Addressing Human Variability in Next-Generation Human Health Risk Assessments of Environmental Chemicals

Background: Characterizing variability in the extent and nature of responses to environmental exposures is a critical aspect of human health risk assessment. Objective: Our goal was to explore how next-generation human health risk assessments may better characterize variability in the context of the conceptual framework for the source-to-outcome continuum. Methods: This review was informed by a National Research Council workshop titled “Biological Factors that Underlie Individual Susceptibility to Environmental Stressors and Their Implications for Decision-Making.” We considered current experimental and in silico approaches, and emerging data streams (such as genetically defined human cells lines, genetically diverse rodent models, human omic profiling, and genome-wide association studies) that are providing new types of information and models relevant for assessing interindividual variability for application to human health risk assessments of environmental chemicals. Discussion: One challenge for characterizing variability is the wide range of sources of inherent biological variability (e.g., genetic and epigenetic variants) among individuals. A second challenge is that each particular pair of health outcomes and chemical exposures involves combinations of these sources, which may be further compounded by extrinsic factors (e.g., diet, psychosocial stressors, other exogenous chemical exposures). A third challenge is that different decision contexts present distinct needs regarding the identification—and extent of characterization—of interindividual variability in the human population. Conclusions: Despite these inherent challenges, opportunities exist to incorporate evidence from emerging data streams for addressing interindividual variability in a range of decision-making contexts.

Human variability underlies differences in the degrees and ways in which people respond to environ mental chemi cals, and address ing these differences is a key consideration in human health risk assessments for chemi cals Hattis et al. 2002;National Research Council (NRC) 2009]. A large array of possible health outcomes is of concern for such assessments, and many sources of variation can influence the sever ity and frequency of the adverse effects at dif ferent exposure levels. These sources may be intrinsic (e.g., heritable traits, life stage, aging), or extrinsic, exogenous, and acquired (e.g., background health conditions, cooccurring chemi cal exposures, food and nutrition status, psychosocial stressors). Interactions between inherent and extrinsic factors create the large range of biological variation exhibited in response to a chemi cal exposure (NRC 2009). Given that biological variability in susceptibil ity is context dependent, so too is the extent to which it needs to be described and quantified to inform any particular environ mental deci sion. The salience of variability information for specific choices is affected by the range of avail able risk management options; the regulatory authority; the available time, resources, and expertise to collect data and conduct analyses; and stakeholder concerns.
Over the past decade, efforts to systemati cally "map" human variability have expanded dramatically, focusing mainly on genetic variation (Schadt and Björkegren 2012). In addition to genetic differences, omics stud ies have examined the impact of epigenetic, transcriptomic, proteomic, and metabolo mic variation on disease susceptibility, prognosis, or options for pharmacotherapy (Chen et al. 2008;Emilsson et al. 2008;Illig et al. 2010;Manolio 2010;Schadt 2009). Tailored chemotherapy treatment based on patient (Phillips and Mallal 2010) or tumor (La Thangue and Kerr 2011) genetics is an example of a significant success in applying such discoveries; however, for many diseases, the substantial nongenetic variation in disease or treatment outcomes has limited their util ity. Thus, the characterization of the broad set of environ mental factors, including those related to chemi cal exposures, that may con tribute to disease is directly relevant to both personal ized medicine and environ mental health protection (Khoury et al. 2011).
In this review, we explore how next generation ("NexGen") human health risk assessments of chemi cals might take advantage of novel data to better characterize and quan tify variability in susceptibility, by using and expanding upon current analytical methods. We begin by describing biological variabil ity through the conceptual framework of the sourcetooutcome continuum. Next, the util ity of that framework is illustrated in a review of current approaches to describing variability in susceptibility in human health risk assess ments. Then, emerging data streams that may be informative in characterizing human vari ability in susceptibility are described. Finally, we consider the opportunities, challenges and methods for using emerging data to help assess inter individual variability in responses to environ mental chemi cals across different decision contexts.

