A Gene Expression Biomarker Identifies Chemical Modulators of the Estrogen Receptor α (ERα) in a MCF-7 Microarray Compendium

Abstract

Identification of chemicals that affect hormone-regulated systems will help to predict endocrine disruption. In our previous study, a 46 gene biomarker was found to be an accurate predictor of estrogen receptor (ER) α modulation in chemically-treated MCF-7 cells. Here, potential ERα modulators were identified using the biomarker by screening a microarray compendium consisting of ~1600 gene expression comparisons representing exposure to ~1200 chemicals. A total of ~170 chemicals were identified as potential ERα modulators. In the Connectivity Map 2.0 collection, 75 and 39 chemicals were predicted to activate or suppress ERα, and they included 12 and 6 known ERα agonists and antagonists/selective ERα modulators, respectively. Nineteen and 8 of the total number were also identified as active in an ERα trans-activation assay carried out in a MCF-7-derived cell line used to screen the Tox21 10K chemical library in agonist or antagonist modes, respectively. Chemicals predicted to modulate ERα in MCF-7 cells were examined further using global and targeted gene expression in wild-type and ERα-null cells, trans-activation assays, and cell-free ERα co-regulator interaction assays. Environmental chemicals classified as weak and very weak agonists were confirmed to activate ERα including apigenin, kaempferol, and oxybenzone. Novel activators included digoxin, nabumetone, ivermectin, and six progestins. Novel suppressors included emetine, mifepristone, niclosamide, and proscillaridin. Our strategy will be useful to identify environmentally-relevant ERα modulators in future high-throughput transcriptomic screens.

Graphical Abstract

INTRODUCTION

Identification of endocrine disrupting chemicals (EDCs) is an important goal of environmental chemical hazard screening. The U.S. Environmental Protection Agency (EPA) has developed a screening program for potential EDCs in which approximately 10,000 existing chemicals would be evaluated for their potential to disrupt the estrogen, androgen, and thyroid signaling systems (The Endocrine Disruptor Screening Program [EDSP]; http://www.epa.gov/endocrine-disruption). Under the guidelines of this program, a battery of Tier 1 in vitro and short-term in vivo screening assays including those that assess nuclear receptor activity were developed which could be followed up with longer term, more definitive in vivo Tier 2 animal tests for endocrine disrupting activity. The EPA’s vision for the EDSP in the twenty-first century (EDSP21) includes utilization of in vitro high throughput screening (HTS) assays coupled with computational modeling to prioritize chemicals, and to eventually replace some or all of the current EDSP Tier 1 screening assays.

Many EDCs interact with and modulate the activity of the two estrogen receptors (ERα and ERβ) which regulate the effects of estrogen hormones including estradiol (E2) on the growth, development, and function of a diverse number of tissues including breast tissue ¹. The two subtypes, like other members of the nuclear receptor family, are similar in overall structure and contain an N-terminal ligand-independent activation domain, a central DNA-binding domain (DBD), and a C-terminal ligand binding domain (LBD) responsible for ligand-dependent transcriptional activation ². The classic mechanism of ER activation involves chemical or hormone binding to the LBD triggering the release of associated complexes of proteins including co-repressors and a number of heat shock proteins, homo- and hetero-dimerization between the two subtypes, and after nuclear localization, binding to estrogen receptor response elements (ERE) in the regulatory regions of target genes. Other mechanisms of activation exist including “non-classical” ligand-independent activation through phosphorylation of ERs by growth factor- and cytokine-dependent signaling cascades ^2,3. Both ligand-dependent and ligand-independent ER activation can lead to different sets of interactions with co-regulator proteins that determine transcriptional initiation at overlapping sets of target genes ⁴.

EDCs can include not only environmentally-relevant industrial chemicals but also drugs that contaminate surface waters. A growing number of structurally and functionally diverse industrial and natural chemicals that modulate ER have been identified. The industrial chemicals include plasticizers, pesticides, polychlorinated biphenyls (PCBs), alkylphenol ethoxylates, organotin compounds, phytoestrogens, and pharmaceuticals such as ethinylestradiol (EE), diethylstilbestrol (DES), and tamoxifen (TAM). These chemicals can bind to ERs and act as agonists and/or antagonists, depending on the tissue-specific context ⁵. Given the costs and time constraints of screening large numbers of chemicals for their ability to cause endocrine disruption in animals (which can occur through many potential pathways that operate at different levels of biological complexity), a number of research organizations are increasingly relying on in vitro test methods to predict in vivo effects. As part of the U.S. Federal government National Institutes of Health Tox21 and EPA ToxCast research efforts, in vitro HTS test methods are being used to screen for EDCs. Eighteen of the Tox21 ER assays have been used to build an ER computational model that can accurately predict effects of chemicals on ER ⁶. The model uses activity patterns across the 18 in vitro assays to predict whether a chemical is an ER agonist or antagonist and was used to screen 1812 chemicals. More recent analysis of the model showed that subsets of assays (with as few as 4) can predict the activity of all chemicals with accuracy equivalent to that of the full assay model ⁷.

A complementary approach to multiple HTS assays is the use of HT transcriptomic (HTTr) screening. HTTr technologies have the potential to examine many more pathways simultaneously and in the near future, could be used in testing programs as Tier 1 assays, defined as those that are carried out prior to more targeted Tier 2 screening to validate predictions based on transcriptomic profiles ⁸. A number of promising techniques are now available that have been adapted to HTS to allow measurement of expression of targeted genes from lysates of treated cells (e.g., ⁹). Parallel computational methods need to be developed to predict modulation of molecular targets and their dose-response characteristics ¹⁰ that can be linked to the network of adverse outcome pathways (AOPs) ¹¹ relevant to chemical-induced toxicity.

One major challenge to interpreting HTTr data is how to predict the primary targets of chemicals. Our group has implemented a computational strategy to build gene expression biomarkers, test their predictive accuracy, and use them in screening programs (e.g., ^12,13). Biomarkers are built utilizing microarray comparisons from cells or tissues in which the factor is known to be perturbed in a predictable manner after chemical exposure or genetic modulation. The biomarker is compared in a pair-wise fashion to microarray profiles of interest using a rank-based method ¹⁴ that allows an assessment of the number of overlapping genes and their degree of correlation. We have used these methods to develop predictive biomarkers for a number of transcription factors that are regulated by xenobiotic chemicals in the mouse and rat liver ^15–20. We have applied this strategy to build and validate a biomarker that can predict ERα modulation in the human breast cancer cell line MCF-7. The method is very accurate resulting in balanced accuracies of 95% or 98% for ERα agonism or antagonism, respectively. The approach can also accurately replicate the results of the ER computational model, allowing the 18 assays to be replaced by one assay that detects both agonists and antagonists ¹³.

To demonstrate the potential of using genomic data to screen for potential EDCs, we used here the ERα biomarker approach to screen for chemicals that modulate ERα in a large compendium of gene expression profiles derived from MCF-7 cells. The profiles include those from the Connectivity Map (CMAP) project in which a collection of genome-wide transcription expression data were collected from cultured human cells treated with ~1200 bioactive small molecules ^21,22. The CMAP database and associated tools have proven useful to identify drug candidates used to treat a number of diseases ²³. Using our biomarker approach, ~170 compounds were identified in the compendium that modulate ERα. While these included known ERα agonists and antagonists/selective ERα modulators, many of these compounds had not been previously shown to have ERα modulating activity. We confirmed the ERα modulating activity of the novel compounds using global and targeted gene expression in wild-type and ERα-null cells, trans-activation assays, and cell-free ERα and co-regulator interaction assays. Novel activators included digoxin, nabumetone, ivermectin, and six progestins. Overall, the results demonstrate that our biomarker screening approach is a useful method to identify ERα modulators in large HTTr screening studies carried out in MCF-7 cells.

EXPERIMENTAL PROCEDURES

Overview of the strategy for identification of chemicals that modulate ERα in MCF-7 gene expression profiles.

