Application of Gene Set Enrichment Analysis for Identification of Chemically Induced, Biologically Relevant Transcriptomic Networks and Potential Utilization in Human Health Risk Assessment

Abstract

The rate of new chemical development in commerce combined with a paucity of toxicity data for legacy chemicals presents a unique challenge for human health risk assessment. There is a clear need to develop new technologies and incorporate novel data streams to more efficiently inform derivation of toxicity values. One avenue of exploitation lies in the field of transcriptomics and the application of gene expression analysis to characterize biological responses to chemical exposures. In this context, gene set enrichment analysis (GSEA) was employed to evaluate tissue-specific, dose-response gene expression data generated following exposure to multiple chemicals for various durations. Patterns of transcriptional enrichment were evident across time and with increasing dose, and coordinated enrichment plausibly linked to the etiology of the biological responses was observed. GSEA was able to capture both transient and sustained transcriptional enrichment events facilitating differentiation between adaptive versus longer term molecular responses. When combined with benchmark dose (BMD) modeling of gene expression data from key drivers of biological enrichment, GSEA facilitated characterization of dose ranges required for enrichment of biologically relevant molecular signaling pathways, and promoted comparison of the activation dose ranges required for individual pathways. Median transcriptional BMD values were calculated for the most sensitive enriched pathway as well as the overall median BMD value for key gene members of significantly enriched pathways, and both were observed to be good estimates of the most sensitive apical endpoint BMD value. Together, these efforts support the application of GSEA to qualitative and quantitative human health risk assessment.

Recently, the field of human health risk assessment has explored using transcriptomic data to make chemical safety testing less time and resource intensive and more cost-effective. These efforts have mainly centered on identifying alterations in gene expression as detected by DNA microarray in the context of a traditional repeated-dose toxicity study (Bhat et al., 2013; Ellinger-Ziegelbauer et al., 2008; Jackson et al., 2014; Labib et al., 2012; Moffat et al., 2015; Powell et al., 2006; Thomas et al., 2011; Thomas et al., 2013). While the potential advantages of these data are clear, regulatory agencies must be able to demonstrate that gene expression profiles can effectively predict biological outcomes of chemical exposure and/or inform molecular mode of action (MOA) (Thomas et al., 2007b; Waters et al., 2010). In addition, consensus regarding best practices in identifying adaptive or adverse molecular events must be reached. Despite these hurdles, organizations such as the U.S. EPA envision that transcriptomic data will play a central role in future chemical risk assessments (Cote et al., 2016; Krewski et al., 2014).

The major challenge in applying transcriptomic data to human health risk assessments lies in understanding how to best distill complex gene expression data into easily comprehensible and linear statements of actionable information. Generally, changes to key transcriptional elements or transcriptional programs are considered to be a primary response to chemical exposure, and it is widely accepted that toxicity is preceded by transcriptional changes that underpin both direct and indirect toxic effects (Auerbach et al., 2010; Bourdon et al., 2013; Fielden et al., 2011; Gatzidou et al., 2007). These events subsequently translate to an alteration in the abundance and activity of specific proteins responsible for mounting the cellular response to external challenge and ultimately drive pathological changes at the cellular, tissue, and organism level. Standard transcriptomics practice has been to analyze individual differentially expressed genes identified via fold cutoffs or analysis of variance (ANOVA)-based methodologies to develop biomarkers of disease initiation or progression. While these methods of assessing the relevance of gene expression changes can be used to identify dose–response relationships and doses corresponding to no- or lowest-observed-transcriptional-effect levels, it is important to transition toward methods of gene expression analyses that are capable of identifying disruption of biologically-relevant signaling networks to better understand and exploit genome-wide signaling events that are ultimately responsible for regulating higher order biological processes.

Gene set enrichment analysis (GSEA) was first described by Subramanian et al. (2005) and Mootha et al. (2003) and has since gained acceptance as a method to characterize both modest and robust, coordinated, biologically relevant changes in molecular signaling pathways. For example, a recent publication describes using GSEA to identify and characterize disease-critical molecular pathways as a biomarker in high-risk prostate cancer patients (Evans et al., 2016). GSEA methodology provides the ability to determine whether an a priori defined set of genes (eg, a molecular signaling pathway) demonstrates a statistically significant difference between two samples or biological states by ranking the entire list of microarray genes according to expression differences (ie, signal-to-noise ratio) between treatment and control. A unique enrichment score for each gene set is then calculated and reflects the degree to which a gene set is overrepresented at either end of this continuum. The methodology works in conjunction with the Molecular Signatures Database (MSigDB), which provides the gene set definitions in the form of eight major collections (13 311 total gene sets at the time of this manuscript). Of these 8 collections, we focused initially on the Hallmark Gene Sets collection because it represents 50 well-defined, large-scale biological processes that are applicable to the widest range of potential cellular responses (Liberzon et al., 2015). Importantly, these gene sets were “biologically derived” through direct experimentation, as many primary published studies were mined for each discrete definition, and computed to remove redundancy and noise while retaining sets of genes that display coordinated expression that can be ascribed to a distinct biological state.

In this study, we employed GSEA to identify significantly enriched transcriptional events following repeated-dose chemical exposure and evaluated the manner in which dose and time affects the enrichment of various biologically relevant gene groups/signaling pathways for potential application in qualitative and quantitative human health risk assessments. Benchmark dose (BMD) modeling of individual genes was performed to compare enriched pathway BMD values to those reported for previously identified noncancer critical effect (ie, apical) points of departure (PODs). For these analyses, we used previously reported (Dodd et al., 2012a; Dodd et al., 2012b; Dodd et al., 2012c; Dodd et al., 2012d; Dodd et al., 2013a; Dodd et al., 2013b) experimental data and publicly available gene expression data sets from liver, thyroid, and bladder tissue generated from repeated-dose, oral exposure of rats to bromobenzene (BRBZ), 1,2,4-tribromobenzene (TRBZ), hydrazobenzene (HZBZ), 2,3,4,6-tetrachlorophenol (TTCP), 4,4′-methylenebis(N,N′-dimethyl)benzenamine (MDMB), and N-nitrosodiphenylamine (NDPA). In this context, patterns of transcriptional enrichment can be examined across increasing dose and time to facilitate both qualitative and quantitative chemical risk assessment. For example, the results demonstrated that GSEA-identified transcriptomic enrichment can be used to inform potential MOAs, as well as provide the knowledge base required to predict long term biological responses, inform weight-of-evidence (WOE) deliberations, and group data-poor chemicals for read-across analysis. Further, when combined with BMD modeling of individual genes, GSEA can facilitate characterization of biologically relevant molecular signaling enrichment dose ranges. In this context, BMD values from GSEA identified genes and most sensitive biologically enriched pathways were shown to be good predictors of the most sensitive apical response BMD values. The methodology described herein constitutes a significant step forward in identifying biologically relevant and dose-dependent molecular signaling events occurring in response to chemical exposure, and we anticipate these efforts will further encourage the incorporation of transcriptomic analyses into human health risk assessment and chemical safety evaluation.

