An initial evaluation of assay performance showed that the assays behaved as expected in terms of biological activity of known agonists and antagonists included as positive control compounds in the screening libraries. The Tox21 compound collection contains a set of 54 known NR ligands assembled by Sigma (St. Louis, MO) (Sigma 2007). lists the ligands with known interactions with NRs included in the present study and their activities observed [phenotype and half-maximal effective concentration (EC50) or inhibitory concentration (IC50)] in the corresponding NR assays. We positively identified all of these known ligands with our assays, with expected potencies confirming the utility of these NR assays. Some of the apparent actives identified from the LXRβ assays indicated a potential problem with the LXRβ cell line (data not shown). We thus excluded data from these assays from further analysis.
NR ligands and their observed assay activities.
shows the distributions of compound activity outcomes in the agonist- and antagonist-mode assays, respectively. In general, more compounds were active in the antagonist-mode assays than in the agonist-mode assays (), in which cytotoxicity might be playing a role. The percentage of compounds classified as active ranged from 0.4% (FXR) to 3.2% (ERα) in the agonist-mode assays and from 3.3% (PPARδ) to 10.9% (AR) in the antagonist-mode assays. When we excluded compounds identified as potentially autofluorescent and/or cytotoxic (see below for criteria applied), the fractions of apparent activators and inhibitors decreased for both the agonist- and antagonist-mode assays (data not shown), but the number of apparent active compounds remained larger in the antagonist-mode assays, although by a smaller margin.
Distributions of compound activity outcomes in the agonist-mode (A) and antagonist-mode (B) NR assays. (C) Distribution of active agonists and antagonists.
Compound reproducibility. There were 130 compounds replicated in the U.S. EPA plate and 66 compounds in the NTP plate, and 416 compounds overlapped between the U.S. EPA and the NTP plates. We calculated compound reproducibility for all NR assays in both agonist and antagonist mode using the ratio readout. We first defined each compound replicate as an agonist, an antagonist, inconclusive, or inactive based on its curve rank [for details, see Supplemental Material, Table 4 (doi:10.1289/ehp.1002952)]. We then made three types of reproducibility calls (match, mismatch, and inconclusive) based on the concordance of each replicate (). Overall, the intraplate replicates showed slightly better reproducibility (88.6%) than did interplate replicates between the U.S. EPA and NTP libraries (85.5%). Both mismatch (4.1%) and inconclusive (10.4%) rates were slightly higher for the interplate than for the intraplate replicates (3.5% mismatch and 7.9% inconclusive). Variations in compound reproducibility were assay dependent (). The overall matching rate, counting both intra- and interplate replicates, ranged from 97.7% (GR agonist) to 74.2% (RXRα agonist); mismatch rates ranged from 0.3% and 0.5% (GR and FXR agonist mode) to 9.0% (RXRα agonist), and inconclusive rates ranged from 2.0% (GR agonist) to 16.8% (RXRα agonist).
Figure 2 Intralibrary (A) and interlibrary (B) compound reproducibility across different NR assays. Intralibrary reproducibility is calculated by comparing the activity of copies of each compound replicated within the U.S. EPA or NTP compound library. Interlibrary (more ...)
The reproducibility of an assay is a good indicator of assay performance and quality. To generate a single measure of reproducibility so that all assays could be easily compared, we scored all NR assays (2× % active match + % inactive match – % inconclusive – 2× % mismatch) and ranked them by this score, sorted in descending order (). We calculated the reproducibility score in a way such that assays with higher concordance rates and lower mismatch rates would be ranked higher. We assigned more weight to active matches and mismatches because we derived these results from more reliable concentration–response curves and to account for the overall low active rate. We assigned each assay an arbitrary grade: A (score, ≥ 90%), B (≥ 80% to 90%), C (≥ 70% to 80%), or D (< 70%), with A being the highest-quality assays in terms of reproducibility and D the lowest. The grades are meant only to serve as a guide for assay prioritization. With this ranking scheme, we ranked the GR and FXR agonist-mode assays as the best-performing assays and the RXRα agonist-mode assay as the worst performing, with the lowest data reproducibility. The agonist-mode assays performed better overall than did the antagonist-mode assays—among the top 50% performing assays, six were agonist-mode assays and only three were antagonist-mode assays. The top five assays were all agonist mode. This outcome was not entirely unexpected because the antagonist-mode assays all required the pre-addition of nonsaturating levels of an agonist compound to stimulate the receptor signal before test compounds could be screened, which introduces an extra source of variance.