Susceptibility as a Function of the Source-to-Outcome Continuum and Biological Variability
The "sourcetooutcome continuum" [U.S. Environmental Protection Agency (EPA) 2007; NRC 2007] is a conceptual frame work for human health risk assessment of environ mental chemi cals in which changes in the sources of chemi cals in the environment volume 121 | number 1 | January 2013 • Environmental Health Perspectives are further propagated within the individual through a series of biological and physiologi cal steps that may ultimately manifest as an adverse health outcome ( Figure 1): Source/media concentrations • are measures of the chemi cal, which may change under specific risk management options being con sidered. A given risk management decision may differentially affect media concentra tions depending on local conditions. External doses • are measures of exposure (e.g., concentration in air × breathing rate per body weight) to or intake (e.g., amount ingested per body weight) of environ mental chemi cals, and are related to source/media concentra tions by exposure pathways. Sources of vari ability that may confer susceptibility include differences in behaviors, such as breathing rates, water consumption, and dietary habits (e.g., the amount of fish consumed), and, in an occupational context, use of personal protective equipment.

Internal doses
• are the amounts/concentrations of environ mental chemi cals or their metabo lites at the target site(s) of interaction with biological molecules, and are related to exter nal doses by pharmaco kinetic (PK) processes. Susceptibility may arise from differences in compartment sizes and composition (e.g., fat concentration in plasma, which rises during pregnancy) (Roy et al. 1994), as well as dif ferences in the rates of uptake (e.g., fraction absorbed from diet or air), metabolism, elim ination, and transport to sites of action (e.g., the blood-brain barrier). Such differences may be due, for example, to genetics (e.g., via poly morphisms in metabolic enzymes, uptake and efflux transporters), other chemi cal exposures (via meta bolic enzyme induc tion and inhibition), and preexisting health conditions and life stage (e.g., via metabolism and mobilization from tissue storage).

Biological responses
• are measures of biological state (e.g., the concentration of gluta thione) altered by interactions with environ mental chemi cals or their metabolites, and are related to internal doses by pharmaco dynamic (PD) processes. Variation leading to differential susceptibility can stem from differences in transport systems, receptors and/or proteins in other toxicity pathways, as well as repair capacity (of, for example, DNA), which in turn are affected by intrinsic and extrinsic factors such as genetics and life stage. Physiological/health status • reflects the overall state, structure, or function of the organ ism and is related to biological responses through systems dynamics, the underlying physiological status of the host to which the chemi calspecific perturbation is added. Examples include maintenance and adapta tion processes (associated with preexisting health conditions, sex hormone levels, for example), and the accumulation of dam age events from past exposures (e.g., loss of alveolar septa from past cigarette smoke Figure 1. Framework illustrating how susceptibility arises from variability. Multiple types of biological variability intersect with the source-to-outcome continuum, either by modifying how changes to source/media concentrations propagate through to health outcomes or by modifying the baseline conditions along the continuum. The aggregate result of all these modifications is variability in how a risk management decision impacts individual health outcomes. The parame ters and initial conditions along the source-to-outcome continuum serve as indicators of differential susceptibility, some of which are more or less influential to the overall outcome (see Figure 2). exposure). Variation in these can confer susceptibility by altering the likelihood of progression from normal function to mild perturbations, early disease, and late disease. Systems dynamics describes the propaga tion of biological perturbations regardless of whether they are due to chemi cal expo sure, thus distinguishing it from pharmaco dynamics, which describes how chemi cal exposure causes biological perturbations. Figure 2 illustrates the distinct effects of different sources of variability on external dose, internal dose, or biological response. The first category of biological variability is indi cated by differences in the parame ters gov erning the relationship of one measureable quantity to the next (e.g., external to internal dose, and internal dose to biological response) (Figure 2A,B). In addition, there may be bio logical variability in the initial conditions for each measureable quantity, as well as the con tribution from the source of environ mental chemi cal exposure under consideration for risk management ( Figure 2C,D). For exam ple, increases in background exposure to the same or a different chemi cal(s) may result in saturation of metabolic activation and/or clearance processes, or temporary depletion of cofactors involved in detoxification, such as glutathione, resulting in either attenuation or amplification of the effect of additional increments of chemi cal exposure on internal dose ( Figure 2C). Nonetheless, a biological response with a low background level may be much less altered by additional exposure than one with a high background because of to the cooperativity associated with a relatively higher baseline internal dose ( Figure 2D).