MATERIALS AND METHODS

Data Acquisition and Characterization

The chemicals tested, animal treatments, and generation and analysis of gene expression microarray data were described previously (Dodd et al., 2012a; Dodd et al., 2012b; Dodd et al., 2012c; Dodd et al., 2012d; Dodd et al., 2013a; Dodd et al., 2013b). Briefly, bromobenzene (BRBZ) (CAS no. 108-86-1; 99% purity), 4,4′-methylenebis(N,N-dimethyl)benzenamine (MDMB) (CAS no. 101-61-1; 99+% purity), 1,2,4-tribromobenzene (TRBZ) (CAS no. 615-54-3; 95% purity), 2,3,4,6-tetrachlorophenol (TTCP) (CAS no. 58-90-2; 98% purity), N-nitrosodiphenylamine (NDPA) (CAS no. 86-30-6; 97+% purity), and hydrazobenzene (HZBZ) (CAS no. 122-66-7) were examined in this study.

For each chemical, the animal sex and strain chosen to test were those that displayed the critical effect identified by the U.S. EPA’s Integrated Risk Information System in their respective toxicological reviews. For NDPA and MDMB, 4- to 5-week-old female F344/DuCrl rats were to examine effects on the bladder (NDPA) and thyroid (MDMB), whereas 4- to 5-week-old male F344/DuCrl rats were employed to examine BRBZ and HZBZ effects on the liver. Finally, 4- to 5-week-old male Sprague-Dawley rats were employed to examine TRBZ- and TTCP-induced liver effects. Group sizes were 10 rats per dose per time point for apical studies, and rats were weighed and randomized into treatment groups.

Rat exposures began at 5–6 weeks of age and continued for 5 days, 2 weeks, 4 weeks, or 13 weeks at the following doses: BRBZ—25, 100, 200, 300, and 400 mg/kg/day; TRBZ—2.5, 5, 10, 25, and 75 mg/kg/day; TTCP—10, 25, 50, 100, and 200 mg/kg/day; MDMB—50, 200, 375, 500, and 750 ppm (5.6, 22.6, 42.4, 56.5, and 84.9 mg/kg/day); NDPA—250, 1000, 2000, 3000, and 4000 ppm (28.3, 113, 226, 339, and 452 mg/kg/day); and HZBZ—5, 20, 80, 200, and 300 ppm (0.5, 2, 8, 20, and 30 mg/kg/day). BRBZ, TRBZ, and TTCP were administered via gavage 7 days per week in 5 ml/kg olive oil or corn oil. MDMB, NDPA, and HZBZ were administered in the feed continuously for 7 days per week. For comparison purposes, BMD values for MDMB, NDPA, and HZBZ in ppm were converted to mg/kg/day in Supplementary Table 1 based on food intake and animal body weight values as reported in U.S. EPA, 1988.

Gene expression microarray analysis for each chemical was performed using a minimum of 6 rats per dose per time point selected at random. Samples were loaded into an Affymetrix GeneTitan System for hybridization with HT Rat230 + PM micro-arrays. Gene expression files were downloaded from the NCBI Gene Expression Omnibus under accession no. GSE45892.

Transcriptional BMD Modeling

Strategies designed to fit dose–response gene expression data to dose–response models were performed as previously described (Thomas et al., 2007a). Briefly, BMDExpress (version 1.41) was employed, and robust multi-array average (RMA) normalized (Irizarry et al., 2003) and log₂-transformed values for individual probes were fit as continuous data to four different dose–response models (linear, 2° polynomial, 3° polynomial, and power models). Each model run was performed assuming constant variance with a benchmark response (BMR) of 1.349 multiplied by the standard deviation observed in control animals. This BMR was used to estimate an approximate 10% shift in the tail area of the distribution (Thomas et al., 2007a). For model selection, a nested likelihood ratio test was performed on all models, with the exception of the power model. A less complex model with a significantly improved fit was preferred. The Akaike’s information criteria (AIC) for the selected polynomial model were compared with the AIC for the power model (Akaike, 1973). The model displaying the lowest AIC was selected as the overall best fitting model and was used to calculate BMD values on an individual gene level. After conversion from Affymetrix probe IDs, BMD values of genes represented by more than one probe set were averaged to obtain a single BMD value for each gene.

Identification of Apical Endpoints

Analysis of apical endpoints and no-observed-adverse-effect level (NOAEL) and lowest-observed-adverse-effect level (LOAEL) determinations following chemical exposure were used as previously described in a series of publications by Dodd et al. (Dodd et al., 2012a; Dodd et al., 2012b; Dodd et al., 2012c; Dodd et al., 2012d; Dodd et al., 2013a; Dodd et al., 2013b). The methods used and the results of BMD modeling of apical endpoint data were as reported by Thomas et al. (2013).

GSEA

Gene expression data were analyzed for enrichment using GSEA software (Broad Institute-version 2.2.0) and MSigDB version 5.1 (Liberzon et al., 2015; Subramanian et al., 2005). GSEA calculates a normalized enrichment score (NES) that reflects any overrepresentation of predefined gene sets in response to chemical exposure as compared to control samples. This software generates a ranked list of all microarray probes according to the expression difference (signal-to-noise ratio) and calculates an enrichment score (ES) by walking down this list and increasing a running sum statistic when it encounters a member within the gene set definition. Conversely, this statistic decreases when encountering a gene not in the gene set. The maximum deviation from zero constitutes the ES and corresponds to a weighted Kolmogorov–Smirnov-like statistic. Once all gene sets have been evaluated, GSEA adjusts the estimated significance level to account for multiple hypothesis testing and adjusts for the respective sizes of the gene sets, ultimately generating a NES. Use of the NES facilitates comparison across gene sets. A false discovery rate (FDR) is also calculated for each NES.

