Expressomal approach for comprehensive analysis and visualization of ligand sensitivities of xenoestrogen responsive genes

Toshi Shiodaa,b,1, Noël F. Rosenthala, Kathryn R. Cosera, Mizuki Sutoa, Mukta Phatakc, Mario Medvedovicc, Vincent J. Careyb,d, and Kurt J. Isselbachera,b,1

aMolecular Profiling Laboratory, Massachusetts General Hospital Center for Cancer Research, Charlestown, MA 02129; bDepartment of Medicine, Harvard Medical School, Boston, MA 02115; cLaboratory for Statistical Genomics and , Department of Environmental Health, University of Cincinnati College of Medicine, Cincinnati, OH 45267; and dChanning Laboratory, Brigham and Women’s Hospital, Boston, MA 02115

Contributed by Kurt J. Isselbacher, August 26, 2013 (sent for review June 17, 2013)

Although biological effects of endocrine disrupting chemicals Evidence is accumulating that the EDCs may cause significant (EDCs) are often observed at unexpectedly low doses with occa- biological effects in humans or animals at doses far lower than sional nonmonotonic dose–response characteristics, - the exposure limits set by regulatory agencies (8, 9). In addition wide profiles of sensitivities or dose-dependent behaviors of the to such low-dose effects, an increasing number of studies also EDC responsive genes have remained unexplored. Here, we describe support the concept of the nonmonotonic EDC effects, whose dose–response curves show U shapes or inverted-U shapes (8- expressome analysis for the comprehensive examination of dose- – dependent gene responses and its applications to characterize es- 10). However, whereas nonmonotonic dose responses are often observed in the endocrine system, this notion conflicts with one trogen responsive genes in MCF-7 cells. of MCF-7 – cells exposed to varying concentrations of representative natural of the Bradford Hill criteria for cause-and-effect relationships that requires stronger effects with greater degrees of exposure and xenobiotic estrogens for 48 h were determined by microarray (11), leading to controversies when biological effects are ob- and used for computational calculation of interpolated approxima- served most strongly with lower rather than higher doses of SCIENCES tions of estimated transcriptomes for 300 doses uniformly distrib- EDCs (12-14). On the other hand, when an EDC does not show uted in log space for each chemical. The entire collection of these significant biological effects at a high dose, that chemical is ENVIRONMENTAL estimated transcriptomes, designated as the expressome, has pro- generally assumed safe at lower doses, and possible nonmonotonic vided unique opportunities to profile chemical-specific distributions dose–response characteristics are seldom appreciated in risk as- of ligand sensitivities for large numbers of estrogen responsive sessment. Thus, as the standard toxicological approaches com- genes, revealing that at low concentrations estrogens generally monly adopted by the industry and regulatory apparatus do not tended to suppress rather than to activate transcription. Gene on- presently assume the low-dose–specific toxicity or nonmonotonic tology analysis demonstrated distinct functional enrichment between dose–response relationship, it is urgently desired to establish high- and low-sensitivity estrogen responsive genes, supporting the solid scientific frameworks for proper handling of data possibly notion that a single EDC chemical can cause qualitatively distinct demonstrating the nonstandard dose-dependent effects. Because biological responses at different doses. Expressomal heatmap vi- most published studies on the EDC effects involve limited num- sualization of dose-dependent induction of Bisphenol A inducible bers of doses for each chemical species, research on mechanisms genes showed a weak gene activation peak at a very low concen- tration range (ca. 0.1 nM) in addition to the main, strong gene Significance activation peak at and above 100 nM. Thus, expressome analysis is a powerful approach to understanding the EDC dose-dependent Cells change their mRNA expression in response to biologically dynamic changes in at the transcriptomal level, active substances in a dose-dependent manner. Because dif- fi providing important information on the overall pro les of ligand ferent genes in a show distinct sensitivities to the same sensitivities and nonmonotonic responses. substance, changes in the genome-wide mRNA expression profile induced by low and high doses of a substance are es- he endocrine disrupting chemicals (EDCs) are environmen- sentially different, but this notion has been commonly over- Ttal pollutants that interfere with the endocrine system to looked in previously published studies. Using a human cell disturb biological processes such as development, reproduction, culture model and microarray, we performed genome-wide and metabolism (1-4). EDCs consist of a wide variety of man- determinations of gene sensitivities to hormonally active sub- made and natural compounds with highly diversified chemical stances with statistically rigorous approaches. Our study pro- structures (4). Some EDCs are agonistic ligands that directly vides a conceptual and methodological framework for the bind to hormone receptors, whereas others inhibit receptor systematic examination of gene sensitivities and demonstrates effective detection of nonmonotonic dose-dependent responses, functions as antagonists. For example, xenoestrogens such as introducing the importance of gene sensitivity analysis to phar- Bisphenol A (BPA) (1) and genistein (5) are EDCs that activate macogenomic and toxicogenomic research. estrogen receptors by direct binding. Tributyltin is an obesogenic EDC that binds to peroxisome proliferation-activated receptor γ Author contributions: T.S. and K.J.I. designed research; T.S., N.F.R., and K.R.C. performed (PPAR-γ) to activate the retinoid X receptor (RXR)–PPAR research; T.S., M.S., and V.J.C. contributed new reagents/analytic tools; T.S., M.P., M.M., and heterodimer nuclear receptor complex, enhancing adipo- V.J.C. analyzed data; and T.S. and K.J.I. wrote the paper. genesis (6). Some other EDCs more indirectly disrupt the en- The authors declare no conflict of interest. docrine system by affecting hormone synthesis, metabolism, Freely available online through the PNAS open access option. sensitivity, or the negative feedback system (4). For example, Data deposition: The data reported in this paper have been deposited in the Gene Ex- pression Omnibus (GEO) database, www.ncbi.nlm.nih.gov/geo (accession no. GSE50705). thyroid hormone EDCs such as polychlorinated biphenyl con- 1To whom correspondence may be addressed. E-mail: [email protected] or geners reduce the circulating thyroid hormone levels, and some [email protected]. of them may also bind to the thyroid hormone receptors as This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. functional ligands (7). 1073/pnas.1315929110/-/DCSupplemental.

