The use of evidence from high-throughput screening and transcriptomic data in human health risk assessments

Abstract

High throughput screening (HTS) and functional genomics (toxicogenomics) have opened new avenues in toxicity testing. Their advantages include the potential for developing short-term in vivo bioassays and in vitro assays in order to keep pace with the growing backlog of chemicals that need to be evaluated for potential human health risk. In addition, these approaches have the potential to address some of the difficulties that arise with interpreting traditional rodent bioassays, such as the relevance of apical outcomes induced by chemical exposure in animals to humans. The wealth of information associated with the HTS and toxicogenomic data can inform human health risk assessment primarily through (i) insight into potential mechanism of action, (ii) prediction of adverse outcomes of chemical exposures, and (iii) dose-response assessment for derivation of toxicity values. In this article we outline current and expected future progress in these three directions and argue for increased role of HTS and toxicogenomic data in chemical risk assessment. We conclude that these approaches can help fulfill the NRC vision for toxicity testing in the 21^st century and we discuss specific examples of chemicals whose health assessments can potentially benefit from available HTS or toxicogenomic data.

1. Introduction

Human health risk assessment (HHRA) is the process that estimates potential adverse health effects in humans who may be exposed to chemicals in contaminated environmental media, now or in the future (EPA, 2016). This process often relies on conventional rodent bioassays (Waters, 2016) that typically evaluate clinical signs, histopathologic findings and clinical chemistry/hematology attributable to the effects of chemicals and assess these apical endpoints in the context of human relevance and dose-dependence.

In its vision for toxicity testing in the 21st century, the NRC recommended a shift away from apical endpoints and toward a more focused use of in vitro data and in silico methodologies, including genomic and post-genomic approaches (NRC, 2007). This recommendation was spurred by the fact that long-term rodent assays were becoming increasingly less viable because of their high-cost, time- and resource-intensiveness, and an inability to keep pace with the backlog of chemicals that need to be evaluated for potential human health risks. In addition, the use of traditional bioassay data in risk assessment involves numerous intricacies, such as, the issue of human relevance in hazard identification and the need to extrapolate effects in dose-response assessment from higher doses in animal studies to low doses typical for human environmental exposures (Ward, 2007).

Approaches that show promise in enabling the 21^st century vision include the use of high-throughput screening (HTS) that allows testing of numerous compounds in parallel by biochemical or cell-based assays, as well as high-density multiplexed (‘omics) methods. The high-density multiplexed methods, such as genomics, epigenomics, transcriptomics, proteomics and metabolomics, study systemic molecular (‘omic) responses in cellular systems or whole organisms. Application of the methods of functional genomics in toxicology, which include transcriptomics, proteomics and metabolomics, has been recognized collectively as toxicogenomics.

Considering the wealth of information associated with the HTS and toxicogenomics data, these approaches can inform human health risk assessment primarily in three ways. First, they can provide insight into potential mechanisms of action (mechanistic toxicogenomics or HTS analysis), which can support the plausibility of exposure-outcome associations found in epidemiological studies or inferences regarding the human relevance of the outcomes of animal studies. Second, these approaches can help predict adverse outcomes and support the hazard identification of chemicals for which toxicological information is limited (predictive toxicogenomics). Third, these approaches can determine exposure levels at which specific ‘omic responses become abnormal and which can be used for estimating points of departure for toxicity value derivation. This paper outlines current progress and future prospects in these three directions.

2. The use of high-throughput screening to support HHRA

High-throughput screening (HTS) allows parallel tests of large compound libraries against selected molecular or cellular targets or whole organisms, such as nematode Caenorhabditis elegans and the zebrafish Danio rerio (Szymański et al., 2011). This approach, which typically allows screening of 10,000 – 100,000 compounds per day (Mayr and Fuerst, 2008), became established over last two decades primarily in drug discovery and development, where it is used for screening of activity against therapeutic targets. More recently, however, HTS approaches have been used to estimate toxicity and to provide understanding of mechanisms of action of a large number of chemicals (Houck and Kavlock, 2008; Seidle and Stephens, 2009; Shukla et al., 2010; Choudhuri et al., 2018).

The Toxicity Forecaster (ToxCast™) program operated by the National Center for Computational Toxicology (NCCT) within the U.S. Environmental Protection Agency (EPA) employs high-throughput screening (HTS) to implement recommendations from the National Research Council ‘s (NRC) report titled “Toxicity Testing in the 21st Century”. This report called for the development of a new toxicity-testing system based on rapid in vitro assays that evaluate biologically significant perturbations of cellular response pathways critical to human health, including dose-response modeling (NRC, 2007). This system, combined with targeted animal testing, which would complement in vitro data and support their interpretation, was envisioned to enable rapid screening, reduce backlog of untested chemicals, and substantially reduce animal use (NRC, 2007). The ToxCast™ program generated, processed and made publicly available data that also include the “Tox21” HTS data produced and analyzed by the NIH National Center for Advancing Translational Sciences (NCATS) and the National Toxicology Program (NTP). The ToxCast™ bioactivity data is also complemented with high-throughput exposure estimates produced by Exposure Forecasting (ExpoCast) program (Wambaugh et al., 2013), all housed currently on the EPA’s CompTox Chemicals Dashboard (https://comptox.epa.gov/dashboard).

As of February 2019, ToxCast™ and Tox21 bioactivity data were generated for 9,076 compounds, which were selected based on availability of animal toxicity data or mechanistic insight, significance of human exposure, and/or regulatory interests (Richard et al., 2016). The criteria used for the selection of chemicals reflect the dual purpose of this program, which includes identification of toxicity pathways and relationships between HTS results and adverse effects, as well as addressing data gaps for chemicals of regulatory concern. While the former purpose was met through sufficient coverage of biological activity space by chemicals with known toxicological properties, the latter called for inclusion of broad range of chemicals with limited data availability and significant exposure or regulatory concerns.

2.1. Description of ToxCast™ /Tox21 bioactivity data

ToxCast™/Tox21 bioactivity data are generated by in vitro cell-free (biochemical) and cell-based assays in human and rodent primary cells or cell lines, which measure a wide spectrum of biological responses to specific chemical exposures, including cell proliferation, cell death, genotoxicity, and activities of enzymes, ion channels, receptors, or transcription factors (Judson et al., 2010). Because the links between individual assays and specific toxicity endpoints are not always defined or validated, a large number of assays have been included to this HTS campaign to ensure extensive sampling of the biological response space and to provide results relevant to major cellular responses, protein targets and key signaling pathways (Houck and Kavlock, 2008). This extensive interrogation of biological space was made possible by advances in cellular and molecular biology and HTS technology, which permitted inclusion of diverse cellular and molecular targets into this highly parallel screening. As of February 2019, the ToxCast™ /Tox21 bioactivity data represent 1192 assay endpoints derived from 763 assay components, which were in turn produced from 360 single readout or multiplex in vitro assays (EPA, 2018a) tested in several high-throughput screening (HTS) platforms (EPA, 2018b).

The raw ToxCast™/Tox21 data generated by diverse HTS platforms are transformed, normalized, modelled for concentration-response (in multiple-concentration screening) and visualized using an open source ToxCast™ Pipeline R package (Watt and Judson, 2018). The raw and processed bioactivity data are available as downloadable files, and the processed data can be also accessed interactively via a web Dashboard (EPA’s CompTox Chemicals Dashboard, https://comptox.epa.gov/dashboard/).

Single-concentration screening identifies potentially active compounds, while multiple-concentration screening can also estimate efficacy, potency and points of departure (PODs) through concentration-response modeling. The concentration-response is modelled using 3 models: (i) a constant model at zero, (ii) a constrained three-parameter Hill model, and (iii) a constrained five-parameter gain-loss model, a product of two Hill functions that allows non-monotonous concentration-response curves. The model with lowest Akaike information criterion (AIC) is selected as the ‘winning’ model; however, if two models have equal AIC values, the model with fewer parameters is selected.

A compound is considered active in a given multiple-concentration assay, if all of the following conditions are met: (i) the winning model is Hill or gain-loss, (ii) the maximum modelled response exceeds efficacy cutoff, and (iii) median response exceeds efficacy cutoff for at least one concentration. AC50 values are provided for the Hill and gain-loss models as a measure of potency and they correspond to concentrations (in μM) at which the activities reach 50% of their maximal values. Other reported data and model summaries include point of departure (POD) estimates: AC10 (concentration at activity equal to 10% of maximal activity), ACB (concentration when activity first reaches 3xBMAD (baseline median absolute deviation for the first two concentrations)), and ACC (concentration at activity cutoff). Additional details are available in the ToxCast™ Owner’s Manual – Guidance for Exploring Data (https://www.epa.gov/sites/production/files/2018–04/documents/toxcastownermanual4252018.pdf). An example result via graphical output for a concentration-response assay is depicted in Figure 1.

Figure 1:

Results of TOX21_p53_BLA_p1_ch2 assay for sodium dichromate dihydrate (SDD; CASRN:7789–12-0). Human colon cancer cells HCT116 were exposed to increasing concentrations of SDD in growth medium for 24 hours in a 1536-well plate format. Assay determines induction of TP53 gene. Winning model: Gain-Loss: (AIC=300.05; Top=63.82%; AC50=6.03 μM; slope=0.91); log ACC=0.406; logAC10=−0.267). Cutoff: 5xBMAD. Color coding: blue: Gain-Loss model; red: Hill model; yellow: constant at zero. Blue vertical line: AC50; Red vertical line: cytotoxicity limit.

A significant issue is the occurrence of potentially false positive or false negative assay results which need to be identified and flagged. At least for this purpose, and thereby to increase the confidence of interpretations and the usability of HTS data in HHRA, it is important to characterize the uncertainties in estimates of concentration-response parameters such as potency and efficacy. This step also improves the use of resources by flagging specific assay-chemical data with high variability that need manual inspection or re-testing. The characterization of uncertainties in parameters related to the concentration-response modeling of ToxCast™/Tox21 assay data has been challenging in the least due to enormous data volume and assay diversity. Consequently, these uncertainties have generally not been explored and they could contribute, among other factors, to lower performance of some predictive models built from ToxCast™/Tox21 data. Recently, smoothed nonparametric resampling has been employed to determine uncertainties in potency parameters, hit calls and model selections and applied to ToxCast™ estrogen receptor activity model that integrates endpoints from 18 assays (Watt and Judson, 2018).

2.2. Interpretation of the ToxCast™ /Tox21 bioactivity data and significance for HHRA

ToxCast™ /Tox21 bioactivity data represent results of a battery of in vitro HTS assays applied to a large set of diverse chemicals. Originally, the major goals for the use of these data included (i) identification of in vitro assays that can reliably detect alterations of biological processes with known toxicological significance in vivo, (ii) development of classifiers to predict toxicity of chemicals based on their response profiles across multiple assays, and (iii) the use of these classifiers, which may also include chemical structure descriptors, to screen large collections of chemicals for prediction of potential toxicity and prioritization for other toxicological studies.