Gene expression data were loaded into GSEA as unfiltered data in a tab-delimited format. Promiscuous probes were collapsed to a single gene vector to prevent genes with multiple probes from inflating the enrichment score. The permutation type chosen was “Gene Set” as directed in the GSEA user guide when having fewer than 7 samples per phenotype, and permutation number was set at 1000 for testing significance. The most current Affymetrix .CSV file was used for mapping HT Rat230 + PM probes. A P-value and FDR of less than .05 was considered significant.

Leading edge genes were identified for each significantly enriched pathway. These genes are considered key drivers of the enrichment signal because they comprise the vast majority of the ES by occurring within the ranked gene list prior to the maximum deviation from zero. Once generated, leading edge gene lists were populated with their corresponding BMD values from the BMDExpress analysis. For each time point, genes within the leading edge of each significantly enriched Hallmark pathway at each dose were compiled to construct a combined leading edge gene list. For detailed information on GSEA methodology, refer to Subramanian et al. (2005) and Mootha et al. (2003).

Workflow for Combining GSEA and BMD Modeling

To identify key drivers of biologically relevant signaling pathways, we investigated the utility of combining GSEA and BMD modeling results to identify chemical exposure doses that would activate signaling through molecular pathways known to be critical for cellular homeostasis or contribute to various disease states. Use of the MSigDB Hallmark gene set definitions was preferred due to its ability to define large scale biological processes. With greater than 13 000 individual gene sets encompassing most biological processes and disease states, there is a great amount of redundancy across and heterogeneity within the vast number of gene sets available for interrogation. To address this, Liberzon et al. (2015) computationally refined large groups of gene sets that shared common biological responses to produce the Hallmark gene sets, which are defined from many founder sets and display coherent expression among associated gene members with low noise. The resultant Hallmark pathways, thus, represent large-scale biological networks or signaling pathways as determined through a variety of experimental outputs, the exact composition of which can be further examined at the MSigDB website (http://software.broadinstitute.org/gsea/msigdb/collections.jsp).

As depicted in Figure 1, raw gene expression data are processed via RMA normalization prior to analysis. Initially, the comparison of each treatment to the corresponding control for each dose at each time point was performed to identify enriched pathways, as well as leading edge gene members for those pathways. Individual genes that were leading edge members of any significantly enriched pathway at any dose for a given time point were then modeled using BMDExpress. To further refine this list of genes and to avoid potential bias from poorly fitting models, any gene with a goodness-of-fit P-value < .1 was removed from further analysis. Likewise, to avoid potential bias introduced from model extrapolation, any gene with a BMD value greater than the highest dose tested was removed from further consideration. Pathways demonstrating enrichment but containing no genes that could be adequately modeled, or those displaying a BMD higher than the highest dose tested, were removed from further analysis. Because pathway signaling contributions cannot currently be solely attributed to individual gene members, no minimum number of genes was prescribed to derive a pathway BMD value. Pathway BMD distributions of genes responsible for driving enrichment of significantly enriched pathways were visualized as box-and-whisker plots generated using GraphPad Prism version 4.0c (GraphPad Prism Software Inc., La Jolla, CA). Whiskers within box and whisker plots represent quartiles. This workflow outlines the application of GSEA to identify biologically relevant molecular signaling events that occur in response to chemical exposure. When combined with BMD modeling of gene expression dose-response data, the identification of doses (or dose ranges) at which these molecular signaling events are activated can be characterized.

FIG. 1.

Flow chart depicting the analysis of microarray dose-response data using GSEA and BMD modeling. GSEA was employed as a means to identify qualitative transcriptional pathway points of departure through characterization of enriched cellular processes and phenotype-associated signaling events in exposed rats as compared to controls. Patterns of significant transcriptional enrichment were observed across dose and time and used to characterize chemical-specific molecular modes/mechanisms of action. Additionally, BMD calculations of dose-response data were paired with individual genes identified as drivers of biological pathway enrichment (ie, the leading edge subset) to determine overall pathway BMD distributions and quantitatively estimate doses at which enriched cellular processes are altered.

RESULTS

GSEA Characterizes Enrichment of Biologically Relevant Transcriptional Programs in a Dose-, Time-, and Chemical-Specific Manner

As depicted in Figure 2A, Hallmark gene sets were grouped into 9 higher order biological functions (stress response, DNA damage, proliferation, metabolism, development, hormone signaling, cellular component, immune, and cellular signaling pathways) for the purpose of forming a legend and aiding in interpreting GSEA outputs.

FIG. 2.

Enrichment of biologically relevant transcriptional programs occurs in a dose-, time-, and chemical-specific manner. A, Wind rose legend depicting NESs of Hallmark gene sets from the GSEA molecular signatures database. These signatures have been derived by aggregating many gene sets to represent well-defined biological states and are grouped in this legend by common biological theme, represented by the individual color groupings and oriented in a clockwise manner. B, Wind rose plots depicting the NES of Hallmark gene sets in male Sprague-Dawley rat livers following oral exposure to 10, 25, 50, 100, or 200 mg/kg/day 2,3,4,6-tetrachlorophenol. NESs are plotted for each dose at 5-day, 2-week, 4-week, and 13-week time points allowing for examination of significantly enriched signature patterns over dose and time. Dotted lines represent boundaries for significantly enriched NES as determined by those NESs having a P-value and FDR <.05. Therefore, any NES outside of the outermost dotted circle can be considered to have a significantly enriched positive NES, and an NES being inside the innermost dotted circle will be considered to have a significantly enriched negative NES. C, Wind rose plots depicting the NES of Hallmark gene sets in male F344 rat livers following oral exposure to 25, 100, 200, 300, or 400 mg/kg/day bromobenzene. Time points and significance levels used are as depicted in panel B. D, Wind rose plots depicting the NES of Hallmark gene sets in female F344 rat following exposure to 250, 1000, 2000, 3000, or 4000 ppm N-nitrosodiphenylamine. Time points and significance levels used are as depicted in panel B.

Using GSEA to analyze dose-response gene expression data from a series of 6 chemicals, distinct patterns of enrichment can be observed across increasing dose and time. Illustrative examples are provided herein. While enrichment scores are presented in the wind rose format to facilitate the comparison across biological function, a complete listing of all significantly enriched pathways across all doses and time points is presented in Supplementary File 1.

