree to which pharmaceutical companies use toxico- genomic screens to identify possible hazards, for the foreseeable future full in- terpretation of risk will be done in the context of additional, required testing information. CHALLENGES AND LIMITATIONS For toxicogenomic technologies to be widely applied to screen compounds for toxicity, it will be essential to demonstrate that toxicogenomic-based screens are sufficiently robust and comprehensive for a wide range of compounds and toxic end points. There are several aspects of this generalizability: • Because the screening of environmental chemicals has different re- quirements than the screening of candidate drugs, algorithms designed for screening candidate drugs will need to be modified and validated for environ- mental chemicals. • The database of toxicogenomic and traditional toxicologic data used for training and validating screening algorithms must be extensive and comprehen- sive enough to allow accurate predictions for a wide range of chemicals. • Costs of toxicogenomic assays must be compatible with implementa- tion in screening. • Throughput of toxicogenomic assays must meet or exceed norms for currently used screens. Each of these factors is described in more detail below. Contrasts Between Screening Needs and Methods for Pharmaceutical Candidates and Industrial Chemicals Pharmaceutical companies and chemical companies face different regula- tory and business requirements that lead to different requirements in screening their products. Drug companies need rapid, inexpensive ways to screen candi- 

Application to Hazard Screening 87 date compounds for toxicity that avoid expenditures on compounds destined to fail at later stages of development. Chemical companies need rapid, inexpensive ways to screen new chemicals and many existing chemicals for toxicity. In both cases, the ability to gain useful information from short-term tests makes toxico- genomic techniques attractive for screening applications. Although the screening needs of pharmaceutical and chemical companies may appear similar, the technical requirements of screening in these two settings are quite different. These differences include dose ranges that must be consid- ered, types of toxicity that must be evaluated, and the degree of sensitivity and specificity required. Evaluation of drug toxicity focuses on therapeutic dose ranges. In contrast, exposures to industrial chemicals can range over many or- ders of magnitude depending on the setting, requiring toxicologic evaluation over a broader dose range. A second major difference between screening candidate drugs and screen- ing chemicals is that candidate drugs undergo a rigorous set of in vivo tests be- fore they are marketed. In contrast, screening tests are likely to be the only sources of new data generated in the assessment of toxicity that occurs for most industrial chemicals before widespread marketing. The redundancy of animal testing after toxicogenomic prescreening for candidate drugs means that the pre- screening test can have relatively low sensitivity and still be valuable. Because the animal tests are believed to be both more sensitive and specific for toxicity, the toxicogenomic screening tests could have a relatively high false-negative rate and still identify enough potentially toxic candidate drugs (thus avoiding more costly animal testing) to be economically useful. Importantly, the animal tests would serve as a backup for identifying those toxic candidate drugs that have been missed by the toxicogenomic screens. In contrast, the absence of any statutory requirement for animal testing for industrial chemicals other than pes- ticides in the United States means that there is no backstop to the screening tests. A third major difference is the breadth needed. A set of toxicogenomic screens limited to liver toxicity would be of relatively little value, as liver toxic- ity is not a primary outcome for exposures to most industrial chemicals. Toxico- genomic screens for industrial chemicals would need to assess a broader array of toxicities, including cancer, developmental and reproductive effects, neurotoxic- ity, and immunotoxicity. Whereas any reliable screening information is of great public health value, care must be taken not to assume that toxicogenomic screens for limited types of toxicity are sufficient to conclude that a screened chemical has no significant hazard associated with it. Ideally, for the screening tests to fully protect public health, they should be highly sensitive, detecting the vast majority of chemicals with potential toxicity. Moreover, they should be able to predict not just acute toxicity but also chronic toxicity; they should also be able to predict toxicity at different developmental stages, because environmental exposures occur throughout the life span. This requires new knowledge of early gene, protein, and metabolite markers of chronic effects so that changes indicative of short- term acute toxicity can be distinguished from changes indicative of more latent 

