shows a heat map of the entire in vitro data set, providing an overview of the data. Generally, the biochemical HTS assays (indicated by red in the top band) had fewer hits than did the cell-based assays, as evident from the increasing density of hits progressing from left to right in the heat map. On the left side of this plot are 87 assays that had no AC50/LEC values identified for any of the chemicals at levels below the highest concentration tested (see for concentration ranges tested). In , all hits are shown, up to where the AC50/LEC occurred at the highest tested concentration. However, some of these values may not be physiologically relevant because in vitro systems can be exposed to concentrations higher than can occur in vivo in relevant tissues under conditions of a bioassay. Supplemental Material, Figure 1 (doi:10.1289/ehp.0901392) shows the number of hits per chemical as a function of the threshold AC50/LEC values used to define a hit. At the comparatively low threshold of 1 μM, there were relatively few hits per chemical. There were 828 chemical–assay pairs (0.5% of pairs tested) with an AC50/LEC < 1 μM (listed in Supplemental Table 2), many of which were related to nuclear-receptor–mediated xenobiotic metabolism. Of the chemicals that had AC50/LEC values < 1 μM in multiple assays, some showed cytotoxicity in one or more of the cell-based assays, which suggests cytotoxicity pathway activation, although in many cases we do not have a specific (cell-free) assay that would indicate which pathway that was. Cytotoxicity may comprise a relevant end point of specific biological process(es) leading to cellular demise (e.g., apoptosis), or it may comprise nonspecific collapse of cellular homeostasis (e.g., necrosis). Both are considered in phase I, and the former may be the result of targeted pathways engaged by specific molecular lesions, whereas the latter may generally follow from nonspecific cell injury. In other chemicals, we only saw specific targeted activities at these low concentrations, without any accompanying cytotoxicity.
Figure 1 Heat map of 624 assay measurements (including multiple time points where available) in ToxCast phase I data set. Assays are arranged left to right, and chemicals are arranged top to bottom. The color bar at the top indicates the assay type: red (cell-free (more ...)
Confidence in the predictive power of in vitro
HTS data builds from many examples that confirm reported mechanisms of action for a number of well-studied chemicals. For example, bisphenol A, a known estrogen receptor (ER) agonist (Chapin et al. 2008
), had AC50
/LEC values < 1 μM for three separate ER (estrogen receptor, ESR1) assays [Supplemental Material, Table 2 (doi:10.1289/ehp.0901392)]. Expected ER activity at concentrations < 1 μM was also found for methoxychlor’s potent metabolite 2,2-bis(4-hydroxyphenyl)-1,1,1-trichloroethane. Similarly, results for the well-known androgen receptor (AR) antagonists linuron, prochloraz, and vinclozolin (Wilson et al. 2008
) showed activity in AR assays (linuron, 57 μM antagonist, 5.1 μM binding; prochloraz, 12.5 μM binding; vinclozolin, 27 μM antagonist, 0.9 μM binding). Expected peroxisome proliferator–activated receptor (PPAR) activators perfluorooctanoic acid (PFOA) and perfluorooctane sulfonic acid (PFOS) (DeWitt et al. 2009
; Lau et al. 2004
), diethylhexyl phthalate (Melnick 2001
), and lactofen (Butler et al. 1988
) were all positive for PPARγ assays, and all but PFOS were also active in PPARα assays. Azoxystrobin, fluoxastrobin, and pyraclostrobin were active mitochondrial poisons in the HepG2 (hepatocellular carcinoma cell line G2) high-content cell-imaging assays, consistent with their pesticidal mode of action (Brandt et al. 1988
). Thus, the redundancy and complementarity of multiple assays allow an integration of data across multiple assay technologies to boost confidence in the results. In some cases, in vitro
results include indications of other biological pathways being activated by these well-studied chemicals, suggesting that other modes of action may be operative as well. To take one chemical as an example, PFOS shows activity against several matrix metalloproteinases, with AC50
values for direct interaction with matrix metalloproteinase (MMP)3 and MMP13 in cell-free HTS assays (14.6 and 32.4 μM, respectively) and perturbation of MMP1 and MMP9 levels in a cell-based assay (13.3 and 4.4 μM, respectively). MMPs are involved in the breakdown of extracellular matrix during development and tissue remodeling. These and other interactions could lead to the formation of specific hypotheses to test regarding toxicity mechanisms of these chemicals.