www.pnas.org/cgi/doi/10.1073/pnas.1315929110 PNAS Early Edition | 1of6 Downloaded by guest on September 30, 2021 of the low-dose and nonmonotonic actions is still in its infancy 120 (15, 16). Obviously, the lack of widely applicable standard approaches to study the low-dose and nonmonotonic EDC 100 effects on gene expression is a critical obstacle to understanding the mechanisms of the genetic and genomic actions of EDCs. 80 Our present study attempts to introduce frameworks for comprehensive analyses and data visualization of the EDC dose- 60 Strong estrogens dependent transcriptomal dynamism and its possible non- monotonic characteristics. We describe the expressome as a li- 40 17-estradiol brary of interpolated approximations of transcriptomal profiles diethylstilbestrol ethynylestradiol for hundreds of doses uniformly distributed in the log space 20 within the range of doses for which the seed transcriptomes are experimentally determined (ACR, analyzed concentration range). Relative cell number (%) 0 Our expressome analysis of representative xenoestrogens has de- 120 termined transcriptome-wide profiles of ligand sensitivities of es- Medium strength trogen responsive genes in MCF-7 cells and provided visualization 100 estrogens of the nonmonotonic aspects of the transcriptomal effects of BPA inducible genes. We propose that expressome analysis is a pow- 80 Bisphenol A erful approach for the comprehensive and statistically rigorous p-nonylphenol examination of the dose-dependent dynamic aspects of tran- 60 scriptomal responses to EDCs. 40 Results Generation of Computational Models of Dose-Dependent Transcriptomal 20

