Genomics and developmental risk assessment

In the aftermath of the thalidomide tragedy, test guidelines were developed to systematically test the potential of pharmaceutical agents to interfere with the reproductive process, especially embryonic development. Similar guidelines, especially for the assessment of developmental toxicity, were adopted for the purpose of testing pesticides and commodity chemicals. To date, a few thousand chemicals have been tested using this protocol. The initial guidelines, published in 1965 by the US FDA, were sent to companies regulated by that agency, accompanied by a letter from Dr. E. I. Goldenthal, the person in charge of evaluating preclinical drug safety data at FDA. In the letter, Dr. Goldenthal indicated that the guidelines were based on the best science at the time, but acknowledged that they were likely to change as experience was gained in running the protocols, and to incorporate scientific advances. Although the guidelines have been modified over the years, it is still largely the same protocol that has been in place since 1965. The test agent is administered to pregnant lab animals, usually from the time that implantation occurs until the end of the embryonic period or term, and fetuses are evaluated for structural manifestations of abnormal development. The procedure has been used to test a few thousand chemicals to this point, and has been useful for supporting regulatory decisions regarding safety. In the meantime, life science has undergone remarkable changes in the last four decades. In 1965, developmental biology and toxicology were largely descriptive sciences. In the ensuing 40 years, major advances in biochemistry, cell biology, and molecular biology have been made that have revolutionized how life science research is done. These techniques have been applied successfully in basic studies in developmental biology, and in mechanistic studies in toxicology. We can credit most of our improved understanding of mechanisms of abnormal development to the technology that teratologists have applied to the fundamental questions of our field. Yet, this revolution has not meaningfully influenced the way we assess toxicity or conduct risk assessment. The principal reason for this inertia lies in the fact that, despite our increasingly impressive understanding of the mechanisms underlying chemically induced birth defects, it is clear that we do not yet know how many of these mechanisms act. Therefore, assessing toxicity by evaluating outcome is a safer bet, as the outcomes that we evaluate—the manifestations of abnormal development—will be influenced by and serve as signals of the interactions that have occurred at a much more fundamental level of biological organization. This apical nature of toxicology testing is thought to maximize the chance of detecting an effect irrespective of mechanism, and to do so efficiently. Apical testing has its disadvantages, too, however. This kind of testing is time-consuming because of the observational, descriptive nature of the assessments, and animal-intensive, as a relatively large number of animals is needed to provide enough statistical robustness to enable one to distinguish effect from background. Accordingly, only a few thousand chemicals have been tested by these methods, far fewer than the number of chemicals estimated to be in commerce. Furthermore, because apical testing is essentially a black box, the process of interpreting the results and extrapolating them for human risk assessment has been the subject of much controversy. The sequencing of the human genome and that of the most commonly used lab animal models, along with the advent of high-quality, commercially available microarrays, has the potential to catalyze a revolution in toxicity testing. It is now possible to evaluate changes in the expression of every gene in a tissue or organ in response to an environmental insult, and it has been found that virtually every type of toxicity that has been investigated produces changes in gene expression. Further investigation has shown that the changes in gene expression are specific for mode of action (e.g., Fielden et al., 2005, others). Comparative analysis of the genomes of species has further corroborated the fact that the evolutionary conservation of biological function across species has its basis at the level of the gene. These findings, when taken together, suggest a number of ways by which genomic analyses can improve applied toxicology and risk assessment (Table 1). Because patterns of altered gene expression are specific for mode of action, it may be possible to use gene expression information for predictive screening. Because changes in gene expression are sensitive measures of biological response, they may lead to improved dose-response assessment, with less uncertainty about the levels that will be without effect on people. The increased availability of mechanistic information will improve our confidence that the animal model used for toxicity testing is relevant for predicting effects in humans. Because of the genomic conservation across species, it may be possible to improve animal-to-human extrapolation (and maybe even assess the validity of the risk assessment by evaluating gene expression, or the lack thereof, in humans [assuming that relevant gene expression occurs in cells that can be obtained by minimally invasive means]). Potential new drugs and pesticides are comprehensively tested for toxicity. Because they are biologically active it is thought they may have a significant potential to produce adverse effects and that their toxicity profile needs to be thoroughly characterized. It hasn't been possible to fully characterize the toxicity of most of the chemicals in commerce, however, because the resources available for testing are limited. For commodity chemicals, decisions on which chemicals to test comprehensively are based almost exclusively on production volume or extent of distribution. To date this has been the only reasonable strategy for prioritizing testing, but it eliminates half of the risk assessment equation—hazard potential—from consideration. Consequently, consumer products ingredients and high production volume chemicals have been well tested, but