Chapter 1: Introduction to Predictive Toxicogenomics for Carcinogenicity

The term “toxicogenomics” (TGx) has been applied to the combined technologies of transcriptomics, proteomics, metabolomics, and epigenomics, as they are used in the field of toxicology to study the expression of genes, proteins, metabolites, and epigenetic modifications, respectively.¹  The goal of this chapter is to place the TGx of predictive carcinogenicity in perspective, with both the relevant conventional toxicological methodologies that have come before genomics and those that coexist at present and quite likely will be used together in the future to study relevant biological processes, including pathways of toxicity and chemical modes of action.

Transcriptomics (commonly referred to as “gene-expression profiling” or “transcript profiling”) involves measuring the relative abundance of potentially thousands of RNA transcripts present in a sample from one tissue extract at one point in time and at a single dose of a chemical. Transcriptomics is typically performed using microarray technology. Microarrays have advantages in speed and ease of sample preparation, low per-sample cost, and well-established protocols and methods for data analysis and data normalization. Microarrays are available for transcriptomics (e.g., messenger (m) ribonucleic acids (RNAs), micro (mi)RNAs and long non-coding (lnc)RNAs) as well as for epigenomics (e.g., DNA methylation microarrays).

Whole-transcriptome next-generation sequencing (RNA-seq) offers an alternative method for estimating transcript abundance in gene expression studies. RNA-seq has the potential to overcome many of the limitations associated with microarrays as it does not rely on predetermined probe sequences for expression measurements, and it is based on simple counting of reads that can be reliably aligned to a reference sequence.²  The use of molecular expression analytical technologies, including microarray or RNA-seq, in TGx is based on the assumption that chemicals exhibiting similar types and degrees of toxicity will induce similar profiles of gene, protein, or metabolite expression. When used in conjunction with conventional toxicity testing methods and/or histopathology, to provide context, these snapshots or images of molecular expression can be assembled sequentially into a series of high-resolution profiles of adaptive and adverse effects, including causal events that occur as a function of dose and time and may reflect particular chemical modes of action.

Another set of analytical technologies, high-throughput screening (HTS) methods, were developed in the mid-1980s in the pharmaceutical industry to rapidly screen libraries of candidate small molecules or drugs for specific types of biological activity or disease processes. This set of analytical technologies has been applied in multi-endpoint toxicity screening for the past 10 years.³  A consortium of US government agencies, referred to as Tox21,⁴  consisting initially of the US Environmental Protection Agency (EPA) ToxCast program,⁴  the National Toxicology Program (NTP), the National Institutes of Health Chemical Genomics Center (NCGC), and later the US Food and Drug Administration (FDA), has used in vitro assays in robotic HTS to study chemical perturbations of biological pathways which can become pathways of toxicity. Chemical–protein interactions have been measured and bioinformatics has been used to delineate the toxic response by making use of known links between genes, proteins, and diseases.⁵  Although some predictive patterns or “signatures” of toxicity have been identified, a comprehensive chemical toxicity predictive capability has been elusive and HTS remains at a developmental stage for many toxicological outcomes.

Toxicogenomics, which originated in name in 1999,⁶  also has been slow to achieve its anticipated predictive and mechanistic analytical capabilities. While microarray and RNA-seq based TGx methods clearly provide useful information about how biological systems respond to chemical compounds, currently they are too expensive to serve in a screening role. Combining HTS and TGx technologies can provide the opportunity both to screen for perturbations in biological, toxicological, and disease pathways and to carefully investigate multiple potential modes of chemical action. Each technology therefore can help to define the dose–response behavior of drugs and chemicals, and, when used together, they ultimately should be able to accurately predict toxicological and disease outcomes and likely modes of action.

The fundamental question to be addressed in this introductory chapter is how best to leverage data and information coming from these technologies to develop integrated strategies for the evaluation of chemicals for potential human health risks. This question may be answered by better understanding the goals and challenges of TGx and HTS.

The stated goal of EPA's ToxCast program³  is to screen large numbers of environmental chemicals using in vitro HTS methods and to prioritize them for further testing based on predictive scores for disease-related signatures and corresponding estimates of human exposure. In the longer term, however, the goal of ToxCast is to improve the efficiency and reliability of current toxicity test methods. This could be accomplished by using both HTS and TGx technologies in in vitro assays to delineate putative modes or mechanisms of action relevant to in vivo chemical toxicity. There are several challenges to be overcome before such a goal can be realized, and its current status is reflected in discussions throughout this book.

Chapter 2 by Li provides details on the evolution and current status of in vitro screening technologies with an emphasis on biomarkers or signatures for that can be further developed for HTS applications in drug safety assessment. There have been several transcriptomic signatures identified for assessing genotoxicity, but the one (TGx-28.65) developed by Li (chapter 2), and also discussed in chapter 3 by Buick and Yauk, shows convincing inter- and intra-laboratory reproducibility and performs robustly and consistently on different transcriptomic platforms.

A common criticism of the in vitro to in vivo prediction approach is that a chemical's toxicity may depend on unique properties of intact tissues and organisms that are not exhibited by genes or cells in vitro. In addition, biotransformation of parent compounds into metabolites that are more or less active than the parent must be considered in each assay or model system. Development of tests in which realistic levels and kinds of biotransformation occur in vitro, with fidelity to what occurs in vivo, is a challenge. Furthermore, understanding the correlation between in vitro lowest-effective-dose values and corresponding chemical concentrations in blood or tissues is crucial in extending either HTS or in vitro TGx studies to inform quantitative assessments of health risk.

In principle, challenges such as those described may be addressed by performing parallel rodent TGx in vitro and in vivo studies with human in vitro assays and by confirming the findings with human clinical and epidemiological studies (i.e., using a parallelogram method).⁷  This approach proposed originally by Sobels^8,9  has the advantage of simultaneously addressing the question of rodent-to-human extrapolation and the issue of internal-tissue dosimetry. Several published TGx investigations appear to validate this thinking, as discussed in chapter 13 by Kienhuis et al. Information garnered from parallelogram-based TGx investigations can benefit both HTS and TGx in the development of quantitative data for use in risk assessment. Furthermore, comparative genomics approaches have improved and extended the parallelogram approach to a variety of species (see chapter 7 by Schaap et al. and chapter 13 by Kienhuis et al.).

Another challenge is how to determine what short-timescale (hours to days) in vitro assays can reveal about long-timescale (months to years) processes that lead to in vivo chronic toxicity outcomes such as cancer. This challenge has been addressed by applying TGx together with (within) standard toxicological and chronic bioassays such that details of changes in molecular expression occurring early in the time course of chronic disease can be defined and used to predict of chronic disease outcomes. Alternatively, technology now exists to retrospectively examine formalin-fixed paraffin-embedded archival tissues to discover the profiles of gene expression that correspond to histopathological phenotypes associated with various disease processes and outcomes. The NTP archives contain tissue blocks from cancer bioassays on hundreds of compounds that could be studied in this way. This approach was successfully applied in TGx studies on furan,¹⁰  as discussed by Webster, et al. in chapter 12.