Responses to Natural and Xenobiotic Estrogens. As an initial step to Relative cell number (%) 0 characterize estrogen dose-dependent responses of animal cells, 120 we classified estrogens into three groups by their strength to Weak estrogens support proliferation of the estrogen-dependent MCF-7 cells 100 (Fig. 1). The strong estrogens that supported MCF-7 cell pro- daidzein β liferation at subnanomolar concentrations included 17 -estradiol 80 genistein (E2), diethylstilbestrol (DES) (17), and 17α-ethynylestradiol (EE2) (18). Estrogens with intermediate strength supported MCF-7 cell 60 proliferation at 1–100 nM and included two representative in- dustrial chemicals BPA and p-nonylphenol (PNP) (1, 19). The 40 isoflavone phytoestrogens genistein and daidzein, which require metabolic activation for their full-strength estrogenic actions 20 (20), were weak estrogens that supported MCF-7 cell proliferation at 0.1–10 μM. Based on these results, we selected estrogen con- Relative cell number (%) 0 centrations for which the seed transcriptomal profiles were de- 13 12 11 10 9 8 7 6 5 none termined (indicated in Fig. 1 by small triangles). MCF-7 cells were -log[ligand (M)] hormone-starved for 24 h and then exposed to estrogens for ex- actly 48 h before total RNA isolation for microarray determination Fig. 1. Estrogen dose-dependent MCF-7 cell proliferation profiles. Cells of transcriptomes. The length of exposure (48 h) was chosen be- were exposed to varying concentrations of estrogens for 10 d followed by cause our previous studies showed that at this timing robust and determination of cell numbers. The three panels share a common x axis in- global changes in mRNA expression were observed with minimal dicating estrogen concentrations. Each datum point represents mean of at SDs (21, 22). Transcriptional responses to the estrogenic chemicals least three independently performed assays (SEM were within 10% of the were confirmed by real-time quantitative PCR (qPCR) de- mean of the replicates). Small triangles indicate ligand concentrations for termination of the mRNA transcripts for WISP2 and BIK, which seed transcriptomes were determined. which were identified as the most robust and quantitative es- trogen inducible and suppressible marker genes, respectively, in our previous study (22). Fig. S1A shows typical BPA dose-de- analysis provided only limited information about the dose-de- pendent expression profiles of the BIK and WISP2 mRNA pendent dynamism of the transcriptomal responses. transcripts. The difference in the effective BPA concentrations To examine the estrogen dose-dependent transcriptomal changes between the 10-d cell proliferation assay (Fig. 1) and the mRNA in more detail, we constructed a computational model that accepts expression assay (Fig. S1A)reflects different estrogen sensitivities arbitrary estrogen concentrations within the ACR as input and es- of these assays as previously reported (22). timates transcriptomes for those estrogen concentrations as output. Unsupervised hierarchical clustering analysis of transcriptomes Table S1 summarizes the experimentally determined transcriptomes of the same samples shown in Fig. S1A identified 140 BPA in- (i.e., the seed transcriptomes) obtained in the present study for ducible genes and 40 BPA suppressible genes (Fig. S1B). The construction of such a model. The collection of the estimated dendrogram of the RNA samples (shown at the left of the heat- transcriptomes for a large number of different estrogen concen- map) indicates minimal transcriptomal changes from vehicle trations is designated as the estrogen expressome. control to up to 60 nM BPA, followed by robust changes between To construct the computational model for generation of the 60 nM and 3 μM. Transcriptomal changes were relatively small at expressome, we first converted normalized and log2-transformed BPA concentrations from 3 μMuntil50μM. The apparent re- microarray intensity values to proportion of the dynamic range of versal of transcriptomal changes of the highest concentration of the dataset for each chemical (i.e., 0∼100%; 0% = minimal ex- BPA (100 μM) was not due to nonspecific cell damage because pression; 100% = maximal expression). By evaluating data var- expressions of the GAPDH mRNA transcripts were not affected iations between and within experiments, we determined the (Fig. S1 C and D), and microscopic observations did not reveal any reproducibility score for each gene (described in detail in SI significant evidence of cell damages. Thus, the standard approach Text). We ranked genes for their response to the 48-h exposure of transcriptomal analysis on experimentally determined micro- to E2 based on the reproducibility score and displayed expres- array data identified ligand-responsive gene clusters; however, this sion trajectories for the best, 500th, 5,000th, and 20,000th ranked

2of6 | www.pnas.org/cgi/doi/10.1073/pnas.1315929110 Shioda et al. Downloaded by guest on September 30, 2021 genes (Fig. S2). The best and the 500th ranked genes exhibited A log(EC50) = -6.2797 (-6.4065 ~ -6.1512) approximately monotonic increases in mRNA expression with an increasing dose of E2, whereas the 5,000th and the 20,000th raw data CELSR2 genes exhibit considerable noise. Based on this reproducibility assessment, we selected the top 500 E2 responsive genes for 100 further analyses. Dose–response transcriptomal data of these selected genes were subjected to interpolation using the gener- 80 EC50 ± 95%C.I. alized additive model (examples are shown in Fig. 2 and Fig. S3, green line graphs), which is appropriate for fitting nonmonotonic relationships (23), followed by estimation of the EC50 and IC50 60 concentrations with 95% confidence intervals. Note that the green line graphs in Fig. 2 and Fig. S3 display fit data only for the concentrations for which the seed transcriptomes were experi- 40 mentally determined. Finally, by applying the B-spline fitting, computational models generating relative mRNA expression (arbitorary units) 20 none levels were constructed for any ligand concentrations within the ACR for each gene and estrogenic chemical. Using these mod- mRNA expression Relative els, smooth and practically continuous dose–response fit curves 0 (indicated as blue dotted lines in Fig. 2 and Fig. S3) were gen- erated for 300 concentrations evenly distributed in the log10 -12 -11 -10 -9 -8 -7 -6 -5 -4 space within the ACR. The expressome is defined as the tran- scriptome-wide collection of the interpolated, smooth fit curves -log10[BPA (M)] of dose-dependent relative mRNA expression. B log(IC50) = -6.8671 (-7.0955 ~ -6.6641) Expressome Analysis Reveals Transcriptomal Distribution of Sensitivities none PRR5 raw data of Estrogen Responsive Genes. As the expressome estimates the EC50 and IC50 50% effective concentrations of estrogen inducible and fi 100

suppressible genes, respectively, with 95% con dence intervals (Fig. SCIENCES 2), we attempted to obtain transcriptome-wide perspective on es- trogen sensitivities of these genes. From the 500 most reproducible 80 ENVIRONMENTAL dose–response curve sets as selected earlier, we further selected fi genes whose 95% con dence intervals of EC50 or IC50 concen- 60 trations were within one log10 range. This data filtration step resulted in selection of minimum 82 genes (for PNP suppressible genes) and maximum 301 genes (for DES inducible genes); the 40 numbers of selected genes for each estrogen are shown as n in