Current Approaches to Addressing Variable Susceptibility
Variability for assumed thresholdlike doseresponse relationships is currently addressed by applying an "uncertainty" or "adjustment" factor (U.S. EPA 2011). The factor to account for inter individual variability in human popu lation has typically been 1, 3, or 10. In some cases, the factor is further divided to separately account for variation in PK and PD (U.S. EPA 2011; International Programme for Chemical Safety 2001). In this context, PD has included both PD and systems dynamics processes described above and in Figure 1. Data permitting, the PK component can be addressed through physiologically based pharmaco kinetic (PBPK) modeling, in which case a factor addressing only PD is applied (U.S. EPA 2011). Occasionally, exposureeffect observations are available for particularly susceptible human populations, such as with ozone and persons with asthma (U.S. EPA 2006), or those sensitive to chronic beryllium disease (U.S. EPA 1998), which allows for a datadriven estimation of the likely impact of inter individual variability on human health risk assessments.
For presumed nonthreshold cancer end points, inter individual variability is not cur rently addressed when risk is estimated from animal studies, with the exception that for mutagenic compounds exposures occurring early in life are weighted more heavily (by a fac tor of 10 between birth and 2 years of age, and a factor of 3 between 2 and 16 years of age). Cancer risk for susceptible populations, such as smokers who have been exposed to radon, may be calculated in addition to that for a general population (U.S. EPA 2003). Alternatively, adjustments may be made to address suscep tible subgroups, such as the sexspecific effects of 1,3butadiene (U.S. EPA 2002). There have been calls to formally account for variability in cancer dose response (NRC 2009).
Over the past 30 years, several strategies to characterize (predominantly PK) variability combining mathematical models and statistical distributions have developed in parallel. The first strategy, mostly used for datarich pharma ceuticals, couples empirical PK models and multilevel (random effect) statistical models to extract a posteriori estimates of variability from clinical data on patients or volunteers. This "population PK" approach (Beal and Sheiner 1982) seeks to measure variability and to dis cover its determinants. The second, the "pre dictive PK," approach takes advantage of the predictive capacity of mechanistic models and assigns a priori distributions to their parame ters (e.g., blood flows, organ volumes). The parame ters having biological meaning can be observed through independent experiments, clinical measurements, or surveillance. Table 1 lists some examples of data sources for developing a priori parame ter distributions. Monte Carlo simulations are used to propagate the distribu tions from model parame ters to model pre dictions (Portier and Kaplan 1989;Spear and Bois 1994). A third approach, the "Bayesian PBPK" approach, offers a synthesis of the other two, applying mechanism or chemi calspecific parame ter variability data from a variety of independent sources while using population observations of relevant biomarkers of internal exposure and effect to further inform parame ter variability (Allen et al. 2007;Bernillon and Bois 2000;Hack 2006). Parameter cova riance can be modeled by multivariate prior distributions (Burmaster and Murray 1998) or joint posterior distributions obtained by Bayesian multi level modeling (Bois et al. 1990;Wakefield 1996). A Bayesian PBPK model based analysis of the population toxicokinetics of trichloro ethylene (TCE) and its metabolites in mice, rats, and humans provides a practical example of how a syste matic method of simul taneously estimating model parame ters and characterizing their uncertainty and variability can be applied to a large database of studies on a chemi cal with complex toxicokinetics (Chiu et al. 2009).
PBPK models have been often used to assess variability on the basis of prior parame ter distributions obtained from in vitro experi ments or the physiological literature Jamei et al. 2009) and can include genetic information regarding variability. For example, PBPK models can inform the  implications of polymorphisms in metabolism genes (Johanson et al. 1999). The effects of such polymorphisms on PK of environ mental toxicants and drugs have been the subject of many empirical studies (reviewed by Ginsberg et al. 2009cGinsberg et al. , 2010. These polymorphisms are of particular concern for xenobiotics whose metabolic fate or mechanism(s) of action is controlled by a particular enzyme (Ginsberg et al. 2010), and in such cases genetic variabil ity can profoundly influence enzyme function with implications for internal dose (Figure 1). However, because enzymatic pathways with overlapping or redundant function and other pharmaco kinetic factors (e.g., blood flow limitation) can also influence metabolic fate (Kedderis 1997), PBPK models are needed to evaluate the implication of genetic polymor phisms in metabolizing enzymes in human health risk assessment (Ginsberg et al. 2010). The situation is somewhat different for PD and systems toxicology models. The biologically based dose-response models describe apical or intermediate end point responses as a function of PKdefined internal doses (Crump et al. 2010). However, models designed purely from our understanding of the disease process, such as the role of cytotoxicity and regenera tive pro liferation in carcinogenesis (Luke et al. 2010b), or the effect of dietary iodide and thyroid hor mones on the hypothalamic-pituitary-thyroid axis (McLanahan et al. 2008), require further development to reliably predict an adverse out come from tissue exposure (the last two arrows in Figure 1), or its variability. Understanding a disease process at the pathway level (i.e., PD and systems dynamics components of the sourceto outcome continuum) is in itself not sufficient to define reliable and informative mechanistic models because of great model sen sitivity to uncertain inputs. Most such models are based on equations derived from the classi cal receptor theory (Csajka and Verotta 2006) and focus on PD rather than system dynam ics elements of the disease process and do not attempt to model the full process from tissue exposure to disease outcome.