One further issue relates to the relative priority being given to HTS screening and TGx analysis of the various chemicals of regulatory concern. Although HTS aims to predict the potential for chemicals to affect human health based on results of in vitro tests, essentially all of the current in vivo data being used to develop prediction models are from high-dose animal testing using conventional methods (as discussed in the following section). For future HTS efforts, priority should be given to chemicals to which humans are exposed where there is evidence of bioaccumulation (i.e., body burdens). Such exposure estimates are routinely developed by the EPA, and blood and tissue analyses are performed by the US Agency for Toxic Substances and Disease Registry. If HTS shows that chemicals activate toxicity pathways at concentrations similar to those found in human biomonitoring studies, they should be given higher priority for more detailed testing using TGx together with conventional test methods.

If challenges such as those previously described can be met, the HTS approach in toxicity testing can become a practical solution in evaluating the backlog of thousands of untested environmental chemicals. Furthermore, the identification of molecular signatures and toxicity pathways with TGx can be used to develop predictive toxicological methods and essential information to substantiate putative modes of action in risk assessment. While it is generally acknowledged that TGx approaches cannot yet replace conventional toxicity tests, data derived through the use of TGx studies within the framework of conventional testing have enabled a more predictive toxicology, provided a better understanding of the mode of action of drugs and chemicals,^11,12  and helped to improve hazard and risk assessment, as discussed by Auerbach, Thomas and Waters, and Guyton and Waters (in chapters 4, 5, and 9, respectively). In subsequent chapters, that which has been accomplished in this regard for several well-studied model rodent carcinogens, including benzene, the conazoles, furan, and acetaminophen (paracetamol), is reviewed by Nesnow, McHale et al., Webster et al., and Kienhuis et al. (chapters 10–13, respectively). Our plan is to work through these prototypic compounds to understand how gene response reflects causal events on pathways of toxicity and disease. This approach is consistent with current regulatory thinking and can facilitate the eventual integration of HTS and TGx data for predictive purposes. When causal events are linked to toxicity and disease pathways, as a function of dose–response, they provide regulatory agencies with essential data with which to determine under what conditions environmental exposures pose quantitative risks to human health.

The US National Research Council report Toxicity Testing in the 21st Century: A Vision and a Strategy¹³  overviewed by Andersen and Krewski¹⁴  proposed that new genomic technologies could both dramatically increase the number of chemicals comprehensively evaluated for toxicity and broaden and improve the human relevance of toxicity endpoints assessed. The use of genomic technologies in toxicity testing holds great promise for informing and significantly improving health risk assessments for a broad range of environmental chemicals. The sections below trace developments in the field of toxicity testing that have led to where we are today. In addition, greater definition of the subjects that will be covered in subsequent chapters of the book is provided.

Since the 1950s conventional toxicity testing involving chemical, physical, or biological agents has relied on high-dose studies in laboratory animals to assess potential human toxicity. The results of such studies are extrapolated to environmental or occupational exposure levels to predict human health outcomes. When this approach was developed, little was known about the modes of action by which chemicals caused toxicological responses in animals. Since that time, in vivo toxicity tests have been standardized and in vitro tests have been developed to characterize toxic effects at the cellular level (e.g., cytotoxicity). As our understanding of the genetic basis of toxicity (in particular, carcinogenicity) has grown, a battery of in vivo and in vitro tests to detect mutations and DNA damage has also been developed and standardized.^15–19 

Conventional toxicity testing currently involves using in vivo and in vitro tests to screen chemicals for defined toxic effects (e.g., neurotoxicity, developmental toxicity, and carcinogenicity) or modes of action relevant to a specific outcome (e.g., mutagenicity, cytotoxicity, and regenerative cell proliferation relevant to genotoxic carcinogenicity). Intact-animal models are the only surrogates for human clinical studies that offer reasonable biochemical, metabolic, and physiological fidelity to the human response. In vitro test systems can represent investigator-selected biological processes with various levels of fidelity under carefully controlled experimental conditions.

To profile the toxicity of a chemical in standard in vivo assays, the chemical typically is tested in rodents for acute toxicity (following a single or short-term exposure); subchronic toxicity (following exposure for 14–90 days); chronic toxicity (including carcinogenicity) following long-term exposure (up to 2 years); eye and skin irritation; reproductive and developmental toxicity (in breeding studies); hypersensitivity (immune response); and, sometimes, phototoxicity (toxicity in response to light). Toxicokinetic studies also are conducted to characterize absorption, distribution, metabolism, and excretion of the chemical following in vivo exposure.

Toxicology studies generally use a series of dose levels, ranging from doses where no effects are expected to doses where frank clinical or histopathologic changes are expected. Typically, the highest dose at which no overt toxicity occurs in a 90-day subchronic toxicity study (the “maximum tolerated dose”) is used to establish dose levels for long-term bioassays, which measure endpoints such as clinical signs of toxicity, body- and organ-weight changes, clinical chemistry, and histopathological responses, and which provide insight into potential latent effects, including cancer, reproductive or developmental toxicity, or immunotoxicity.

The relationship between exposure to toxicants and adverse health effects is also examined through epidemiological studies. These studies attempt to assess the relationship between exposure levels and the likelihood of adverse effects (“exposure–response relationships”) by comparing estimated exposure levels and disease distributions in human populations. However, exposure can be difficult to estimate, and such studies are complicated by confounding factors such as co-exposure to other potential toxicants such as cigarette smoke or pesticides.

Most of our understanding of chemical toxicity has come from data obtained through conventional in vivo testing. However, in vivo toxicity testing is costly and time consuming, requiring large numbers of animals. A 2-year rodent study requires the use of more than 800 mice and rats and the histopathological examination of more than 40 tissues. Full characterization of the potentially hazardous properties of the many thousands of environmental chemicals through conventional toxicity testing is thus severely constrained by financial and other resource limitations. In practice, a tier-based testing framework generally has been used for evaluation of environmental chemicals, whereby the choice of specific tests for a given chemical is guided by observations of toxicity in early-stage tests and information on potential human exposure (routes and levels).

Another limitation of conventional toxicity testing is that even extensive animal testing does not fully explain the mechanisms of toxicity in humans. Detailed knowledge concerning risk to humans is still inadequate for many chemicals. The need for more mechanistic data and a “theoretical framework for rational decision making” was identified in the early 1980s.²⁰ 

The discovery in 1944 that genetic information was transmitted by deoxyribonucleic acid (DNA), the discovery in 1953 of the structure of DNA, and the cracking of the genetic code in 1966—showing how the sequence of chemical bases in DNA codes for the synthesis of proteins—opened the door to an explosion of knowledge in molecular biology. In the nucleus of the cell, a gene's DNA serves as a template for synthesis of mRNA; mRNA then travels from the nucleus to the cytoplasm and serves, in turn, as a template for the synthesis of a protein molecule. Proteins are involved in virtually every biological function, including cell growth, differentiation, and death, regulation and coordination of physiological processes, metabolism, immune function, and disease processes. Gene expression or transcription refers to the construction of mRNA molecules (transcripts) from the template provided by a specific gene. Knowledge of how gene expression is controlled and coordinated, as well as the roles of specific proteins in cellular functions and in disease processes, has informed our understanding of the modes and mechanisms of action by which chemicals produce toxic effects in cells and whole organisms.