Table S2, and the list of genes are provided in Dataset S1.The (arbitorary units) estimated EC50 or IC50 concentrations were sorted and visualized 20 fi mRNA expression elative

as the sensitivity distribution curves (SDCs) with their 95% con - R A dence intervals. Fig. 3 shows SDCs for E2, BPA, and genistein; IC50 ± 95%C.I. complete sets of SDCs for all estrogenic agents examined in the 0 present study are shown in Fig. S4. Numeric parameters charac- terizing these SDCs are summarized in Table S2. The shape of the -12 -11 -10 -9 -8 -7 -6 -5 -4 SDCs indicates that the majority of the estrogen responsive genes -log [BPA (M)] belong to the compound-specific “typical” sensitivity group con- 10 stituting the intermediate part of the SDCs with modestly in- Fig. 2. Curve fitting of BPA dose-dependent mRNA expression. MCF-7 cells creasing EC50 or IC50 concentrations. However, the presence of were exposed to the indicated concentrations of BPA for 48 h followed by highly sensitive genes (with small gene numbers) and relatively transcriptomal determination by Affymetrix microarray. Each datum point insensitive genes (with large gene numbers) is readily perceived by (red dot) represents relative amount of mRNA expression determined by an deviations of the edges of the SDCs from their overall trends in the individual microarray chip. Green lines are interpolated fit line graphs drawn middle parts. Thus, the SDC analysis provides a unique opportunity for the 31 BPA concentrations for which the seed transcriptomes were ex- to categorize genes into distinct groups of different sensitivities. perimentally determined. Blue dotted lines are expressome-generated fit The range of the EC50 and IC50 concentrations shown in Fig. curves drawn for 300 concentrations distributed evenly in the ACR on log fi 3G and Table S2 classifies estrogens into three groups based on scale. EC50 and IC50 concentrations with 95% con dence intervals are in- the sensitivities of their responsive genes. From the three strong dicated for (A) BPA inducible gene CELSR2 and (B) BPA suppressible gene estrogens identified in Fig. 1 for their comparable potencies to PRR5, respectively. support MCF-7 cell proliferation, DES is now separated from the other two for its exceptionally strong genomic efficacy. The in- termediate-strength and weak estrogens shown in Fig. 1 form distributions of estrogen inducible and suppressible genes may a single group, although genistein showed slightly weaker efficacy cause “imbalance” between the gene inducing and suppressing than the others. These results demonstrate significant discrep- effects exclusively at certain low concentration ranges, where ancies in relative estrogenic strengths of the same set of xenoes- inducible genes are not yet effectively induced but where sup- trogens when evaluated based on their MCF-7 cell proliferation pressible genes are already suppressed. Such imbalance is not effects and transcriptomal effects, corroborating our prior obser- observed at high concentrations, where both the inducible and vations made with limited numbers of estrogen responsive genes suppressible genes show maximal responses. (22). Fig. 3G also indicates that the estrogen suppressible genes Based on the results of the SDC analysis, we hypothesized that are always more sensitive than estrogen inducible genes for all genes with the highest and lowest ligand sensitivities may have xenoestrogens tested (P values for the differences in sensitivities of distinct biological importance. To obtain initial insights into this the estrogen inducible and suppressible genes are shown in Table possibility, we performed Gene Ontology analysis for (i) all of A ii S2, and histograms of EC50 and IC50 concentrations are shown the 113 E2 inducible genes shown in Fig. 3 ,( )36mostsensitive in Fig. S5 A–G). A significant discrepancy in the sensitivity E2 inducible genes, and (iii) 36 least sensitive E2 inducible genes