there are many other compounds that have not been. Environmental activists consider this situation to be unacceptable, and the European Union is within weeks of approving sweeping legislation (called REACh) that will require massive testing efforts. While the sentiment behind REACh is understandable, the proponents of the legislation have not considered the practical limitations of the testing establishment. The reason that only a few thousand chemicals have been tested for developmental toxicity over the past 40 years is not due to industry resistance or regulatory neglect but to the fact that testing resources are limited and have been directed towards evaluating chemicals for which the consequences on public health, should they prove to be toxic, would be the most significant. It seems obvious now that a genomic analysis of gene expression can provide specific information about mechanism of action of a toxicant, but this conjecture has only been proven recently. Much of the research on the value of gene expression analysis has been done by or for the pharmaceutical industry, as a means of rapidly determining whether new drug candidates have the potential to cause toxicity that would be likely to make them unusable in the clinic. Much of this work has been focused on target organ toxicity (e.g., liver and kidney), and has proven to be effective in identifying modes of toxicity (e.g., Fielden et al., 2005, others). Gene expression analysis also provides specific information about mode of action for reproductive and developmental toxicants. We began our research in genomics by addressing the question of whether estrogens elicit a specific pattern of gene expression in developing and mature reproductive organs. We believed that gene expression analysis of estrogen responses was likely to be successful because gene expression is integral to the signal transduction process for this and all other steroid hormones. Changes in gene expression after estrogen treatment are a direct effect of the interaction of the ligand with its receptor, and subsequent binding to response elements on DNA. Therefore, our expectation was that the pattern of expression of genes would be specific and diagnostic of estrogenic activity. This expectation is true. We assessed gene expression in the uterus and ovaries of rats at two different life stages: late fetal and juvenile periods. These tissues have estrogen receptors at both life stages but the biological response is different. There is no visible immediate response to estrogens by the fetal reproductive system, but high doses of potent estrogens produce latent, permanent effects on reproductive function later in life. The juvenile and adult uterus goes through a proliferative (uterotrophic) response after treatment with physiological/ pharmacological levels of estrogens, but this response is transitory and comparable to normal estrous cycling in the adult rat. We identified characteristic transcript profiles for estrogens at both life stages (Naciff et al., 2002, 2003). Although literally hundreds of genes are affected by estrogens, we were able to identify a set of approximately 60 genes in the fetus and 120 in the juvenile rat that responded in a robust, dose-responsive manner to a variety of estrogens. Subsequent work in our lab and in others suggests that one needs only a handful of these, perhaps 10 or fewer, to confidently distinguish estrogenic activity from other modes of action. There are a number of applications for this kind of information in toxicity testing. First, it is possible to use gene expression analysis as a sensitive and specific screen for specific types of toxicity. A number of regulatory agencies, including the U.S. EPA, would like to screen large groups of chemicals for their ability to affect endocrine function. While first tier screens are likely to be simple receptor binding or reporter gene assays, the proposed assays for subsequent tiers include in vivo tests that may not have the resolving power or specificity of gene expression analyses. Approaches such as the one we have used may provide more definitive information. Second, it is possible to use gene expression analysis as an adjunct to traditional toxicity testing. The analysis that we carried out to identify the fetal transcript was done in term fetuses at the same stage at which they are analyzed for morphological development in a standard developmental toxicity study. The tissue requirement for analysis is small (there is enough RNA in a single fetal uterus, for example) so major modifications in study design would not be necessary. The added value of information could be considerable, particularly as we gain more information about gene expression changes elicited by a number of toxic mechanisms. Third, information about mechanism of action can be used to tailor subsequent toxicity testing. Currently, toxicity testing generally involves a survey of toxicity—an overview of the toxicity to organs (i.e., subchronic toxicity tests), reproduction and development, and of carcinogenic potential—but not in-depth assessments. If it were possible to identify the subset of all potential mechanisms through which a chemical is likely to act, it would be possible to better select the toxicity tests that are used to assess that chemical. To use estrogens as an example, tailored testing might focus on reproductive effects and possible carcinogenicity in estrogen-responsive tissues, with perhaps a secondary focus on long-term immune or neurological effects. Dose-response assessment is another aspect of risk assessment that has not changed appreciably, and is another area in which gene expression analysis may be useful. In traditional toxicology studies, a no-observed adverse effect level (NOAEL) is determined, and an acceptable intake level is estimated by applying uncertainty factors to the NOAEL. Animal toxicology studies are conducted at exaggerated dose levels, because the intent of the study is to maximize the chance of detecting a hazard if one exists, in a study design that is limited logistically to a relatively small number of animals (20–25 per dose group) compared to the potentially large number of people who could be exposed to the chemical once it