The genome is the entire complement of genetic information found in each cell of an organism, and genomics refers to the analysis of the entire genome of an organism (or cell or tissue). The introduction of high-throughput dideoxynucleotide DNA sequencing in 1977²¹  made it feasible to determine the sequences of entire genomes, ushering in the era of genomics. This technology was used in the Human Genome Project, which was proposed in the late 1980s and begun in 1990. The reference human genome sequence—consisting of about 23 000 genes that code for proteins, made up of 3 billion chemical base pairs—was made available in 2000.²²  Since then, high-throughput genome sequencing has become faster and less expensive.²³  By 2010 a whole human genome could be sequenced on the Illumina platform for less than $14 500²⁴  and today the basic cost approaches one-tenth that amount.

To capitalize on the enormous potential of genome-wide DNA sequence information, new molecular technologies and bioinformatics tools (to be discussed later) have been developed that make it possible to generate and analyze biological datasets of unprecedented magnitude and complexity.¹³  Beyond DNA-sequence analysis, postgenomic technologies enable analysis of gene expression or the production of mRNA transcripts (transcriptomics); the production of the proteins themselves (proteomics); the endogenous production of metabolites (metabolomics); and the transmissible epigenetic modifications of chromatin and DNA (epigenomics).

Simply stated, TGx is the application of genomic technologies to the study of adverse effects of toxicants. It aims to study the response of the entire genome to toxicants or environmental stressors. In 2002, scientists at the National Institute of Environmental Health Sciences (NIEHS) National Center for Toxicogenomics (NCT) proposed that patterns of induced gene expression changes are characteristic of specific classes of toxic compounds and that their distinctive molecular expression fingerprints can help in classifying agents with different mechanisms of action.^25,26  Since then, TGx technologies have made it possible to use genotypes and toxicant-induced gene expression, protein, and metabolite profiles to screen chemicals in hazard identification, to monitor exposure, to measure dose–response curves at the cellular level, to elucidate mechanisms of action, and to predict individual variability in sensitivity.

Whereas genomes are relatively static, transcriptomes, proteomes, metabolomes, and epigenomes are dynamic—displaying moment-to-moment changes in response to diet, stress, disease processes, and exposure to toxicants and stressors—and their analysis must be linked to the state and condition of the biological system under investigation. Toxicity coincides with changes in specific mRNAs, proteins, metabolites, and epigenetic modifications. These changes observed under defined conditions of cellular location, dose level, time, and biological context can provide meaningful information about biological responses to toxic insult. Typically, adverse effects can be detected at subcellular and molecular levels at time points before they are manifested at the level of tissues, organs, or the whole organism. Toxicant-specific alterations in gene expression, protein synthesis, and metabolite production correspond with observable or phenotypic responses of cells, tissues, and organisms. The process of relating molecular expression data to toxicity and pathology observed in conventional toxicology tests and using histopathological evaluation is referred to as “phenotypic anchoring”.^27,28  This subject is covered in greater detail later.

Transcriptomics, the measurement of the relative abundance of mRNA transcripts in TGx investigations, has typically been measured using microarrays—glass slides or silicon chips on which arrays of thousands of DNA (copy (c)DNA) probes are attached, which react (hybridize) with RNA transcripts in the biological sample. Hybridization is detected by labeling of the RNA with fluorescent dye, and the intensity of fluorescence indicates the level of gene expression. The results of microarray assays have been shown to correlate well with measurements of the expression of single genes by methods such as quantitative real-time (qRT) polymerase chain reaction (PCR, a technology that “amplifies” DNA segments, creating large numbers of copies to facilitate analysis).

Modern microarray technology uses oligonucleotide probes (shorter DNA sequences), allowing over a million probes per chip.^29–31  Microarrays corresponding to essentially all known human genes and representing the genomes of animal models used in conventional toxicology are available from commercial vendors (such as Affymetrix, Agilent, and Applied Biosystems). However, only those mRNAs, miRNAs or lncRNAs for which there is a complementary sequence on the microarray can be hybridized and measured in a given experiment. As discussed earlier, microarrays have advantages in speed and ease of sample preparation, low per-sample cost, and well-established protocols and methods for data analysis and data normalization. However, microarrays also have limitations, including a finite dynamic range due to the limitations of fluorescence and saturation of binding sites within the probe sets printed on the microarray. At low levels of expression, microarrays often have poor resolution of transcript binding due to high background effects and nonspecific binding.

RNA-seq offers an alternative method for estimating transcript abundance and has the potential to overcome many of the limitations associated with microarrays: it does not rely on predetermined probe sequences for expression measurements and is based on simple counting of reads that can be reliably aligned to a reference sequence. In providing count data, RNA-Seq has effectively no limit to the dynamic range of signal detection, and, in theory, can provide a higher degree of accuracy and precision in estimating relative expression levels.²  However, RNA-seq data also has potential challenges which remain less well explored. Reliable quantification of expression levels appears highly dependent on read depth, and low transcript abundances are characterized by high variance. Thus, the effective or useful lower limit of the dynamic range of RNA-seq is not clear. Additionally, methods for normalization and statistical analysis of RNA-seq data are less mature, and no established best practices exist for RNA-seq data analyses.² 

Proteomics is the study of collections of proteins in living systems (i.e., the complete set of proteins found in a particular cell type at a given time). Because a given protein may exist in several different forms, proteomes are more complex than the genomes and transcriptomes that code for them. Exposure to toxicants can result in two types of proteome changes: (1) changes in protein levels due to changes in gene expression, mRNA stability, protein stability, or some combination of these; and (2) changes in the relative levels of two or more modified forms of a protein, which may be more critical to function than the absolute protein levels.³² 

Proteomics relies on a combination of technologies to separate and identify proteins and measure their levels. Because there is no technology analogous to PCR that can amplify proteins, they must be analyzed at their native concentrations, which span more than six orders of magnitude. As a result, it is difficult to detect less-abundant proteins in complex mixtures. Another challenge to proteome analysis is that proteins may be damaged by reactive chemical intermediates produced from toxic chemicals or as a result of oxidative stress. Most analytical proteomic methods are not truly parallel, but involve elaborate serial analyses. Two major approaches are gel-based proteomics and “shotgun” proteomics.

In gel-based proteomics, proteins are separated by electrophoresis (e.g., two-dimensional (2D)) sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS-PAGE), and selected proteins are identified by digestion to peptides, mass spectrometry (MS) analysis, and database searching. Comparative 2D SDS-PAGE with differential fluorescent labeling (for example, differential gel electrophoresis) allows effective quantitative comparisons of proteomes.^33,34  Modified and unmodified forms of proteins often can be resolved for separate characterization and quantitative analysis. Although 2D gel electrophoresis has been used mostly for global analyses of complex proteomes, the method is also useful for comparative analyses of smaller subproteomes.

In shotgun proteomic analysis, protein mixtures are digested to form complex mixtures of peptides, which are analyzed by liquid-chromatography-coupled MS.³⁵  Databases are then searched to match the resulting peptide tandem mass spectrometry spectra with corresponding peptide sequences, and software is used to reassemble the collection of peptide sequences into proteins. Shotgun proteomics is the most effective technology for automated analysis of complex peptide mixtures.^36,37 

A key issue in proteomics is the standardization of data analysis methods formats for data representation and reporting. Another unresolved issue is the fact that different database search algorithms can yield different identifications of proteins and peptides.

Metabolomics is the study of collections of molecules (intermediates and products) generated through metabolic processes in an organism or cell at a given point in time. (Some authors distinguish between “metabolomics” and “metabonomics”; however, the distinction in the literature is not consistent, and this chapter uses the term “metabolomics”.) Metabolites reflect the actions of proteins in biochemical pathways; metabolomes thus represent biologic states (phenotypes) analogous to proteomes.