The use of these data for human health risk assessment is presently under active development. The International Agency for Research on Cancer (IARC) used ToxCast™ /Tox21 bioactivity data for the first time in its Monograph program volume 112 (IARC, 2017) to evaluate whether chemicals may act through the 10 key characteristics of human carcinogens (KCC) defined by Smith et al. (Smith et al., 2016) For this purpose, the IARC Working Group and IARC Monographs staff independently explored 821 then-available ToxCast™ /Tox21 assay endpoints and mapped them to one or more KCC. Consensus was achieved for 263 endpoints that could be aligned with 7 of the 10 KCCs (Table 1).

Table 1.

Numbers of ToxCast™/Tox21 assays^* and genes aligned with key characteristics of human carcinogens (KCC).

Key characteristic #	Number of endpoints	Genes
KCC #1 (Is electrophilic or can undergo metabolic activation)	31	-CYP inhibition: CYP19A1, CYP1A1, CYP1B1, CYP2A6, CYP2B6, CYP2C18, CYP2C19, CYP2C8, CYP2C9, CYP2D6, CYP2E1, CYP2J2, CYP3A4, CYP3A5, CYP4F12, Cyp1a1, Cyp1a2, Cyp2a1, Cyp2a2, Cyp2b1, Cyp2c11, Cyp2c12, Cyp2c13, Cyp2c6v1, Cyp2d1, Cyp2d2, Cyp2e1, Cyp3a23/3a1, Cyp3a2
KCC #2 (Is genotoxic)	9	-Assays for TP53 activity (9 endpoints)
KCC #3 (Alters DNA repair or causes genomic instability)	0	-
KCC #4 (Induces epigenetic alterations)	11	-DNA binding (4 endpoints) GLI1, PAX6, SOX1, SP1 -Histone modification (7 endpoints) HDAC3, HDAC6, SIRT1, SIRT2, SIRT3, MTHFR, Comt
KCC #5 (Induces oxidative stress)	18	-Metalloproteinase activity (5 endpoints) MMP1, MMP9, TIMP2 -Oxidative stress (7 endpoints) ATF6 -Oxidative stress markers (6 endpoints) HIF1A, MTF1, NRF1, NFE2L2, HSF1
KCC #6 (Induces chronic inflammation)	45	-Cell adhesion (14 endpoints) SELE, HLA-DRA, ICAM1, VCAM1, SELP -Cytokines (29 endpoints) CXCL8, CCL2, CXCL9, CCL26, CCL2, IL1A, CXCL10, IL6, CXCL8, IL1A, CD40, TNF, CD38, CD40, CD69 -NF-kB activity (2 endpoints) NFKB1
KCC #7 (Is immunosuppressive)	0	-
KCC #8 (Modulates receptor-mediated effects)	81	AHR, AR, Ar, ESR1, ESR2, NR1H4, NR1I3, NR1H2, ESRRA, ESRRG, NR3C1, HNF4A, NR1H3, NR4A2, RORB, RORA, RORC, RXRA, RXRB, THRA, PGR, Nr3c2, THRB, PPARA, PPARD, PPARG, VDR, NR1I2, CYP24A1**, RARA, RARB, RARG
KCC #9 (Causes immortalization)	0	-
KCC #10 (Alters cell proliferation, cell death, or nutrient Supply)	68	-Cell cycle (16 endpoints) MYB, MYC, TGFB1, XBP1, KDR, CSF1, EGFR -Cytotoxicity (41 endpoints) -Mitochondrial toxicity (7 endpoints) -Cell proliferation (4 endpoints)

^*Details on specific assays are available through iCSS ToxCast™ Dashboard at https://actor.epa.gov/dashboard/

^**Gene encoding enzyme that modifies a ligand of nuclear receptor VDR

This alignment allowed linking respective chemicals to mechanisms associated with carcinogenesis through their activity in ToxCast™ /Tox21 HTS system. The activities of the chemicals were scored as “active” or “inactive” for each of these aligned endpoints through evaluation of raw data as previously described (Sipes et al., 2013). Cumulative score ToxPi (Toxicological Priority Index) (Reif et al., 2010) was determined for each chemical as a measure of its association with “key characteristics” relative to 178 other compounds that have been previously evaluated by IARC, and for which ToxCast™ /Tox21 assay data are available. Of these 178 compounds, 8 were classified as carcinogenic to humans (Group 1), 16 as probably carcinogenic to humans (Group 2a), 58 as possibly carcinogenic to humans (Group 2b), and 95 were not classifiable (Group 3). One chemical was classified as probably not carcinogenic to humans (Group 4), and this group includes caprolactam as the only chemical classified by IARC as probably not carcinogenic to humans. The ToxPi approach, which has been developed for transparent integration of data across multiple sources, has been used to integrate multiple assay endpoints within KCC using an interactive ToxPi graphical user interface (GUI) (Reif et al., 2013).

This approach has been applied to malathion and its toxic metabolite malaoxon. Malathion was found to be active in 21 assay endpoints related to KCC #1; one endpoint related to KCC #4; three endpoints related to KCC #5; 17 endpoints related to KCC #8, and inactive across assay endpoints related to KCCs #2, #6, and #10. In contrast, its metabolite malaoxon was found inactive in endpoints related to KCC#4 and KCC #8, but active against three endpoints related to KCC #6. Interestingly, in assays for KCC #6, malaoxon ranked second in activity among 178 chemicals included in the comparison, while malathion showed no activity for any included assay endpoint. For KCC #5, both malathion and malaoxon showed intermediate activity among chemicals tested. Taken together, of 7 “key characteristics” covered by aligned HTS assay endpoints, malathion and malaoxon displayed activity in at least one assay endpoint aligned to 5 “key characteristics” of human carcinogens (Figure 2). The use of ToxCast™ /Tox21 data in the HHRA of malathion by IARC underlines the importance of the activity of its metabolite malaoxon in assays aligned with KCC#6 (induces chronic inflammation), where malaoxon ranks second among 178 IARC evaluated chemicals, due to its activity in selective assay endpoints related to cytokine and cell-adhesion activity.

Figure 2:

ToxPi ranking of 185 chemicals based on activity across 273 assay endpoints aligned to “key characteristics” of human carcinogens. Colored slices correspond to 7 of 10 “key characteristics”, to which IARC Working Group mapped ToxCast™/Tox21 assay endpoints. Horizontal axis: rank of chemicals; vertical axis: cumulative (ToxPi) score. ToxPi charts are shown for malathion, its metabolite malaoxon and for clomiphene citrate, which represents the highest-ranked chemical for activity across KCCs. Color coding: Lime: characteristic 1, blue: characteristic 2, yellow: characteristic 4, grey: characteristic 5, black: characteristic 6, tomato: characteristic 8, turquoise: characteristic 10. Blue x-symbols in the scatter plot depict malaoxon, malathion and clomiphene citrate

On the other hand, both malathion and malaoxon were found to be inactive in the 9 assay endpoints aligned with KCC #2 (genotoxicity), whereas IARC assessed evidence for genotoxicity of malathion to be strong based on studies in exposed humans, human cells in vitro and animal systems that showed effects ranging from gene mutations to chromosomal aberrations and micronuclei formation. The absence of activity on endpoints aligned to genotoxicity is likely related to the limited coverage of the 9 assays included in this group, all of which test for the expression of the TP53 gene. Probably due to this limitation, the nine TP53-relevant assays were not included in the KCC #2 in the IARC assessments published in volume 113 (June 2015). Concerns about possible low stability of malaoxon in assay systems were not substantiated, since it was found to be active in several assay endpoints that included both, cell-free and cell-based assays. As a result, the lack of positive results in assays mapped to KCC #2 are not attributable to instability of malaoxon in solutions.

Interpretation of ToxCast™/Tox21 data in the context of 10 key characteristics of human carcinogens (KCC) have also been used in IARC monograph volume 115. In this volume, IARC updated the alignment of ToxCast™/Tox21 assays to the “key characteristics” of human carcinogens that included 281 assay endpoints (ref 115). Nevertheless, grouping ToxCast™/Tox21 assays to KCCs by IARC failed to serve as the basis for development of predictive models for discrimination between carcinogens and non-carcinogens, which raised doubts about the usability of this approach for carcinogenic hazard identification (Becker et al., 2017). Consequently, the use of KCCs in the context of ToxCast™/Tox21 data needs additional development to reach a more mature stage in which it could support chemical risk assessment.

A similar approach was used by Rager et al. to interpret 113 Tox21 assay endpoints available for sodium dichromate dihydrate (SDD) (Rager et al., 2017) where they complemented their analysis with interpretation of data on hexavalent chromium-gene/protein interactions from in vitro studies in the Comparative Toxicogenomics Database (CTD) (Davis et al., 2019). The highest number of positive assay endpoints mapped to KCC# 2 (is genotoxic) and KCC#10 (cell proliferation, cell death and nutrient supply). Querying of the CTD filtered for relevant entries identified by in vitro studies produced lists of 3502 molecules supported by at least one study. Pathway enrichment analysis identified canonical pathways enriched by these genes, including apoptotic signaling and p53 signaling pathways, which supported the results of ToxCast™/Tox21 assays.

In addition to supporting ToxCast™/Tox21 results, the Comparative Toxicogenomics Database (CTD) was employed to identify associations between in vitro assays and adverse outcomes, which can potentially support the use of ToxCast™/Tox21 results to inform HHRA. Hu et al. (Hu et al., 2015) mapped 1438 chemicals tested in ToxCast™/Tox21 to 16,588 unique chemical-disease associations in the CTD. The diseases were grouped using a hierarchical disease vocabulary into cardiovascular, liver, nervous system and kidney toxicity categories, and 429 ToxCast™ chemicals were associated with these 4 toxicity categories that encompassed 639 unique disease conditions. Strong associations with the 4 toxicity categories were found for 88 out of 821 ToxCast™/Tox21 assay endpoints.

The activity of a chemical in a specific assay indicates the ability of the chemical to perturb one or more in vitro targets, which may be associated with in vivo endpoints that could lead to human disease. However, perturbation of one or several of these in vitro targets is not sufficient on its own for determination that a specific chemical causes a given adverse outcome. This is, at least in part, due to the fact that the ultimate effects of chemicals in vivo are further modulated by numerous factors, such as exposure, pharmacokinetics, metabolism, and the ability of animals to mount adaptive responses to the perturbations observed in vitro (Auerbach et al., 2016).