enters commerce, albeit at a much lower exposure level. For chemicals that exert some developmental toxicity, there is often some residual level of effect at the NOAEL, but it is below the statistical resolving power of the assay. It is hoped that the acceptable intake level of the compound is below a threshold for any toxicity, but whether there is any residual level of risk at these ambient exposure levels is unknown, and at present unknowable. One of the potential benefits of gene expression analysis is that it is sensitive: it is possible to detect sub-attomole concentrations of individual messages. Assuming that some of the gene expression elicited by the compound is related to its toxicity, then it should be possible to extend observations much farther down the dose-response curve to see if there are any effects on these precursors to toxicity below the NOAEL. Of course, before this approach can be practically applied, it will be necessary to 1) identify which genes are involved in the biological processes that lead to an adverse outcome; and 2) determine the quantitative relationship between expression of these genes and changes at higher levels of biological organization at which the adverse effect is manifested. This research will take time, but is feasible by using bioinformatic approaches that group genes according to the biological processes they control, and make it possible to identify which biological processes are affected by treatment (see below). The culmination of this research will be the development of biologically-based dose-response models that take into account the dosages at which critical changes are elicited at each level of biological organization, from molecular to cellular to tissue/organ, to better describe the dose-response relationship and better estimate a level of exposure that has negligible risk. There have already been elegant attempts to construct biologically based dose-response models (e.g., Lau et al., 2001) but these have been limited by less-than-perfect knowledge of the underlying biochemical changes that lead to abnormal development. Gene expression analysis at a genomic level puts far more information in the hands of the modeler. While we don't yet have the information we need to apply gene expression analysis to biologically based dose-response model construction, it is still possible to use the methodology to address dose-response controversies. One of these controversies over the past decade concerns the possible existence of unusual effects of estrogens at doses orders of magnitude below the NOAEL from traditional toxicology studies. Attempts to replicate these studies have failed to do so despite the fact that the protocol of these latter studies followed that of the original ones as closely as possible, albeit with more animals to achieve greater statistical power. A panel of scientists convened by the U.S. National Toxicology Program was unable to reach a consensus as to which sets of studies were valid (Melnick et al., 2002). The panel noted that, although the repeat studies had greater statistical power, they could not definitively dismiss the results of the studies reporting low-dose effects because of the many factors that may not have been adequately replicated. Furthermore, there is sufficient variability in the response being measured—organ weight—that it may not be possible to resolve the dilemma simply with replicate studies. It was our premise that, should these low dose effects be real, they must be preceded by changes in gene expression, as gene expression is integral to the signal transduction for estrogens. This would be just as true at low dose levels as at high dose levels. Therefore, we conducted a study to evaluate gene expression over five to six orders of magnitude of exposure to three estrogens—ethynyl estradiol, bisphenol A, and genistein—in the fetal rat testis. Gene expression was assessed using microarrays that cover more than 15,000 genes in the rat genome. We had no preconceived notion as to which genes might be affected. It could be the same ones that were affected at high dosages, either in the same direction but at a different level of expression, or in the opposite direction, or a completely different set of genes, or no expression. We found that there were no observable changes in gene expression at the low dose levels that had been reported by some labs to affect male reproductive development (Naciff et al., 2005). Dose-response curves for gene expression were monotonic: robust expression of a subset of genes at the highest dose levels used, diminishing in expression to non-detectable levels within two to three orders of magnitude from the maximally tolerated dose. This experiment demonstrates the power of gene expression analysis to provide information about biological response at levels far below traditional NOAELs. It provided a novel approach to addressing a controversy that could not have been solved by more replicate studies. More importantly, it demonstrated the potential for gene expression analysis to support investigations into biological responses at the low end of the dose response curve and indicate that biologically based dose-response modeling will be feasible. Perhaps the greatest contribution of genomic analysis will be in rapidly accelerating the pace at which we understand the mechanisms that underlie abnormal development and developmental disease. Risk assessment has generally made little use of mechanistic information because little information has been available. However, there is great potential for mechanistic information to improve the risk assessment process. A clear understanding of mechanism increases the confidence that potential hazards have not been missed in the toxicity testing process. Mechanistic understanding will also improve dose-response assessment by better informing the shape of the dose-response curve and by confirming or refuting the existence of a threshold. Mechanistic information, especially at the molecular level, is critical in reducing the uncertainty in extrapolating animal results to predict human risks. It may eventually be possible, at least in some instances, to confirm the lack of biological response at the acceptable intake level