NR assays ranked by their compound reproducibility.
Using single-channel readouts of
bla assays to assess autofluorescence and cytotoxicity.
Compound autofluorescence and cytotoxicity can interfere with assay readouts and produce artificial results, because fluorescent compounds could appear as activators and show up as false positives in agonist-mode assays, and cytotoxic compounds could appear as inhibitors because of reduced cell viability and show up as false positives in antagonist-mode assays. The green channel (530-nm readout) is the control channel of the bla
assay. Increased or decreased fluorescence activity in this channel can be interpreted as an indicator of compound autofluorescence or cytotoxicity (Xia et al. 2009b
). Blue fluorescent compounds may not be detected in the 530-nm readout but could still interfere with the 460-nm readout, which is the blue, reporter-gene–dependent signal channel of the bla
assay. Therefore, pan-activation in the blue channel across multiple NR assays would also indicate compound autofluorescence. We then identified a compound as autofluorescent if it showed activation in > 10 of 20 agonist-mode assay readouts (counting both the 530-nm and 460-nm readouts separately) or if it showed activation in the 460-nm readout in more than four agonist-mode assays and was identified as fluorescent (activity > 10% fluorescent control compound) at 460 nm in the autofluorescence spectra scans (Simeonov et al. 2008
). Using these criteria, we identified 25 compounds (11 from NTP and 14 from U.S. EPA) as autofluorescent and excluded them from further analysis [see Supplemental Material, Table 5 (doi:10.1289/ehp.1002952)]. The criteria chosen for identifying autofluorescent compounds (as well as cytotoxic compounds) were empirical and were used to minimize false positives. Most of the compounds identified as autofluorescent by this method were well-known fluorophores, partially validating the approach. Further experimental studies are needed to fully confirm the apparent artifacts.
The cell viability assay has been widely used as a measure of compound cytotoxicity (Xia et al. 2008
). We identified active compounds (non-class 4) identified by the cell viability assay in parental HEK 293 cells [for a description of the cell viability assay, see Supplemental Material (doi:10.1289/ehp.1002952)], or compounds that reduced activity in the 530-nm readout of more than four antagonist-mode assays, as cytotoxic and excluded them from further analysis of the antagonist-mode data. We identified a total of 323 compounds as potentially cytotoxic, 152 of which were from the NTP collection and 171 from the U.S. EPA collection (Supplemental Material, Table 6). Cytotoxicity interference is less of a concern for agonist-mode assays than for antagonist-mode assays because agonists that are cytotoxic at higher concentrations generate bell-shaped (inverted-U) concentration–response curves in the 460-nm channel. For more discussion on how activities in the 460-nm and 530-nm reads correlated with the cell viability assay results, see Supplemental Material, “Using single channel readouts of bla
assays to assess cytotoxicity” and Supplemental Material, .
Chemical genomics: compound activity-pattern similarity and NR LBD sequence homology.