Importantly, not all positive calls detected represent compounds that are truly active in vitro. When single compounds are evaluated using ToxCast™/Tox21 bioassay data, the quality of the assay data need to be examined to rule out false positive calls. This examination includes assessment of the shape of concentration-response curves, effect-sizes, and consistency in concentration-response trend. It needs to be considered that the data processing package for raw ToxCast™/Tox21 bioassay data was developed to minimize false negatives trading-off for higher tolerance to false positive findings, as is a typical approach in high-throughput screening. Flags, which can help to identify false positive or false negative findings, are generated by the processing pipeline and reported so that anomalies can be visually examined during quality control (for details see ToxCast Data Pipeline Overview (https://www.epa.gov/chemical-research/toxicity-forecaster-toxcasttm-data).

Interpretation of the data needs to consider specificity with respect to the number of assays scoring positive for a specific tested compound and the number of compounds that test positive in specific assays. Information that can help in this consideration, such as % positive results for a specific assay among all tested compounds, is available in the downloadable ToxCast™/Tox21 summary files (EPA, 2018b). However, there are no guidelines recommending at which percentage of positive results an assay would be considered promiscuous and non-informative, and several factors needs to be considered on a case-by-case basis. These factors would include assay types and properties of the chemicals. In general, chemicals, which induce cell damage through high chemical reactivity or through their physico-chemical properties, or chemicals that non-covalently interact with numerous targets can display pleiotropic effects and score positive in multiple bioassays.

If cytotoxicity is detected, there is often a need to distinguish a non-specific cytotoxicity due to general cell stress from the more targeted cytotoxicity triggered by disruption of specific molecular functions. Likewise, certain targets can be activated non-specifically at higher exposure levels due to general cytotoxicity, which needs to be distinguished from specific chemical-induced activation of these targets. This problem can be addressed using data available in downloadable ToxCast™/Tox21 summary files (EPA, 2018b), where 35 “burst assays” potentially responding to non-specific cell stress are indicated. For a given chemical, the median AC50 value of all active “burst assays” is calculated and its lower 95%-confidence interval is considered as cytotoxicity limit (see Figure 1, red vertical line). The most informative results are those that display modulation of activity of specific targets or pathways at concentrations below the levels at which cytotoxicity is observed. Indeed, statistically significant inverse association has been found between number of pathways disrupted at low concentrations of specific chemicals and their in vivo LOAEL values. Assays scoring positive at concentrations above this cytotoxicity limit should be interpreted with care (Judson, 2016). Interpretation of ToxCast™/Tox21 data in IARC in monographs 112, 113 and 115 did not include formal adjustment of assay results to account for the influence of cytotoxicity on some assay results. Because cytotoxicity was shown to influence results particularly for assays mapped to the key characteristic of oxidative stress, adjustment for cytotoxicity is warranted prior to the interpretation of ToxCast™/Tox21 assay results in the context of cancer HHRA (Becker et al., 2017).

Comparison of ToxCast™/Tox21 data with traditional toxicology studies in vivo for two data-rich pesticides displayed concordance in some and non-concordance in other assay endpoints (Silva et al., 2015). Interestingly, discordances mainly consisted of results that were negative in vitro and positive in vivo (false negatives), which were likely caused by a lack of metabolic activation or limitations in assay design for some endpoints. This finding demonstrates that interpretation of negative results in vitro needs to include consideration of confounding factors potentially affecting interpretation of some HTS results.

Another study, compared Tox21 assay results for 130 chemicals with corresponding rat liver transcriptomics data (Klaren et al., 2018) from public database DrugMatrix(Roter, 2005). in vivo transcriptomics data were limited to 24-hour exposures, and when not available, data from 3–5-day exposures were used. For in vivo data, differentially expressed genes were identified and analyzed by pathway enrichment analysis to provide significantly enriched canonical pathways from Ingenuity® System Knowledgebase. Tox21 assay results were mapped to the canonical pathways based on modulated receptors or type of biological responses measured in these assays. This approach allowed the alignment of 40 Tox21 assay endpoints to 18 canonical pathways and subsequently the evaluation of concordance between corresponding in vitro and in vivo data. The results showed 41–100% agreement on a per-chemical basis between canonical pathways inferred for in vitro and in vivo data. Tox21 assay endpoint results had better negative predictive value for in vivo transcriptomic results than positive predictive value, with 89% agreement for negative Tox21 results, but only 13% agreement for positive Tox21 results with corresponding in vivo results. Taken together, both studies discussed above found appreciable degree of concordance between ToxCast™/Tox21 results and in vivo results, although their conclusions differ concerning the ability of negative ToxCast™/Tox21 results to predict negative in vivo results correctly.

Models that have been developed to predict bioactivity from ToxCast™/Tox21 assay results display variable performance. Thomas et al. reported modeling 60 in vivo endpoints using data from ToxCast phase I chemicals. The models were built using 84 different statistical methods using data from more than 600 ToxCast assay endpoints with and without inclusion of chemical structural descriptors, and in vivo data available in the ToxRefDb database. The study found the predictive models to perform poorly overall (Kleinstreuer et al., 2012). Similarly, predictive models for carcinogenicity that incorporated assay endpoints aligned to 7 key characteristics of carcinogens failed to discriminate better than chance between substances determined to have human cancer hazard potential and substances not posing cancer hazard (Becker et al., 2017). In contrast, others reported better performance of models based on ToxCast™/Tox21 assay data. For instance, a model predicting rodent carcinogenicity, which was trained on 232 chemicals using results from 672 assay endpoints and in vivo data from the ToxRefDB database was found successful in discriminating between “possible/probable/likely carcinogens” and “not likely carcinogens/evidence of no carcinogenicity” classes in an external set of 33 compounds (Kavlock et al., 2012). In addition, assays associated with rodent cancers in this model were also associated with genes, pathways or biological processes relevant to cancer development and progression (Kavlock et al., 2012). Promising predictive signature models have been developed for rat liver cancer, reproductive toxicity, developmental toxicity, and developmental toxicity from vascular disruption (reviewed in (Richard et al., 2012)).

The ToxCast™/Tox21 program has been conceptualized to screen the activity of large sets of chemicals for biological activity in model systems in vitro; therefore, extrapolations of these results to the level of whole in vivo organisms have inherent limitations. For this reason, classification of chemicals as active by predictive multi-assay models developed from ToxCast™/Tox21 data does not constitute conclusive evidence for specific toxicity. Nevertheless, their use for prioritization of chemicals for toxicity studies is largely accepted and the robustness and applicability of predictive models is expected to increase with time (Richard et al., 2012). Responsible use of ToxCast™/Tox21 data is contingent on careful consideration of the complexity of data generated and some intricacies that include: quality control of chemicals; flagged assay endpoints due to autofluorescence and other interferences; and cytotoxicity burst interference and propagation of uncertainties in data processing that may influence assay endpoint hit calls. The incorporation of AOP or MOA frameworks have been recommended for interpretation of ToxCast™/Tox21 results and for feature selection prior to building statistical or machine learning predictive models (Richard et al., 2012).

ToxCast™/Tox21 bioassay data, when available, can serve as a valuable resource supporting chemical risk assessment through increased mechanistic evidence. For instance, these data can help to interpret the results of other in vitro studies, whose results may be found less reliable because of identified risk of bias or due to other identified problems in study design or execution. Likewise, ToxCast™/Tox21 bioassay data can help reconcile conflicting results of published mechanistic studies. This is supported by the fact that several biological endpoints are covered by multiple ToxCast™/Tox21 bioassays, and consistency among the results of these orthogonal assays and/or biologically related targets can further strengthen confidence in conclusions about specific activity and mode of action of chemicals tested in vitro.

In addition to hazard identification and validation of results of other mechanistic studies, ToxCast™/Tox21 data can be examined to ascertain whether adverse effects observed at high exposure levels of a given chemical can also be expected at lower exposure levels through consistency of bioactivity at concentrations below levels at which “burst assays” display activity for this chemical.

Likewise, ToxCast™/Tox21 concentration-response data can be used to estimate in vivo POD values (PODs) as previously mentioned (Wetmore et al., 2013). In vitro bioactivity data, in conjunction with in vitro to in vivo extrapolation (IVIVE), have been used to derive oral equivalent doses, which were found to be consistent with lowest observable effect levels (LOEL) identified from in vivo studies, supporting the potential of HTS in vitro data in quantitative risk assessment (Silva et al., 2015). Another study used in vitro PODs derived from ToxCast™/Tox21 data for several hundred chemicals together with in vivo PODs for murine and rat livers from the Toxicity Reference Database (ToxRefDB) (Martin et al., 2009) to build a robust regression model to predict in vivo PODs from ToxCast™/Tox21-derived in vitro PODs. The predicted PODs were found to be within a 10-fold difference from in vivo PODs for both animal species (Wang, 2018). Likewise, integration of ToxCast concentration-response data with chemical structure features using nonparametric Bayesian modeling improved concentration-response curves over parametric models (NAS, 2017). This model assumed that two responses are a priori correlated if induced by chemicals with similar structures at similar concentrations, and this allowed improvement of dose-response prediction by “borrowing” information on chemicals that have similar structures. Furthermore, this modeling can be used to predict dose-response curves for chemicals with known structures that lack dose-response data (NAS, 2017).

3. The use of toxicogenomic data to support HHRA

Toxicogenomics represents an umbrella term for application of a diverse range of functional genomic studies in toxicology. Owing to the advances in microarray technology, which allowed whole-genome analysis of gene expression on mRNA level, transcriptomics has become the most mature toxicogenomic area over the past 20 years since this field emerged. Therefore, for the purpose of this paper, toxicogenomic data are considered to be whole-genome expression (transcriptomic) data produced by expression microarray studies or RNA sequencing (RNA-seq) studies.

Transcriptomic data have shown a considerable promise for human health risk assessment. Toxicogenomic approaches offer a deeper understanding of molecular effects produced by toxicants and have matured to facilitate prediction of adverse health outcomes for hazard assessment (Webster et al., 2016), and to determine transcriptomics benchmark doses (BMDs) from microarray or RNA-seq data for use in dose-response assessment (Thomas and Waters, 2016).