by evaluating whether gene expression has occurred in a responsive tissue that can be non-invasively sampled, or the metabolic products of altered gene expression that can be sampled in urine or blood (e.g., Nicholson et al., 2002). The process by which gene expression data can be put into mechanistic context involves statistical and bioinformatic analysis that takes advantage of the power of 21st century computers. It is possible to organize and filter the massive data sets from microarray experiments in a variety of ways that can help shed light on the possible ways by which changes in gene expression drive a biological response. While these computational approaches are insufficient by themselves to demonstrate mechanisms of action, they are essential for organizing data and as support for generating a limited number of hypotheses regarding mechanism that can be experimentally tested. One such way of organizing data is to cluster affected genes according to their function. There are a number of ways to do this, many of which are available free through the National Center for Biotechnology Information (NCBI) at the National Library of Medicine. Programs available at NCBI include the Gene Expression Omnibus, which, among other things, clusters genes according to the biological process in which they are involved, or the location at which they are expressed. Figure 1 shows clustering of data from an experiment in which estrogens at various dose levels were given to juvenile rats and gene expression in the uterus analyzed (Naciff et al., 2003). The data are sorted according to gene ontology (GO) terms for biological processes. Detailed GO descriptions, organized into data structure that links genes to their function on a molecular, cellular, and organismal basis can be found on the Gene Ontology website (http://geneontology.org). set of a dose level, and within each each a different level of statistical a biological process, in a For example, is a large which with steroid being a level Changes in the expression of genes with each GO term are as a on the and the the the more genes that are affected. This kind of one to all biological processes and focus on processes that are affected genes affected in the estrogen-responsive rat uterus after treatment with dosages of an estrogen from et al., 2003). are according to the biological process with which they are set of the data from a and each within a set a different level of statistical are gene ontology (GO) (see for more Figure is a that most GO of this one to focus on of activity and Figure is an of one of the identifying the GO terms for which gene expression has been There are a number of genes altered for each of the GO terms It is then possible to the data for the individual genes in a The from the of and are particularly and are Figure is an of a with data from the juvenile uterus In this genes are in in and the robustness of expression indicated by help which genes are which are likely to be from and which genes may affect the same sets of of data sets such as the et al., have the ability to across data sets in different different developing tissues, at different developmental There are many more The is that both the power and are available to make of high information data and even to do across data A of genes involved in cell control, responses in the rat uterus to estrogen genes are in in with robustness of expression indicated by For the of mechanistic understanding in the value of these approaches is to support the of a number of of such hypotheses have been given and et al., 2003). of these hypotheses will take time, but the information provided by microarray experiments has made it possible to hypotheses that take into account the interactions of genes a over the approaches that have been taken in toxicology. of this kind of and testing are already available in the et have identified a set of target and development that are changed in the same direction by a variety of in and This work not only suggests a mechanism of action for a number of it also a that, although there is likely to be a large number of immediate molecular for there are a number of by which these initial interactions affect development. these developmental of signal transduction genes, or a is not yet However, if the conjecture is it suggests that we may be able to make in identifying the of that, when lead to developmental and in so identify the mechanistic in toxicity when are the of an a mechanistic understanding of normal and abnormal development will be the to developing better assays for detecting developmental toxicity, including that can be used for more screening. There are a few models in that are but that will either be for in vivo or that have the to be used for the large number of chemicals that are in or being The purpose of this has been to the power of genomics and other approaches to the way that applied toxicology and risk assessment is done. It is already feasible to screen chemicals in to determine if they elicit a pattern of gene expression that is characteristic of a and biological response. It should be feasible to use this information to animal test so that more information can be obtained from these studies about toxicity. It is also feasible to use the information from level assays to tailor the testing of compounds to in-depth investigation of the effects of the chemicals on biological processes that are to be responsive to the mechanisms we more about mechanisms of action, it will be possible to and less to screen large of and perhaps eventually to much of the animal testing that is done. and other techniques will also be useful in other of risk assessment, reducing the uncertainty in dose and animal-to-human We already have in which the high resolving power of gene expression data has been used to shed light on a controversy about possible effects of dose estrogens. this with a better understanding of the quantitative between the expression of genes and adverse tissue will be the of development of biologically-based dose-response models to that on uncertainty the mechanistic understanding that these will will our understanding of the that are at the of abnormal development, whether the cause is an or metabolic