Shioda et al. PNAS Early Edition | 3of6 Downloaded by guest on September 30, 2021 1 e-10 3.5 e-11 in color heatmaps. Fig. 4 shows examples of such expressomal A 9 e-11 17β-Estradiol 3.0 e-11 17β-Estradiol heatmaps for E2 and BPA in two-dimensional (Fig. 4 A–D) and 8 e-11 inducible suppressible E–H 7 e-11 2.5 e-11 three-dimensional (Fig. 4 ) spaces. Expressomal heatmaps 6 e-11 of E2 inducible genes (Fig. 4 A and E) demonstrate a monoto- 5 e-11 2.0 e-11 nous increase in gene expression with increasing E2 concen- 4 e-11 1.5 e-11 3 e-11 trations up to 100 pM. The EC50 concentrations, which are 1.0 e-11 Concentration (M) 2 e-11 Concentration (M) displayed in green, show modest differences throughout the E2 50 2 e-11 50 inducible genes, agreeing with the SDC analysis results shown in IC 0.5 e-11 EC 1 e-11 A 0 0 Fig. 3 . Expressomal heatmaps of E2 suppressible genes (Fig. 4 1 20 40 60 80 100 120 1 20 40 60 80 100 120 B and F) also demonstrate monotonous decreases in mRNA Gene Number Gene Number expression with increasing E2 concentrations. However, the IC50 5.0 e-6 7 e-7 concentration line (displayed in green) is strongly bent at the 4.5 e-6 Bisphenol A Bisphenol A 6 e-7 boundary between the most sensitive genes (gene numbers 1–10) 4.0 e-6 inducible suppressible B F fl 3.5 e-6 5 e-7 and other genes (asterisk in Fig. and ), re ecting the presence of a group of high-sensitivity genes. Interestingly, expressomal 3.0 e-6 4 e-7 2.5 e-6 heatmaps of BPA inducible genes show weak but readily per- 2.0 e-6 3 e-7 ceived increases in mRNA expression at about 10 pM BPA 1.5 e-6 2 e-7 C G Concentration (M) Concentration (M) concentration (orange rectangles in Fig. 4 and ), whereas 50 50 1.0 e-6

IC 1 e-7 much stronger mRNA expression was detected in the same EC 0.5 e-6 0 0 heatmaps at 5 nM and greater concentrations. This weak mRNA 1 20 40 60 80 100 120 140 160 1 20 40 60 80 100 induction at the extremely low BPA concentration was observed Gene Number Gene Number for genes with the highest BPA sensitivities (gene numbers = 1 e-10 4.0 e-6 1∼40), whereas genes with lower BPA sensitivities (e.g., gene 9 e-11 Genistein 3.5 e-6 Genistein = ∼ 8 e-11 inducible suppressible number 100 140) showed minimal evidence of such mRNA 7 e-11 3.0 e-6 induction. In contrast, expressomal heatmaps of the BPA sup- 6 e-11 2.5 e-6 pressible genes showed generally monotonous decreases in 5 e-11 2.0 e-6 D 4 e-11 mRNA expression with increasing BPA concentrations (Fig. 4 3 e-11 1.5 e-6 and H). The statistical significance of the weaker peak of mRNA Concentration (M) Concentration (M) 2 e-11 1.0 e-6 induction by low concentrations of BPA was confirmed by 50 50 2 e-11

IC 0.5 e-6 EC 1 e-11 Bonferroni-adjusted generalized random set analysis (24) (Fig. 0 0 S5H, asterisk), whereas mRNA suppression by the same low 1 20 40 60 80 100 120 1 10 20 30 40 50 60 70 80 90 fi Gene Number Gene Number BPA concentrations did not reach statistical signi cance (Fig. S5H, sharp). These results demonstrate the usefulness of the B expressomal heatmap for detection and characterization of inducible genes Genistein nonmonotonic transcriptomal responses. suppressible genes Daidzein Discussion p-nonylphenol An increasing number of studies suggest that the traditional practice of toxicological risk assessments overlook important Bisphenol A characteristics of EDCs such as low-dose effects and nonmono- 17β-estradiol tonic dose–response (3, 4, 14). Because animal studies often demonstrate significant biological effects of EDCs at concen- Ethynylestradiol trations far lower than the safety limits set by regulatory agen- Diethylstilbestrol cies, appropriateness of the use of “no observable effects limit” for EDC risk assessment has been questioned (8, 25). The -15 -14 -13 -12 -11 -10 -9 -8 -7 -6 -5 -4 nonmonotonic dose–response may result in paradoxical obser- -log[estrogen] vations that a lower EDC dose can be more toxic than a higher dose of the same chemical (8, 10). Accumulation of transcriptome- Fig. 3. Distributions of EC50 and IC50 concentrations of estrogen responsive genes. (A) SDCs of estrogen responsive genes. Genes were ranked by their wide data on EDC sensitivities of individual EDC responsive genes will be critical for the mechanistic understanding of these ex- expressome-estimated EC50 or IC50 concentrations (red curves) displayed with 95% confidence intervals (blue curves). Note that y axis is in linear scale. traordinary phenomena. However, there have been no conceptual