Emerging Data Streams on Biological Variability
Experimental populationbased paradigms to address intrinsic variability in response to exposure comprise multiple levels of biological organization, from molecules to whole bodies. Published examples, reviewed by Rusyn et al. (2010), include animal models and large scale in vitro screening platforms to study population based genetic determinants. Those studies have also aided in the identification of genetic susceptibility factors that underlie toxi city phenotypes. Complementary to these are genomewide (Hutter et al. 2012) and expo surewide (Patel et al. 2010) association stud ies for assessing human population varia bility.
Experimental in vitro data on genetic variability. Human cell lines obtained from genetically diverse subjects and multiple popu lations (Durbin et al. 2010) hold the promise of providing data for assessing genetic deter minants of different components of toxic response. Many recent studies have used human lymphoblastoid cell lines, represen tative of the genetic diversity in populations of European, African, Asian, and North and South American ancestry, to quantify inter individual and interpopulation variability in response to drugs (Welsh et al. 2009). Dozens of studies published in the past 5 years have profiled the cytotoxicity of single to as many as 30 drugs (mostly chemotherapeutics) in hundreds of cell lines. Diverse applications for such a populationbased cell model has been suggested. Drug class-specific signatures of cytotoxicity, which could indicate possi ble shared mechanisms, have been identified and replicated in both cell lines from different populations and for additional compounds . Furthermore, such stud ies may potentially inform the prioritiza tion of chemotherapeutic drugs with a sizable genetic response component for future investigation ) and assist in identifying germline predictors of cancer treatment out comes (Huang et al. 2011).
The utility of such in vitro models to toxicol ogy, especially for exploring the extent and nature of genetic components of inter individual variability in PD and systems dynamics, was recently demonstrated (Lock et al. 2012;O'Shea et al. 2011). Quantitative highthroughput screening (qHTS) pro duced robust and reproducible data on intra cellular levels of adenosine triphosphate and caspase3/7 activity (i.e., biological response) indica tive of general cytotoxicity and activa tion of apoptosis (i.e., physiological status), with utility for variability assessment as fol lows. First, standardized and highquality concentration-response profiling, with repro ducibility confirmed by comparison with pre vious experiments, enables prioritization of chemi cals based on inter individual variability in cytotoxicity. Second, genomewide associa tion analysis of cytotoxicity phenotypes allows exploration of the potential genetic determi nants of that variability. Finally, the highly significant associations between basal gene expression variability and chemi calinduced toxicity suggest plausible modeofaction hypotheses for followup analyses.