Metabolomic analysis relies mainly on nuclear magnetic resonance (NMR) spectroscopy and MS (both gas and liquid chromatography MS). NMR-based technology has been shown to be robust and reproducible in laboratories that follow similar analytical protocols,³⁸  and consensus standards for analytical standardization and data representation in metabonomic analyses have been agreed.³⁹  NMR is valuable for identifying patterns of spectra reflecting global metabolic changes, while MS-based analyses offer the advantage of greater sensitivity. Both technologies can detect differences in metabolic profiles that correspond to various modes of toxicity, but integration of the technologies will allow a more comprehensive approach.

In conventional toxicology, the term “biomarker” generally refers to biological indicators of exposure to specific toxicants. However, in TGx, the term “biomarker” has been used to refer to a gene-expression, protein, or metabolic profile that serves as an indicator of toxicity. Such a “signature” or “fingerprint” represents the results of complex interactions within the organism. If shown to be reproducible and correlated with a toxicity endpoint, the molecular signature potentially can serve as a predictive biomarker. Several review articles explore issues related to biomarker assay development and provide examples of the biomarker development process.^40–42  TGx biomarkers or gene signatures in vitro and in vivo are discussed extensively in chapters 2, 3, and 4 of this volume.

The magnitude and complexity of the data generated by genomic technologies, as well as the need to integrate genomic data with data from other disciplines, requires the use of advanced computational techniques. “Bioinformatics” is the branch of computational biology focused on the collection, management, analysis, and integration of numerical biological data and it is essential to genomic analyses. Bioinformatics encompasses the integration of data across genomic technologies, the integration of genomic data with data from other observations and measurements (including HTS data), and the integration of all these data in databases and related information resources. At a basic level, bioinformatics is represented by information resources such as GenBank, i.e., repositories of gene sequence data and associated information, structured for easy retrieval. At an intermediate level are tools, such as BLAST or SAGEmap, that perform insightful sequence alignment and function and structure analysis. Finally, sophisticated information systems (e.g., expert systems) integrate data from numerous sources to solve multifaceted problems. An excellent illustrated introduction to bioinformatics and various bioinformatics methods applicable to TGx and predictive carcinogenicity is provided by Bushel in chapter 14.

As previously discussed, microarray- or RNA-seq-based TGx methods provide useful information about how thousands of genes in biological systems respond to chemical compounds, but they are too expensive to serve as high-volume screening tests for the effects of thousands of chemicals in numerous cell types at a range of dose levels and time points. In contrast, individual HTS methods look at only one or a few genes at a time, but allow thousands of chemicals to be screened in one day over a wide range of concentrations. Compared with more extensive conventional toxicology and TGx tests, in vitro screening tests as used in the triage of molecular libraries for particular molecular or biological activities typically employ higher and fewer doses of the chemical, fewer test subjects, shorter observation periods, and less extensive evaluation of the toxic outcomes. HTS thus provides a practical method to investigate more than 100 000 compounds per day in miniaturized in vitro assays in order to identify those with the potential to cause adverse effects.⁴³  For safety evaluation and toxicity testing “hits” in the screening assays correspond to biological pathways that are known to lead to adverse outcomes. With sufficient accumulated data, it may be possible to use structure–activity analysis to predict HTS hits, so that potential targets can be predicted prior to screening. The application of robotic HTS as a useful complement to conventional toxicology has been expanding.^3,44–46 

HTS assays fall into two broad categories: biochemical assays and cell-based assays. Biochemical (cell-free) assays generally measure direct effects on specific molecular targets of interest. These assays have been used to measure enzymatic activity,^47–49  binding of substances to receptors,⁵⁰  ion-channel activity,⁵¹  nuclear receptor activity,⁵²  and protein–protein interactions.⁵³  Because they involve homogenous reactions, biochemical assays are readily miniaturized. However, not all targets can be prepared satisfactorily for biochemical testing. Furthermore, a chemical's activity measured in cell-free assays does not necessarily correspond to its activity in the intact cell, which may be affected by the presence of intracellular cofactors, issues of membrane permeability, cytotoxicity, and other influences on the target molecule. In contrast, cell-based assays measure the effects of chemicals on pathways of interest in the physiological environment of a cell, without the need to specify a molecular target. Examples include functional assays,^54,55  reporter-gene assays (which use “marker” genes to signal activation of target genes),^56,57  and phenotypic assays for processes such as cell migration⁵⁸  or division.⁵⁹  Because cell-based assays measure effects on entire pathways, perturbations can be assessed at more than one step in a pathway. Cell-based HTS in 1536- or even 3456-well plate formats is not uncommon.^60–62  Unfortunately, space does not permit more extensive coverage of HTS methodology in this volume.

Conventional toxicity testing has served regulatory risk assessment reasonably well since its inception in the early 1970s. However, while conventional tests have become more refined and more definitive, they have not been able to meet the demand for higher throughput at lower cost. Recent innovations in HTS technologies promise higher throughput and lower cost; however, a key question is whether they can significantly improve risk assessment in support of regulatory decision-making. TGx technologies potentially can improve predictive toxicology (and thus risk assessment) by improving our understanding of dose–response relationships, cross-species extrapolation, exposure quantification, mechanisms of action, and variations in individual susceptibility. Although current risk assessment processes use mechanistic information in hazard identification and in understanding dose–response relationships, the information on mechanisms of toxicity often is incomplete or inconclusive. TGx is well suited to the classification of compounds by mode of action and to prediction of toxicological outcomes based on perturbation of known pathways of toxicity and disease. Applied in parallel with standard toxicological assays, TGx is quite useful in defining the details of mechanisms of toxicity and in providing early predictions of chronic disease outcomes (e.g., cancer). TGx technologies also hold promise for improving exposure assessment. Because exposure to environmental chemicals is rarely measured directly, exposure is estimated typically by mathematical models based on factors that would affect exposure, such as typical water consumption, respiratory rates, and activity patterns. Following are specific roles that have been proposed for TGx studies and HTS tests in risk assessment:

• Exposure assessment: the high level of information inherent in gene-expression analysis may enable various types of exposure to be distinguished. In addition, proteomic analysis of biofluids may provide a means for noninvasive identification of biomarkers of exposure.

• Hazard screening: HTS tests are ideally suited for rapid testing of large numbers of compounds and may also be useful for preliminary hazard classification based on the perturbation of well-understood pathways of toxicity.

• Hazard classification and mode-of-action triage: identification of TGx signatures and corresponding toxicity pathways can allow various toxic and nontoxic endpoints to be distinguished. Mathematical modeling can be used to identify signatures for known toxicity outcomes and then to predict the effects of unknown toxicants in medium-throughput assays. Signature genes can be used, for example, to identify potential genotoxic and nongenotoxic carcinogens, aneugens, and cytotoxic agents.^63,64 

• Mechanistic information: TGx studies offer the opportunity to screen, or to evaluate in detail, molecular mechanisms of toxic action. When applied to the study of large classes of chemicals, TGx information can be used to globally define modes or mechanisms of toxic action. Significant changes in biologic mechanisms may occur with increasing dose, so that the TGx profiles are altered qualitatively as well as quantitatively.⁶⁵  Although some mechanistic information can be garnered from results of HTS in vitro, this is more readily accomplished at present by TGx methods, because of the density and highly parallel nature of TGx datasets. Furthermore, the mechanistic insights provided by TGx data can help refine the models used to predict target-organ doses, thus increasing the accuracy of dose–response assessment.