Most of the methods used in the study are described in Ryan et al. ¹³, including identification of differentially expressed genes in microarray datasets, details of the assembly of a compendium of gene expression experiments carried out in MCF-7 cells, construction of the ERα biomarker, and comparison of the biomarker to statistically-filtered gene lists in the compendium. In short, a screen for ERα chemical modulators required a gene expression biomarker of ERα-dependent genes and an annotated database of gene expression profiles of statistically-filtered genes (here called biosets). The ERα biomarker is a list of 46 differentially expressed genes that are consistently altered in expression after exposure to ERα modulators (Figure 1). The biomarker includes fold-change values associated with each gene, derived from the average differences in expression across treatment by 7 ERα agonists. A commercially available gene expression database (Illumina BaseSpace Correlation Engine (BSCE); formally called NextBio) was used to assemble the gene expression compendium for chemical screening. The database contains over 139,000 lists of statistically-filtered genes from over 22,800 microarray studies carried out in 16 species (as of June, 2019). Available information about each bioset was extracted from BSCE and used to populate a spreadsheet of the experimental parameters for each comparison. To facilitate analysis, each bioset was annotated for the general category of the perturbant (e.g., hormone) and the specific name of the perturbant examined for each category (e.g., 17β-estradiol). Only biosets generated from experiments in the human breast cancer cell line, MCF-7 were used in the present analysis. Biosets from the CMAP study derived from the prostate cancer cell line PC3 and the human promyelocytic leukemia cell line HL-60, and biosets from the De Abrew et al. study ²⁴ carried out with Ishikawa, HepG2, and HepaRG cells were used to highlight cell line specificity of the biomarker. The ERα gene biomarker within the BSCE environment was compared to each bioset in the database using the Running Fisher algorithm ¹⁴. The method allows an assessment of the overlap in regulated genes between the biomarker and the bioset and whether those overlapping genes are significantly (anti)correlated. Biosets which exhibit expression of biomarker genes that are positively correlated with the biomarker would be predicted to exhibit ERα activation. This activation could be due to direct agonism or occur indirectly (e.g., increasing pools of estradiol). Biosets which exhibit expression of biomarker genes that are negatively correlated to the biomarker would be predicted to exhibit ERα suppression through direct or indirect mechanisms. Due to endogenous ERα activators in the growth media (e.g., ²⁵), MCF-7 cells exhibit some constitutive ERα activity that allows the genes in the biomarker to be regulated in an opposite manner in the presence of ERα antagonists. Results of the comparisons were exported and used to populate the annotated compendium with a Running Fisher test p-value from each comparison and direction of correlation. We have previously used this analysis strategy to accurately identify chemicals that activate or suppress other transcription factors (AhR, CAR, PPARα, STAT5b) ^{15–17,19,26}.

Figure 1.

Strategy for screening an MCF-7 microarray compendium for chemicals that modulate ERα. The ERα biomarker was uploaded into BSCE and compared in a pairwise fashion to all biosets in the database through a series of steps including ranking genes based on fold-change and assessment of correlation using the Running Fisher test. Results were exported from BSCE and used to construct an MCF-7 microarray compendium. Putative modulating chemicals were tested in a number of orthologous assays.

Identification of differentially expressed genes in BSCE microarray datasets.

All differentially regulated genes were identified using the criteria in the BSCE analysis pipeline and are described in detail in Kuperschmidt et al. ¹⁴. Briefly, following platform-appropriate processing and normalization, statistical analysis to identify differentially expressed genes involved Welch or standard t-tests with a p-value cutoff of 0.05 (without multiple test correction) and a minimum absolute fold-change cutoff of 1.2. The CMAP database was downloaded as CMAP 2.0 build01 into BSCE. Even though there was only one biological replicate per chemical exposure (i.e., one Affymetrix .cel file per treatment), statistically significant genes were identified by comparing each treatment with a group of control samples using a t-test to calculate the p-value with an assumption of equal variance between case and controls. Chemicals were excluded from the analysis if there was an insufficient number of corresponding controls (<3 replicates) matched to a treated sample.

Assembly of a compendium of gene expression experiments carried out in MCF-7 cells.

Information in the BSCE database was used to build an annotated compendium of gene expression biosets derived from experiments carried out in MCF-7 cells. First, annotated information from BSCE about human-derived biosets was used to populate a master file with information about each bioset including Biodesign, Biosource, Chemical Name, Gene and Gene Mode (if the comparison examined effects of gene modulation), Phenotype, Tissue, and Study ID. Approximately 150 biosets were removed from subsequent annotation because the full name of the bioset was represented more than once in the database. The table was then filtered for biosets derived from MCF-7 cells, and these biosets were used to populate a separate table. Each bioset was annotated for category and name of the perturbant examined based on the name of the bioset. For example, the bioset called “MCF-7 cells + hexestrol, 14.8µM _vs_ DMSO vehicle” is in the category “Chemical” and the specific perturbant is “Hexestrol”. Biosets that examined more than two perturbants at one time (e.g., exposure to three chemicals vs. control) or that could not be interpreted were not used in any further analyses. The final compendium which included the effects of hormones and specific genes in addition to chemicals contained ~2300 biosets. Only the biosets examining the effects of chemicals were utilized in the present study.

Comparison of the ERα biomarker to biosets in the MCF-7 compendium.

The strategy for comparison of a biomarker to collections of biosets has been described in previous studies ^15–17. Using the Running Fisher algorithm, the ERα biomarker was compared to each bioset in BSCE. The p-value and direction of the correlation were exported. P-values were converted to –log(p-value)s, and those with negative correlations were converted to negative numbers. The final list of –log(p-value)s was used to populate the master table containing the study characteristics of each bioset. This final master table enabled the determination of effects on ERα by categories of perturbants (e.g., chemical) as well as individual perturbants in that category (e.g., genistein). Only chemical effects are described in the present study.

Comparison between ERα biomarker predictions and the Tox21 ERα quantitative HTS assays.

ERα biomarker predictions were compared to two sets of assays that were carried out by the NIH National Center for Advancing Translational Sciences Chemical Genomics Center as part of their chemical screening program ²⁷. Data was taken directly from the following websites on Nov. 30, 2016: https://ntp.niehs.nih.gov/sandbox/tox21-curve-visualization/ and https://ntp.niehs.nih.gov/sandbox/tox21-activity-browser/. The data was derived from the reporter gene assay, ER-luc, run in both agonist and antagonist modes in a quantitative HTS format. The VM7 cell line used was originally thought to be the human ovarian cancer cell line, BG1 but is now recognized to be a variant of MCF-7 cells (https://ntp.niehs.nih.gov/iccvam/methods/endocrine/bg1luc/bg1luc-vm7luc-june2016-508.pdf; accessed Sept. 2, 2019). The cell line was stably transfected with an estrogen-responsive luciferase reporter gene plasmid (pGudLuc7ere) containing the estrogen responsive element (ERE) and luciferase reporter gene ²⁸. Results of the agonist and antagonist assays were filtered prior to comparison with the ERα biomarker predictions to exclude chemicals:

• with marginal activity (i.e., hit calls of 0.5),

• classified as resulting in weak or noisy data,

• assayed in duplicate and gave inconsistent results, i.e., were positive in one assay but not the other,

• that did not meet purity standards as specified in https://tripod.nih.gov/tox21/samples to eliminate substances with suboptimal F-graded purity (Fns, Fc, and F) when either 4-month purity or 0-month purity information was available (when both were available, 4-month purity information was used),

• that exhibited activity that significantly overlapped with concentrations that induced cytotoxicity. Chemicals flagged as cytotoxic were identified as described in ²⁹.

In addition, chemicals were excluded if the concentration of the chemical in the microarray study was below the EC50 value in the HTS assay and was not active using the biomarker approach. This criteria excluded experiments that were examined at concentrations of chemical likely to be insufficient for inducing activation or suppression of ERα.