RNA-Seq, a more contemporary technology, offers promise for overcoming some limitations related to expression microarrays, such as their limited dynamic range and dependence on pre-designed probes for quantitation of mRNA transcripts. Consistent with expectations, RNA-Seq was shown to outperform microarrays in detection of differentially expressed genes in toxicological studies, mainly due to improved accuracy of low-abundance transcripts (Wang et al., 2014; Rao et al., 2019). Although potential of RNA-seq to generate more mechanistic insight has been noticed (Rao et al., 2019), classifiers predicting modes of action displayed similar performance for both, microarray and RNA-seq data (Wang et al., 2014). Furthermore, this technology is more costly and continues posing algorithmic and logistic challenges in data analysis (Zhao et al., 2014). Higher maturity of data analysis and lower cost of microarray expression studies is likely to ensure their continuous popularity and substantial role in future toxicological whole-genome expression studies. Nonetheless, increasing interest in using RNA-seq in toxicogenomics, and fever limitations of this method over time, will likely result in its more frequent use, at least in projects that need to quantify low-abundant transcripts, splice variants, and/or non-coding RNAs. Analysis of extensive data by the Sequencing Quality Control Consortium (SEQC) project found high reproducibility of identification of differentially expressed genes across multiple centers and RNA-seq platforms, further supporting confidence in this method (Xu et al., 2016).

3.1. Description of toxicogenomic data and their quality

Expression microarray data generally include (i) probe identifiers to identify probes that quantify mRNA transcripts, (ii) measure of the abundance of transcripts, and (iii) additional information for all transcripts that reflect data quality, (such as detection calls, which can be used to assess whether or not a transcript is reliably measured for all quantified transcripts).

These data have been used extensively to study alterations of gene expression in various biological systems ranging from human or animal tissues, in vivo or ex vivo, primary cell cultures, and immortalized cell lines. These alterations are induced by various conditions that include chemical exposures, gene perturbations, and disease. The data are available in public repositories, such as GEO (Soboleva et al., 2012) and ArrayExpress (Tikhonov et al., 2014), which represent large-scale and general-purpose repositories of user-uploaded expression data across species, tissues, experimental conditions and analytical platforms. In contrast, more specialized repositories such as Connectivity Map (Lamb et al., 2006), DrugMatrix (Ganter et al., 2006), and TG-GATEs (Igarashi et al., 2015) link expression data to chemically-induced biological perturbations and toxicity. Notably, specialized repositories of in vivo data DrugMatrix and TG-GATEs provide also associated phenotypic data that can be used for phenotypic anchoring of gene expression perturbations.

A critical step in the analysis of expression microarrays is the pre-processing of raw fluorescence signals to expression estimates that are proportional to the amounts of transcripts in the profiled specimens. This pre-processing, also referred to as “normalization”, addresses non-biological variability introduced by sample preparation, labeling, hybridization, fluorescence reading and other technical issues. Preprocessing approaches have undergone considerable development and numerous methods are currently available in the field; for instance, pre-processing of Affymetrix expression microarrays can be performed with several methods such as the Microarray Affymetrix Suite version 5.0 (MAS 5.0), Robust Multichip Analysis (RMA), GeneChip Robust Multichip Analysis (GCRMA) and PLIER (Probe Logarithmic Intensity Error Estimation), all of which include background adjustment, normalization and probe summarization steps (Bolstad, 2008). Due to their differences in underlying assumptions (Jaksik et al., 2015), these steps are implemented differently in different preprocessing methods, which was shown to influence the results of downstream analyses, such as identification of differentially expressed genes (Stafford and Tak, 2008), clustering of genes or specimens (Freyhult et al., 2010), development of gene expression-based classifiers, or building gene networks (Lim et al., 2007). Considering its potential influence, pre-processing of microarray data is likely of critical importance for the use of these data in support of regulatory decisions.

Critical components for consideration of toxicogenomic studies and data in HHRA is the evaluation of quality of studies that employed gene expression, quality gene expression data and performance of data pre-processing. For microarray data, reporting quality of gene expression studies was recommended to meet the requirements defined as the Minimum Information About a Microarray Experiment (MIAME). The MIAME requires that adequate information be supplied for (i) experimental design, (ii) array design, (iii) samples, (iv) hybridizations, (v) measurements, and (vi) normalization controls (for details, see (Brazma et al., 2001)). Criteria to assess quality of toxicogenomic microarray animal studies were developed by Bourdon-Lacombe et al., including criteria that are essential to experimental integrity and reproducibility (for details, see (Bourdon-Lacombe et al., 2015)).

Assessment of the quality of microarray data is closely associated with pre-processing of the raw microarray data, because some quality control methods are performed as a part of microarray data pre-processing. Moreover, certain quality control measures inform about the results of microarray pre-processing and may identify specific microarrays that do not meet quality criteria and need to be removed from the analysis, which may subsequently require new pre-processing of the “cleaned” data. Pre-processing is dependent on specific microarray platforms. Whenever available, raw microarray data should be considered for the use in HHRA, because it is possible to exercise more transparency and control over their analysis and interpretation starting from raw data, as opposed to data pre-processed by study investigators. Depending on the microarray platform, appropriate software can be used for data pre-processing, such as Expression Console (Affymetrix, Santa Clara, CA, USA) or appropriate R packages from Bioconductor (https://www.R-project.org). Plots that allow for examination of pre-processing performance and quality control of microarray data include the normalized unscaled standard error (NUSE) plot and/or the relative log expression (RLE) plot; these allow determination of whether the data meet assumptions underlying normalization approaches on comparable median intensities and spreads (inter-quantal ranges). Additional QC plots may include MA plots that allow identification of intensity-based bias in microarray data, and pseudochip images of residuals from probe level models fits to identify special biases (more details are available at affyAnalysisQC module description: URL: http://www.arrayanalysis.org/). Failed results of QC can be sometimes resolved by using alternative pre-processing pipelines; however, if alternative pre-processing does not improve diagnostic plots, microarrays that introduced QC problems need to be removed from analysis.

Additional analyses of pre-processed microarray data can include evaluation of reproducibility through similarity among replicated microarrays or among microarrays corresponding to similar experimental conditions. This evaluation usually employs unsupervised data exploration using clustering (e.g., hierarchical clustering) and/or multivariate projection methods (e.g., principal component analysis). Use of these methods can identify outliers or anomalous patterns, such as “batch effect” that need to be properly addressed before further use of microarray data. The description of details affecting quality of microarray data, their diagnosis and corrective actions are beyond the scope of this report (Gohlmann and Talloen, 2009).

3.2. Interpretation of toxicogenomic data and significance for HHRA

Depending on the purpose of the study, microarray data can be analyzed by one or more of the following methods that can be broadly classified as: (i) class discovery, (ii) class comparison and (iii) class prediction (Simon, 2003). Class discovery serves for identification of natural structure in the data, and one of its applications includes identification of outliers or unusual patterns during the microarray data quality control. Class comparison is extensively used to detect differences in gene expression among classes of data, such as the data corresponding to exposed and control animals, or the data corresponding to animals exposed at different exposure levels. The major output of class comparison analysis is a list of differentially expressed genes or a list of biologically meaningful groups of genes (gene sets) that correlate with specific classes of data. Class prediction includes various classification methods, such as, e.g., neural networks and support vector machines, which enable the development of classifiers (e.g. “gene signatures”) for prediction of correct class membership of samples within biologically defined classes (such as genotoxic vs non-genotoxic carcinogens).

3.2.1. Mechanistic toxicogenomics

Lists of differentially expressed genes from class comparison analysis of microarray data serve as an input for mechanistic toxicogenomic analyses. Mechanistic toxicogenomics attempts to derive mechanistic understanding of adverse outcomes induced by toxicants using functional genomics, particularly transcriptomics, data. Approaches include pathway enrichment analysis and construction of gene interaction networks that are more efficient tools for elucidation of mode of action than interpretation of expression changes on the level of single genes (Pennie et al., 2004). This is, at least in part, due to the reduction of data size (e.g., from thousands of genes to dozens of pathways) and the use of systems-level annotations to inform data interpretations.

Pathway enrichment analysis provides insight into the underlying biology and reduces data complexity through the arrangement of differentially expressed genes into pathways or other biologically meaningful groups, such as, genes sharing target sites for specific miRNAs or genes under control of specific transcription factors. This allows identification of a priori defined pathways or processes that may be affected by a condition, such as chemical exposure, which induced differential expression.

Methods used in pathway enrichment analysis have evolved over time. The first-generation methods are based on statistical evaluation of over-representation of differentially expressed genes in particular pathways. As an input to over-representation analysis (ORA), the following needs to be used: (i) the list of up-regulated genes, list of down-regulated genes, and/or list of all differentially expressed genes; (ii) annotated gene lists, such as pathways from repositories such as GO (Biological Process, Molecular Function or Cellular Component), KEGG or Reactome (Fabregat et al., 2018), (iii) background list and (iv) an appropriate statistical test for over-representation and/or depletion, such as one- or two-sided Fisher’s exact test (Rivals et al., 2006) with or without multiplicity correction. Enriched categories are subsequently biologically interpreted in the context of possible mechanism. Limitations of the ORA reflect arbitrary character of thresholds used to identify differentially expressed genes and the loss of quantitative information on the changes in gene expression. In addition, the ORA assumes independence among genes and ignores correlation structure of gene groups. These limitations are addressed by functional class scoring methods (FCS), such as Gene set Enrichment Analysis (GSEA) (Subramanian et al., 2005), that do not depend on an arbitrary selection of differentially expressed genes and account for effects of small but coordinated changes in gene expression. However, the use of FCS methods still results in a loss of part of the quantitative information on differential gene expression, as it considers genes by ranks rather than by their values of expression changes. Likewise, FCS methods do not overcome limitation of ORA related to the assumption that each pathway (or gene set) is independent of any other pathway. The most advanced third-generation methods consider pathway topology to extend the enrichment analysis. In these methods, information from annotated pathways is not limited to pathway membership, but also includes interactions, directions of interactions (activation and inhibition) and/or sub-cellular localizations of interacting genes or their products (reviewed in (Khatri et al., 2012)).

Analysis of gene interaction networks can employ differentially expressed genes together with published information on functional gene interactions, such as transcriptional regulations and physical interactions or influences on gene expression, to build gene networks using various algorithms (e.g., direct interactions, shortest paths, and transcriptional regulation implemented in MetaCore (Shi et al., 2010b)). A promising approach has been developed based on mapping of differentially expressed genes onto curated protein-protein interaction network from the MetaCore knowledge base, and identification of genes topologically relevant to sets of differentially expressed genes (Dezso et al., 2009). The importance of this and other network-based approaches is in the identification of highly relevant hub genes, which may not be necessarily differentially expressed, but which interact with several downstream genes and regulate their expression, and likely drive the perturbations induced by toxic agents. These genes can inform the molecular initiating event or downstream key events and help elucidate the mode of action (Bourdon-Lacombe et al., 2015).