Several extensions of these studies can be envisioned to advance the identification of determinants of genetic susceptibility and vari ability in toxic response. Opportunities include the testing of additional, and more diverse, chemi cals (including major metabolites) and concentrations (to account for lower meta bolic capacity of these cells). Other specific end points could also be assessed. Further, these studies could be expanded to include larger panels of lymphoblasts and other cell types from genetically and geographically diverse populations. Development of related assay sys tems to monitor differences in susceptibility to perturbation of communication between cells (e.g., neurotransmission or differentiation sig nals) could address other aspects of variability not present in cultures comprising only one kind of cell. The development and use of these and other types of in vitro assays would be fur ther informed by quantitative comparisons of the PD inter individual variability measured in vitro with observable human pharmaco dynamics variability in vivo. Candidate chemi cals for this comparison would be selected environ mental toxicants (such as ozone) and pharmaceuticals that have been tested for responses in appreciable numbers of human subjects at different known exposure levels. The extent of inter individual variability in response that was observed for different chemi cals in in vitro assays could also be compared with previously collected sets of in vivo human PD variability data . Experimental in vivo data. Several proof ofconcept studies that utilized a "mouse model of the human population" have demon strated the potential for translation to clinical appli cations and for addressing both PK and PD components of variability (Guo et al. 2006(Guo et al. , 2007Harrill et al. 2009b;Kleeberger et al. 1997;Prows et al. 1997). For example, the extent and nature of TCE metabolism is an important consideration in relating adverse health effects in rodents to humans. Bradford et al. (2011) measured variability in PK for TCE using a panel of inbred mouse strains, revealing marked differences among indi vidual mice (e.g., a greater than 4fold differ ence in peak serum concentrations of TCE metabolites). These experimental data on intra species differences in TCE metabolism may be used to calibrate the variability in outputs of PBPK models, and thus inform quantitative assessment of variability in TCE metabolism across species.
With regard to PD variability, genetically diverse mouse strains can be used to under stand and predict adverse toxicity in hetero geneous human populations. For example, Harrill et al. (2009a) evaluated the role of genetic factors in susceptibility to acetaminophen induced liver injury in a panel of inbred mouse strains and two cohorts of human volunteers. The authors identified genes associated with differential susceptibility to toxicity in a preclinical phase. This finding has the potential to focus further toxico genetics research, overcome the challenges of studies in small human cohorts, and shorten the validation period. The data acquired with this model may be used in analyses of individual risk to toxicants. Furthermore, when combined with omics data collected on an exposed population of individual strains, it may be possible to explore underlying genotype dependent and independent toxicity pathways involved in PD response (Bradford et al. 2011;Harrill et al. 2009a).
Experiments such as these afford the opportunity to quantitatively understand the interplay between genetics, PD, and systems dynamics. In addition, genetically defined mouse models may be used to supplement the limited data from human studies to not only discover the genetic determinants of susceptibility and understand the molecular underpinnings of toxicity (Harrill et al. 2009a;Koturbash et al. 2011) but also to develop descriptions of variability for use in dose-response and mechanistic evaluation components of human health risk assessments.
Such rodent systems can also be used to assess the role of epigenetics, as well as its potential interplay with the genetic back ground, in susceptibility. For example, Koturbash et al. (2011) demonstrated that interstrain differences in susceptibility to 1,3 butadiene-induced genotoxicity may be due to strainspecific epigenetic events that are also part of a PD response.
Practical use of this type of experimental information is possible mainly when the mech anistic pathways to human adverse responses are better established. More general application will also depend on the development of suites of rodent models that more fully represent human diversity in both genetics and other factors, such as age (Hamade et al. 2010). Such studies can, in turn, provide important insights concerning the identity and extent of sources of variability that may arise in the sourceto outcome continuum for a given chemi cal class, physiologic state, or adverse response.
Human clinical and observational data. Genomewide association studies (GWAS) with disease severity as the phenotypic trait are used to associate genetic loci with risk for complex diseases (Rosenberg et al. 2010). Even though GWAS approaches have uncov ered numerous genomic loci that may affect the risk of human disease (Manolio 2010), the identified variants explain only a small proportion of the heritability of most complex diseases (Manolio et al. 2009). Some have sug gested that unexplained heritability could be partly due to gene × environment interactions, or complex pathways involving multiple genes and exposures (Schadt and Björkegren 2012).
The GWAS concept is now being applied to identify additional genotypedependent metabolic phenotypes and to gain insight into nongenetic factors that contribute to the effects of xenobiotics on system dynamics. In animal studies, metabolic phenotyperelated quantitative trait loci were shown to be use ful in understanding genome × phenotype relationships and how extended genome (microbiome) perturbations may affect dis ease processes through transgenomic effects (Dumas et al. 2007). In a series of human studies (Gieger et al. 2008;Illig et al. 2010;Suhre et al. 2011), serum collected from two large European cohorts (2,820 individuals in total) was analyzed with nontargeted metabo lomics, focusing on endogenous metabolites and covering 60 biochemi cal pathways. Ratios of metabolites to parent chemi cal concentra tions served as surrogates for enzymatic rate constants. Thirtyseven genes were associated with blood metabolite concentrations and, in some cases, explained a substantial fraction of the variance. Endogenous and xenobiotic metabolites (mostly of drugs) were studied.