88 Applications of Toxicogenomic Technologies or chronic effects. Use of more limited toxicogenomic assays must be accompa- nied by recognition that the screening is only partial. From the industry stand- point, for the screening tests to be economically feasible, they must also be highly specific and avoid falsely identifying benign chemicals as potentially toxic. To be acceptable to public health and industry stakeholders, toxicoge- nomic screening assays for industrial chemicals need to demonstrate greater accuracy than traditional tests as well as the ability to detect a broader array of outcomes than they currently do. Database Size and Quality A second challenge to the development of toxicogenomic approaches to screening is the availability of sufficiently robust and comprehensive data col- lected into a database. Many of the issues of database quality and comprehen- siveness are relevant to predictive toxicology in general and are discussed in detail in Chapter 10. Issues of specific relevance to screening applications in- clude comprehensiveness of end points and modes of action assessed. As dis- cussed elsewhere in this chapter, to be useful for the broad screening of envi- ronmental chemicals for many potential effects, toxicogenomic assays must be able to signal the broad array of end points of concern, including cancer, repro- ductive and developmental toxicity, neurotoxicity, and immunotoxicity. Datasets are used only to classify compounds that act via those modes of action that are represented in their training sets of model compounds. False-negative rates from incomplete datasets may be acceptable in some settings but not in stand-alone screening programs for environmental chemicals. In addition to being informative about a useful range of end points, the data need to include information useful in predicting responses at relevant doses and exposure timing. This is because responses to chemicals can involve multi- ple modes of action, depending on dose and time course of exposure. A popula- tion of databases with data on dose dependence and timing dependence of well- characterized compounds is required to validate the use of toxicogenomic assays for environmental chemical screening. Reducing Costs of Toxicogenomics Screening Tests A basic requirement of an effective screening test is that its cost is low enough to allow widespread application. At present, the cost of toxicogenomic technologies is expected to present a significant barrier to their widespread ap- plication in screening programs, especially for environmental chemicals. Toxi- cogenomic methods are currently being used primarily as adjuncts to standard animal tests. Whereas this may result in selective use of expensive animal as- says, the actual costs of screening tests using toxicogenomic techniques in this manner include costs of the short-term animal assays plus the additional costs of toxicogenomic analyses. If toxicogenomic screens with model organisms such 

Application to Hazard Screening 89 as Caenorhabditis elegans can replace rodent assays, then cost will become less of a barrier to using toxicogenomic approaches as screening tools. Achieving High Throughput The pharmaceutical and biotechnology industries have developed highly automated assays to screen numerous chemicals for relatively simple properties, such as target receptor binding (Walum et al. 2005). Similar assays are being developed to assess toxicity; they range from receptor binding assays linked to toxicity mechanisms to automated cytotoxicity assays (Walum et al. 2005). Such assays are capable of assessing tens or hundreds of thousands of compounds per week. Current toxicogenomic assays are “high throughput” only in the sense that they analyze thousands of genes or proteins simultaneously. To the extent that they can predict outcomes of standard animal tests in a shorter time frame, toxi- cogenomic technologies allow faster screening of compounds. In vitro screening assays using toxicogenomic techniques, combined with reproducible, validated automation techniques will be needed to truly achieve high-throughput toxico- genomic-based screening. CONCLUSIONS Toxicogenomic technologies can facilitate the screening of chemical com- pounds for their ability to cause toxicity. In vivo and in vitro screens currently are used to detect well-characterized toxicities (hepatotoxicity, nephrotoxicity) in settings where false-negative results can be tolerated, such as drug develop- ment. The use of toxicogenomic technologies in the validation of chemical cate- gories for screening programs that also employ traditional toxicologic assays is quite feasible in the near term. However, broad application of toxicogenomic technologies to screen environmental compounds will require demonstration of false-negative rates comparable or superior to those of traditional testing meth- ods. For using toxicogenomic assays to screen for a battery of end points, reli- able assays and their associated algorithms for a broader set of end points will need to be developed. These end points include those toxicities of greater inter- est for environmental chemicals (for example, cancer, neurodevelopmental tox- icity) than are currently used in screening candidate drugs (for example, hepato- toxicity, nephrotoxicity). Although it seems unlikely the full range of toxicities assessed in traditional animal screening studies could soon be assessed by toxi- cogenomic screening, the availability of assays for specific end points (for ex- ample, cancer) would help alleviate the dearth of publicly available information on the potential toxicity of many chemical agents. The pharmaceutical industry has developed the most advanced toxicoge- nomic screening applications. This reflects incentives to screen out undesirable properties and more efficiently identify drug candidates with the safest and most efficacious profiles. The business case for using toxicogenomic technologies 