• Cross-species extrapolation: TGx methods can be used to evaluate the degree to which the biologic response pathways responsible for toxicity are conserved across species. Comparative TGx studies in two different species can identify orthologous genes (i.e., genes in different species that have evolved from a common ancestral gene by speciation and generally retain a similar function).

• Dose–response relationships: TGx studies are usually performed at several different doses and in a series of time points, enabling construction of dose–response curves over dose ranges that may be relevant to human exposures. There have been significant advances in this area of application of TGx methods.^66–68  In addition, HTS offers a significant potential advantage by enabling studies over an extensive range of doses in each individual assay.

• Developmental exposures: TGx techniques can be used in conjunction with conventional developmental toxicology methods to examine the highly time- and stage-sensitive effects of exposure on developmental processes in vivo.

• Variability in susceptibility: TGx can examine variability in gene expression due to mutations in genes or in the regions of chromosomes that regulate gene expression, and other factors that modify expression, such as epigenetic effects (reversible heritable changes in gene expression caused by mechanisms other than changes in DNA sequence).

• Mixtures: environmental exposure of humans typically is to complex mixtures, rather than to single agents. TGx and HTS technologies can be used to evaluate typical environmental mixtures, albeit with special requirements for sample preparation.

• Gene–environment interactions: TGx can provide insight into the responses of biological pathways that reflect the interactions between genetic makeup and environmental conditions and exposures.

The various potential roles of TGx has been discussed in numerous scientific reviews and commentaries over the past decade and a half,^{10–12,66–92}  which attests to the evolving reality that TGx is enhancing the ability of scientists to study and estimate the risks that chemicals pose to human health and the environment. Furthermore, it is becoming clear that HTS approaches using in vitro and computational methods can reduce the use of animals in toxicity testing, perhaps eventually replacing some uses. An example of a computational systems model currently being developed is the EPA's Virtual Liver project.⁹³  The aim is to develop models for predicting liver injury resulting from chronic chemical exposure by simulating the key chemical-induced events that result in changes in liver cells and tissue. Many scientists question whether this approach can completely replace the use of animals in regulatory testing. These scientists see the pathway-based screening approach as useful mainly to prioritize chemicals for further testing and perhaps to verify the relevance of animal toxicity testing methods to humans. However, the limited gain in knowledge from conventional animal studies and significant uncertainties regarding hazard identification hamper appropriate risk evaluation without efforts in pathway-based screening. Programs such as the ToxCast project and other alternative testing strategies were initiated to obtain more meaningful data that are beyond the limited information gathered from EPA or Organisation for Economic Co-operation and Development (OECD) guideline studies.⁹⁴  By 2010, nearly 500 assays that focus particularly on pathway perturbations had been developed under ToxCast.⁹⁵ 

A very difficult challenge for in vitro testing is that it ideally must encompass the full range of toxicokinetic and toxicocodynamic phenomena that contribute to whole-animal toxicity.⁹⁶  However, combining the new technologies with refined approaches to whole-animal toxicity testing (e.g., eliminating the need for 2-year rodent bioassays) could be one of the major decision-making opportunities in implementing the National Research Council's vision for the 21st century.⁹⁷  In any event, the goal in incorporating new technologies into risk assessment must be better informed decisions concerning potential adverse human health outcomes than is possible through conventional toxicity testing and risk assessment practices.

Predictive toxicology is the study of how toxic effects observed in model systems (or humans) can be used to predict pathogenesis, assess risk, and prevent human disease. Improving risk assessment is an essential aim of predictive toxicology. Information gaps and inconsistencies⁹⁸  include:

• the need for more toxicity screening data, data on effects in humans and on human exposure levels;

• information on the relevance of animal data to humans;

• exposure–response data (especially at low, environmental exposure levels);

• data on different routes of exposure;

• data on the effects of co-exposure to more than one chemical;

• data on human variability in susceptibility to toxicants; and

• information to resolve or explain inconsistencies in data from various animal models.

TGx methods are capable of addressing many, if not all, of these information gaps. However, fully integrating TGx technologies into predictive toxicology will require a coordinated effort to identify key datasets and to extract the essential data and information that can be used for predictive purposes. In this regard, there is a continuing need for better bioinformatics, statistical, and computational approaches and software to analyze TGx datasets as described in subsequent chapters. The resulting datasets should be deposited into curated public databases^99–101  with appropriate documentary information so that they can be readily analyzed and compared. There is also a need for public databases to facilitate sharing and use of TGx data and for tools to mine these databases.¹  The concerted efforts of government and the private sector are necessary to address these needs and propel the field forward.¹⁰² 

A long-term goal of TGx is to achieve a systems-level understanding of the responses of organisms to toxicants.¹  This “systems toxicology” approach attempts to synthesize many different types of data in order to more completely understand of the biological response of a cell, organ, or organism to a particular toxicant, leading to the creation of predictive biomathematical models. Figure 1.1 illustrates the sequence of events between initial exposure to a toxicant and final disease outcome. After exposure, the absorption, distribution, metabolism, and excretion systems of the body control local concentrations of a chemical stressor in various body compartments. The internal dose in a given body compartment is affected by genetics through the involvement of specific forms of genes that code for various transporters and metabolizing enzymes. Mathematical models such as exposure models, physiologically based pharmacokinetic (PBPK) models, and biologically based dose–response (BBDR) models can be used to approximate these processes. PBPK models are a set of differential equations structured to provide a time course of a chemical's mass-balance disposition (wherein all inputs, outputs, and changes in total mass of the chemical are accounted for) in preselected anatomical compartments. BBDR models are dose–response models that are based on underlying biological processes. Once the target tissue is exposed to a local stressor, the cells respond and either adapt or undergo a toxic response; this process can be modeled using systems toxicology approaches. Finally, the disease outcome itself can be mimicked by genetic or chemically-induced models of particular diseases.¹ 

An advantage of HTS technologies is their ability to rapidly and efficiently test chemicals over a wide range of doses and to describe the dose–response curve at concentrations that better reflect real-world human exposure to environmental chemicals. Low-dose effects observed in HTS systems will need to be interpreted in the context of no-effect and effect-level doses reported in whole-animal toxicity tests. Preferably, low-dose effects would be addressed through the use of TGx methods in concert with conventional animal testing.^84,103 

Conventional toxicity studies of environmental chemicals generally express dose simply as external exposure levels, whereas the biologically relevant doses are those delivered to the target tissues. Recent advances in analytical technologies provide an opportunity to capture internal dosimetric information in conventional toxicology studies.¹⁰⁴  Integration of animal internal dosimetry findings with human biomonitoring data and with dose–response information from HTS and TGx technologies will provide information of great value in health risk assessment. Such opportunities are illustrated by recent efforts to develop “biomonitoring equivalents” (benchmarks for relating safety to concentrations of chemicals in biological specimens, rather than to estimated daily intakes). Efforts to relate human exposure to doses in toxicity test systems and regulatory exposure standards will aid in development of approaches for efficient selection of environmental chemicals for detailed toxicological assessment.^105,106  Biomonitoring equivalents can inform selection of concentrations used in TGx and HTS, to ensure that they correspond to meaningful human risk assessment benchmarks, such as reference doses.¹⁰⁷  These concepts are discussed in detail by Thomas and Waters in chapter 5.