As can be observed in Figure 2B, GSEA was capable of clearly detecting biologically relevant molecular signaling enrichment in rat livers following oral exposure to TTCP for 5 days, 2 weeks, 4 weeks and 13 weeks. At 5 days, all doses tested displayed a significantly enriched pro-proliferative signature and a generic immune response centered on interferons γ and α. Interestingly, significant enrichment of cellular stress sensing and response pathways was observed at ≥25 mg/kg/day suggesting that these doses were eliciting a unique stress not observed at 10 mg/kg/day. Significant pro-proliferative and interferon γ and α signaling were detected at all doses of TTCP at 2 weeks, while additional proliferative pathways (G₂M checkpoint and Myc targets V2) were enriched at ≥25 mg/kg/day. A combined DNA damage/p53 and apoptosis signal was observed at ≥100 mg/kg/day, suggesting that higher exposure levels are required to drive significant cell death signaling at 2 weeks. At the 4-week time point, enrichment of proliferative signaling pathways was only observed at higher doses (≥100 mg/kg/day); however, immune response signaling through interferon γ and α persisted at all 4-week doses. MTORC1 signaling was observed to be elevated at ≥50 mg/kg/day, along with the cellular stress pathways responsible for responding to hypoxia and unfolded proteins. Interferon gamma- and alpha-mediated signaling persisted at all doses at the 13-week time point, thus, representing a pathway that was significantly enriched across all doses and time points examined following exposure to TTCP. MTORC1 signaling, TNF-α (via NF-κB) signaling, and pro-angiogenic signaling were elicited at ≥25 mg/kg/day, while doses ≥50 mg/kg/day promoted cellular stress and death pathways (apoptosis, reactive oxygen species [ROS], hypoxia, unfolded protein response, and p53 signaling). Finally, a pro-proliferative signaling network of E2F targets and G₂M checkpoint targets was only enriched at the highest dose tested (200 mg/kg/day) at 13 weeks. All together, these data suggest that at 5 days, TTCP doses ≥25 mg/kg/day can drive significant cell stress and death, and that the response of the affected liver tissue is to upregulate pro-proliferative pathways in order to regenerate and replace damaged tissue. These effects largely carry over into the 2-week time point, wherein doses that previously elicited stress and death (25–50 mg/kg/day) are now refractory to these effects, presumably due to compensatory (metabolic or otherwise) mechanisms taking place over time. These compensatory responses carry over into the 4-week time point at which point the apoptotic response has largely been mitigated, but stress pathways remain detectable (hypoxia and unfolded protein response) at doses ≥50 mg/kg/day. MTORC1 signaling is enriched at these same doses and is known to be a common liver response mechanism for maintaining overall tissue size and function. Proliferative signaling is enriched at only ≥100 mg/kg/day, suggesting that the need for the liver to regenerate healthy tissue mass may only be necessary at elevated doses. Finally, at the 13-week time point, MTORC1 signaling seems sufficient to maintain liver stasis at TTCP doses below 50 mg/kg/day, which are observed to promote significant amounts of stress in the form of ROS, hypoxia, unfolded protein response, and ultimately cell death. Again, pro-proliferative signaling pathways induced in an attempt to mitigate the apparent TTCP-induced stress responses are only observed in the highest dose tested.

In the case of BRBZ (Figure 2C), characterizing molecular signaling events following oral exposure to 25, 100, 200, 300, and 400 mg/kg/day also reveals another seemingly liver-specific molecular response. At 5 days, all doses examined are enriched for pro-proliferative signaling (E2F targets, G₂M checkpoint, and Myc targets V2) as well as prominent cell stress and death signaling pathways (ROS, unfolded protein response, p53 signaling, and apoptotic signaling). Direct upregulation of xenobiotic metabolism pathways were detected at doses ≥100 mg/kg/day and DNA damage repair pathways were enriched at ≥300 mg/kg/day, suggesting that high levels of acute BRBZ exposure may be capable of promoting direct DNA damage. Notably, significant negative enrichment of interferon γ and α-mediated immune responses were observed at ≥100 mg/kg/day and ≥200 mg/kg/day, respectively. These findings are contrary to those observed in response to TTCP, wherein interferon γ and α signaling were both significantly positively enriched at all doses and time points. All 2-week, BRBZ-exposed samples displayed significant enrichment of p53-mediated signaling as well as unfolded protein-response signaling indicating that a detectable amount of stress was still being induced. Additionally, xenobiotic metabolism pathways were enriched as well as MTORC1 signaling, both of which likely represent the attempt of BRBZ-exposed liver tissue to mitigate chemically-induced damage and stress. These mitigation mechanisms seem generally effective, as livers exposed to BRBZ doses <200 mg/kg/day no longer display significant enrichment of cell death signaling. The increased cell death at higher doses coincides with elevated enrichment of pro-proliferative pathways at ≥200 mg/kg/day, again suggesting that hepatocellular proliferation is being employed to cope with BRBZ-induced cell death at doses ≥200 mg/kg/day. At 4 weeks of BRBZ exposure, doses ≥100 mg/kg/day are observed to elicit p53-mediated responses in addition to the unfolded protein response stress pathway. MTORC1 signaling is concurrently enriched, suggesting that this may provide the liver tissue a mechanism through which the organ can retain healthy tissue mass and overall function. Upon analysis of doses ≥200 mg/kg/day, significant upregulation of ROS response pathways as well as hypoxia response pathways occurs. Enrichment of these pathways is observed concurrently with robust enrichment of multiple pro-proliferative pathways (ie, E2F targets, G₂M checkpoint, mitotic spindle, Myc targets V1 and V2), suggesting that doses ≥200 mg/kg/day provide a dose beyond which MTORC1 signaling can no longer compensate for the loss of healthy tissue mass, and pro-proliferative measures are required to maintain normal liver function and size. These general liver phenomena continue out to the 13-week time point wherein rats exposed to ≥200 mg/kg/day demonstrate significantly enriched p53-mediated signaling in combination with ROS response pathways, suggesting that these exposures result in sustained levels of cellular damage, likely through generation of ROS. While not consistent, all doses ≥200 mg/kg/day display significant enrichment of multiple pro-proliferative pathways. Taken together, the hepatocellular response to BRBZ exposure is one, wherein MTORC1 and proliferation pathways share the burden in maintaining and restoring healthy liver tissue mass and function after chemically induced stress or damage. Compensatory and/or metabolic effects can be observed to display dose-dependent enrichment and ultimately result in mitigation of cell death by 13 weeks, despite persistent DNA damage and cell stress signaling occurring via p53 and generation of ROS.