Clinical (Brown et al. 2008;Hernandez et al. 2010) and epidemiological (Jia et al. 2011;Wood et al. 2010) studies of acute and chronic effects of ambient air exposures have long had important roles in quantifying human variability in the risks of exposures to widespread toxicants such as ozone and airborne particulates. The addition of GWAS to these established tools has the potential to widen the capability for quantification of effects on susceptibility of many individual genotypic variants that individually have rela tively modest effects (Holloway et al. 2012). Establishing the roles of individual pathways in affecting susceptibility via genetic analysis, in turn, has the potential to advance the assess ment of effects of other exposures during life that also affect the same pathways. Elucidating these determinants for prominent toxicants, however, requires a very considerable research effort. Nonetheless, this research paradigm provides opportunities to explore variability in adverse responses that is due to physiological states for which in vitro and experimental ani mal models are lacking.
Variability in human response to an agent stems in part from differences in the under lying exposures that contribute to a given dis ease response prevalence within the population. A person's internal "chemi cal environment" may be as important for possible disease asso ciations as exposures to the variety of chemi cals in the external environment. Under this "expo some" concept (Wild 2005), exposures include environ mental agents and internally generated toxicants produced by the gut flora, inflamma tion, oxidative stress, lipid peroxidation, infec tions, and other natural biological processes (Rappaport and Smith 2010).

Advances in in Silico Methods to Address Human Variability
Modeling of variability is expected to be needed for both datarich and sparse chemi cals. Recent advances in software, publicly available data and ongoing computational activities in biomedical research should facili tate the development and use of the results of this type of modeling.
Modeling the PK dimension of human variability. Commercial software prod ucts [e.g., by Simcyp (http://www.simcyp. com), Bayer Technology (http://www.pksim. com)] are available to explicitly address vari ability for pharmaceutical or human health risk assessment applications to, for example, adjust dosing for different target patient popu lations (Jamei et al. 2009;Willmann et al. 2007). Several of these offer generic PBPK models, applicable to "any" substance; how ever, their substancespecific parame ters have to be obtained from in vitro experiments (particularly on metabolism) or quantitative structure-property relationships. The variabil ity of subjectspecific physiological parame ters can be informed by compiled databases (see above) and literature searches Ginsberg et al. 2009c), and could include volume 121 | number 1 | January 2013 • Environmental Health Perspectives adjustments or protocols to address limitations in data availability. Quantitative structureproperty relationship models or in vitro data can also be used to derive substancespecific parame ters. These models are being applied in an exploratory fashion in in vitro-based assess ments (Judson et al. 2011;Rotroff et al. 2010).
Using a Bayesian multilevel population approach, some of the key parame ters of these generic models could be calibrated by integrating human observational data with data from lower levels of biological organization. This presents a computational challenge on a chemi calspecific basis, because those models are neither particularly parsimonious nor quickly evaluated. Yet an extensive calibration of a complex generic model for a selected number of datarich environ mental or pharma ceutical chemi cals could be used as support to develop generic approaches for PK varia bility treatment in human health risk assessment. For example, generalizations could be made about the extent to which particular enzymes may contribute to overall human PK. Extensions of the approach of Hattis et al. (2002) can also be developed to construct "bottom up" quantitative descriptions of PK variability that can be applied as defaults across classes of chemi cals.
Modeling the PD dimension of human variability. Semiempirical PD models can include observed biomarkers of susceptibil ity as covariates. Such models are increas ingly applied in predictive toxicity and human health risk assessment. Environmental epidemiol ogy also routinely models quantal types of biomarker data in logistic regressions. Harmonizing the tools and models of toxico logical risk assessment with those of epidemio logical risk assessment, and reconciling their data and results, should facilitate the develop ment of better approaches for background and variability descriptions in NexGen human health risk assessments.