Cell culture and chemical treatment.

The human breast cancer cell line MCF-7 was purchased from American Type Culture Collection. The C4-12 cells are a variant of human MCF-7 cells, selected for loss of ER expression by long-term estrogen withdrawal. C4-12-ERα cells were stably transfected with ERα to restore ER activity ³⁰. The cells used were between passages 10 and 18. All the three cell lines were maintained in phenol red–free DMEM:F12 medium (Invitrogen) supplemented with 10% FBS , 4 mM L-glutamine and 1X Antibiotic-Antimycotic (Sigma). For all treatments, cultures were changed to 10% charcoal/dextran stripped FBS medium (sFBS; Gemini Bio Products). The chemicals used were provided as stock solutions in DMSO from the USEPA ToxCast inventory. Chemicals were tested at the following concentrations: cyproterone acetate (100 µM), ethisterone (15 µM), lynestrenol (15 µM), megestrol acetate (15 µM), mestranol (15 µM), mifepristone (15 µM), norethindrone (15 µM), pregnenolone (15 µM), progesterone (15 µM), apigenin (100 µM), chrysin (100 µM), daidzein (100 µM), kaempferol (100 µM), gibberellic acid (100 µM), spironolactone (100 µM) and oxybenzone (100 µM) and ivermectin (15 µM) with or without ICI (1 uM). All treatments were for 18 h in stripped FBS medium. 17β-estradiol (E2, 10 nM) was used as the positive control. To test the effect of aromatase inhibition, MCF-7 cells were pretreated with letrozol (200 nM) for 30 min, followed by exposure to ethisterone (15 µM), lynestrol (15 µM), megestrol (15 µM), norethindrone (15 µM), pregnolone (15 µM), progesterone (15 µM), 4’-Androstene-3,17-dione (10 µM), and testesterone propionate (10 µM) for 18 h. For the microarray study, MCF-7 cells were exposed to crotamiton (20uM), bendroflumethiazide (20uM), ketorolac (20uM), nitrendipine (20uM), triprolidine (20uM), DEHP (10uM), flutamide (10uM), phenobarbital (10uM), thyroxine (0.1uM), and E2 (0.1uM) for 6 hrs. Solvent controls were either 0.1% DMSO or 0.01% ethanol. There were three biological replicates carried out on separate days for each treatment group.

Microarray analysis.

Total RNA was isolated from treated cells homogenized in RNAzol^®RT (Molecular Research Center, Cincinnati, OH), purified with the RNeasy MinElute column protocol (Qiagen GmbH, Hilden, Germany), evaluated for integrity using an Agilent 2100 Bioanalyzer (Agilent Technologies GmbH, Berlin, Germany), and quantitated using the NanoDrop spectrophotometer (NanoDrop Technologies, Wilmington, DE). Samples were hybridized onto Human Ref-12 v2.0 Expression BeadChip arrays (Illumina, San Diego, CA) in the National Health and Environmental Effects Laboratory (NHEERL) Genomics Research Core Laboratory using standard Illumina protocols. Arrays were scanned and raw data (.idat files) were obtained using Illumina iScan software (v3.3.28) and analyzed using Illumina GenomeStudio^® Data Analysis Software. Raw gene expression intensities were quantile normalized using Illumina GenomeStudio^®. ANOVA analysis was performed in Partek Genomics Suite with false discovery rate testing (Benjamini–Hochberg test) and a P ≤ .05 for significance. Array data is publicly available at Gene Expression Omnibus, accession number GSExxx. (Reviewers: to be made public upon acceptance of the manuscript.)

ERα trans-activation assays in AroER tri-screen cells.

The chemicals used were provided as stock solutions in DMSO from the USEPA ToxCast inventory. Examination of effects of chemicals on ERα were performed essentially as detailed in Chen et al. ³¹. AroER tri-screen cells were cultured in maintenance medium. Prior to the assay, cells were incubated for 2 days with phenol red-free MEM containing 10% charcoal-treated-FBS and then seeded in 96-well plates at 2 × 10⁵ cells per well. After overnight incubation, as reference controls, cells were treated with E2 (0.5nM), E2 plus ICI (100nM), or DMSO for 24 h. For the ERα assay’s agonist mode, cells were incubated as before and then treated with chemicals at two concentrations (2uM and 20uM). For the ERα assay’s antagonist mode, cells were treated with E2 (0.5nM) together with chemicals at the same two concentrations for 24 h. Ivermectin and oxybenzone were not included in this screen. Luciferase signal was measured using the One-Glo Luciferase Assay System (Promega Corporation), adjusted by protein concentration, and then normalized to each plate’s DMSO and 0.5nM E2 values as internal controls. At least three independent experiments were performed, and in each independent experiment, there were three technical replicates for each treatment. The DMSO background was subtracted for each of replicates.

Assessment of interactions between ERα and coregulatory peptides using MARCoNI (Microarray Assay for Real-time Coregulator-Nuclear receptor Interaction).

This peptide microarray platform contains a set of 154 immobilized (CoR)NR-box sequences of which the interaction with a GST-tagged ERα ligand-binding domain can be characterized by measuring the level of bound fluorescently-labelled GST-antibody as a function of an added compound ³². The method has been previously described in detail elsewhere ^33,34. In agonist mode, all compounds were analyzed at 100 uM in DMSO (final assay mix: 2% V/V), using solvent only and 10 uM 17β-estradiol as negative and positive controls, respectively. In antagonist mode, all compounds were analyzed at 100 uM in the presence of 5 nM E2, using 5 nM E2 in absence or presence of 10 uM 4-hydroxytamoxyfen as negative and positive controls, respectively. Each condition was tested using three technical replicates (arrays). Modulation of ERα interaction with each coregulator motif on the array was calculated as compound-mediated log2 fold change of binding vs. negative control. Significance in alteration of binding vs. negative control was assessed through Student’s t-test and resulting p-values were corrected by false discovery rate analysis ³⁵. Data analysis and visualization were performed in R (R Development Core Team, 2008).

RT-qPCR studies.

Cells were harvested 18 h after the treatment in Trizol. Total RNA was extracted and reverse-transcribed into cDNA by SensiFAST cDNA synthesis kit (BIOLINE, UK). A Bio-Rad CEX 384 real-time PCR system (Bio-Rad, California, USA) and SsoAdvanced Universal SYBR Green Supermix (Bio-Rad, California, USA) were used for quantitative qPCR amplification using gene-specific primers (Supplemental File 1). Initial experiments were carried out with 16 biomarker genes (GREB1, SGK1, MYBL1, RET, EGR3, CXCL2, JAK2, NMRK, PLAUR, EPHA4, SULT1E1, CCNG2, ESR1, ESR2, GPER, PGR). The first 5 genes were found to be the most sensitive and reproducible for an initial set of ER activators. We then performed all studies with the first 3 genes and in one analysis used all 5 genes. The Ct values were normalized with ACTB of the same sample, and the 2^–ΔΔct method was used to analyze the relative changes in gene expression. Student’s t-test was used to determine significance between chemical and chemical + ICI groups, or between C4-12 ERa cells and C4-12 cells. The data are mean ± SEM of at least three separate batches of cells. The criteria of significance was set at p < 0.05.

RESULTS AND DISCUSSION

Using a gene expression biomarker to screen for chemicals that modulate ERα.