(B) Summary of the lowest, highest, and mean EC50 or IC50 concentrations of or methodological frameworks for guiding systematic investigations estrogen inducible genes (red) and estrogen suppressible genes (blue), re- on this subject. In an effort to develop such a framework, in the

spectively. The x axis is estrogen concentrations in the log10 scale. present study we propose the expressome analysis along with its two practical applications—namely, the SDC analysis of sensitiv- ities of EDC responsive genes (Fig. 3, Fig. S4,andTable S2)and (Dataset S1). Groups i and ii showed strong enrichment of genes the expressomal heatmap visualization of nonmonotonic dose– involved in DNA replication and cell cycle regulation, whereas response (Fig. 4). These unique tools will provide not only the no statistically significant enrichment was observed with group iii. research communities but also the governmental regulatory ap- These results suggest that the cell proliferation effects of E2 may paratus with opportunities to perform comprehensive, objective, be primarily carried on by a relatively small number of the most and precise dose-dependent effects of toxic substances on mam- sensitive genes, whereas relatively low sensitivity genes might be malian genomes. As the costs and the accuracy of transcriptomal relevant for the other biological estrogen effects not readily analyses are rapidly improving due to the technological revolu- detectable by the Gene Ontology analysis. Thus, the SDC anal- tions such as deep sequencing, it is practical for toxicological risk ysis of the expressome provides useful foundations to examine assessments to take the anomalous dose–response relationships the dose-dependent transcriptomal effects of EDCs based on and changes in sensitivities to natural or other exogenous sub- sensitivities of their responsive gene sets. stances into consideration. As our present study demonstrates, the nonmonotonicity of the dose–effect relationship possibly charac- Expressomal Visualization of Nonmonotonic Gene Expression. To terizing certain toxic substances may not be clearly evident when obtain insights into the nonmonotonic aspects of the estrogen dose- each series of dose–response data are independently examined, dependent transcriptomal responses, we visualized the expressome due to the relatively stochastic nature of this phenomenon. To

4of6 | www.pnas.org/cgi/doi/10.1073/pnas.1315929110 Shioda et al. Downloaded by guest on September 30, 2021 A B C D 110 110 140 100 100 100 120 90 90 80 80 80 100 70 70 80 60 60 60 50 50 60 40 40 40 40 Gene Number Gene Number Gene Number 30 Gene Number 30 20 20 20 20 10 10 * 1 1 1 1 -9 e -10 -10 e-9 -13 -12 e-14 2.2 5.2 e-7 3.4 e-8 8.0 e-6 5.2 e-7 3.4 e-8 8.0 e-6 2.2 9.1 e-12 5.9 e-13 1.4 e-10 1.6 e-12 1.0 e 2.5 e-14 1.0 e-13 4.0 e-13 6.3 e-12 2.5 e-11 5.9 e-13 1.4 e 9.1 e-12 1.0 e-13 1.0 e-10 2.5 4.0 e 1.6 e 6.3 e-12 2.5 e-11 log[17-estradiol (M)] log[17-estradiol (M)] log[Bisphenol A (M)] log[Bisphenol A (M)] E F G H

1.0 e-10 8.0 e 110 140 2.5 e-11 5.2 100 90 130 -6 6.3 e-12 120 G 80 G 110 3.4 e-7 ene Num70 1.6 e-12 ene Num10090 60 80 2.2 e 70 -8 50 4.0 e-13 110 1.4 e-10 ] 60 e-9 100 40 * 100 50 30 1.0 e-13 -14 40 9.1 e 90 30 90 20 2.5 e-14 e-13 ber 80 ber 20 5.9 80 SCIENCES 10 diol (M)] 1 a 2.5 e e-13 70 10 1 -12 ol A (M) -12 70 60 e 5.9 e 60 log[171.0 50 -13 50

40 hen 9.1 ENVIRONMENTAL -estr -12 log[Bisphenol-9 A (M)] 40 ber 4.0 e-13 30 e 30  20 1.4 e-10 -8 20 1.6 e e-12 10  2.2 e 10 -estradiol (M)]-11 6.3 1 3.4 log[17 ene Number log[Bisp 1 ene Num 2.5 e G 5.2 e-7 G 1.0 e-10 8.0 e-6 0 10 20 30 40 50 60 70 80 90 100 Relative mRNA Expression (%)