Analysis of dose–response gene expression changes in bladder tissue of rats orally exposed to NDPA reveals unique patterns of transcriptional enrichment in the context of proliferative signaling. Specifically, lower doses (≥1000 ppm) of NDPA exposure promoted enrichment of proliferative pathways only at later time points (4 weeks), implying a potential requirement for accumulation of NDPA or a metabolic product thereof. In contrast, treatment with ≥2000 ppm NDPA promoted robust enrichment of proliferative transcriptional programs (E2F, G₂M checkpoint, and mitotic spindle) at all time points as well as a distinct inflammation transcriptional program. While these pro-proliferative transcriptional pathways are known to reflect the act of tissue regeneration within the liver, these same pathways are known to synchronize and connect the cell cycle machinery to DNA damage and repair genes and response pathways in the bladder. These events facilitate cell fate decision-making and are best interpreted in the context of additional signaling events. As such, apoptotic transcriptional programs are significantly enriched at doses ≥3000 ppm (5-day and 2-week), ultimately suggesting that damage occurring at 2000 ppm would likely be repaired, while increasing the exposure level of NDPA to ≥3000 ppm drives robust enrichment of cell death signaling. These patterns of transcriptional alteration are largely carried through the 2-week time point. At 4 weeks, all doses display significant enrichment of proliferative pathways (E2F targets, G₂M checkpoint, and mitotic spindle), while MTORC1 signaling is enriched at doses ≥2000 ppm, and p53-mediated transcriptional programs remain enriched at ≥3000 ppm. Finally, at the 13-week time point, cell motility transcriptional programs (EMT and myogenesis) are enriched at the lowest doses (≤1000 ppm). Importantly, at this time point exposure levels ≥2000 ppm were shown to promote sustained enrichment of proliferative signaling pathways (E2F, G₂M checkpoint, and mitotic spindle), inflammation and p53-mediated transcriptional signaling. This combination of transcriptional events is suggestive of sustained cellular damage following subchronic-duration exposure to NDPA.

Identification of GSEA-Derived Transcriptional Enrichment Patterns

While identifying unique transcriptional signaling events is useful for summarizing and characterizing molecular MOAs, identification of chemical doses capable of promoting sustained transcriptional signaling enrichment can assist human health risk assessors in characterizing which apical responses are likely to exhibit adaptation and which may persist in a chronic-duration exposure setting. Additionally, understanding the transcriptional enrichment pattern of single (or few) administered doses across multiple time points could inform regulatory decisions when existing toxicity data are limited.

At ≥50 mg/kg/day TTCP exposure, pathways indicative of cellular stress (unfolded protein response), cellular proliferation (Myc targets V2) and xenobiotic metabolism are significantly enriched across all time points suggesting sustained levels of hepatocellular stress at and above this exposure level (Figure 3A). Upon examination at higher doses (≥100 mg/kg/day), enrichment of pathways associated with apoptosis, p53 signaling, ROS, inflammatory response, and cholesterol homeostasis can be observed at all time points, demonstrating a clear increase in severity of the molecular response to TTCP exposure with increasing dose. In this context, 50 mg/kg/day could be considered a molecular tipping point, and a clear departure from the enrichment patterns observed at doses <50mg/kg/day. Exposures at or above this dose promotes consistent enrichment of molecular signaling pathways that are known to be liver-specific responses to extracellular stress. In comparison to enrichment patterns observed at doses ≤25 mg/kg/day, ≥50-mg/kg/day TTCP exposure represents an increase in the severity of the molecular response that occurs across all time points examined, suggesting it is the lowest dose examined to elicit this molecular response.

FIG. 3.

Identification of GSEA-derived transcriptional enrichment patterns. Venn diagrams depicting the number of significantly enriched Hallmark gene sets at each dose for 5-day, 2-week, 4-week, and 13-week time points. Lists below Venn diagrams describe significantly enriched gene signatures relevant to the biology associated with previously described apical endpoints. Numbers within Venn diagrams represent shared or individual significantly enriched pathways. A, 2,3,4,6-tetrachlorophenol exposure in male Sprague-Dawley rat livers demonstrating dose-specific enrichment consistent across time, B, bromobenzene exposure in male F344 rat livers demonstrating transient cellular stress and compensatory proliferative transcriptional pathway enrichment at early time points after exposure to less than 200 mg/kg/day, and C, N-nitrosodiphenylamine exposure in female F344 rat bladders demonstrating a biphasic proliferative response after exposure to less than 200 ppm.

Enrichment observed in response to BRBZ exposure (Figure 3B) describes a similar pattern. However, at lower doses (≤100 mg/kg/day) an immediate molecular signaling enrichment pattern consisting of pro-proliferative (E2F targets, G₂M checkpoint, mitotic spindle) and cellular stress/death signaling (apoptosis, ROS, p53, hypoxia, and inflammatory response) was characterized at the 5-day and 2-week time points only. These data suggest that the immediate/early response to low-dose BRBZ is mitigated at later time points, as these events are undetectable beyond 2 weeks. Only upon reaching exposure levels ≥200 mg/kg/day can these events be observed as significantly enriched beyond the early time points, identifying 200 mg/kg/day as a clear departure from the transient signaling events that are enriched at ≤100 mg/kg/day. While doses >200 mg/kg/day elicit transient, early time course effects that would be considered canonical cellular responses to extracellular stresses (ie, apoptosis, DNA repair, xenobiotic metabolism, etc.), the group of enriched pathways that is observed at all time points for exposures ≥200 mg/kg/day remains generally unchanged, suggesting that these molecular signaling pathways constitute the core of the sustained, long-term molecular response to higher dose (≥200 mg/kg/day) BRBZ exposure.

Lastly, NDPA-mediated transcriptional enrichment patterns described in Figure 3C depict that enrichment of proliferative responses identified at lower doses (≤1000 ppm) occurs in a biphasic manner that is detected at both 5 days and 4 weeks but is absent in the subchronic-duration (13-week) setting. Proliferative transcriptional program enrichment (E2F targets, G₂M checkpoint, and mitotic spindle) was uniform at all time points at doses ≥2000 ppm, suggesting that the presence of NDPA-mediated stress may require entry of the bladder cells into active cell cycle to rectify stress-induced damage. Additionally, apoptotic and DNA damage and stress sensing (p53) pathways were frequently enriched in the majority of time points at doses ≥2000 ppm, reinforcing an association between induction of pro-proliferative transcriptional program enrichment and cell death in response to NDPA exposure and providing a clear delineation in the doses producing these molecular effects in a transient versus chronic-duration manner.