The adage that “the dose makes the poison”, credited to Paracelsus, implies distinguishing a nontoxic or subtoxic dose from a toxic dose of a chemical. The increasing focus on toxicity evaluations at low, environmentally relevant, doses must be accompanied by greater attention to what constitutes an “adverse” response. Initially, “toxicity pathways” are normal physiological and biochemical pathways through which the cell or organism is attempting to compensate for chemically induced perturbations. If chemicals tested are simply described as affecting toxicity pathways, it may be inferred that such effects must be adverse. However, such perturbations may represent normal homeostatic responses to interactions of chemicals with cellular systems at low doses.^106,108  Indeed, it has been suggested that the “exposure-adverse effect continuum” is in reality a ‘‘discontinuum”,¹⁰⁸  because responses that remain within homeostatic tolerance limits will not result in adverse effects, even at sensitive life stages and in highly susceptible individuals. Research is needed on the range and tolerance limits of normal homeostatic responses and on the dose-dependent transitions that result in shifts from a normal state to an adaptive state to a state that reflects adverse effects.

An important aspect of risk assessment is the extrapolation of results from high-dose animal toxicity studies to human health risks, taking into account differences among animal species, strains, or sexes. In the absence of information on the biological basis of toxicity, risk assessors must rely upon default assumptions or “uncertainty” factors; for example, it may be assumed that the response to a chemical is linear, with no threshold. Debates surrounding the use of default assumption may delay regulatory decisions.¹⁰⁹  An important role for TGx, therefore, may be in testing the validity of currently used default assumptions and replacing or refining them,¹¹⁰  as discussed by McMullen et al. in chapter 6.

Phenotypic anchoring^27,28  relates specific changes in gene-expression profiles to specific typically adverse effects observed in conventional toxicity testing. This process generally is based on histopathology (microscopic anatomical changes in tissues) or clinical chemistry (chemical analysis of fluids or tissues). Tissue histopathology is regarded as the most precise conventional indicator of toxicity. Ideally, tissues are examined by a board-certified pathologist, whose interpretation is recorded through the use of a controlled vocabulary that matches an image, as may be available in an online atlas of histopathological images. Alternatively, clinical chemistry indicators are appropriate for other parameters of toxicity, such as oxidative stress. Phenotypic anchoring is necessary in order to clearly distinguish between gene-expression changes that are associated with adverse effects and changes that are incidental or are associated with normal homeostatic responses (i.e., physiological adjustments to maintain the equilibrium of internal systems).

The same level of exposure to a chemical may produce different biologic effects in different individuals. TGx offers opportunities to more fully characterize genetic variation in human susceptibility to toxic effects of chemicals, through studies on how individual variations in gene sequence or epigenetic modifications influence the response to chemicals. Variability in gene expression reflects individual variability as a result of genetic polymorphisms (the existence of alternative forms of a particular gene, which produce different phenotypes) or mutations in genes or in DNA that regulate gene expression, as well as other factors, such as epigenetic alterations. The bases for human variability in responses to toxicants can be explored by studying differences in levels of expression of a given gene, as well as differences in which genes are expressed.

For any screening test to be useful, it must be sensitive (able to detect the state being tested where it truly exists) and specific (responding only to the specific state being tested). A challenge for the design and validation of TGx screening tests is identification of a “gold standard”—an indicator of the true state of toxicity against which the screening test's sensitivity and specificity can be measured. Setting a relatively low threshold for a positive result provides greater sensitivity and lower specificity (that is, fewer false-negative and more false-positive results), whereas setting a high threshold provides lower sensitivity and higher specificity (that is, more false-negative and fewer false-positive results). In the case of environmental chemicals, false-positive results can lead to inappropriate regulatory restrictions on their use, and may divert resources from other public health issues, while false-negative results may delay or prevent responses needed to protect public health.

Therefore, caution is required in any transition from in vivo to predominantly in vitro testing. The focus must be on identifying true human health risks with a higher degree of confidence than is possible with conventional test systems. Confidence in HTS and computational profiling methods will depend on evaluation of the tests’ relevance, reliability, sensitivity, and specificity. The National Research Council¹¹⁰  has suggested approaches for validation of TGx technologies, and the OECD has provided principles and guidance for the validation of quantitative structure–activity relationships¹¹¹  and evidence-based toxicology.^112,113  The OECD Guidance Document on the Design and Conduct of Chronic Toxicity and Carcinogenicity Studies was released in September 2014¹¹⁴  and current OECD activities related to molecular screening and toxicogenomics are described on the OECD website (www.oecd.org/chemicalsafety/testing/toxicogenomics.htm, 2015).

In 2012, the OECD launched a new program on the development of adverse outcome pathways.¹¹⁵  An adverse outcome pathway is defined as “an analytical construct that describes a sequential chain of causally linked events at different levels of biological organization that lead to an adverse health or ecotoxicological effect”. Adverse outcome pathways are the central components of a toxicological knowledge framework being constructed to support chemical risk assessment based on mechanistic reasoning, and are equally applicable to TGx or HTS investigations.¹¹⁶ 

The selection of classes of chemicals and prototype compounds for study by industry and government in TGx and HTS studies has been driven largely by the investigative focus. Thus the NIEHS NCT targeted NTP-tested hepatotoxicants and nephrotoxicants, including environmental chemicals and generic drugs for which there are substantial in-life, clinical, and histopathology data, as well as human exposure information.¹¹⁷  The decision to do so was taken because of the requirement for phenotypic anchoring of the TGx data and the anticipated direct applicability of the resulting datasets to public health.

In candidate drug screening and subsequent TGx investigations, pharmaceutical companies have addressed mainly libraries of candidate drugs from their development pipelines, and have compared them to prototypic industrial chemicals and failed drugs for purposes of drug safety assessment.

For HTS, the EPA National Center for Computational Toxicology and ToxCast selected food-use pesticides and other chemicals for which there exist animal toxicology and human exposure information. The corresponding NTP compound collection for HTS consists of solvents, fire retardants, dyes, preservatives, plasticizers, therapeutic agents, inorganic and organic pollutants, drinking water disinfection byproducts, pesticides and natural products, partly based on the availability of toxicological data from standard tests of carcinogenicity, genotoxicity, immunotoxicity and/or reproductive and developmental toxicity. The NIH NCGC Pharmaceutical Collection (Ruili and colleagues, unpublished data) contains small molecules or compounds that have been evaluated in clinical trials and approved by the FDA in the United States or elsewhere.

Both TGx and HTS efforts to date have emphasized studies on hepatotoxicity and hepatocarcinogenicity. Prototypic compounds have included direct- and indirect-acting genotoxins, nongenotoxic carcinogens, and noncarcinogens that target the liver.¹¹⁷  The NTP has identified and studied numerous hepatoxicants and hepatocarcinogens in rats and mice. The rationale for the selection of the liver is that it is the principal metabolic organ of the body and that tissue and primary cells are readily available from mice, rats, and humans. The availability of several metabolically active rodent and human liver cell lines make in vitro TGx and HTS assays feasible. In chapter 3 Buick and Yauk address specifically the issue of metabolic competency of mammalian and human cell systems. Subsequent chapters in this book discuss predictive TGx in substantial detail with the liver as a model target organ.