In our previous study ¹³, an ERα biomarker was characterized as a potential screening tool to identify putative EDCs that interact with ERα. We proposed that this approach could be used as an alternative to a screening strategy using a battery of 18 HTS assays that measure different aspects of ERα modulation ⁶. The ERα biomarker consists of 46 genes commonly regulated by 7 structurally-diverse ERα agonists. (The genes are listed in Supplemental File 1.) In our approach outlined in Figure 1, putative modulators of ERα were identified by comparing the biomarker gene expression pattern to those in lists of significantly altered genes after chemical exposure using the fold-change rank-based Running Fisher algorithm ¹⁴. This pattern matching approach finds, in an unsupervised manner, biosets with expression patterns of ERα biomarker genes with statistically significant positive or negative correlation. The biomarker has the advantage of simultaneously identifying biosets that have either ERα agonist-like or antagonist-like activity. The term “activation” is used in the present study to encompass all mechanisms that lead to increased activity of ERα including agonism. ERα “suppression” includes true antagonism as well as decreases in activation through other mechanisms. For 114 chemicals originally evaluated, the balanced accuracies were 95% and 98% for activation or suppression, respectively ¹³.

While our first study was designed to determine if the biomarker approach could accurately identify known ERα agonists and antagonists, the present study was carried out to determine the practical use of the biomarker in a HTTr screening setting. Unlike our first study, we used a number of validation techniques to confirm the activity of novel ERα modulators. The validation techniques were those that are typically used to investigate ERα modulators and include both genomic and nongenomic techniques.

We performed a virtual screen for ERα modulators in ~1600 microarray comparisons derived from MCF-7 cells which included ~1200 chemicals. These included 1513 comparisons from the CMAP study and additional comparisons from smaller studies that were annotated in our compendium. Biosets from the CMAP study that exhibited positive or negative correlation to the ERα biomarker (indicated by a Abs(-Log(p-value)) ≥ 4 from the Running Fisher test) were preliminarily classified as potential modulators. To visualize the relationship between the expression of genes in the biomarker as a function of the Running Fisher algorithm p-value, 1513 biosets were compared to the ERα biomarker and ranked by p-value. Figure 2 (left side) shows that the more significant the p-value, the greater the similarity between the bioset and the biomarker in both the direction and the relative magnitude of the expression changes. These biosets with positive correlation included those from cells treated with typical ERα agonists including E2. The drug with the greatest similarity to the ERα biomarker was ketorolac, a non-steroidal anti-inflammatory drug (NSAID). Ketorolac was shown to weakly activate ERα, and the activation was inhibited by cotreatment with ICI (Kanaya et al., 2015). Figure 2 (right side) shows biosets that exhibited negative correlation to the biomarker, with the biosets on the farthest right exhibiting the lowest p-values for negative correlation. In general, these biosets exhibited a pattern of expression that was opposite to that of the ERα biomarker. These biosets included those derived from cells treated with ERα antagonists including ICI 182,780 (also known as fulvestrant) as well as selective estrogen receptor modulators (SERMs). It should be noted that chemicals that inhibit aromatase such as letrozole did not have any activity in our study due to low expression of endogenous aromatase (discussed below). The screening strategy was selective for ERα modulators in MCF-7 cells, as the same chemicals tested in different cell lines showed remarkably different responses. Out of the 1025 CMAP chemicals tested in the human prostate cancer cell line PC3, only one chemical induced an effect marginally similar to the biomarker (activation by profenamine, 11.4 µM; p-value = 1E-4). None of the 1045 CMAP compounds tested in the human promyelocytic leukemia cell line HL-60 had a significant similarity to the biomarker (p-value > 1E-4) (Supplemental Figure 1). The cell line specificity was expected as the ER biomarker was able to identify ER modulators in only a subset of ER-expressing cell types but not other cell types ¹³. The entire dataset is provided in Supplemental File 1.

Figure 2.

Identification of ERα modulators in a virtual screen. Chemicals examined in 1513 individual microarray comparisons (biosets) were examined for significant positive or negative correlation to the ERα biomarker. The −log(p-value)s were rank ordered and superimposed on the heat map that shows the behavior of the ERα biomarker genes in the ranked comparisons. The scale represents fold-change from red (positive) to blue (negative). The dashed lines represent the cutoffs for determining significance Abs(−log(p-value)) = 4. The positions of ketorolac, the aromatase inhibitor letrozole, and the antiestrogen ICI are indicated. The ERα biomarker is indicated on the left (B).

Identification of ERα chemical modulators in the CMAP study.

In the MCF-7 compendium, there were 1226 biosets from CMAP including 1153 unique compounds used to treat MCF-7 cells for 6 hrs. This short time of exposure was selected by Lamb et al. ²¹ to increase the probability of identification of primary targets. Around 10% of the compounds (114 of the 1153 compounds) led to putative modulation of ERα (biomarker correlation p-value ≤ 1E-4). Seventy-five compounds in 78 biosets were found to activate ERα (Figure 3A). ERα activators included many known ERα agonists (17beta-estradiol, 17alpha-estradiol, dienestrol, diethylstilbestrol, estriol, estrone, estropipate, equilin, mestranol, and the nonsteroidal estrogen hexestrol). There were also compounds that fell into a number of categories including progestins (ethisterone, lynestrenol, pregnenolone, ethynodiol diacetate, levonorgestrel, norethindrone, norethynodrel), flavonoids (apigenin, kaempferol, genistein, resveratrol), steroids (danazol, dehydroepiandrosterone, epitiostanol, flurandrenolone, testosterone), and a number of chemicals (the non-opioid analgesic drug, nefopam hydrochloride, cyclosporin, luteolin, and naringenin). Many of these chemicals have been previously found to be ERα activators ^36–40. Chemicals not previously recognized to activate ERα and that were examined using other assays (described below) included 4 strong activators (based on significance of correlation): bendroflumethiazide, crotamiton, nitrendipine, triprolidine and the relatively weak activators, ivermectin and oxybenzone.

Figure 3.

Compounds in the CMAP data set predicted to modulate ERα. A, B. (Top) Heat map showing the behavior of the genes in the ERα biomarker after treatment with the indicated chemical and concentration for 6 h. The scale represents fold-change from red (positive) to blue (negative). (Bottom) −log(p-value)s of the comparisons between the bioset and the biomarker. In (A) chemicals that are ERα activators are shown. Arrowheads indicate the agonist or antagonist treatments used to create the ERα biomarker in Ryan et al. (13) Arrows indicate the five chemicals selected for further study. In (B), the chemical treatments that led to putative suppression of ERα are shown.

There were 39 chemicals in 41 biosets that suppressed ERα. Chemicals that acted as ERα suppressors included a known antagonist (ICI 182,780 (ICI)) and SERMs (clomiphene, raloxifene, tamoxifen) (Figure 3B). The phosphatidylinositol 3-kinase (PI3K) inhibitor LY-294002 also suppressed ERα. ERα is activated by growth factors through cascades regulated by PI3K ⁴¹. In an examination of compounds by drug class annotated in Berninger et al. ⁴², the greatest number of the ERα actives included those that were “sex steroid receptor agonists” and “sex and other steroid receptor antagonist” (data not shown). Other than the sex steroid receptor-associated drugs and the chemicals mentioned above with at least some evidence for interaction with ER, there are many chemicals identified in the CMAP library that are potential agonists or antagonists.

Identification of additional chemicals in the MCF-7 compendium with ERα activity.

We also analyzed a smaller set of chemicals examined in MCF-7 cells not in the CMAP 2.0 dataset, but derived from separate Gene Expression Omnibus (GEO) submissions. Figure 4A (left) shows the behavior of genes in the biomarker after exposure to chemicals predicted to activate ERα as well as the correlation -Log(p-value)s. There were 10 chemicals in 18 biosets predicted to activate ER, most of which were known ERα agonists. These included the phytoestrogen genistein and another phytoestrogen daidzein that is metabolized by bacterial flora in the gut to two other activators that were identified, S-Equol and the (+/−)-isomer of Equol. An environmentally-relevant compound, 4-nonylphenol, was classified as a weak activator in the OECD reference chemical list ⁶. Si-Wu-Tang, a mixture of chemicals used in traditional Chinese medicine, was identified here as an activator of ERα at two different concentrations from the same study ⁴³, supporting our hypothesis that this mixture has phytoestrogenic activities. Tamoxifen and its metabolite 4-hydroxytamoxifen (4-HT) were found to be activators of ERα; both chemicals can act as partial agonists in some cellular contexts and antagonists in the presence of E2⁴⁴ as shown below. The cholesterol metabolite 27-hydroxycholesterol (27HC) was found to be a partial agonist of ERα and promote the growth of MCF-7 cell–derived breast xenografts when propagated in ovariectomized mice ⁴⁵. The proteasomal inhibitor bortezomib, used to treat breast cancer, appeared to be a weak activator of ERα.