Unlike gene networks that employ prior knowledge on gene-gene interaction data, correlation networks, such as weighted gene co-expression network analysis (WGCNA) are driven entirely by data and are free of bias related to knowledge from published literature (Zhao et al., 2010). WGCNA can be used to identify clusters of correlated genes that usually have coordinated regulation and biological functions, as well as relationships between these clusters (modules) and intramodular hub genes. The use of WGCNA on bisphenol A (BPA) dose-response data demonstrated that, for low doses of BPA, the correlation network contained modules that did not show enrichment for estrogen receptor-alpha (ESR1) target genes, but displayed clear signatures of other transcription factors suggesting that the mode of action at low doses included events other than direct estrogen receptor activation (Maertens et al., 2018). This example demonstrates the utility of transcriptomic data for inferring the modes of action at different exposure levels, including those relevant for low environmental exposures.

Upstream regulator analysis is an inference of a cascade of upstream regulatory molecules responsible for observed gene expression changes. This analysis uses prior knowledge about expected effects of regulators, such as transcriptional regulators, miRNAs, kinases or chemicals, and their gene targets. This analysis is implemented in the Ingenuity Pathway Analysis (IPA), which employs knowledge from the Ingenuity Knowledge Base (Krämer et al., 2013) and produces the overlap p-value for each potential upstream regulator to test significance of overlap between dataset genes and target genes known to be regulated by this upstream regulator. In addition, IPA provides activation z-scores that account for activation status and directional relationships between genes and infers the activation status of upstream regulators (“activated” or “inhibited”). Furthermore, IPA provides additional causal analytical algorithms that include Mechanistic Networks (MN), which connects upstream regulators if they represent members of the same signaling pathway. These approaches have been successfully used to study molecular events contributing to the development of forestomach tumors in mice exposed to benzo(a)pyrene (Labib et al., 2013). The identified upstream regulators included TP53 and IFNG genes, which implicated the role of DNA damage and immune response in the development and progression of forestomach tumors.

3.2.2. Predictive toxicogenomics

Predictive toxicogenomics uses class prediction approaches to predict adverse outcome, such as carcinogenicity or hepatotoxicity from toxicogenomic data. This approach typically employs feature (gene) selection and application of classification algorithms to toxicogenomic data for samples with known class membership, in order to extract molecular signatures that can be used to predict class membership of samples with unknown class membership. This concept of supervised prediction is not limited to binary categorical classes (e.g., the presence and absence of a specific toxicity), as it can be generalized to predict values of continuous variables (such as the IC50¹ which also extends to the field of quantitative toxicogenomics). Feature selection methods consider features either individually or in combinations to reduce data dimensionality to manageable subsets of relevant features, and this process can be independent or integrated with classification algorithm. There is a wide variety of classification algorithms that include statistical models and machine learning methods, such as linear discriminant analysis (LDA), nearest neighbor classifier (k-NN), logistic regression, support vector machines (SVM), neural networks, decision trees, as well as numerous “ensemble” methods that combine many individual models to improve classification performance (e.g., random forests and nearest shrunken centroids). Although some classification algorithms can perform better than others depending on the data, no algorithm is expected to outperform others in all situations.

Gene expression biomarkers have been developed to predict endocrine disrupting chemicals (reviewed in this journal issue by Corton et al. “Identification of potential endocrine disrupting chemicals using gene expression biomarkers”). The use of whole-genome expression data that included more potential gene targets of endocrine disrupting chemicals than ToxCast™/Tox21 data allowed to develop from microarray data classifiers with high predictive accuracy for both estrogen receptor α (ERα) (Ryan et al., 2016) and androgen receptor (AR) (Rooney et al., 2018) agonists and antagonists. Classifiers that predict modulation of NRF2 and NFkB genes have also been recently developed (Corton et al., in this issue).

Classification of chemical carcinogens as genotoxic or non-genotoxic is a critical component of HHRA, and influences considerations regarding low dose extrapolation in quantitative risk assessment. In this context, genotoxic carcinogens are those that display direct DNA-reactivity and induce mutations on gene or chromosome level, while non-genotoxic carcinogens display diverse modes of action that impact cell proliferation, differentiation and survival without direct DNA-reactivity (Mishima, 2017). For genotoxic carcinogens, currently adopted quantitative risk assessment employs linear low-dose extrapolation of dose-response. Gene expression data generated in vitro and in vivo and various classification algorithms have been explored for development of classifiers that can discriminate between genotoxic and non-genotoxic carcinogens or discriminate among non-genotoxic carcinogens with different MoAs (reviewed in (Waters et al., 2010)). The discrimination among non-genotoxic carcinogens may inform the assessment of human relevance of rodent tumors (for example, the role of cytotoxicity-mediated proliferation and the activation of nuclear receptors) (Felter et al., 2018).

Development of toxicogenomic classifiers that could discriminate genotoxic from non-genotoxic carcinogens was intended to overcome limitations of short-term in vitro bioassays, such as the Ames test. However, numerous classifiers developed so far have shown differences in applicability and predictive accuracy (Waters et al., 2010). A major concern is related to secondary genotoxicity displayed by non-genotoxic carcinogens, which induce tissue damage, inflammation and ROS-mediated stress. A robust “general” 32-gene signature of genotoxicity in vivo has been developed by Auerbach (Auerbach, 2016) through mining expression data across various studies involving many tissues and different species. This signature includes activation of the p53-gene and its down-stream partners Mgmt, Ccng1, Cskn1a, Bax and Btg2. The use of this signature depends on the experimental dose selection and length of study, because its activation can also be seen following significant tissue damage and inflammation. In addition, in order to use this qualitative signature in HHRA specific criteria need to be developed to support decisions on whether the signature is activated or not when only a few but not all genes are found to be up-regulated.

It is recognized that non-genotoxic carcinogens can induce secondary genotoxicity which needs to be distinguished from primary genotoxicity mediated by DNA-reactivity of tested chemicals. This can be a complicated issue due to the integrated nature of biological responses. An approach that reduces false positivity caused by cytotoxicity-mediated secondary genotoxicity uses 65-gene TGx-DDI biomarker (Li et al., 2015; Cho et al., 2019). This biomarker was developed from expression data on 13 genotoxic and 15 non-genotoxic chemicals in human TK6 cell line in vitro, using a nearest shrunken centroid algorithm. The assay protocol includes a concentration-finding pre-test, which is critical to avoid transcriptional attenuation and interference by cytotoxicity. The interpretation of expression data for tested chemicals is simplified and transparent through availability of web-interface that can use Affymetrix, Agilent or generic microarray data input and reports probability of chemical being DNA-damaging, as well as hierarchical clustering with queried chemical and known genotoxic and non-genotoxic compounds used for the development of this biomarker (https://manticore.niehs.nih.gov/tgxddi/tool).

Presently, this biomarker is being considered for genotoxicity evaluation in drug development. Although not considered to have been validated for this purpose at this time, TGx-DDI biomarker has been found to be promising for its potential to complement panel of assays that includes gene mutation in bacteria, and chromosomal aberration/micronucleus tests in vitro and in vivo. More specifically, its potential value has been recognized with respect to reconciliation of conflicting results for compounds that tested negative in an Ames test and chromosomal aberration/micronucleus formation in vivo and positive for chromosomal aberration/micronucleus formation in vitro (FDA, 2017). The interpretation of the results of this biomarker in the context of other more traditional genotoxicity assays and other findings should not be limited by pending formal analytical validation and additional consideration of sample stability, which have been identified as the key remaining issues.

Another method for the prediction of adverse outcomes, which has been increasingly reported in published toxicogenomic literature, relies on the comparison of expression profiles induced by chemicals of interest with other expression profiles in data repositories (Bourdon-Lacombe et al., 2015). This allows one to associate the queried chemical with (i) other chemicals with known hazards, (ii) gene perturbations, or (iii) diseases and conditions through positively or negatively correlated gene expression. Identification of genes, whose silencing, mutations or knock-out induced gene expression correlated to the changes induced by perturbing chemicals can also provide mechanistic insight through identification of potential gene targets that are affected by chemical treatment and drive observed changes in gene expression. Software suites that allow this type of analysis usually contain normalized gene expression data from public and/or private repositories as well as tools for data analysis and visualization. Examples of available software suits include BaseSpace Correlation Engine (Illumina) and Genevestigator (Nebion) (Figure 3). BaseSpace CorrelationEngine has been used to predict adverse outcomes and their human relevance from gene expression profiles of lungs of mice exposed to carbon black nanoparticles (Bourdon et al., 2013).

Figure 3.

Illustration of the predictive toxicogenomics using Genevestigator tool. The tool identified 50 perturbation studies (Affy_U133Plus_2 platform) with gene expression most similar to the expression profile of duodena of mice exposed to sodium dichromate dihydrate in drinking water at 520 mg/L for 8days. Gene signature was derived from dataset GSE75214 (Gene Expression Omnibus) normalized by study authors (264 human genes corresponding to the top differentially expressed genes between duodena of exposed and control mice). This approach identified numerous biologically relevant studies, including human celiac disease-related gene expression. Exposure of mice to hexavalent chromium in drinking water was previously found to induce histopathologic changes consistent with tissue remodeling in celiac disease (Rager et al., 2017).

3.2.3. Quantitative toxicogenomics

Analysis of dose-dependent changes in expression on the level of individual genes, pathways and other biologically-defined groups of genes, enables the determination of transcriptomic benchmark doses (BMDs) and their lower 95% confidence limits (BMDLs) that could potentially serve as estimates of biological points of departure (BEPOD) for deriving reference values for use in quantitative human health risk assessment (Thomas and Waters, 2016). Compared to traditional quantitative risk assessment, which typically employs apical outcomes from animal-based toxicity studies, transcriptomic BMD modeling using short microarray expression studies in vivo offers higher throughput, shorter time to results and reduction of animal use (Farmahin et al., 2017). Furthermore, this approach benefits from the generally higher sensitivity of gene expression to low, environmentally-relevant chemical exposures as compared to most sensitive apical outcomes often observed at higher exposure levels.

BMD modeling using pre-processed (normalized) gene expression data is a multistep process that includes (i) identification of probesets (genes) responsive to chemical exposure, (ii) BMD model fitting and filtering of probeset (gene)-level BMD values, (iii) identification of biologically-defined gene sets or pathways for determination of pathway-level BMDs, including selection of gene set (pathway) repository and statistical considerations, (v) summarization of gene-level BMDs to pathway-level BMDs, and (vi) estimation of BEPODs using pathway-level BMDs (BMDLs). These steps can be implemented with various methods that are based on different assumptions. As a result, there are numerous alternative approaches, which combine these methods in multistep pipelines from normalized expression data to BEPOD estimates. These alternative approaches to transcriptomics BMD modeling have been extensively studied and their evaluations were reported in several peer-reviewed articles (Webster et al., 2015; Farmahin et al., 2017). The approach presently used by the NTP uses data from 5-day dose-response studies in rodents, which are pre-processed by RMA method and analyzed using BMDExpress software (https://github.com/auerbachs/BMDExpress-2/releases ) (Yang et al., 2007). The details on the implementation of BMD modeling steps have been previously reported (NTP, 2018b). This and other approaches previously used for transcriptomic BMD modeling (Table 2) can differ from each other in (i) determination of differentially expressed genes, (ii) selection of parametric dose-response models and criteria for identification of winning models, (iii) selection of repositories of biologically-relevant gene sets (e.g., GO:BP, IPA pathways, MetaCore pathways, or MSigDB (Subramanian et al., 2005)), and (iv) selection of gene sets (pathways) for determination of pathway-level BMDs. In spite of these differences, published reports display overall concordance of BEPOD values derived from transcriptomics and apical outcome data for both, cancer and non-cancer outcomes, with usual difference within a factor of 10 (Table 2).