90 Applications of Toxicogenomic Technologies over other testing methods for pharmaceuticals has not yet been clearly demon- strated, but even the touted economic advantage of using toxicogenomic screen- ing to avoid more detailed and expensive testing of pharmaceuticals does not exist for industrial chemicals. This is because regulatory requirements for gener- ating toxicity data for industrial chemicals under the TSCA are far less stringent than those for drugs. Toxicogenomic technologies for hazard screening have relied primarily on transcriptome profiling assays. New studies have shown the usefulness of pro- teomics and metabolomics for classifying biologic activity. Because protein and metabolite changes may be more closely linked to pathologic changes, they may ultimately prove to be valuable screening indicators of toxicity. Additional proof of concept research is required to assess their value. Of the remaining critical challenges to implementing toxicogenomic tech- nologies for screening, most fundamental is the development of comprehensive databases suitable for screening purposes. These databases should include toxi- cologic end points relevant to environmental chemicals and examples of all known toxic modes of action. RECOMMENDATIONS Intermediate 1. Convene an expert panel to provide recommendations for which model compounds, laboratory platforms, and specific data elements are necessary for building toxicogenomic databases for screening applications. Assessment of in vitro approaches for specific toxic end points should be emphasized. All proc- esses, toxicogenomic data, and outcome data must be publicly accessible. 2. Expand the toxicogenomic component in existing testing programs such as the National Toxicology Program. The toxicogenomic data, collected in conjunction with the highest quality traditional toxicology testing, can then be used to help build public databases and increase understanding of the role of toxicogenomics in predicting chronic toxicity. 3. Develop databases and algorithms for using proteomic and me- tabonomic data in screening. 4. Regulatory agencies (including EPA and the Food and Drug Admini- stration) should continue to develop and refine guidance documents for their staff on interpreting toxicogenomic data. In particular, guidance for environ- mental chemicals must ensure that screening protocols address the types of end points most relevant for the general population, including sensitive subpopula- tions. This effort will help define database, basic research, and training needs for the agencies. 5. Develop mechanisms to improve the quantity and quality of data avail- able for deriving screening profiles: 

Application to Hazard Screening 91 a. Establish a dialog with entities holding currently inaccessible toxicogenomic data to evaluate options for increasing the availability of data. b. Regulatory government agencies (including EPA and the Food and Drug Administration) should consider appropriate ways to address the following disincentives to industry generation and public submission of data that could be used to populate pub- lic databases: additional costs of testing; concerns about re- porting requirements of the TSCA and the Federal Insecticide, Fungicide, and Rodenticide Act; concerns about competitors’ use of toxicity data if they are made public; and concerns about the use of toxicogenomic data in tort liability cases. c. Integrate data relevant to toxicogenomic screening of envi- ronmental chemicals into ongoing biomedical initiatives such as the National Institutes of Health Molecular Libraries initia- tive. Such data may include physicochemical characteristics, in vitro assay results such as cytotoxicity and receptor binding, and other screening-level types of data. Long Term 6. Ensure that the regulatory framework provides incentives, or at least removes disincentives, for premarket testing of chemicals. 7. Upon validation and development of adequate databases, integrate toxi- cogenomic screening methods into relevant current and future chemical regula- tory and safety programs.