Transcriptional BMD Value Distribution of Gene Subsets Responsible for Driving Biological Enrichment

For each significantly enriched pathway characterized via GSEA, leading edge genes responsible for driving enrichment were identified. Biologically, these genes are considered to be key drivers of enrichment activity for the corresponding pathway. Upon resolving the leading edge gene lists for each pathway, a three-tiered filtering process was implemented to ensure that the genes ascribed to a given pathway were (1) demonstrated to be in the leading edge of at least one dose comparison at the time point being evaluated; (2) sufficiently BMD modeled through possessing a goodness-of-fit P-value >.1; and (3) plausibly biologically activated by displaying a BMD value that was below the highest dose examined. After filtering, the remaining genes are considered to be key drivers of biologically relevant molecular signaling. Once this gene list was generated, the distributions of BMD values were displayed as box-and-whisker plots for BRBZ (Figure 4), TRBZ (Supplementary Figure 1), HZBZ (Supplementary Figure 2), TTCP (Supplementary Figure 3), MDMB (Supplementary Figure 4), and NDPA (Supplementary Figure 5). For brevity, only results obtained from BRBZ will be discussed.

FIG. 4.

BMD values of genes driving the enrichment of biologically relevant transcriptional programs. Box-and-whisker plots of BMD values derived for leading edge genes associated with significantly enriched (P < .05, FDR <.05) pathways. Bars are colored by grouping into larger scale biological process as depicted in Figure 2. Hatched bars represent pathways that demonstrate negative NES values following chemical exposure. Red arrows on the x-axis denote overall median of enriched gene BMD values. Distribution of BMD values for gene members of significantly enriched pathways identified in male F344 rat livers exposed to 25, 100, 200, 300, or 400 mg/kg/day bromobenzene at A, 5 days, B, 2 weeks, C, 4 weeks, or D, 13 weeks.

The distribution of BMD values for genes identified as drivers of enriched liver transcriptional pathways following exposure to BRBZ exposure was initially investigated (Figure 4). Analysis of the transcriptional response to BRBZ revealed a distinct response at 5 days, at which a large number of Hallmark gene sets were significantly enriched (Figure 4A), many of which were no longer enriched at subsequent time points (Figs. 4B – D). The initial 5-day response to BRBZ exposure consisted of significant enrichment of proliferative pathways (blue bars) as well as DNA damage (red bars), suggesting a canonical liver stress-response was occurring. This pattern of enrichment was continued through the remaining time points and displayed a similar distribution of BMD values at each time point, demonstrating that enrichment of these molecular pathways was not only consistent but occurred at similar dose ranges over time. Interestingly, the median BMD value of enriched cellular stress response pathways (orange bars) generally decreased over time. At 2 weeks (Figure 4B), the number of significantly enriched pathways was decreased appreciably and the median BMD value for all significantly enriched genes was decreased from 221 mg/kg/day to 157 mg/kg/day (indicated by red arrows). The median BMD values for stress response pathways occur at a dose lower than that for DNA damage pathways, suggesting that DNA damage events may be secondary to or a product of the cell stress pathways. Many prominent signaling cascades (yellow bars) that were enriched at 5 days are no longer enriched at 2 weeks, or display a negative NES, suggesting that these pathways may be responsible for the immediate/early molecular response to BRBZ exposure and potentially driving an adaptive response. Overall median BMD values are similar at 183 mg/kg/day and 191 mg/kg/day for the 4-week (Figure 4C) and 13-week (Figure 4D) time points, respectively. The majority of significantly enriched pathways at these time points are stress response constituents, or damage or proliferative cellular processes, and continue to display patterns of median pathway BMD values wherein stress response BMDs < DNA damage BMDs < proliferation BMDs. Ultimately, while identifying significantly enriched pathways can provide insight into the molecular consequences of chemical exposure, identifying and BMD modeling gene sets responsible for driving individual pathway enrichment can identify specific exposure dose ranges associated with transcriptional enrichment of individual pathways.

Comparison of GSEA-Identified Transcriptional BMD Values and Apical Endpoint BMD Values

To assess the association of biologically relevant transcriptional BMD values and those identified for apical endpoints, median transcriptional BMD values were calculated for the most sensitive enriched pathway and the overall median BMD value for all leading edge gene members of significantly enriched pathways. These values were compared to apical endpoint BMD values previously identified as the most sensitive for each chemical tested (Thomas et al., 2013). With respect to the liver toxicants tested, the median BMD value of all filtered, leading edge genes comprising significantly enriched pathways (filled symbols) generally served as a good estimator of the most sensitive apical endpoint BMD value at early (5 day) time points only (Figure 5A). At 2, 4, and 13 weeks, liver BMD values identified for the most sensitive significantly enriched molecular pathway for each chemical (open symbols) were observed to be more relevant for approximating the BMD values of the most sensitive apical endpoints (Figs. 5B – D). For thyroid-specific responses following exposure to MDMB, BMD values associated with the most sensitive enriched transcriptional pathway were more predictive of most sensitive apical endpoint at all time points. Conversely, the overall median of gene BMD values of leading edge genes best approximated the most sensitive apical bladder endpoint at all time points following NDPA exposure. The values represented throughout Figure 5 are also reported in Supplementary Table 1. Importantly, only 4 out of 48 calculated transcriptional BMD values (which include the most sensitive pathway median BMDs and overall leading edge gene median BMDs) displayed an order-of-magnitude difference from the previously identified most sensitive apical BMD value.

FIG. 5.

Relationship between biologically relevant transcriptional BMD values and identified most sensitive apical endpoint BMD. Scatterplots of most sensitive apical response BMD values and GSEA-identified transcriptional values for liver-specific chemicals at A, 5-day, B, 2-week, C, 4-week, and D, 13-week time points (TRBZ (∎), BRBZ (●), and TTCP (▴)). Open symbols indicate median BMD value for the most sensitive significantly enriched pathway. Filled symbols represent the median BMD value of all filtered, leading edge genes comprising significantly enriched pathways and as depicted in Figure 1. Thyroid-specific MDMB responses (E) and bladder-specific NDPA responses (F) are also depicted. For each chemical, apical endpoint BMD values previously identified as the most sensitive endpoint as determined by Thomas et al. (2013) are represented. Solid lines represent a 10-fold difference in BMD values. Dashed lines indicate equal apical and transcriptional BMD values.