Fig. 4. Expressomal visualization of estrogen dose-dependent transcriptomal changes. Relative amounts of mRNA expression of each estrogen responsive

gene are indicated by 2D or 3D heatmaps. Estrogen inducible genes (A, C, E, and G) and estrogen suppressible genes (B, D, F, and H) were ranked by their EC50 or IC50 concentrations (indicated in green) for 2D (A–D)or3D(E–H) displays. (A, B, E, and F) Effects of E2. (C, D, G, and H) Effects of BPA. In A–D,thex axes indicate the molar concentrations of estrogens in log10 scale, and the y axes indicate the rank of genes for their estrogen sensitivity (number 1 is the most sensitive gene). The horizontal lines represent dose-dependent transcriptional changes of individual genes. Asterisks in B and F denote the points where the

IC50 lines strongly bend. Note the weak transcriptional activation of the high-sensitivity BPA inducible genes by low concentrations of BPA at about 1.0 × 10−11∼21.0 × 10−10 M(C and G; indicated by orange rectangles). In E–H,thez axes represent the relative expression level identical to the color keys.

detect the nonmonotonicity, data obtained from multiple repeated heterogeneities among genes with differential EDC sensitivities experiments should be processed without subjective removal of (Dataset S1). Because genes responsive to a single EDC have apparently “outlier” data and summarized using an appropriate significantly varying sensitivities, distinct subsets of EDC re- statistical approach. The expressomal analysis provides unique sponsive genes are induced or suppressed by different concen- opportunities to perform these tasks with proper objectivity trations of an EDC. Consequently, genomic responses to a low- and sensitivity. dose EDC exposure are presumably distinct from responses to To perform expressomal analyses, a sufficient number of dose- a higher dose, and this may cause low-dose–specific biological dependent transcriptomal changes (seed transcriptomes) have to effects of EDCs. Future studies should examine presumable bi- be experimentally determined. Credibility and resolution power ological consequences of the differentially expressed EDC re- of expressomal analyses are primarily determined by the number sponsive genes at different EDC doses. and distribution of seed transcriptomes, which should be exam- Molecular mechanisms determining sensitivities of estrogen ined in more detail by future studies. In the present study, the responsive genes can be very diverse, as depicted in Fig. S6. MCF-7 exposure time to estrogens was fixed at 48 h to focus on Transcriptional activation or suppression by estrogen receptors the dose-dependent aspects of gene expression. Responses of the (ERs) involves differential recruitments of specific coregulators so-called “early responder genes” (e.g., MYC) were no longer on different target genes (26). Specific DNA base sequences at detected after 48 h of exposure to estrogens (21, 22). Future and around the ER binding sites serve as allosteric modulators of studies should attempt to examine time-dependent changes in the ER conformation and thereby affect selection of coregulator expressomes, which may be visualized in the 3D space whose x, y, recruitment and transcriptional activities (27). Multiple other and z axes are dose, mRNA expression, and time, respectively. transcription factors and chromatin-modifying mechanisms also The expressomal SDC analysis introduced in this study pro- affect the genomic actions of ERs, which may bind to chromatin vides unique opportunities to obtain the following information: hundreds of kilobases away from the transcription initiation sites (i) transcriptome-wide profiles of sensitivities of EDC responsive (28). As a type I nuclear receptor, ERs do not interact with DNA genes (Fig. 3, Fig. S4, and Table S2), (ii) statistical significance of until ligand binding. Therefore, it is possible that different ligands, differential sensitivities between two or more groups of genes including endogenous ligands such as 27-hydroxycholesterol (29), (such as EDC inducible genes and suppressible genes; Table S2), may alter the affinity of ERs to various DNA targets and (iii) identification of genes with distinct EDC sensitivities (Fig. 3, coregulators, resulting in a highly complex, and possibly Fig. S4, and Dataset S1), and (iv) demonstration of functional nonmonotonic, dose–response relationship. The expressomal