Integrating PK and PD into a systems biology framework. The link between toxicity pathway and "normal cell physiology" models of systems biology could also be further devel oped and used as the basis to explore poten tial ranges of human variability. The potential of publicly accessible and curated biomodel and database repositories will be increasingly exploited as familiarity increases in the risk assessment and risk management communi ties. Importantly, systems biology models can describe background biological processes and the impact of their perturbation and provide a framework for exploring human variability and identifying susceptible populations for targeted assessment and management efforts. Although they come at the price of tremendous complex ity, their development can leverage the consid erable ongoing effort by the biomedical and pharmaceutical research community to support applications other than toxicant risk evaluation. Further, because of these largescale efforts, the necessity of sharing and standardization is well understood in the United States. The systems biology markup language (Hucka et al. 2003), for example, is a highlevel language developed explicitly to provide a common intermediate format for representing and exchanging sys tems biology models. Predictive toxicology will benefit from these developments.
The frontier for both PK and PD is in the integration of the rapidly growing informa tion about metabolic networks, receptors, and their regulation with toxicity pathways. The models so far most amenable to quantitative predictions are differential equation models. PBPK models will likely be merged with systems biology and virtual human models. The boundary between PK and PD actually tends to blur as metabolism becomes more and more integrated into detailed models of toxicity pathways when, for example, model ing enzymatic induction by xenobiotics (Bois 2010;Luke et al. 2010a). The variability of the different components of those models will be directly informed by time series of genomic, proteomic, metabolomic data on the chemi cal species considered. This may provide a framework for assessing the variability in susceptibility to chemi cally induced effects as influenced by possible metabolic interac tions as well as preexisting disease. In time this may facilitate computing the impact of, for example, single nucleotide polymorphisms on the reaction rates of enzymes and recep tors and translating these calculations to esti mates of human variability (Mortensen and Euling 2011). Ongoing work on simula tions of enzymatic reactions or receptor binding at the atomic level (e.g., the potassium channel pore) shows the way forward for predicting fundamental reaction rates by physical chem istry approaches. Prediction of the quantitative impact of sequence or aminoacid variation on the function of the reactive species involved in systems biology models is coming within reach (Giorgino et al. 2010;Sadiq et al. 2010).
Biologically based PD models, such as the systems biology models of response networks (Schuster 2008), models of toxicity pathway perturbations, and biologically based doseresponse models proposed to link biochemi cal responses to apical effects, clearly hold promise (Csajka and Verotta 2006;Jonsson et al. 2007;Nong et al. 2008) but face challenges similar to those that hampered the use of biologically based cancer models (Bois and Compton Quintana 1992;Chiu et al. 2010). To explore the extent of human variability in response to toxicant and stressor exposures, the various steps in the relevant causal path need to be modeled quantitatively and on a population basis. A problem is that the quantitative link ing of omics biomarkers to risk is missing. For many markers (e.g., of apoptosis, cell divi sion), the linkage to risk is highly uncertain (Woodruff et al. 2008), so the ranges of pos sible variability may be very large. Further, the ability to reinforce information by linking with the impact of injury on multiple targets is also limited because such links are generally not well understood.

Implications for NexGen Human Health Risk Assessments
Multiple "tiers" of human health risk assess ment needs, requiring different levels of precision, can be envisioned. These include screeninglevel analyses of multiple chemi cals to inform the prioritization of management and enforcement actions across communities, ensuring protection across the population to widespread exposure to legacy contaminants, or identifying subpopulations for which differ ing risk management options might be applied.