Predictive carcinogenicity focuses on cancer as the disease entity and conventional animal testing has dominated the field for more than four decades. It may be surprising to learn that the highly resource intensive 2-year rodent cancer bioassay was originally established by the National Cancer Institute (NCI) as a screening test to identify potential carcinogens for further analysis in human epidemiological studies.¹¹⁸  The 2-year rodent NTP bioassay has evolved as the primary means and gold standard for determining the carcinogenic potential of a chemical by providing the dose–response information that is required for risk assessment. However, it is so costly and time consuming that as of 2007 only 1547 chemicals had been tested, including about 560 tested in the NTP carcinogen bioassay program.^119,120  As of early 2015 there are 582 technical reports available from the NTP carcinogen bioassay program.¹²¹ 

Initially, chemicals selected for testing in the NCI/NTP rodent carcinogenicity bioassay program were those suspected to be carcinogenic based on expert knowledge. Later, the selection process changed to include chemicals based on widespread human exposure and high production volume. To determine the dependence of positive results in the carcinogenicity bioassay on these chemical selection criteria, Fung et al.¹²²  analyzed the results from bioassays of 400 chemicals tested since 1995. Of 267 (67%) chemicals selected on the basis of suspect carcinogenicity, 187 (70%) were carcinogenic. Of 133 chemicals selected based only on exposure or production volume, the majority (80%) were not carcinogenic in animals, even when tested at the maximum tolerated dose. Of the total 400 chemicals, 210 (53%) caused tumors in at least one organ of one sex of one species (among male and female mice and rats). Only 92 chemicals (23%) were carcinogenic in both species, the criterion used by the International Agency for Research on Cancer (IARC) to determine whether a chemical is likely to be carcinogenic in humans. Based on this analysis, the authors concluded that fewer than 5–10% of the 75 000 chemicals in commercial use might be reasonably anticipated to be carcinogenic to humans. In fact, to date IARC has classified 116 agents as group 1 “known human carcinogens” (http://monographs.iarc.fr/, 2015).

The likelihood that a relatively small percentage of the hundreds of thousands of chemicals in industrial use will cause cancer favored the development and application of highly predictive short-term genotoxicity screening tests beginning in the late 1960s and early 1970s. Publications in the 1980s on short-term tests resulting from EPA's GENE-TOX program^15–19  defined the basic protocols and performance characteristics of these tests and helped to begin the gradual process of reducing their numbers from more than 200 to the handful of guideline tests used today.

Among the early short-term bioassays, the Ames test for gene mutation (point mutation) in bacteria held the greatest promise as a predictor of carcinogenicity.^123,124  This test emerged as the screening test of choice for potential genotoxic carcinogens. However, the Ames test could not be used blindly, because the bacterial standard tester strains did not respond positively to certain classes of carcinogens (e.g., inorganic metal and halogenated organic compounds) and displayed poor specificity for others (e.g., nitrogen- and sulfur-containing organic compounds).¹²⁵  Despite these limitations, the test was so successful that in the early 1990s, some scientists tended to categorize as nongenotoxic (“Ames negative”) chemicals that gave positive results in other assays for various types of genetic damage.¹²⁶ 

Butterworth¹²⁷  emphasized that the “primary biological activity of the [genotoxic] chemical or a metabolite is alteration of the information encoded in the DNA … point mutations, insertions, deletions or changes in chromosome structure or number … Nongenotoxic chemicals … may yield genotoxic events as a secondary result of other induced toxicity, such as forced cellular growth, but their primary action does not involve reactivity with the DNA”. From a regulatory perspective, the gradual recognition of nongenotoxic mechanisms of carcinogenesis¹²⁸  complicated the established relationship between genotoxicity and carcinogenicity and also challenged the conventional interpretation of rodent carcinogenicity results in terms of relevance to human cancer.⁶³  Because of the default assumption in regulatory decision-making regarding the presumed linearity of the dose–response curve for genotoxic carcinogens, the classification of carcinogens as genotoxic or nongenotoxic became an essential but highly debatable component of cancer risk assessment. As science advances with new technologies, it is important to recognize that the makeup of the reference set of chemicals and databases can dramatically influence chemical classification. Because of the purely statistical approaches frequently used in interpreting the results of highly parallel and high-content assays, these technologies can be very susceptible to misinterpretation due to inappropriate composition of the training sets of chemicals.¹²⁹ 

We have learned over the years that of all compounds tested in rodents, about half are carcinogenic,¹³⁰  and roughly half of these are putatively nongenotoxic.¹³¹  Rat liver appears to be particularly sensitive to nongenotoxic carcinogens, and has been studied extensively to understand the potential mechanisms involved. However, it is clear that both genotoxic and nongenotoxic chemicals induce cancer in a variety of target sites in rodents.

The eight most frequent target sites in both rats and mice are liver, lung, mammary gland, stomach, vascular system, kidney, hematopoietic system, and urinary bladder. There are some species differences in results for particular target organs, including the liver, Zymbal's gland, and kidney.¹³²  However, there is no support for any systematic inter-species differences in tissue distribution and pharmacokinetics between genotoxic and nongenotoxic agents, nor for the idea that these two categories of agents induce tumors in different target organs.¹³²  The central dogma in the etiology of carcinogenesis is that DNA damage resulting in mutation and other events that enhance cell proliferation are both required. Many carcinogens are cytotoxic and induce compensatory cell proliferation. Thus in predicting carcinogenicity it is appropriate to further study this phenomenon in vitro and comparatively in vivo using the appropriate TGx methods to identify relevant and predictive pathways.

The investigations in vivo, reviewed in Waters et al., 2010¹²⁹  and updated in chapter 4 by Auerbach, have identified cancer-relevant gene signatures or biomarkers that discriminate between direct and indirect genotoxic carcinogens, nongenotoxic carcinogens, and noncarcinogens. In an early series of studies by Ellinger-Ziegelbauer and colleagues,^63,133,134  a strong DNA-damage response at the gene-expression level suggested direct DNA modification, whereas increased expression of genes involved in cell-cycle progression appeared characteristic of indirect-acting agents. Metabolism genes were prominently represented among gene-expression signatures that discriminated nongenotoxic modes of action (e.g., cytotoxicity and regenerative proliferation, xenobiotic receptor agonists, peroxisome-proliferator-activated receptors, or hormone-mediated processes^77,135 ). The preponderance of accumulated evidence suggested that gene-expression profiles reflect underlying modes or mechanisms of action and are therefore useful in predicting chemical carcinogenicity in rodents, especially in conjunction with conventional short-term tests for gene mutation and other forms of DNA damage.¹³⁶ 

A more difficult task is distinguishing nongenotoxic carcinogens from noncarcinogens. Evidence to date suggests that some modes of action of nongenotoxic carcinogenicity such as the induction of oxidative stress exhibit definitive signatures as early as 24 h following single dosing in animals.^77,135,137  Furthermore, it seems clear that false-positive TGx signatures may be resolved by repeat dosing up to 28 days. Because of multiple modes of action, the resolution of classification of nongenotoxic carcinogens vs. noncarcinogens cannot be accomplished without efforts to clarify what combinations of marker gene sets related to specific pathophysiological processes of carcinogenesis. This will take time and highly directed efforts.