DISCUSSION

Transcriptomics provide a potential avenue for cost-effective and efficient hazard evaluation of data-poor chemicals. In the context of gene expression microarray data, evaluation of expression levels of virtually all identified protein-coding genes across dose and time can provide insight into the molecular underpinnings of the biological response to chemical exposure. These studies generate large amounts of data and provide a challenge in deciphering relevant biological signals from molecular noise. In fact, this has been the primary hindrance to incorporating transcriptomic data into human health risk assessment, and current efforts have been directed at increasing confidence in the ability to use this information in a regulatory setting (Bourdon-Lacombe et al., 2015).

Genome-wide transcriptional changes elicited across increasing dose and time provide a large number of potential critical effect and BMD values (eg, individual gene modeling, pathway modeling, etc.) from which a single POD could be identified. Previous efforts directed at applying gene expression data to human health risk assessment have described applicability in using the median BMD value of the most sensitive activated transcriptional pathway following repeated-dose chemical exposure (Thomas et al., 2011; Thomas et al., 2013). The efforts reported in this study by Thomas et al. demonstrated that BMD values of the most sensitive transcriptional responses were generally within 2-fold of the most sensitive apical endpoint identified. In the Thomas et al. (2013) study, although the concordance among transcriptional and apical BMD values was demonstrated, the biological relevance of the identified most sensitive transcriptional pathways was elusive and the potential association between the etiologies of the apical responses to the identified transcriptional events was uncertain.

GSEA seeks to identify biologically relevant transcriptional events by identifying deregulation of biologically derived gene sets. There are two important distinctions in the principles underlying GSEA compared to traditional gene expression microarray data filtering methodologies. First, the strength in using experimentally derived gene set definitions lies in the ability to define the gene sets based on an observed phenotype or apical endpoint/outcome. Previous definition databases have relied heavily on manual curation of literature sources by knowledge base experts, and while this results in the generation of highly inclusive pathway definitions, generating aggregate transcriptional patterns directly from many experimental sources allows for the capturing of only the most robust and coordinated signals that drive a specific apical endpoint. In addition to the unique definition structures, GSEA uses the entire unfiltered data set and does not require fold cutoff filtering or ANOVA-based significance as it seeks to identify modest, coordinated transcriptomic changes. Given the complexities of understanding individual gene contributions as well as tissue-specific transcriptional regulation, it is prudent to retain the entire data set when interrogating transcriptomic information. This substantially reduces the chances of missing potentially important findings. In total, GSEA, in combination with the MSigDB Hallmark definitions, provides a mechanism through which modest but coordinated biologically relevant transcriptional events can be reliably and impartially characterized with no prior knowledge of apical outcomes.

In this study, GSEA was used to evaluate transcriptomic data generated in response to chemical exposure. The subsequent application of GSEA-identified patterns of biologically relevant transcriptional enrichment to chemical risk assessment was evaluated in three ways. First, we examined the ability of GSEA to identify enriched transcriptional events that could be used to inform MOA and WOE approaches and evaluated whether enriched signatures corroborated each other within the overall transcriptional milieu in a dose- and time-dependent manner. Second, we evaluated the utility in using GSEA as a stand-alone methodology for identifying and characterizing molecular enrichment patterns to identify doses capable of sustained deregulation of biologically relevant transcriptional programs. Lastly, upon pairing GSEA outputs with BMD modeling of individual genes responsible for driving identified biological enrichment, we investigated the distribution of BMD values and the ability to approximate previously determined apical PODs.

For gene expression analysis to become accepted for routine use in human health risk assessment, it is necessary to demonstrate that the mRNA profiles captured reflect the biological outcome of exposure. As depicted in Figure 2, GSEA-identified patterns of enrichment can be grouped into larger, broad-scale, biological processes that allow comparison of enrichment among grouped transcriptional processes. For example, enrichment of proliferative pathways and cellular stress pathways can be observed at all doses following 5 days of exposure to BRBZ (Figure 2C). After 2 weeks of exposure, four distinct stress response pathways (apoptosis, reactive oxygen species, hypoxia, and unfolded protein response) remained enriched, but only doses ≥200 mg/kg/day displayed proliferative enrichment. By 4 weeks, exposures ≥100 mg/kg/day induce enrichment of stress response pathways, a trend that is continued out to 13 weeks of exposure. Proliferative enrichment is observed at ≥200 mg/kg/day and also continues out to 13 weeks of exposure, but is slightly reduced in enrichment while remaining statistically significant. While these are conclusions regarding only two of the nine larger Hallmark groupings outlined in Figure 2A, the concordance between enrichment of two highly interconnected pathways provides increased confidence that GSEA is identifying transcriptional events relevant to the biology that underlies the response to exposure. We can observe similar results over time, wherein patterns of enrichment appear at consecutive time points and depict time-dependent molecular adaptation. In the case of proliferative pathway enrichment following exposure to BRBZ, all doses elicit a proliferative burst following 5 days of exposure, while only the higher doses are still driving this response at 13 weeks. These results suggest that compensatory or adaptive events are occurring between 5 days and 13 weeks, and that a proliferative component is likely to be relevant only at exposures ≥200 mg/kg/day in a subchronic- or chronic-duration setting. In the context of stress response enrichment, a similar pattern can be observed wherein all doses elicit an initial stress response, and by 13 weeks the lowest dose (25 mg/kg/day) does not. Again, these results characterize a transcriptional adaptation to the stress response being elicited by chemical exposure over time and demonstrate that GSEA can help identify distinct transcriptional phases dependent on dose and time.

A direct application of GSEA is to use identified patterns of transcriptional enrichment to determine which molecular events are transient and which are sustained and incorporate this information to potentially support adverse outcome pathway (AOP) development or general mechanistic conclusions. Characterization of doses capable of driving sustained molecular alterations can help to identify those capable of driving long-term biological alterations. Conversely, doses that elicit exclusively transient molecular effects may not be expected to drive significant long-term biological changes. Taken together, the ability to differentiate among doses that promote significant long-term molecular alterations and those whose effects are mitigated through compensatory molecular activity can help to identify molecular tipping points and differentiate distinct biological states. While an optimally designed study would include a range of doses and time points, identification of individual doses that are capable of promoting significant long-term enrichment has potential use in situations where limited dose response data exists, or when existing legacy data are of limited scope.