Activity against human genes and pathways
Most of the ToxCast assays use human proteins and cells because our ultimate aim is to predict human toxicity. Assays probed 231 human genes either through direct interactions with the relevant protein or using a variety of indirect, downstream readouts of mRNA or protein levels. These genes were mapped to 143 published pathways from the KEGG (Kyoto Encyclopedia of Genes and Genomes) (Kanehisa et al. 2002
) and Ingenuity Systems (http://www.ingenuity.com
). From these human-based assays, composite gene and pathway perturbation scores were calculated. We computed “gene perturbation scores” for the subset of genes for which we had one or more assays, and these were assigned an LEC value for each chemical. The LEC is the minimum AC50
/LEC value for that chemical in any assay that was mapped to that particular gene. We also computed “pathway perturbation scores,” which were assigned the minimum AC50
/LEC value for a chemical in any assay that was mapped to a gene in the pathway. For a chemical to be considered active in a pathway, it had to have shown activity in at least five assays that mapped to that pathway. A total of 122 pathways had at least one chemical hit. [The chemical-by-pathway assay LEC values are given in Supplemental Material, Table 3 (doi:10.1289/ehp.0901392).] This collection of published pathways show significant overlap, so we also compiled a minimal set of 33 pathways inclusive of all genes represented in the total pathway set. Although this is a small subset of the total number of human pathways that could lead to toxicity, it allows us to sample the range of potential activities across phase I chemicals. Supplemental Figure 2 shows a network diagram of the minimal set of pathways linked to the genes for which we have assays. From this one can see redundancy between pathways in the down-selected target set.
shows the distribution of hits across all assays, direct assays, and gene and minimal pathway perturbation scores, as a function of the minimum AC50/LEC value used to define a hit. Direct assays are those measuring perturbation of chemical–target activity in an optimized biochemical assay (). The balance of the assays are cell based and mostly measure up- or down-regulation of particular genes or proteins through direct or indirect mechanisms of chemical activity. Because indirect effects can arise from multiple direct chemical–target interactions, chemicals logically show broader activity in these assays. The number of direct assay and total assay measurements for human targets are 130 and 425, respectively. In general, the ratio of hits between direct and indirect is much less than the overall ratio of the number of direct to indirect assays. Some chemicals show a large number of hits against direct targets. At a 30-μM cutoff for activity, nine chemicals have at least 20 direct hits: emamectin benzoate, fentin, imazalil, mancozeb, maneb, metiram-zinc, milbemectin, oxytetracycline dihydrate, and PFOS. Mancozeb, maneb, and metiram-zinc are different salts of the same parent, and emamectin benzoate and milbemectin are related macrocyclic antibiotics. Overall, however, these nine chemicals are structurally diverse. shows the same distribution of hits for the gene and minimal pathway assays. Note that the scale for the pathways is significantly smaller because of the requirement that chemicals hit at least five pathway-mapped assays to be considered to have a positive pathway perturbation score. Except at the lowest cutoff of 1 μM, the median number of hits for genes or minimal pathways is > 5, and a number of chemicals show much broader activity than this. The chemicals that hit ≥ 20 of the minimal pathways with a 30-μM cutoff are fluazinam, mancozeb, maneb, metiram-zinc, and pyraclostrobin.
Figure 2 Distribution of number of hits per chemical as a function of AC50/LEC cutoff used to define a hit. (A) Distributions for all human assay measurements (out of 425) and the “direct” measurements from the cell-free HTS assays. The other assays (more ...)
This broad range of activity is not seen universally across chemical classes. shows the distribution of hits against the minimal pathway set with chemicals parsed by chemical class (limited to classes with at least 10 chemicals). The conazoles and triazoles (many of which overlap) and pyrethroids show the broadest activity spectrum, with median number of pathway hits of around 10 of the 33 minimal pathways. In contrast, the sulfonylurea and phenoxy compounds are active in only a few pathways on average. However, even across the broadly active chemical classes, there is a spectrum of activity. These findings show that environmental chemicals are active across multiple human genes and pathways.
Figure 3 Distribution of number of hits against the 33 minimal pathways by chemical class (active at concentrations of < 30 μM). Only chemical classes with at least 10 chemicals are included. In each box and whisker plot, the heavy bar indicates (more ...)
We next examined the consequence of the multiplicity of pathways perturbed by these chemicals. A simple analysis is to see if the likelihood of cytotoxicity increases with the number of pathways in which a chemical is active. The data set includes 15 cytotoxicity assays using 11 primary human cell types or cell lines. We found a strong correlation between the number of pathways in which a chemical is called active and the minimum concentration at which cytotoxicity is observed across 15 cytotoxicity assays. shows the correlation between the number of pathway hits and the minimum AC50/LEC for cytotoxicity across the 15 assays. The p-value for the association is < 2.2E-16, and R2 = 0.55 for linear correlation.
Figure 4 Plot of the minimum concentration at which a chemical caused cytotoxicity as a function of the number of minimal pathways in which the chemical was active at concentrations < 30 μM. Chemicals for which no cytotoxicity was observed were (more ...)