Figure 4.

Identification of miscellaneous chemicals in the compendium that modulate ERα. A. Chemicals in the compendium that activate or suppress ERα by themselves. The biosets were derived from MCF-7 cells treated with the compound vs solvent control from the indicated study. The two biosets indicated as Genistein_0.3μM_GSE9936 are from 4-h (left) and 24-h (right) treatments. The scale represents fold-change from red (positive) to blue (negative). B. Chemicals that suppress the ability of ERα to activate in the presence of E2. The comparisons are between E2 plus compound vs E2 alone.

Compounds which act as suppressors were also identified (Figure 4A, right). These included known antagonists (ICI) and SERMs (bazedoxifene, raloxifene). Additional chemicals have not been previously recognized to suppress ERα including the nucleoside analogue and transcriptional inhibitor 4-amino-6-hydrazino-7-beta-D-ribofuranosyl-7H-pyrrolo(2,3-d)-pyrimidine-5-carboxamide (ARC), the p53 inducer BMH-21, metformin, and the laxative oxyphenisatin acetate. Oxyphenisatin acetate and metformin inhibit growth or induce apoptosis of MCF-7 cells ^46,47 similar to known ERα antagonists. The liver X receptor (LXR) activator GW3965 also suppressed ERα. Evidence indicates that LXR controls estrogen homeostasis by regulating the basal and inducible expression of estrogen sulfotransferase (EST, SULT1E1), an enzyme involved in metabolic estrogen sulfation and deactivation. Pharmacological or genetic activation of LXR resulted in SULT1E1 induction, which in turn inhibited estrogen-dependent uterine epithelial cell proliferation and gene expression, as well as breast cancer growth in a nude mouse model of tumorigenicity ⁴⁸. More recent studies indicate that LXR activators can act as both antiproliferative and lipogenic factors in breast cancer cells, but the antiproliferative effect of LXRs is independent of lipid biosynthesis ^49,50, consistent with the hypothesis that LXR activators suppress ERα through depletion of E2 pools. LXR ligand treatment resulted in a significant decrease in estrogen receptor positive breast cancer cell proliferation ⁵¹.

In the compendium, there were a subset of compounds examined in GEO datasets for ability to act as antagonists in the presence of E2 in comparisons with E2 treatment by itself. The chemicals identified as antagonists included antiestrogens (ICI) and SERMs (4-hydroxy-N-desmethyltamoxifen (also called endoxifen), bazedoxifene, lasofoxifene, raloxifene, tamoxifen, 4-hydroxytamoxifen) (Figure 4B). BI2536 was identified as an antagonist-like compound and is an inhibitor of the polo-like kinase 1 (PLK1) protein, a key regulator of cell division which mediates ERα-regulated gene transcription in MCF-7 cells ⁵². While further work is needed to determine how the novel chemicals modulate ERα, our computational approach using the gene expression biomarker correctly identified known agonists and antagonists consistent with our earlier study ¹³.

Comparison of the biomarker to a set of 34 chemicals examined in four cell lines.

After our analysis of the MCF-7 compendium was completed, a large microarray study was published consisting of 34 reference chemicals including ERα actives examined at 6 hrs at 3 concentrations in 4 cell lines ²⁴. The biomarker was compared to the biosets generated from this study. Figure 5A shows the cell type specificity of the responses; biosets achieved significance for activation or suppression in MCF-7 cells but not HepG2, HepaRG, or Ishikawa cells. This finding of cell type specificity of the biomarker response is in line with lack of responses in HL60 and PC3 cells (discussed above). There was significant activation of ERα by the known agonists ethinyl estradiol, genistein, bisphenol A, and dehydroepiandrosterone (DHEA) as well as suppression by the antagonist tamoxifen (Figure 5B). Other potentially novel agonists exhibited activation including clofibrate, di-2-ethylhexyl phthalate (DEHP), farnesol, flutamide, phenobarbital, retinoic acid, thyroxine (T4), and trenbolone acetate (Figure 5C). Except for flutamide, none of these chemicals were examined in other microarray studies in the MCF-7 compendium. We examined the effects of a subset of these chemicals in independent experiments described below.

Figure 5.

Screening 34 chemicals in four cell lines for ERα activity. The 34 chemicals examined at three concentrations at 6 h in four cell lines came from the De Abrew et al. study. (24) A. Comparison of activity by cell line. The −log(p-value)s of the Running Fisher test between each of the biosets and the biomarker were rank ordered and separated by cell line. B. Activity of known ERα agonists and an ERα antagonist. Numbers indicate the concentrations used in μM. C. Activity of eight putative activators of ERα.

Comparison of predictions to Tox21 NIH 10K chemical library screens.

The ERα biomarker results were compared to the results of a screen conducted by the National Center for Advancing Translational Sciences (NCATS) which examined the ERα activity of ~10,500 chemicals (~8,300 unique) in the Tox21 library of compounds. The screen was carried out in VM7 cells containing a stably integrated luciferase gene under control of 4 concatamerized EREs (ER-luc) ²⁸. ERα agonist or antagonist activities were measured in separate assays after 18–24 hrs of treatment. The 750 unique chemicals that overlapped between the CMAP set of drugs and the 10K library were filtered to remove chemicals based on a number of criteria including poor assay performance and inadequate chemical purity (see Methods). Of the 34 chemicals that had agonist activity in the ER-luc assay, 19 (56%) of the chemicals were also predicted to be activators using the ERα biomarker (Figure 6A, red). These included estrogens (17alpha-estradiol, dienestrol, diethylstilbestrol, estriol, estrone, estropipate, mestranol), progestins (ethisterone, lynestrenol, pregnenolone), flavonoids (apigenin, kaempferol, genistein, resveratrol), various steroids (danazol, epitiostanol, testosterone), and nefopam hydrochloride. Out of this group nefopam hydrochloride was the only chemical that is a novel ERα activator. In general, as the potency of the compounds in the agonist assay increased (lower effective concentration (EC)50), there was greater similarity between the gene expression pattern in MCF-7 cells and the ERα biomarker (Pearson correlation coefficient = 0.652) (Supplemental Figure 2). Of the 14 compounds that were found to be agonists in the trans-activation assays but negative in MCF-7 cells using the biomarker (“false negatives”), all but one (furazolidone, EC50 ~ 0.1 uM) were weakly potent (EC50 > ~10 uM), and none of these compounds were known ERα agonists (Figure 6A, yellow). The ERα biomarker identified 20 “false positive” chemicals that were putative ERα activators in MCF-7 cells but were inactive in the ER-luc agonist assay (Figure 6A, green). None of these chemicals have been previously associated with ERα activation. Five of these chemicals are tested in independent assays described below.

Figure 6.

Comparison of ERα biomarker predictions to those from Tox21 ERα high throughput screening assays. A. Agonist predictions. The CMAP chemicals were rank ordered by −log(p-value) and compared to the findings of the Tox21 ER-luc agonist screening assay carried out in MCF-7 cells. Using the ER-luc agonist assay data as the reference, the chemicals were classified as true positives, true negatives, false positives, and false negatives as indicated. B. Antagonist predictions. The CMAP chemicals were rank ordered by −log(p-value) as above and compared to the findings of the Tox21 ER-luc antagonist screening assay. Using the ER-luc antagonist assay data as the reference, the chemicals were classified as true positives, true negatives, false positives, and false negatives as indicated.