As demonstrated in Figure 3, doses that displayed sustained enrichment of cell cycle/proliferation pathways and/or cell stress response pathways could be readily identified and signify a molecular tipping point, below which, enrichment of these pathways was transient and largely nonexistent at later time points tested. Doses above the identified molecular tipping points demonstrated consistent enrichment of the core set of pathways across all time points, while doses below did not. These data reinforce the notion that GSEA is identifying coordinated transcriptional changes that are occurring at doses associated with previously identified adverse apical endpoints. Also, through identifying doses capable of driving sustained enrichment of prominent cellular signaling pathways, GSEA could serve as a stand-alone means to identify doses likely to result in adverse outcomes, and potentially inform chemicals with little-to-no available hazard information.

While there is apparent use in the application of GSEA-identified patterns of enrichment to aspects of human health risk assessment, the ability to supplement the results of GSEA with BMD modeling of dose-response gene expression data have direct implications for quantitative chemical risk assessment. Combined, these two methodologies identify a dose range responsible for activating transcription of genes identified as key drivers of individual enriched, biologically relevant pathways, and provide insight into the molecular events occurring at individual doses. While it is intriguing to place emphasis on individual genes identified as being more sensitive (ie, display lower BMD values) within an enriched pathway, current knowledge surrounding the relative contribution of individual gene members to composite pathway signaling remains very much incomplete, but anticipated to be the focus of future investigations. Such efforts are likely to lead to identification of genes that could serve as pathway biomarkers in specific contexts or in response to specific chemicals and uncover any associations between individual gene expression sensitivity and overall pathway signaling.

In comparing GSEA-identified most sensitive pathway BMD values and overall median gene BMD values to previously identified most sensitive apical BMD values, it is notable that these values are all generally within an order of magnitude. Interestingly, these data may suggest tissue-specific differences in the most sensitive GSEA-identified pathway versus the overall median gene BMD value, as the former was shown to be a more accurate estimate of the MDMB-induced apical response in the thyroid, and the latter more accurate in estimating the NDPA-induced apical bladder response. Additionally, it was observed in the context of the liver that the overall median gene BMD value more accurately approximated the most sensitive apical response BMD only after 5 days of exposure, while longer exposure durations tested displayed a preference for GSEA-identified most sensitive pathway BMDs. While these findings demonstrate that both GSEA-identified most sensitive pathway median BMDs and overall enriched gene median BMD values are reasonable for estimating the most sensitive apical endpoint BMD values, the predictive ability of identified transcriptional BMDs should be improved when comparing them to BMD values from a suite of apical endpoints (irrespective of sensitivity) driven by enrichment of potentially causal molecular pathways.

While GSEA has the ability to uncover and characterize coordinated changes in gene expression, the enrichment scoring statistics are only as informative as the robustness of the pathway definitions and the quality of the experimental design that generate the data. Also, as the use of GSEA grows, there will be a need to continue to expand the library of gene set definitions to encompass a more diverse range of apical endpoints. Currently, existing computationally derived Hallmark gene sets are limited to large-scale cellular processes. Gene set definitions that define highly specific physiological states that may be tissue-specific or chemical class-specific currently lack the experimental support necessary to compute robust gene set lists. Over time, as additional gene expression data sets encompass a more diverse range of cellular responses, highly specific gene set definitions will become available for secondary analyses and will support characterization of transcriptional enrichment from both broad, multi-pathway transcriptional events as well as highly specific, small-scale alterations. Given the impetus for incorporating high-throughput technologies into human health risk assessments, future efforts must also be directed toward reaching consensus as to which molecular events can be considered adverse. Once refined, these efforts have potential use in chemical screening and prioritization by identifying chemicals likely to drive uniquely adverse events, or informing read-across techniques for surrogate-type analyses by grouping like chemicals based on molecular or predicted apical responses (Wang et al., 2012). As additional molecular characterization of chemical-specific responses are performed, transcriptional enrichment patterns can be employed as a molecular fingerprint of response and used to identify and group chemicals by the specific biological response that they elicit. As a screening methodology, GSEA-identified transcriptional enrichment could be used as a first-pass screening procedure to identify candidate in vitro assays to be used to functionally corroborate key events within a MOA or AOP.

The results described herein demonstrate that GSEA is capable of detecting enrichment of large-scale transcriptomic events following chemical exposure, and that these events can be plausibly linked to the etiology of the chemically induced biological response through patterns of enrichment apparent across increasing dose and time. Identification of these molecular events has clear implications within both qualitative and quantitative human health risk assessment. Given the biologically driven nature of the gene set definitions and increased confidence in results obtained from independent cohorts across a dose-response, information obtained from GSEA can be used to inform molecular MOAs and WOE approaches, as well as fill in data gaps or support existing conclusions within AOP frameworks. We demonstrate that these events occur in patterns that can be discerned with no prior knowledge of apical endpoints and used to identify molecular tipping points representing distinct molecular states and responses. Furthermore, using GSEA to identify candidates of enriched groups of genes for BMD modeling facilitated characterization of BMD ranges for identified pathways. While this information essentially identifies the dose range responsible for promoting the transcriptional enrichment of GSEA-identified pathways, it also allows for visualization of how different doses are capable of promoting biological enrichment of diverse pathways and combinations of pathways and allows for comparison of relative potencies in the context of individual molecular pathways. Importantly, the ability to query the enrichment of user defined gene lists allows for follow-up analysis to be conducted and finer mapping of transcriptional events within highly specialized molecular pathways.

Taken together, the concepts presented herein outline a framework in which initial GSEA is performed using the Hallmark gene sets, with patterns of enrichment informing subsequent rounds of analysis, ultimately promoting a comprehensive understanding of the molecular events occurring on a large multi-pathway scale, in addition to the more specialized linear signaling events that comprise them. Much additional work will be required before achieving complete integration and comprehensive understanding of molecular events as they pertain to chemical exposure. However, we conclude that the use of GSEA to analyze transcriptomic data in response to chemical exposure represents a significant step toward applying genomic analyses to human health risk assessment. Through characterization of coordinated disruption of large-scale cellular processes and employment of biologically relevant gene set definitions, these technologies may facilitate more rapid and efficient risk assessments while also increasing our working knowledge of the molecular underpinnings of responses to chemical exposure.