Table 2. Summary of previously reported studies that reported or enabled comparisons of transcriptomic and apical PODs (cancer and/or non-cancer health outcomes).

Toxicogenomics studies were performed with rats and/or mice. Tissues profiled for gene expression included livers, kidneys, lungs, thyroid and bladder. Exposure times ranged from 1 day to 90 days. Apical outcomes included both cancer and non-cancer health effects. Gene expression profiling was performed by microarrays and in one study (Zhou et al. 2017) by RNAseq.

Study	Tested article	Species Route of exposure Transcriptomics platform Profiled tissue	Apical outcome	BMDL§ concordance
Bhat VS et al.(Bhat et al., 2013)	Cyprocinasole Epoxyconasole Propiconazole Triadimefon, Myclobutanil	CD1 Mouse 30 days Oral gavage Affymetrix mouse 430 2.0 Liver	Liver weight (BMR=1SD) At 30 days Liver tumors (BMR=10%) at 18 months	BMDLt/BMDLa 1–2.4 BMDLt/BMDLa 0.21–4.3
Dong H et al.(Dong et al., 2016)	Furan	F344 rats (M) 90 days/5 days per week Oral gavage Agilent G4853A SurePrint G3 Rat GE 8 × 60 K Liver	Most sensitive: hepatocyte apoptosis	Lowest BMDLt: ERK/MAPK: 0.04 mg/kg-day Lowest BMDLa: 0.02 mg/kg-day
Jackson et al. (Jackson et al., 2014)	Furan	Mice B6C3F1 3 weeks/daily Oral gavage	Hepatocyte apoptosis Hepatocellular adenoma Hepatocellular carcinoma	BMDLa=0.11 mg/kg-day BMDLa=0.92 mg/kg-day BMDLa=1.57 mg/kg-day BMDLt=1.89 mg/kg-day
Moffat I et al.(Moffat et al., 2015)	Benzo[a]pyrene	Mice B6C3F1, Muta or C57BL/6J 3 days/daily 28 days/daily 6 days/4expo Agilent, Affymetrix Liver, lung and forestomach tissues	Liver tumors Liver mutations Lung tumors Lung mutations Forestomach tumors Forestomach mutations	BMDLa=1.2 mg/kg-day BMDLa=4.8 mg/kg-day BMDLt=0.2 or 1 mg/kg-day BMDLa=0.8 mg/kg-day 1.4 mg/kg-day BMDLt=2.1 or 3.7 mg/kg-day BMDLa=0.5 mg/kg-day 0.3 mg/kg-day BMDt=4.5 or 7.4 mg/kg-day
Dunnick JK et al.(Dunnick et al., 2017)	N,N-dimethyl-p-toluidine (DMPT)	F344/N rats (M) 5 days Gavage Affymetrix Rat Genome 230 2.0  Liver tissue	Liver tumors induced by DMPT	Liver tumors DMPT LOAEL: 20 mg/kg-day BMDt = 2 mg/kg-day for DMPT
Rager JE et al.(Rager et al., 2017)	Sodium dichromate dihydrate	B6C3F1 mouse 7 days 90 days Exposed through drinking water ad libitum Agilent mouse 4 × 44 K whole-genome oligonucleotide microarrays Duodenum	Not included In this report BMDLa=17.1 mg/L$ (chronic exp, male mice, histiocytic infiltration of duodenum)	7d Lowest BMDLt (eiF2 signaling)= 9.2 mg/L 90d Lowest BMDLt (HIF1A signaling)= 23.2 mg/L
Thomas RS et al.(Thomas et al., 2013b)	1,2,4-Tribromobenzene (TRBZ) 2,3,4,6-tetrachlorophenol (TTCP) Bromobenzene (BRBZ) 4,4′-Methylenebis(N,N-dimethyl)benzenamine (MDMB) Hydrazobenzene (HZBZ) N-Nitroso diphenylamine (NDPA)	SD Rats F344 Rats 5 days 2 weeks 4 weeks 13 weeks Liver: TRBZ, BRBZ, TTCP, HZBZ Bladder: MDMB Thyroid: NDPA	For 5-day assay most sensitive apical endpoints: -TRBZ (abs. liver weight) -BRBZ (inflammation) -TTCP (abs. liver weight) -MDMB (follicular cell hypertrophy) -NDPA (absolute bladder weight) -HZBZ (none)	BMDt/BMDa Non-cancer TRBZ:1.16 BRBZ: 0.31 TTCP:0.61 MDMB:0.96 NDPA:0.71 BMDt/BMDa Cancer: (2-year assay) MDMB:0.46 NDPA: 0.77
Thomas RS et al.(Thomas et al., 2012)	1,4-Dichlorobenzene (DCBZ) 1,2,3-Trichloropropane (TCPN) Propylene glycol mono-t-butyl ether (PGBE) Naphthalene (NPTH) Methylene chloride (MECL)	B6C3F1 mice (F) 13 weeks DCBZ, TCPN: gavage PGBE, MECL, NPTH: inhalation Lungs or livers	DCBZ, PGBE: Relative liver weight TCPN, NPTH: bronchiole epithelial degeneration MECL: periportal liver vacuolization	BMDLt/BMDLa Non-cancer DCBZ:0.54 PGBE:0.22 TCPN:0.27 MECL:0.91 NPTH:0.48 Cancer DCBZ:0.39 PGBE:0.43 TCPN:0.35 MECL(lung):1.04 NPTH:0.06
Thomas RS et al.(Thomas et al., 2013a)	β-Chloroprene (2-chloro-1,3-butadiene)	F344 Rats (F) B6C3F1 Mice (F) 5 days 15 days Inhalational exposure Lungs	From 2-year rodent bioassays: Female rat lung tumors Female mouse lung tumors=2.9 ppm	5-day exposure Rats lungs BMDLt=24.14 ppm lung tumors BMDLa=76 ppm Mice lungs BMDLt=1.12 ppm lung tumors BMDLa=2.9 ppm
Zhou et al.(Zhou et al., 2017)	Trichloroethylene (TCE) Tetrachloroethylene (PCE)	B6C3F1/J mice (M) 1 day (single intragastric exposure) RNAseq	Noncancer --TCE liver: relative liver weight kidney: relative kidney weight --PCE liver: increased angiectasis kidney: nuclear enlargement Cancer Liver cancers Kidney cancers	TCE Liver: Non cancer BMDLa and BMDLt: distance of ~2 orders of magnitude Cancer BMDLa and BMDLt: Distance ~ 1 order of magnitude TCE Kidney: Non cancer BMDLa and BMDLt: distance of 1–2 orders of magnitude PCE Liver: Non cancer BMDLa and BMDLt: distance < 1 orders of magnitude Cancer BMDLa and BMDLt: distance < 1 order of magnitude PCE kidney: Non cancer BMDLa and BMDLt: distance < 1 order of magnitude
Farmahin R et al.(Farmahin et al., 2017)	1,2,4-Tribromobenzene (TRBZ) 2,3,4,6-tetrachlorophenol (TTCP) Bromobenzene (BRBZ) 4,4′-Methylenebis(N,N-dimethyl)benzenamine (MDMB) Hydrazobenzene (HZBZ) N-Nitroso diphenylamine (NDPA)	SD Rats F344 Rats 5 days 2 weeks 4 weeks 13 weeks Liver: TRBZ, BRBZ, TTCP, HZBZ Bladder: MDMB Thyroid: NDPA	The same as in Thomas RS et al.(Thomas et al., 2013b)	Cancer** For 5-day assay BMDLt and BMDLa within an order of magnitude. Best agreement found for 90-day BMDL values (for all approaches 3-fold difference) Non-cancer: timepoint-matched ratios BMDLt/BMDLa for all approaches within order of magnitude
NTP(NTP, 2018a)	Triphenyl phosphate (TPHP)	SD rats 4 days Oral gavage Livers	Increased serum HDL	BMDLa=39 mg/kg-day BMDLt=11 mg/kg-day

^§BMDL=statistical lower 95% confidence limit of a benchmark dose (BMD). Subscripts t=transcriptional, a=apical

^*Values derived from graphical presentation of results. BMDs were converted to human effective doses using multispecies PBPK models

^**Used the same data/tested articles as Thomas et al. 2013 et al. (Thomas et al., 2013b) across 11 approaches to derive BMDt

^$Pre-computed from NTP database

Notably, transcriptomic BMD modelling groups genes (probesets) with gene-level BMD values into biologically-relevant gene sets or pathways, subsets of which are subsequently selected for pathway-level BMD determination. This selection can be based on appropriate enrichment metrics, such as minimal number of mapped genes per pathway with BMD values, and/or percentage of pathway genes with determined BMD values, or statistical significance of enrichment of pathways with genes that have BMD values. Alternatively, selection of pathways can be based on prior mechanistic knowledge related to the toxicity of an assessed chemical. While the latter approach requires mechanistic insight about the tested chemical, the former approach does not require any prior knowledge of mechanisms or adverse outcomes induced by tested chemicals, which offers a substantial advantage in case of chemicals with limited toxicological knowledge. Most studies presented in Table 2 estimated BEPODs without considering prior mechanistic knowledge, and employed enrichment-based selection of gene sets (pathways) from general repositories, such as Gene Ontology, IPA or MetaCore. A slightly different approach was used in a study reported by (Bhat et al., 2013) that utilized 330 probesets corresponding to genes previously identified as responsive in livers of rodents exposed to 3 carcinogenic conazoles. Comparison of transcriptomic and apical BEPODs reported in these studies suggest that mechanistic insight is not necessary for estimation of BEPOD values from dose-response gene expression studies, at least for the level of concordance observed in these studies. In fact, this concordance implies that dose-response dependence of relevant molecular events is implicitly reflected by approaches used for estimation of transcriptomic points-of-departure. Consequently, transcriptomic BMD modeling has the potential to provide mechanistic insight through identification of responsive molecular events and its significance extends to mechanistic toxicogenomics.