Of the 14 chemicals which had antagonist activity in the ER-luc assay, 8 were predicted to be ERα suppressors using the biomarker (Figure 6B, red). These included ICI, SERMs (raloxifene and tamoxifen), cardiac glycosides (proscillaridin), the emetic emetine, the anthelmintic niclosamide, and the antiprogestin mifepristone. The 6 compounds which were found to be antagonists in the ER-luc assay but negative using the biomarker (“false negatives”), had potencies which ranged from EC50 = 0.1~5 uM (Figure 6B, yellow). None of these compounds were known ERα antagonists or SERMs. An additional 14 “false positive” chemicals were identified as ERα suppressors using the biomarker that did not overlap with those from the ER-luc assay (Figure 6B, green). None of these chemicals were known ERα modulators. Thus, the fact that 4 compounds (emetine, mifepristone, niclosamide, proscillaridin) act as suppressors in the biomarker screen or as antagonists in the ER-luc assay supports that these chemicals are antagonists. Recently, emetine was shown to induce the degradation of the ERα protein ⁵³. Niclosamide and another potential novel suppressor parthenolide identified in the biomarker screen were shown to act as inhibitors of breast cancer stem-like cells ⁵⁴.

Using the ER-luc assays as the reference data set, the balanced accuracies of predicting both agonism and antagonism using the biomarker were 77% (Figure 6A,B). In both tests, the specificity was excellent (97% and 98% for agonism and antagonism, respectively). However, the sensitivity was poor (58% and 57% for agonism and antagonism, respectively). Differences between the two assays in predicting ERα actives could be due to a number of factors. There were deficiencies in the CMAP 2.0 data set. Statistically-significant gene lists were derived using a t-test comparison between one treated replicate vs. multiple control replicates at only one time point and one dose level (~10–20 µM for most chemicals). As described below, there were 5 strong activators from the CMAP collection called positive using the biomarker approach but not the ER-luc approach, and these could not be confirmed using other methods. Using a HTTr approach encompassing multiple replicates of one or two high doses at one time point may be appropriate as an initial screen to identify potential positives that can be more completely analyzed using a full dose response thus minimizing the number of samples that need to be screened.

It should be noted that while the biomarker predictions were compared to the ER-luc assay used as the reference set, there is evidence that the ER-luc assay itself does not result in 100% accuracy in a 39 chemical reference set. At best, the balanced accuracies for agonism and antagonism excluding compounds with inconclusive activity outcomes was 93% and 91%, respectively ²⁷. Furthermore, the Tox21 ERa_Luc_VM7_Agonist assay was ranked 7^th out of the 16 ER assays evaluated as to the frequency of use of assays in different models with balanced accuracies >0.92 in efforts to identify the minimal numbers of assays that had the same predictive power as the full set of assays ⁷. The assay was not one of the four that could be used to replace the current set of assays. Thus, we reason that differences between the biomarker results and the ER-luc assay does not mean that the biomarker approach is less accurate at predicting ERα actives but may be due to differences in the assays themselves which lead to identification of different sets of compounds. The two assays are likely complementary in that the ER-luc assay can identify compounds that affect the ability of ER to interact with the ERE and the biomarker may be able to identify not only those compounds but additional compounds that act through other mechanisms that might include ERE-independent interactions.

Assessment of ERα activity using trans-activation assays.

In order to fully assess the predictive ability of the computationally-derived biomarker, a series of experiments were carried out. Putative ERα actives were examined using a number of approaches. First, we utilized transactivation assays in the MCF-7aro-ERE.2 cell line, a luciferase-based reporter useful in identifying ERα modulators with different mechanisms of action ³¹. A number of known agonists were tested and shown to increase luciferase activity including 17alpha-estradiol, dehydroepiandrosterone, diethylstilbestrol, equilin, estriol, estrone, and mestranol (Figure 7A, left. A number of flavanoids were tested including apigenin, genistein, and kaempferol, all of which activated ERα and were predicted to be ERα activators in the Tox21 assay (Figure 7A, middle). The high activation of luciferase by genistein may be due to stabilization of the luciferase protein noted earlier ⁵⁵. Chrysin, digoxin, nabumetone, and nordihydroguaiaretic acid acted as ERα activators (Figure 7A, right). These chemicals were not previously known to act as ERα activators. Digoxin has been shown to bind to ERα ⁵⁶, but no studies have shown that digoxin acts as an activator.

Figure 7.

Examination of ERα activity by transactivation and RT-qPCR assays. In A. and B. chemicals were examined for the ability to either activate the ERE-linked luciferase promoter in AroER Triscreen cells or to act as antagonists to suppress the activation of 500pM E2. Cells were exposed to the chemicals for 24 h at the indicated concentrations. The treatments that showed no values were those in which the assay results were the same or lower than the DMSO control values. A. (Left) Activity of known agonists of ERα. (Middle) Activity of three flavonoids. (Right) Activity of four potentially novel ERα activators. B. Activity of chemicals in the antagonist assay. C, D. Measurement of activation of ERα biomarker genes by known and putative ERα activators. The indicated cells were exposed to the chemicals at the indicated concentration for 18 h, and expression of the genes was measured by RT-qPCR. C. Expression of the three ERα biomarker genes in MCF-7 cells after exposure to seven putative activators. Cells were exposed to the chemicals with and without ICI. D. Activation of ERα by ivermectin. Cells lacking ERα (C4-12) and C4-12 cells expressing ERα were exposed to ivermectin (15 μM for 18 h).

Chemicals were examined for antagonist activity in the MCF-7aro-ERE.2 cell line by assessing the ability of the chemical to suppress luciferase activity in the presence of E2. Tamoxifen and its metabolite 4-hydroxytamoxifen acted as antagonists as expected (Figure 7B). Niclosamide acted as a suppressor in the biomarker studies and was confirmed to act as a suppressor in the MCF-7aro-ERE.2 cell line (Figure 7B). All findings for the luciferase studies as well as the other orthologous assays is summarized in Supplemental File 1.

Activation of ERα measured by RT-qPCR of ERα biomarker genes.

ERα reference activators and a number of test chemicals were examined for ability to affect the expression of four ERα biomarker genes in MCF-7 cells. E2 exposure led to increases in the expression of all four genes that was abolished in the presence of ICI (Figure 7C). We tested a number of chemicals that were classified as weak or very weak agonists by OECD ⁵⁷ including daidzein, kaempferol, apigenin, and chrysin. All of these chemicals activated most or all of the ERα biomarker genes. Activation was abolished in the presence of ICI. Oxybenzone identified in the CMAP collection as a putative ERα activator was found to activate all 4 of the ERα biomarker genes examined. Activation was blocked by ICI. Gibberellic acid was inactive.

We also tested the ability of ivermectin to activate the ERα biomarker genes. Ivermectin is a medication used to treat a number of types of parasite infections. We examined responses in two cell lines useful for determining the contribution of ERα: a MCF-7 derived cell line lacking ERα responsiveness (C4-12) and the C4-12 cell line in which the ESR1 gene is exogenously expressed (C4-12 ERα cell line) ⁵⁸. The C4-12 ERα cell line exhibited a ~110-fold increase in ESR1 expression over the parent cell line as determined by RT-qPCR (Supplemental Figure 3). Ivermectin at 15 uM activated 5 of the ERα biomarker genes as well as RET after 18 hrs of treatment (Figure 7D). Activation was abolished in the C4-12 cell line. Ivermectin is an importin inhibitor that may indirectly result in ERα activation by decreasing the nuclear translocation of HE4, a protein which suppresses ERα activity ⁵⁹.

Activation by a number of putative strong agonists could not be confirmed.