We tested the hypothesis that the lower the concentrations at which a chemical shows activity in vitro, the lower will be the doses at which in vivo toxicity will be observed for that chemical. This hypothesis is based on three assumptions: a) Pathways perturbed by a chemical in vitro will also tend to be perturbed in vivo, although the magnitude may be very different because of tissue-specific feedback or adaptation not active in vitro. b) Pathway perturbations in vivo arising from specific chemical–target interactions require chemical concentration at the target site to be in the range where effects on the in vitro assay are seen; hence, lower in vitro AC50 values imply lower concentrations at which in vivo effects are seen. c) There are combinatorial pathways that, when perturbed, can lead to a given observed toxicity, and the AC50 values for the toxicity-related pathways for a chemical will be distributed randomly through the total distribution of AC50 values, including some in the low concentration tail of that distribution.
To test this hypothesis, we first looked for direct correlations between low in vitro pathway perturbation score AC50 values for the minimal pathway set and the lowest dose at which toxicity was seen in vivo. Because we have only sparsely sampled the space of direct targets (e.g., enzymes, receptors), we used the number of pathways perturbed below some concentration threshold as a surrogate estimate for minimum concentration at which a chemical significantly perturbs pathways. This is based on the assumption that each chemical shows a distribution of AC50 values across the complete set of pathways and that this distribution has a long tail going toward low concentrations. More pathway hits below a defined cutoff will correlate with the entire distribution shifting toward lower concentrations. For each chemical and each in vivo study type in ToxRefDB, we tabulated the lowest dose at which any treatment-related effect occurred. A linear regression fit between the number of pathway hits at concentrations < 30 μM (trend and significance is relatively insensitive to this cutoff) and the lowest dose at which toxicity was observed yielded p-values of 0.0031 (chronic rat), 0.0007 (chronic mouse), 0.037 (developmental rat), 0.053 (developmental rabbit), and 0.019 (multigenerational rat). Except for the developmental rabbit study, all study types showed a significant association at the 0.05 level. In addition, the sign of the association was correct in all cases: The higher the number of low-concentration in vitro pathway hits, the lower the observed lowest toxic dose in vivo. Therefore, these results show a significant association between low in vitro concentrations for pathway perturbations caused by a chemical and the lowest dose at which treatment-related effects are first seen in vivo.
We also performed the association calculation using the short-term half-maximal lethal dose (LD50
) (International Programme on Chemical Safety 2005
) as a covariate. LD50
has a strong correlation with the lowest dose at which other toxic effect occurs and can help correct for factors not included in the pathway parameter, including pharmacokinetics. In models including both terms, the p
-values for association between the number of pathway hits at concentrations < 30 μM and the lowest dose at which toxicity was observed were 0.0019 (chronic rat), 0.00015 (chronic mouse), 0.00049 (developmental rat), 0.011 (developmental rabbit), and 0.00063 (multigenerational rat). We see stronger correlations between in vitro
activity and the threshold of toxicity after adjusting for LD50
, and the sign of the effect was as hypothesized in all cases. The example in shows the results of the model fit for prenatal developmental toxicity in rats, which resulted in the highest correlation across the five study types (R2
Figure 5 Association between the number of minimal pathway hits (which we assume is inversely correlated with the minimum concentration at which significant pathway activity occurs for the chemical) and the lowest dose in vivo at which a significant toxicity end (more ...) Rat liver tumors and PPAR signaling
Almost half of the tested chemicals caused tumors in either rats or mice in high-dose 2-year chronic/cancer bioassays (Martin et al. 2009a
), with most of these having been determined by the U.S. EPA to be nongenotoxic tumorigens (U.S. EPA 2009
). Of the 309 chemicals tested, 248 have rat 2-year chronic/cancer bioassay data entered into ToxRefDB, and 21 of these are liver tumorigens [chemicals shown in Supplemental Material, Figure 3 (doi:10.1289/ehp.0901392)]. These 21 are a subset of the 97 chemicals that are rat tumorigens of any tissue type. All rat liver tumors caused by this set of chemicals were hepatocyte derived. We tested for univariate associations of all in vitro
assays and gene perturbation scores against all rodent liver in vivo
end points, and identified a total of five in vitro
assays with a significant association with rat liver tumors (Fisher’s exact test p
-value < 0.01). Results for these five assays and for the 21 chemicals that are rat liver tumorigens are illustrated in Supplemental Material, Figure 3. Three of the five assays are associated with the nuclear receptor pathway genes PPARA
, one is associated with the cytokine chemokine (C-C motif) ligand 2 (CCL2
), and the last with the AR. The PPARA
transcription reporter assay shows high specificity (0.99) but low sensitivity (0.19) (Fisher’s exact p
-value = 0.0005). The relative risk of causing rat liver tumors for chemicals being positive for this assay was 9.5. The PPARG
assay shows high sensitivity (0.86) but low specificity (0.53) (Fisher’s exact p
-value = 0.0009). Also associated with rat liver proliferative lesions is hydroxymethylglutaryl-coenzyme A synthase 2 (HMGCS2
), which is a gene regulated by PPARA
, providing indirect evidence that the human PPARα pathway has been activated by this group of chemicals.