The agonist-mode and the antagonist-mode NR assays were both hierarchically clustered using the correlation of the compound curve ranks (from the ratio readout) as the similarity metric. We excluded compounds identified as potentially autofluorescent from the clustering exercises and also excluded compounds identified as potentially cytotoxic from the clustering of the antagonist-mode assays. shows results for agonist and antagonist mode, respectively. The agonist-data–based clustering of NRs () matched nearly perfectly with the NR LBD sequence homology [; see also Supplemental Material, (doi:10.1289/ehp.1002952)] (Laudet 1997
; Zhang et al. 2004
), where the agonist data clustering segregated the nine NRs into two major branches (): the ER-like branch, with ERα, AR, and GR, and the thyroid hormone receptor (TR)-like branch, with RXRα and the rest of the NRs. The clustering again further divided the TR-like branch into two subgroups, one containing TRβ, FXR, and VDR and the other containing the two PPARs and RXRα. The only difference from the sequence clustering () was that PPARδ clustered more closely with RXRα than with PPARγ in the assay-data–based clustering ().
Figure 3 Comparison of the human NR LBD similarity and compound activity-pattern similarity. (A) and (B) Hierarchical clustering (Spotfire DecisionSite, version 8.2; Spotfire Inc., Cambridge, MA) of the agonist-mode (A) and antagonist-mode (B) NR assays using (more ...)
The antagonist-mode data–based clustering of the NRs () showed poor concordance with the NR LBD sequence homology (). The ER-like subfamily members were segregated into two different branches, with ERα and GR in one branch and AR clustered into the other branch (). This was surprising because GR is more closely related to AR than to ERα in terms of sequence similarity (). The clustering grouped the TR-like subfamily into two clusters as well, with VDR, PPARδ, and PPARγ in one cluster and TRβ and FXR in the other (). The two PPARs are the most closely related by sequence (), but this was not reflected in the antagonist-mode–based phenotype clustering ().
We clustered all compounds in the NTP and U.S. EPA libraries based on structural similarity (2,048-bit Daylight fingerprints; Daylight Chemical Information Systems, Inc., Laguna Niguel, CA) using the self-organizing map (SOM) algorithm (Kohonen 2006
), yielding 336 clusters. Despite of the diversity in response [for details, see Supplemental Material, “Structural diversity assessment of apparent NR agonists and antagonists,” and Supplemental Material, (doi:10.1289/ehp.1002952)], we identified 16 classes of compounds with consistent NR activity patterns by examining the activity patterns of compounds within each structure cluster. shows the structure scaffolds and NR activities of these compound classes. Among these are many known ligands or disruptors of NRs whose activities observed in our study was consistent with their known NR activities. Examples include the known ERα-active classes of compounds, such as the estradiol and tamoxifen analogs, parabens (Harvey and Everett 2004
), bisphenyls (including bisphenol A, bisphenol B, and methoxychlor), and flavonoids; steroid hormones and analogs and flutamides with known AR activity; and corticosteroids with known GR activity. We also observed that subtle changes in structures of compounds belonging to the same class led to variations in their NR activity. For example, the class of steroid hormones was clustered into several subclasses based on their NR activity patterns, where the sex hormones appeared as agonists of AR and ERα and the corticosteroids showed activities mostly against AR and GR. Another example is the ERα activity of the flavonoids, where the isoflavones (e.g., genistein) were identified as more conclusive/potent ERα agonists than the normal flavonoids (e.g., kaempferol). Several classes of compounds, including the lactofen analogs (Butler et al. 1988
) and dicarboximide fungicides (Kelce et al. 1994
), have been reported to induce liver toxicity. The lactofen analogs appeared primarily as PPARγ and AR antagonists, and the dicarboximide fungicides as AR antagonists, consistent with literature reports (Gray et al. 2001
); however, the AR activity of the lactofen analogs has not been reported before. Of the chloroacetanilide herbicides, alachlor (Klotz et al. 1996
), and acetochlor (Rollerova et al. 2000
) have been reported to have weak estrogenic effects, consistent with their weak activities observed in our ERα assays. However, we consistently identified this class of compounds as PPARγ antagonists in our NR assays as well. The NR activity of these compounds may be related to their liver toxicity.
Figure 4 Example structure classes with consistent NR activity patterns or signatures. Compounds were clustered by structure similarity using the SOM algorithm. Compounds in the same cluster belong to the same structure class. The structure classes shown contain (more ...)