There were 5 compounds in the CMAP collection with very significant positive correlation with the biomarker (-Log(p-value)s = ~11 to ~29) selected for validation (bendroflumethiazide, crotamiton, ketorolac, nitrendipine, triprolidine). Three of the five chemicals were examined in the Tox21 analysis (crotamiton, bendroflumethiazide, triprolidine) and all three were negative for agonist activity (described above). None of the compounds had activity in the MCF-7aro-ERE.2 cell line (Supplemental Figure 4A). We characterized the ability of the chemicals to modulate receptor interactions with coregulators through the MARCoNI assay. The reference agonist E2 induced significant modulation of a large number of ERα-coregulator interactions (Supplemental Figure 4B) including enhanced interactions with coactivator proteins PRGC1, NCOA1, NCOA2, NCOA3, and EP300, and decreased interactions with corepressor proteins DHX30, NCOR2 and CHD9. None of the 5 compounds tested at 100 uM concentrations displayed modulation of any receptor-coregulator interaction. In addition, none of the compounds acted as ERα antagonists in the same assay (Supplemental Figure 4B).

We speculated that the lack of activity of the 5 chemicals could be due to mistakes in processing of the microarray data. We determined if the filtered gene lists for the 5 chemicals derived using a different method would give the same results. The statistically-filtered gene lists were recalculated from the original archived microarray data using a Z-score methodology (see Methods). Comparison of those lists to the biomarker gave similar -Log(p-value)s when compared to the -Log(p-value)s derived from the BSCE lists for known agonists (E2, equilin, alpha-E2), putative antagonists (mifepristone, corticosterone), and three of the five test chemicals (bendroflumethiazide, crotamiton, nitrendipine) (Supplemental Figure 4C). However, there was no longer significant correlation to the biomarker for ketorolac and triprolidine.

We next determined if we could obtain similar gene expression profiles for the 5 chemicals using similar exposure conditions. MCF-7 cells were exposed to the five chemicals at 20 uM or E2 at 100 nM for 6 hrs, and Illumina microarrays were used to examine global gene expression. Statistically-filtered gene lists were derived as described in the Methods. While the profile from E2 had significant correlation to the biomarker, the profiles of the 4 chemicals with significantly altered genes lacked any significant correlation (Supplemental Figure 4D). There was also no significant correlation between the individual filtered gene lists derived in our study and the list of genes from the same chemical in our compendium. The exception was triprolidine that had a significant correlation to that in MCF-7 cells (p-value=6.2E-5) and PC3 cells (p-value=3.0E-13) but not HL60 cells from the CMAP study (data not shown).

We also attempted to confirm the activation of a number of ERα activators identified in the De Abrew et al. study ²⁴. MCF-7 cells were exposed to DEHP, flutamide, phenobarbital, and thyroxine for 6 hrs and then examined by Affy arrays. Using the derived statistically filtered gene lists, there was no significant correlation to the biomarker (Supplemental Figure 5A). Clofibrate, DEHP, and phenobarbital were also examined by qPCR and none of them activated any of the four ERα biomarker genes (Supplemental Figure 5B). Similar to our approach above, we ruled out mislabeling of the original Affymetrix files or analysis procedures in BSCE as the source of the discrepancies by regenerating the microarray profiles from the publicly available raw data. Each list when compared to the ERα biomarker resulted in similar -Log(p-value)s independent of the method used to generate the filtered gene list (Supplemental Figure 5C). Only the flutamide gene list exhibited any significant correlation to two flutamide gene lists from the De Abrew et al. data set (100 uM flutamide from MCF-7 cells (p-value=7.1E-7) or HepaRG cells (p-value=6.0E-5).

In summary, our validation studies could not confirm the activity of a number of ERα actives identified using the biomarker approach. We are not the first group to note that a subset of the 5 CMAP chemicals exhibit similarities to ERα agonists, and ERα activity could not be confirmed. Iskar et al. ⁶⁰ used the CMAP 2.0 dataset to identify drug-induced transcriptional modules based on gene co-expression; one module called “MCF7-9” was enriched for ERα agonists and antagonists. Out of the total of 49 genes in this module, 23 overlapped with the 46 genes in the biomarker (data not shown). All 37 of the drugs identified in this module were also identified using our methods as ERα activators (31 drugs) or ERα suppressors (6 drugs). (It should be noted that using our biomarker we also identified additional chemicals with estrogenic or antiestrogenic activity some of which are well known to activate (estradiol, resveratrol, kaempferol) or suppress (clomiphene) ERα that were not identified in the Iskar et al. study.) Out of the five chemicals that we examined in detail, triprolidine, nitrendipine, and bendroflumethiazide were included in the MCF7-9 group. Nitrendipine and bendroflumethiazide were examined for binding to ERα and both were found to be negative ⁶⁰ consistent with our results.

The CMAP and De Abrew et al. studies have provided the scientific community with a valuable resource to perform chemical classification using various bioinformatic techniques, given the large number of reference compounds with known targets that were examined. Highlighting their usefulness, our lab used CMAP study biosets to generate the ERα biomarker ¹³. The inability of our lab and Iskar et al. to confirm novel hits highlights the fact that high throughput transcriptomic assay results should be viewed with caution until results are confirmed. The basis for our lack of confirmation could be due to a number of factors including mistakes in annotations in GEO or BSCE, differences in culturing conditions that are not readily obvious, or possibly cross contamination between treatment wells. For the subset of the DeAbrew et al. study that we reanalyzed, there were essentially no differences in biomarker correlations between filtered gene lists derived using the BSCE protocols or independently by Partek while for the CMAP study there appeared to be at least some discrepancies between the BSCE-derived lists and lists derived using the independent Z-score method. It is possible that the likelihood of deriving inaccurate sets of genes increases when the gene sets are based on only one biological replicate vs. multiple controls, as is the case for the CMAP dataset. Further work is needed to determine the basis for the differences between our findings and these other two studies.

Activation of ERα by progestins.

Crosstalk between the estrogen receptor and the progesterone receptor pathways are well documented and include induction of the PGR gene by estrogens ⁶¹. Our screen of the CMAP drugs identified progestins that activated ERα (ethisterone, ethynodiol diacetate, levonorgestrel, lynestrenol, megestrol, norethindrone, norethynodrel, pregnenolone) while cyproterone and progesterone itself were not active. The PR antagonist mifepristone (RU486) suppressed ERα using the biomarker approach. We examined a subset of these progestins in the MCF-7aro-ERE.2 cells and confirmed that ethisterone, lynestrenol, mestranol, norethindrone, pregnenolone, and progesterone activated ERα, and mifepristone suppressed ERα in the trans-activation assays (Figure 8A). MCF-7 cells were exposed to 9 of the compounds for 18 hrs, and the expression changes of four ERα biomarker genes were examined by qPCR assays (Figure 8B). Six of the compounds (ethisterone, lynestrenol, mestranol, norethindrone, pregnenolone, and progesterone) consistently activated one or more of the 4 genes while cyproterone, megestrol acetate, and mifepristone had little if any activity. The increases in expression could be abolished by cotreatment with ICI. Similar to the results in wild-type MCF-7 cells, the expression of one or more of the ERα biomarker genes was increased in the C4-12 ERα cells but not in the C4-12 cells by all of the tested compounds except megestrol acetate and mifepristone (Figure 8C). Lastly, of the 6 progestins examined in the Tox21 study, four of the compounds activated ERα (ethisterone, lynestrenol, mestranol, pregnenolone) and two were negative (megestrol acetate, mifepristone) corresponding exactly to the results of the qPCR analysis (data not shown).

Figure 8.

Modulation of ERα by progestins. A. Examination of ERα activity in transactivation assays in AroER triscreen cells. Cells were treated for 24 h. (Left) Activation by a subset of the progestins tested. (Right) Suppression by mifepristone. The treatments that showed no values were those in which the assay results were the same or lower than the DMSO control values. B. Expression of ERα biomarker genes after exposure to the progestins in MCF-7 cells with and without treatment with ICI. The expression of the four biomarker genes was measured by RT-qPCR. C. Expression of ERα biomarker genes in the C4-12 and C4-12 ERα cell lines.