PPAR activation is a well-studied mechanism or mode of action for chemically induced liver tumors in rodents (Abbott 2008
; Klaunig et al. 2003
; Lai 2004
; Peters 2008
; Takeuchi et al. 2006
). The primary role of PPARs is in lipid and fatty acid metabolism; however, xenobiotic compounds may activate PPAR in hepatocytes, leading to induction of xenobiotic metabolizing enzymes as well as peroxisome proliferation and hepatocyte hypertrophy. During prolonged exposure to PPAR activators, rodent hepatocytes can become hyperplastic, necrotic, or apoptotic, and in some cases neoplastic. The relevance of PPAR-mediated rodent tumors to human toxicity and disease is an active area of research and debate (Desvergne et al. 2009
; Guyton et al. 2009
; Klaunig et al. 2003
). Nonetheless, based on the carcinogenic potential of PPAR-activating compounds, current U.S. Food and Drug Administration (FDA) guidance on PPAR agonists requires 2-year carcinogenicity evaluations in rats and mice before initiation of human clinical studies longer than 6 months (U.S. FDA 2008
CCL2 levels have been shown to be associated with severity or progression in a number of tumor types (Roca et al. 2008
). CCL2 helps drive angiogenesis (Kuroda et al. 2005
). There is also evidence linking CCL2 with up-regulation of bile acids, cholestatic liver injury, and fibrogenesis in rats (Ramm et al. 2009
). Studies have discovered linkages between AR and androgen levels and hepatocellular carcinoma in humans and animals [reviewed by Kalra et al. (2008)
There is extensive evidence that perturbing androgen signaling activity is associated with increased risk of liver tumors. AR is expressed in the liver of rats (Konoplya and Popoff 1992
) and humans (Iqbal et al. 1983
), and hepatic tumor development is likely influenced by androgens, as indicated by the fact that males have a greater prevalence of liver neoplasms in humans (Curado et al. 2007
) and rodents (Kemp and Drinkwater 1989
). Elevated levels of testosterone (Grange et al. 1987
) are associated with increased risk of hepatic adenomas in men. In male rats, testosterone (Morris and Firminger 1956
) promote rat liver tumor development. The hypothesized mode of action for the liver tumorigenicity of AR antagonists such as vinclozolin and linuron is as follows: The antiandrogens block AR function and negative feedback of the pituitary, so more luteinizing hormone is produced, which in turn leads to increased production of androgens by testicular Leydig cells. Whereas androgen homoeostasis may eventually reset, animals can have significantly elevated androgen levels, which can in turn promote liver tumor development.
We also investigated associations between in vitro
assays and the progression of liver disease in rats. Chemicals were categorized according to rat liver disease progression: those causing no liver lesions (122 chemicals) or causing any type of liver lesion (126 chemicals). Chemicals causing liver lesions could be classified further into subsets of those causing preneoplastic or neoplastic liver lesions (58 chemicals), or those causing just neoplastic liver lesions (21 chemicals). All assays were correlated against these three rat liver lesion categories. shows associations with a p
-value < 0.01 (either t
-test or Fisher’s exact test), in which the genes linked to assays statistically associated with the three rat liver lesion categories, as well as human disease categories assigned through the Online Mendelian Inheritance in Man (OMIM) database (Goh et al. 2007
, and CCL2
are all associated with preneoplastic and neoplastic levels in the liver disease progression, and PPARA
is additionally associated with neoplastic lesions.
Figure 6 Network of genes associated with the progression of rat liver tumor end points. Associations were calculated using Fisher’s exact test, with assay AC50/LEC values ≤ 100 μM set to 1 and those with > 100 μM set to (more ...)
More than half of genes with any association were involved with xenobiotic metabolism in the liver (9 of 15), with most of these being cytochrome P450 enzymes. Many of these XME genes are regulated by PPAR or other nuclear receptors, and other assays indicated direct associations with rat and human pregnane X receptor (NR1I2). Preneoplastic and neoplastic liver lesions are also associated with PPARG activation. These data suggest that induction of liver neoplasms by these chemicals is PPARA dependent, and potentially coupled with PPARG and CCL2, whereas a variety of xenobiotic metabolism and other pathways can lead to more general liver lesions.