Quantitative toxicogenomics based on transcriptomics BMD modeling is an area of toxicogenomics that is seemingly the most mature for use in HHRA. Its wider use in HHRA is possible, at least in part, due to increased understanding of the influence of most parameters used in the analysis, which offers their informed and transparent selection. The significant advantage stems from concordance of results of short-term (5-day) dose-response in vivo studies with long term bioassays (Table 2), which might facilitate derivation of reference values with substantial reduction of time and animal use. Future refinement of transcriptomic BMD modeling is possible through suggestions that include nonparametric differential gene expression analysis and dose-response modeling (NTP, 2017). We have also identified the need to explore sensitivity of estimation of BEPOD to the selection of methods for pre-processing of the raw microarray data. Selection of methods for microarray data pre-processing has been shown to considerably influence results of some analyses of gene expression data and this potential influence should be considered in the application of transcriptomic BMD in HHRA.

4. Case studies for the use of toxicogenomics and/or HTS in human health risk assessment

As a proof of concept for the notion that toxicogenomics and/or HTS may be routinely used in human health risk assessments along the various directions outlined in this manuscript, we propose developing case studies for various chemicals, for which adequate conventional animal or epidemiological data as well as toxicogenomic or HTS data are available. To that end, Table 3 summarizes HTS and/or toxicogenomic data publicly available for selected IRIS assessments for chemicals, for which competing carcinogenic modes of action have been suggested in the literature. In each case, possible approaches are mapped out below to conceptually illustrate the potential of these data towards adding to the weight of evidence for a particular mechanism or mode of action. Our conclusion is that the available data offer a unique opportunity to resolve some uncertainties even if not always sufficient for clearly elucidating mode of action. (Table 3)

Table 3.

Availability of ToxCast/Tox21 screening and curated toxicogenomic (transcriptomic) studies from BaseSpace inventory (Illumina) for several chemicals in the IRIS database for which competing modes of action have been presented in the literature.

Chemical CAS	Tumor sites in animal studies (chronic exposure)1	Modes of action2	Potential for HTS and omics data to contribute further3
Acrylamide 79–06-1	Endocrine (rats), dermal (mice), respiratory (mice), CNS (rats), reproductive (rats)	Mutagenic MOA suggested based on a variety of evidence for genotoxic effects, including the formation of DNA adducts in mice and rats. The strongest support for this MOA is based on mutations occurring at the target tissue. Alternative MOA for thyroid tumors, scrotal mesotheliomas, tunica vaginalis, and mammary tumors in rats is based on disruption of hormone levels via dopamine agonist activities or alteration of signal transduction pathways leading to thyroid cell proliferation.	ToxCast: 881 assay endpoints (18 active) Expression data available from (i) testicular tissue of mice exposed to acrylamide in drinking water at 3 dose levels/3 exposure times, and (ii) human TK6 cells exposed to 3 different concentrations in vitro.
1,4-dioxane 204–661-8	Multiple sites in rats (hepatic abdominal endocrine, respiratory zymbal gland, dermal); hepatic tumors in mice and guinea pigs.	Generally, the MOA, including a mutagenic action, is not conclusive. Does not cause point mutations and does not affect DNA repair. Proposed MOAs for liver tumors involve increased cell proliferation of spontaneously transformed cells either in the presence of cytotoxicity or due to tumor promoting action and stimulation of DNA synthesis; either case is proposed at doses exceeding metabolic saturation. Proposed MOA for nasal cavity tumors involves sustained cell proliferation of spontaneously transformed cells due to cytotoxicity or chronic irritation effects. However, data to support MOAs are conflicting for both tumor types.	ToxCast data not available Expression data available for livers of rats exposed to 1,4-dioxane for 2 exposure timepoints
Tetrachloroethylene (perc) 127–18-4	Urinary (mice and rats), hematologic (mice and rats), CNS (rats), reproductive (rats), dermal (mice)	Hypothesized but controversial MOAs exist for rat kidney and mouse liver tumors (only). For rat kidney, MOAs include genotoxicity, peroxisome proliferation, alpha 2u nephropathy, and cytotoxicity unrelated to alpha 2u accumulation. Of these, peroxisome proliferation and alpha 2u nephropathy (male rats) have been given prominence by various scientists with the argument that these mechanisms are not relevant for humans. For mouse liver, MOAs include mutagenicity, DNA hypomethylation, oxidative stress, and PPAR alpha receptor activation; some have asserted a key role for the PPAR related MOA and argued that this MOA lacks relevance for humans. IRIS assessment concludes that available evidence is inconclusive in that either tumors are driven solely by one of these MOAs. The role of mutagenicity in perc carcinogenesis is uncertain. The human relevance of rat mononuclear cell leukemias has been questioned.	ToxCast: 113 assay endpoints (2 active) Expression data available for human HepG2 cells treated with tetrachloroethylene for two timepoints.
Methylene Chloride 75–09-2	Hepatic (mice), respiratory (mice), endocrine (rats)	Mutagenic MOA via GST-mediated metabolic conversion to formaldehyde; however, the relevance to low human exposure scenarios has been controversial on grounds of very low fraction metabolized through this pathway.	ToxCast: 113 assay endpoints (1 active) Gene expression data available for lungs and livers of mice exposed to methylene chloride at 5 exposure levels/single exposure timepoint
Trichloroethylene (TCE) 79–01-6	Urinary and reproductive (rats); hepatic and respiratory (mice); and hematologic (rats and mice)	Metabolites produced via CYP- and GST-dependent pathways are central to kidney and liver MOA. Interspecies differences in metabolism, and therefore extent to which key events appear in humans, is a controversial issue. It is generally accepted that kidney tumors can be ascribed to a mutagenic MOA due to metabolites from GST-dependent pathway. Involvement of cytotoxicity and compensatory cell proliferation at high exposures, leading to nephrotoxicity following GSH conjugation, has been proposed as a causal link but the weight of this key event is controversial. Evidence of a mutagenic MOA for liver tumors is weak. Activation of PPAR alpha has been proposed as the key event in MOA for liver tumors in rodents, but the weight of this as a sole operant MOA is controversial with multiple lines of evidence suggesting that it is not necessary for tumor induction in mice.	ToxCast: 113 assay endpoints (4 active) Gene expression data available for (i) livers of rats and mice exposed to a single dose/single exposure timepoint; (ii) kidneys of rats exposed at single exposure level/3 timepoints; (iii) rat hepatocytes exposed at single concentration/single timepoint; and (iv) human TK6 cells exposed at 3 concentrations/single timepoint.
Trichloroacetic acid (TCA) 76–03-9	hepatic (mice)	MOA is not established. There is limited evidence of genotoxic potential. Human relevance of tumors seen in B6C3F male mice, which are particularly susceptible to liver tumors, is controversial. Proposed mechanisms include growth inhibition of normal cells with proliferation of selected cell populations. PPAR alpha agonism has been proposed as the most likely MOA, but this hypothesis does not appear to be sufficient to account for TCA tumorigenicity (see TCE above)	ToxCast: 886 assay endpoints (5 active) Livers of rats exposed subcutaneously to trichloroacetic acid at single exposure level/3 exposure timepoints
Hexachlorethane (HCE) 67–72-1	Renal (rats & mice), endocrine (rats)	Kidney tumors in male rats is thought to involve male rat specific alpha 2u mediated MOA, with concerns regarding a lack of human relevance; however, EPA concluded that the evidence for this MOA was insufficient. Liver tumors in mice have been ascribed to cytotoxicity, inflammation, and regenerative cell proliferation due to binding of HCE metabolites to liver macromolecules and free radical generation during metabolism. EPA concludes that data were insufficient to evaluate this MOA. Human relevance of rodent pheocromocytomas has been a controversial topic, but there is some evidence of similarity in tumors and mechanism of action.	ToxCast: 113 assay endpoints (0 active) Expression data available for (i) kidneys of rats exposed by oral gavage to a single dose/3 timepoints; (ii) rat primary hepatocytes in vitro at single concentration/single exposure timepoint; (iii) human TK6 cells in vitro exposed at 3 concentrations/single timepoint.
Chloroprene 126–99-8	Many sites, primarily hepatic (rats & mice), endocrine (rats), respiratory (mice), hematologic (mice)	Mutagenic MOA involving metabolism to reactive epoxide intermediates that form DNA adducts. This MOA is presumed to apply to all tumor types. This MOA is supported by analogy with 1,3 butadiene. Concerns have been raised that there are substantial species differences in metabolism. Liver tumors increase was statistically significant only in female mice at high exposures (but not in other rodents). The occurrence of lung tumors was statistically significantly elevated in mice but not in rats.	ToxCast data not available Expression data available for lung tissues of rats and mice exposed to chloroprene by inhalation at 4 different concentrations/single timepoint
Carbon Tetrachloride 56–23-5	Hepatic (rats, mice, hamsters), endocrine (mice)	At high exposures, CYP2E1 metabolism to peroxyl radical leading to hepatocellular cytotoxicity and sustained regenerative cell proliferation is a hypothesized MOA for liver cancers. However, the reactive nature of CCl4 metabolites and the observation of liver tumors at non-cytotoxic exposures make it problematic to view this as the sole MOA. There is paucity of data to evaluate genotoxicity at sub-cytotoxic doses.	ToxCast: 113 assay endpoints (0 active) Expression data available for: (i) fibrotic livers of mice exposed at single dose/3 timepoints; (ii) livers, kidneys and bone marrows of rats exposed at several different concentrations for several different timepoints; (iii) murine embryonic stem cells, primary murine and rat hepatocytes (1 concentration/2 timepoints); (iv) rat primary hepatocytes (4 concentrations/single timepoint) human hepatocytes; and (v) TK6 cells exposed at several concentrations/several timepoints
1,2,3- trichloro-propane 96–18-4	Hepatic and gastrointestinal (rats & mice); endocrine, urinary and reproductive (rats); ocular (mice)	A mutagenic MOA is generally accepted and includes bioactivation leading to DNA adduct formation. DNA adduct formation observed in organs that showed increased tumor incidence. However, DNA adducts also formed in other organs where tumors were not observed, which has been raised as an uncertainty with the mutagenic MOA. Cytotoxicity with tissue repair due to DNA degradation or disruption of cell signaling have been proposed as competing MOAs.	ToxCast: 898 assay endpoints (7 active) Expression data available for lung tissues of mice exposed at a single dose level/single timepoint
Chlordecone (Kepone) 143–50-0	Hepatic (rats and mice)	MOA not clear but compound appears to be non-genotoxic based on published reports. Proposed MOAs include proliferation of pre-initiated cells and induction of cytochrome P450 with subsequent ROS production and induction of oxidative stress.	ToxCast: 901 assay endpoints (216 active) Expression data not available
Nitrobenzene 98–95-3	Endocrine and hepatic (rats & mice); Urinary and reproductive (rats); respiratory (mice)	The compound appears to be non-genotoxic or weakly genotoxic. MOA is not mutagenic. Human relevance of endocrine and kidney tumors not clear.	ToxCast: 877 assay endpoints (3 active) Expression data available for kidneys of rats exposed at single concentration/2 timepoints by oral gavage, and for human HepG2 cells exposed in vitro at single concentration/2 timepoints