Because some of these compounds have the potential to be metabolized by aromatase to estrogens, we determined whether the activation of ERα could be inhibited by the aromatase inhibitor letrozole. Chemical activation of ERα by the chemicals was not inhibited by 200 nM of letrozole (Supplemental Figure 6). In retrospect, this result was not surprising given that MCF-7 cells express the aromatase enzyme at low levels, and the role of aromatase can only be assessed in MCF-7 cells after exogenous expression ³¹.

In the present study, we confirmed the ERα activation by 6 progestins predicted to be ERα activators by the biomarker using trans-activation assays and by qPCR. The activation of ERα biomarker genes could be abolished in the presence of ICI or in cells lacking activity of ERα. There were indications that progestins have the ability to activate ERα from a study published over 25 years ago in which the effects of 19-nortestosterone derivatives on ERα activity in MCF-7 cells were examined ⁶². These authors found that two of the same progestins which activated the ERα biomarker in our study (levonorgestrol (also known as norgestrel) and norethindrone) were strong ER activators whereas two other progestins (medroxyprogesterone acetate, R2050) were either inactive or weakly estrogenic in their ERα trans-activation assay. Although R2050 was not examined in the MCF-7 compendium, ERα was activated by 10 nM R5020 in T47D-MTVL breast cancer cells treated for 6 hrs (Running Fisher p-value = 2.8E-5) but not in MCF-7 cells treated for 24 hrs (from GSE25077). Jordan et al. ^62,63 concluded that exposure to the active progestins led to ERα activation and did not involve the progesterone receptor as the activation could be blocked by ERα antagonists (ICI 164,384, 4-hydroxytamoxifen) but not by the PR antagonist, RU486 (mifepristone). Although these authors state that they could not block the ERα activation with aromatase inhibitors, “(data not shown)”, they could not rule out that these progestins were converted to estrogens by steroidogenic enzymes. It should be noted that another study showed that RU486 is a weak estrogen receptor activator ⁶³. It appears unlikely that the progestins are metabolized by aromatase to ERα activators as MCF-7 cells express aromatase to only low levels and cotreatment by the aromatase inhibitor letrozole had no effect on the ability of the chemicals to activate ERα. While the mechanism(s) by which these compounds activate ERα remains to be determined, these data provide additional evidence for cross talk between the estrogen and progesterone receptor pathways.

SUMMARY

Our previous study ¹³ described the construction and testing of a gene expression biomarker that could be used for screening microarray profiles and to identify chemicals that modulate ERα including those that are environmentally-relevant. In the present study, we used the biomarker to screen for ERα active compounds in a MCF-7 compendium of ~1600 microarray comparisons of ~1200 chemicals. Our screen of the compendium which included the CMAP 2.0 collection in MCF-7 cells identified chemicals not previously recognized as ERα modulators. Of the 1153 chemicals in the collection ~10% (114) were found to act as ERα modulators with 75 and 39 chemicals predicted to activate or suppress ERα, respectively. Many of the identified chemicals were also detected as active in ERα trans-activation assays carried out in VM7 cells used to screen the Tox21 10K chemical library in agonist or antagonist modes. Novel chemicals predicted to modulate ERα in MCF-7 cells using the biomarker were examined further using global and targeted gene expression, trans-activation assays, and cell-free assays which measure interactions between ERα and co-regulator proteins.

Known ERα agonists and antagonists were readily identified using the biomarker approach. The agonists included 17beta-estradiol, 17alpha-estradiol, dienestrol, diethylstilbestrol, estriol, estrone, estropipate, mestranol, and hexestrol. The antagonists/SERMs included ICI, clomiphene, raloxifene, and tamoxifen. We confirmed the activity of putative ERα activators, ivermectin and oxybenzone as well as a number of progestins. The activation of ERα by the progestins is not widely appreciated. Additionally, our approach could readily identify relatively weak environmentally-relevant ERα agonists (e.g., daidzein, kaempferol, apigenin, and chrysin) as observed earlier ¹³. While many of the chemicals that we identified in this study have been previously shown to modulate ER, most of the drugs are potentially novel ERα modulators as no evidence exists that they act as ERα agonists or antagonists. Overall, the evidence supports the assertion that our biomarker approach will be useful in future HTTr screens of chemicals such as those presently carried out as part of the ToxCast screening program. We envision that in the near future we can simultaneously screen for not only ERα actives but those that modulate other transcription factor targets using a panel of characterized biomarkers (for example, the TGx-DDI biomarker that detects p53 activators ⁶⁴). This approach will enable comprehensive identification of interactions between environmental chemicals and molecular targets important in various toxicities ¹⁰.

There were a number of ERα activators identified in two large microarray studies that were not previously associated with ERα activity and which stimulated initial interest in determining how these chemicals activate ERα. However, 5 very strong putative activator chemicals from the CMAP study and 4 chemicals from the De Abrew et al. study could not be confirmed to be ERα activators using one or more techniques. The basis for why these chemicals were positive using the biomarker approach but not using other techniques remains to be determined but could include mistakes in sample handling or file annotation. There is also the possibility that our methods of culturing and treatment of the MCF-7 cells is different in some aspect resulting in these divergent responses. These results highlight that predictions derived from HTS data including HTTr data should be validated using other methods.

There are a number of relatively minor weaknesses using the biomarker approach for chemical screening. The biomarker cannot distinguish between those chemicals which regulate by classical agonism or antagonism mechanisms and those chemicals that work through alternative mechanisms. For example, overexpression of wild-type or constitutively active mutants of ERα lead to increased activation of ERα based on the ERα biomarker while knockdown of ERα expression suppresses the biomarker ¹³. Thus there could be chemicals that modulate ERα through regulation of ERα expression levels similar to chemicals that modulate the androgen receptor through effects on AR gene expression ⁶⁵. Thus, if active chemicals are found with our approach, further work would be required to identify the precise mechanism of action. Another weakness of screening in MCF-7 cells is that aromatase inhibitors are not identified. None of the chemicals identified as aromatase inhibitors in a screen in the MCF-7aro cell line ⁶⁶ were identified in the present study (data not shown), presumably because MCF-7 cells do not express aromatase to levels appropriate for screening. For screening to be effectively carried out, aromatase is overexpressed and cells are co-treated with testosterone ³¹. Lastly, the MCF-7 cells lack most of the metabolic machinery to metabolize parent chemicals to active metabolites. Despite these drawbacks, the biomarker approach coupled with an ERα positive breast cancer cell line is an effective approach for identification of putative ERα actives and will be useful in future screens of HTTr data.

ABBREVIATIONS

27HC

27-hydroxycholesterol

4-HT

4-hydroxytamoxifen

AOP

adverse outcome pathways

ARC

4-amino-6-hydrazino-7-beta-D-ribofuranosyl-7H-pyrrolo(2,3-d)-pyrimidine-5-carboxamide

BSCE

basespace correlation engine

CMAP

connectivity map

DBD

DNA binding domain

DEHP

di-2-ethylhexyl phthalate

DES

diethylstilbestrol

DHEA

dehydroepiandrosterone

E2

17beta-estradiol

EC

effective concentration

EDC

endocrine disrupting chemicals

EDSP

Endocrine Disruptor Screening Program

EDSP21

EDSP in the twenty-first century

EE

ethinylestradiol

EPA

U.S. Environmental Protection Agency

ER

estrogen receptor

ERE

estrogen responsive element

EST

estrogen sulfotransferase

GEO

Gene Expression Omnibus

HTS

high throughput screening

HTTr

high throughput transcriptomic

ICI

ICI 182,780

LBD

ligand binding domain

LXR

liver X receptor

NCATS

National Center for Advancing Translational Sciences

NHEERL

National Health and Environmental Effects Research Laboratory

OECD

Organization for Economic Cooperation and Development

SERM

selective estrogen receptor modulator

T4

thyroxine

TAM

tamoxifen