In the lowest (simplest) tier of assessments, evaluations are expected to primarily rely on the results of high and medium throughput in vitro screening tests in mostly human cell lines, as well as complementary in silico predic tive methods. The Tox21 collaboration (Collins et al. 2008) is leading the field in exploring how a broad spectrum of in vitro assays, many in qHTS format, can be used to screen thousands of environ mental chemi cals for their potential to disturb biological pathways that may result in human disease (Xia et al. 2008). Such data on toxicologically relevant in vitro end points can be used as toxicity based triggers to assist in decision making , as predic tive surrogates for in vivo toxicity Zhu et al. 2008), to generate testable hypotheses on the mechanisms of toxicity (Xia et al. 2009), and to develop screening assays based on pathway perturbations. The extent of inter individual variability in toxic response to be estimated from these types of assays can be informed by empirical data and PK/PD models that address multiple factors in the source to outcome continuum as described in Figure 1. The genomic component of variability may be partially informed by test data from geneti cally diverse but welldefined human cell lines, such as from the HapMap (http://hapmap. ncbi.nlm.nih.gov/) and 1000 Genomes (http:// www.1000genomes.org/) projects. For exam ple, emerging data based on standardized and highquality concentration-response profiling can help inform characterizations of the extent of inter individual variability in cytotoxicity. When chemi calspecific estimates are lack ing, the range of inter individual variability for structurally related compounds may be infor mative, in a readacross approach. Quantitative data characterizing the range in response (e.g., size and variance) may be integrated with probabilistic default distribu tions addressing the remaining key sources of inter individual variability. Quantitative estimates of PK vari ability would be also incorporated. In addition, factors such as life stage and background expo sures may be particularly important consider ations for approaches accounting for baseline differences in the spectrum of the "chemi cal environment" (Rappaport and Smith 2010), in interpreting results from the omics assays, and in evaluating the potential contributions of non genetic variability factors.
At these lower tiers, a probability distri bution may best acknowledge the many uncer tainties involved in making inferences with limited data. Systematic analyses of chemi cal sets will be needed to refine distributions for the chemi calspecific and general case. For instance, external comparisons of in vitro mea sures based on genetic variability in pharmaco dynamics to in vivo observations may inform the choice of distribution used for a particular chemi cal or chemi cal category. Standard cate gories, comprising different size and variance distributions for multiple variability factors that can then be applied to other chemi cals, may emerge from these analy ses. The ranking and grouping of chemi cals for the applica tion of these distributions may be based on structural class, the relative extent of observed varia bility, a common determinant of vari ability (e.g., as identified in GWAS analysis of cyto toxicity pheno types), or other factors (e.g., likelihood of coexposures or confounders). Compounds demonstrated or predicted to have highly variable toxic responses may also be given a higher priority for further study, in combination with chemi cal and other expected modifiers of susceptibility.
At higher tiers of NexGen human health risk assessments, animal and in some cases human data are available for evaluating doseresponse relationships, major pathways for some of the critical toxicities for risk assess ment can be reasonably well understood, and some in vivo human data relevant to those pathways may be available. For some chemi cals, sensitive populations may have been iden tified and studied using omics technologies. In the case of ozone, for example, gene expression data and genomic markers may be collected on individuals of high and average sensitivity. Toxicity pathways exhibited in cultured air way epithelial cells exposed to ozone may also be compared with those in humans exposed in vivo to ozone. Such data will aid a better characterization of the dose-time-response severity relationships at low doses. In other cases, where individuals are studied epidemio logically, the current bioinformatics analyses lack power and require pooling of subjects to detect trends, losing variability estimation in the process. In such cases, there will be a need to couple default descriptions of PD variabil ity with PBPK modeling to obtain an overall prediction of variability. In the future, new hypothesisbased molecular clinical and epide miological approaches that integrate emerging biological knowledge of pathways with obser vations of physiological disease status, mark ers of early biological response, and genetics are likely to provide the way forward with populationbased descriptions of variability.

Conclusions
Emerging data streams can inform multiple aspects of biological variability, be used in different modeling approaches addressing PK and/or PD variability, and have applica tion across different chemi cal screening and evaluation schemes. Successful examples of addressing PK variability include the devel opment and application of a Bayesian PBPK model-based analysis syste matically estimat ing model parame ters and characterizing their uncertainty and variability for TCE, a chemi cal with complex toxicokinetics (Chiu et al. 2009). Additionally, data from animal models and largescale in vitro screening platforms that have incorporated population based genetic determinants (reviewed by Rusyn et al. 2010), have provided insight into the extent of genetic variability in response to a diversity of toxicants, as well as aided in the identi fication of genetic susceptibility factors that underscore the development of toxic pheno types. Hypothesisbased molecular clinical and epidemiological approaches to integrating genetics, molecular pathway data, and clini cal observations and biomarkers are likely to contribute to populationbased descriptions of variability. Complementary to these are genomewide (Hutter et al. 2012) and expo surewide (Patel et al. 2010) approaches for assessing human population variability in toxic response. Opportunities exist to employ these emerging data streams in the development of in silico predictive models for application in a range of decisionmaking contexts.