The majority of in vivo studies reviewed in Waters et al., 2010¹²⁹  were performed in the liver, with the notable exception of studies by Thomas and colleagues^79,138  in mouse lung. It is important to extend these investigations to other target organs and to identify within these organs the target cell populations from which tumors develop. For TGx studies to be broadly predictive, in vivo studies should be performed simultaneously in several relevant metabolically active target organs. In such studies, it is important to distinguish between a tissue carcinogen and a tissue toxin, since not even all hepatotoxicants cause liver cancer. Five tissue sites (liver, lung, mammary gland, kidney, and the hematopoietic system) account for the positive responses for about half of the chemicals identified by the NTP as carcinogens and positive results have been observed at 24 tissue sites for five or more chemicals in at least one species and sex.¹³⁸  Developing gene-expression biomarkers for each of the top five tumor sites in mice and rats should provide an efficient means to prioritize chemicals for further testing. As suggested by Thomas et al.,¹³⁸  it may be useful in the longer term to develop biomarkers for each of the 24 main target tissues, which may facilitate replacement of the rodent cancer bioassay. It may also be of value to use TGx methods to better understand the mechanistic basis for species differences between rats and mice in particular target organs, including the liver, Zymbal's gland, and kidney.¹³² 

Thomas et al.¹³⁸  advocated applying the TGx approach in the preclinical phase of drug and chemical development to discriminate compounds likely to be human carcinogens. Such testing could provide an assessment of product safety earlier in the development pipeline, leading to substantial monetary savings and reduced time to market. The prevalence of potential nongenotoxic carcinogens in the drug-development pipeline has been one of the primary motivators for the pharmaceutical industry to develop TGx approaches for predictive carcinogenicity. The work of Fielden et al.¹³⁵  and Nie et al.⁷⁷  suggests that transcription profiling in appropriate target organs in vivo after short-term treatment (up to 14 days) has the potential to predict putative non-DNA-reactive mechanisms. Indeed, it may be possible to use TGx methods to exclude from further consideration DNA-reactive mechanisms for compounds for which positive results are observed only at high concentrations in in vitro gene mutation or chromosome damage assays. However, the results of the Fielden and Nie studies also raise the question of why the signatures of nongenotoxic carcinogenicity were so different between the studies and yet apparently predictive in both. Clearly when such predictive approaches are combined with standardized test procedures in prospective interlaboratory validation studies, their accuracy and potential utility in carcinogenicity evaluation can be better ascertained.^139,140 

Many types of commercial and industrial chemicals are not required to be tested for carcinogenicity unless evidence for adverse health effects is obtained. For those that do require further testing, TGx approaches would seem particularly valuable when used together with range-finding toxicity (14-day and 90-day) studies, as currently performed in conjunction with the rodent carcinogenicity bioassay. A report by Auerbach et al.¹⁴¹  focused primarily on defining classification models with the best cross-validation based on results from 2-, 14-, and 90-day studies. Most of the TGx studies reviewed by Waters et al.¹²⁹  identified liver carcinogenicity signatures from animals exposed to chemicals for 28 days or less, a duration that is referred to in traditional toxicology as “subacute”. Auerbach et al.¹⁴¹  hypothesized that exposures up to 90 days would accentuate the expression of genes related to carcinogenic activity and therefore allow the models to achieve a higher degree of certainty when making predictions. Auerbach et al.¹⁴¹  suggested that longer exposure durations would limit the influence of “mode of action” genes and allow for better identification of predictive genes with biology related to processes involved in the formation of neoplasms that are typically seen subsequent to the primary toxicity. Auerbach et al.¹⁴¹  also noted that the idea of a shared precancerous biology (that is independent of a specific chemical challenge) is not unreasonable, since the process of cancer manifestation is a continuum, and most types of cancer share a degree of universal biology that is manifested in their gene expression.¹⁴²  Auerbach et al.¹⁴¹  concluded that the duration of exposure is the primary factor affecting the utility of the models, and that a 90-day exposure provides superior data.

Indeed, carcinogenesis is a complex and protracted multi-step process. It could easily be argued that a focus on transcription profiling with exposures of up to 14 days or even 90 days cannot hope to predict the outcome of a 2-year rodent carcinogenicity bioassay. To better understand what transpires in the complex process of carcinogenesis, it would seem most appropriate to “profile to the phenotype”; i.e., to perform transcription profiling in conjunction with the pathological evaluation of target tissues (even if this were to be done using archival material, such as is available in the extensive NTP archives). Admittedly, this is a diagnostic (or retrospective) approach, as opposed to prognostic or predictive carcinogenicity methodology. However, the information potentially to be gained could lead to the ability to predict much earlier based on “shared precancerous biology” whether neoplastic lesions will progress to cancer over the lifetime of the animal. If data are encouraging, similar information developed based on studies of archived human tissues could readily be performed.

A multi-laboratory project coordinated by the Health and Environmental Sciences Institute (HESI) Committee on the Application of Genomics in Mechanism-based Risk Assessment evaluated gene-expression profiles of TK6 cells treated with model genotoxic agents using a targeted high-density RT-PCR. This study adds to the increasing body of evidence indicating that TGx analysis of cellular stress responses provides insight into mechanisms of action of genotoxicants.⁶³  The reproducibility of data across collaborating laboratories indicates that expression analysis of a relevant gene set is capable of distinguishing compounds that cause DNA adducts or double strand breaks from those that interfere with mitotic spindle function or that cause chromosome damage as a consequence of cytotoxicity.

TGx studies of chemicals that are both rodent and human carcinogens could identify biomarkers with more direct relevance to human health.¹³⁸  Compounds that do not produce positive test results in the conventional genotoxicity assays and that do not exhibit biomarkers of genotoxicity in TGx methods are very unlikely to pose a genotoxic carcinogenic risk to humans. The same cannot be said for putative nongenotoxic carcinogens that are identified through the use of TGx methods. However, it should be possible in such cases to use TGx methods to characterize their likely modes of action by comparison with previously well-studied chemicals, as demonstrated by Fielden and colleagues^90,135  and Uehara and colleagues,^87,137  and, with more experience, to predict relevance to humans.

The potential of -omics technologies to explore transcriptional regulation (including epigenetics and miRNA) as well as downstream events (proteomics and metabolomics) in evaluating mechanisms of genotoxicity and carcinogenicity must also be investigated.^75,143  No single organization has the resources to accomplish all of this independently. Therefore, collaborative efforts that include scientists from academia, industry, and regulatory agencies, such as the HESI Genomics Committee, the Critical Path Initiative in the United States, and the Innovative Medicines Initiative in Europe, are essential for developing standardized testing protocols and critically needed reference data.⁶³  If the TGx approach proves to be more broadly applicable through such efforts, it has the potential to become an efficient and economical alternative to the rodent cancer bioassay, potentially reducing the use of experimental animals while increasing the efficiency of predictive carcinogenicity. Clearly, the investigations discussed here and to be discussed in subsequent chapters have opened the door to a paradigm shift in chemical safety testing and health risk assessment.

The author acknowledges funding from Health Canada Existing Substances Risk Assessment Bureau, to Integrated Laboratory Systems (ILS), Inc., and significant contributions of colleagues at ILS, including Marcus Jackson, Isabel Lea, H. Frank Stack, and Susan Dakin, in the preparation of background information (Interim Report, Approach to Development and Integration of Toxicogenomics and High-Throughput Screening in Support of Human Health Risk Assessment, submitted January 24, 2011, (Health Canada Contract Reference Number 4500247514)) included in part in this chapter by permission.