¹The reader is referred to references to studies in the study summary section of the toxicological review for the chemical in the IRIS database (https://cfpub.epa.gov/ncea/iris/search/index.cfm)

²Unless otherwise indicated, the reader is referred to references in the cancer MOA discussion in the toxicological review for the chemical in the IRIS database (https://cfpub.epa.gov/ncea/iris/search/index.cfm)

³Toxcast_dashboard_v2 for ToxCast and Tox21 assay annotations and data (released on 10/01/2014)

The expression data from TK6 cells exposed in vitro to acrylamide at several concentrations and timepoints may allow determining exposure levels that do not induce apparent cytotoxicity. These data could be analyzed by the transcriptomic TGx-DDI biomarker to classify the observed effects as DNA damaging or DNA non-damaging, thus providing additional evidence to assess genotoxicity of acrylamide in vitro. For 1,4-dioxane, expression data from livers of exposed rats can be used to explore effect on cell proliferation, DNA damage response and DNA repair and to interpret them in the context of the presence or absence of evidence for cytotoxicity. Large and diverse set of available toxicogenomic data for trichloroethylene can help examine the molecular basis for the ability of this chemical to induce liver tumors in mice but not in rats. The analysis can be complemented by comparison of these data with expression data for rats exposed by subcutaneous injection to trichloroacetic acid, because these cross-species differences have been suggested to reflect different rates of metabolic conversion of trichloroethylene to trichloroacetate in mice and rats. Additional insight into genotoxicity in humans is unlikely to be derived from available expression data on TK6 cells, because metabolic activation of trichloroethylene is not expected in this cell type.

Likewise, the analysis of expression data available for chloroprene can help elucidate differences in lung carcinogenicity between mice and rats. Analysis of the extensive expression dataset for carbon tetrachloride may provide details about molecular processes at non-cytotoxic exposure levels and inform their relevance for liver carcinogenesis. Furthermore, expression data available for human, murine and rat hepatocytes may provide insight into cross-species differences. The value of expression data for TK6 cells for inferring the DNA damaging effect of carbon tetrachloride is limited because metabolic activation is not expected to occur in this cell type.

Comparisons of a given expression dataset with other expression datasets in public repositories is emerging as a potent approach (Bourdon-Lacombe et al., 2015). Such a comparative analysis may be carried out between the expression data from kidneys of male rats exposed to hexachloroethane and other expression datasets corresponding to kidneys of male rats exposed to compounds known to induce alpha(2u)-globulin (a2uG) nephropathy. This analysis may help in evaluating the role of this MOA in the development of kidney tumors in male rats. Due to negative results for all assay endpoints, probably caused by the high volatility of this chemical, ToxCast™/Tox21 screening in this case is not informative for health risk assessment.

Similar to the conceptual analysis suggested for hexachloroethane, analysis of expression data for kidneys of male rats exposed to nitrobenzene may provide insight regarding the role of a2uG-mediated MOA and potentially inform the human relevance of the observation in rats. Furthermore, expression profiles of HepG2 cells can be used to assess the role of DNA damage in liver carcinogenesis through various approaches. For instance, these expression profiles can be compared with publicly available expression profiles for HepG2 cells exposed to numerous genotoxic and non-genotoxic chemicals. The analysis presented in Figure 4 demonstrates that expression profiles for HepG2 cells exposed to carbon tetrachloride at two timepoints positively correlate with expression profiles of the genotoxic agents 1-ethyl-1-nitrosourea (ENU), 2-chloroethanol (2Cl), bromodichloromethane (BDCM), and furan (FU) and negatively correlates with profiles of non-genotoxic compounds diclofenac (Diclo) and D-mannitol (DMAN). An opposite conclusion is however supported from examining the correlations of expression profiles with two other chemicals: acetaminophen (APAP), reported as genotoxic in vitro, displayed negative correlation and, caprolactam, reportedly non-genotoxic, displayed positive correlation with nitrobenzene expression data (Figure 4)(Magkoufopoulou et al., 2012). The negative correlation with acetaminophen has lower evidential value, since acetaminophen has not been found to induce gene mutations, but only chromosomal damage in mammalian cells exposed to high concentrations in vitro (Bergman et al., 1996).

Figure 4.

Results for meta-analysis of gene expression data for Hep2G cells exposed to nitrobenzene (NBZ) in vitro using bioset-to-bioset correlation implied in BaseSpace correlation engine (Illumina). Correlated biosets’ scores have value of 100 (orange; maximum) or −100 (green; minimum) for all presented chemicals. Green bars: negative correlations between biosets; orange bars: positive correlations between biosets. Top 12 results are presented. Diclo = diclofenac; APAP=acetaminophen; Dman= D-mannitol; ENU= 1-ethyl-1-nitrosourea; 2-Cl = 2-chloroethanol; BDCM = bromodichloromethane; CCl4 = tetrachloromethane; CAP = caprolactam; FU = furan.

Other data listed in Table 3 may provide limited mechanistic insight and may have limited value for addressing competing modes of actions. Among these 13 compounds, ToxCast™/Tox21 data were not available for 3 compounds, and 5 compounds scored positive in only 0–5 assay endpoints. Low number of hits for at least some of these 5 compounds could have been caused by their volatility under conditions of assay and/or a lack of metabolic activation in some specific assays. Nevertheless, the finding of 216 positive calls among 901 assay endpoints for chlordecone (kepone) may open an avenue to expand the presently limited mechanistic insight at least towards a more complete picture in vitro.

5. Conclusion:

Toxicogenomics and HTS have been successfully used in various areas of biological and medical sciences, and most notably in drug discovery and development; however, their use in assessing human health risk arising from environmental exposures to various toxicants is still at an incipient stage. Barriers, which have been identified as most limiting for wider use of toxicogenomics in HHRA include lack of guidelines for quality assessment and interpretation of toxicogenomic data, but also the lack of familiarity on the side of risk assessors (Vachon et al., 2017). Nevertheless, advances in Microarray/Sequencing Quality Control Project (MAQC/SEQC)(Shi et al., 2010a), a growing body of relevant literature, including seminal articles, such as (Bourdon-Lacombe et al., 2015), (Farmahin et al., 2017), (Chepelev et al., 2015), and the availability of refined tools such BMDExpress (Sciome, Research Triangle Park, NC) (Yang et al., 2007) address many of the perceived limitations and allow for some optimism regarding future use of these modern approaches in HHRA.

The EPA has already used conclusions from published toxicogenomics studies to inform chemical risk assessment (reviewed in (Wilson et al., 2013)), but the analyses of available toxicogenomic data conducted specifically to inform chemical risk assessment have not yet been reported. These analyses would empower risk assessors with transparency and control over data processing from raw data to interpretation, which is usually not possible with published phenotypic studies considered in HHRA. The wealth of information embedded in the HTS and toxicogenomics data have the potential to inform HHRA in many ways, including prediction of adverse outcomes (hazard assessment) and estimation of points of departure (dose-response assessment). Other critically important areas of HHRA, which can benefit from toxicogenomics and HTS data, include mechanistic insight (mechanistic toxicogenomics or HTS analysis) and relevance of the predicted adverse outcomes to human exposure scenarios. Mechanistic insight can support causality of associations between exposures and adverse outcomes found in epidemiological studies, and human relevance of adverse outcomes found in animal studies.

An example of a robust method to identify human relevance is a parallelogram approach that compares in vivo and in vitro toxicity pathways in rodents as well as in vitro rodent and human toxicity pathways to infer in vivo toxicity pathways in humans (Kienhuis et al., 2009). This approach, which corrects for in vitro-specific and rodent-specific effects, is highly promising in addressing a critical question of human relevance in HHRA.

The studies reviewed in this report provide grounds for optimism in regards developing a health assessment based solely on short-term animal toxicogenomic dose-response studies combined with in vitro expression studies performed on human cells in the absence of other conventional data streams involving apical outcomes. In assessments for which other data streams are available, toxicogenomic data may help resolve conflicting results of non-toxicogenomic mechanistic studies, including genotoxicity studies, as well as address concerns regarding the relevance of inferred mechanisms for low exposure scenarios. Furthermore, the concordance noted from Table 2 indicates that quantitative estimates derived from such data can plausibly be used to support reference values derived from apical outcomes.

The results from the ToxCast™/Tox21 system, if available, can complement the mechanistic insight derived from toxicogenomics studies. With further methodological development, it is possible that these data can also help predict adverse outcomes. The use of ToxCast™/Tox21 HTS data in IARC monographs to complement mechanistic insight represents an appreciable early development, which has the potential for further refinement and wider use in carcinogenic hazard identification, likely in the context of MoA/AOP approaches or integration with toxicogenomic data.

It is reasonable to expect that whole-genome expression data will be routinely available in the future for most chemical exposures, and these may be the only data available for the assessment of data-poor chemicals. For data rich chemicals, analysis of expression data can improve confidence in assessments based on non-genomic studies by providing independent support for hazard and dose-response inferences and their extrapolation to the human context. Future development in the field will likely extend beyond the whole-genome expression on mRNA level, and advances are expected to expand the use of proteomics and metabolomics, but also profiling of non-coding RNA (ncRNAs) in human health risk assessment. Among ncRNAs, microRNAs (miRNAs) (Lema and Cunningham, 2010), and more recently long non-coding RNAs (lncRNAs) (Dempsey and Cui, 2016), have been implicated as biomarkers or key regulators of gene expression in response to chemical exposures. Their potential to inform chemical assessment stems from the ability of individual ncRNAs to regulate expression of many protein-coding genes (Selbach et al., 2008), which makes their profiles more biologically informative than profiles of one or a few mRNAs. Considering emerging role of circulating miRNAs as biomarkers of various diseases (Ghai and Wang, 2016), we envision miRNA profiling in circulating blood in short animal studies to allow for prediction of many adverse outcomes and to replace mRNA profiling across multiple tissues.

Highlights.

• Functional genomics and HTS can substantially support chemical risk assessment

• Their most immediate benefit lies in dose-response assessment

• Their use for prediction of toxic effects is being actively developed

• They can inform mechanistic insight and human relevance of animal studies

• They may replace traditional animal studies in future chemical assessments