Shioda et al. PNAS Early Edition | 5of6 Downloaded by guest on September 30, 2021 SDC analysis may provide important clues to recognize involved in the high-sensitivity BPA activation of transcription and characteristic sets of coregulators and other molecular events their possible roles in the nonmonotonic responses. differentially involved in the high-sensitivity and low-sensi- In summary, the present study introduced the expressome, tivity transcriptional responses to estrogens. On the other a collection of transcriptome-wide approximations of continuous hand, Kortenkamp et al. reported that the cell culture effects dose-dependent mRNA expression curves generated based on of estrogenic chemical mixtures are well predictable from the limited numbers of experimentally determined transcriptomes. strength of estrogenicity of each component by the concen- Analyses of the expressome provide unique information about tration addition model (30), implying a single or relatively genome-wide profiles of gene sensitivities to transcriptionally limited number of mechanisms involved in the estrogenic active substances and nonmonotonic aspects of dose-dependent actions of xenoestrogens. However, they also observed the gene expression. Expressomal analyses may contribute to the “effect modulation” phenomenon, in which inclusion of low development of computational handling of genomics data rele- estrogenicity chemicals results in overall reduction in the vant to gene sensitivities of EDCs for pharmaco- and toxico- estrogenicity strength of the mixture (30). The expressomal genomics research and risk assessment. SDC analysis may be able to determine detailed tran- Methods scriptomal effects of EDC mixtures, providing clues to un- Cell Culture and Determination of Transcriptomes by Affymetrix Microarray. derstand mechanisms of the concentration additive features MCF-7 cell culture, 10-d cell proliferation assay for estrogenicity, determination and the effect modulation phenomenon. fi – of transcriptomal pro les using Affymetrix Human Genome U-133 plus 2.0 Mechanisms involved in nonmonotonic dose response of microarray, and hierarchical clustering and heatmap visualization were de- EDC effects are largely unknown. In the whole animal, the scribed in our preceding studies (21, 22). Outlier transcriptomal data were negative feedback mechanisms of the endocrine system and identified by box plot and Q–Q plot analyses and eliminated from the study. other factors affecting metabolism or distributions of hormones All experimentally determined transcriptomal data are available from the may contribute to the nonmonotonicity. Our present study pro- National Center for Biotechnology Information Gene Expression Omnibus vides evidence that nonmonotonic EDC dose–response can be database (accession no. GSE50705) and our website (http://mplwebserver. observed even in cell culture models (Fig. 4 C and G and Fig. partners.org). TaqMan real-time qPCR determination of BIK and WISP2 mRNA S5H). Because MCF-7 cells express only minimal amounts of expression was also performed as previously described (22). All estrogenic ERβ (31), the observed nonmonotonicity was unlikely due to agents were highest grade chemicals purchased from Sigma-Aldrich. differential transcriptional regulations by the two isoforms of Computational Generation and Analyses of the Expressome Models. Compu- estrogen receptors. Because the low-dose gene expression peaks – (orange rectangles in Fig. 4 C and G) of BPA inducible genes tational tasks were performed using R/CRAN package mgcv 1.6 1 (23) and homemade scripts written in Python 2.7 or Ruby 1.9 programming lan- were observed only for genes showing the highest BPA sensi- guages. Details are described in SI Text. tivities in the major mRNA expression peaks, mechanisms of the nonmonotonicity may be related with the high BPA sensi- ACKNOWLEDGMENTS. This study was supported by Susan G. Komen for tivity. Future studies are needed to determine transcriptional Cure Grants FAS0703860 and KG090515 (to T.S.) and NIH Grants U41- coregulators, epigenetic marks, and other transcription factors HG004059 and R01-HL086601 (to V.J.C.).

1. Vom Saal FS, Nagel SC, Coe BL, Angle BM, Taylor JA (2012) The estrogenic endocrine 17. Goodman A, Schorge J, Greene MF (2011) The long-term effects of in utero exposures— disrupting chemical bisphenol A (BPA) and obesity. Mol Cell Endocrinol 354(1-2): The DES story. N Engl J Med 364(22):2083–2084. 74–84. 18. Kidd KA, et al. (2007) Collapse of a fish population after exposure to a synthetic es- 2. Schug TT, Janesick A, Blumberg B, Heindel JJ (2011) Endocrine disrupting chemicals trogen. Proc Natl Acad Sci USA 104(21):8897–8901. and disease susceptibility. J Steroid Biochem Mol Biol 127(3-5):204–215. 19. Soto AM, Justicia H, Wray JW, Sonnenschein C (1991) p-Nonyl-phenol: An estrogenic 3. Zoeller RT, et al. (2012) Endocrine-disrupting chemicals and public health protection: xenobiotic released from “modified” polystyrene. Environ Health Perspect 92:167–173. A statement of principles from The Endocrine Society. Endocrinology 153(9):4097–4110. 20. Shor D, Sathyapalan T, Atkin SL, Thatcher NJ (2012) Does equol production determine 4. Diamanti-Kandarakis E, et al. (2009) Endocrine-disrupting chemicals: An Endocrine soy endocrine effects? Eur J Nutr 51(4):389–398. Society scientific statement. Endocr Rev 30(4):293–342. 21. Coser KR, et al. (2003) Global analysis of ligand sensitivity of estrogen inducible and 5. Cederroth CR, Zimmermann C, Nef S (2012) Soy, phytoestrogens and their impact on suppressible genes in MCF7/BUS breast cancer cells by DNA microarray. Proc Natl Acad reproductive health. Mol Cell Endocrinol 355(2):192–200. Sci USA 100(24):13994–13999. 6. Chamorro-García R, et al. (2013) Transgenerational inheritance of increased fat depot 22. Shioda T, et al. (2006) Importance of dosage standardization for interpreting tran- size, stem cell reprogramming, and hepatic steatosis elicited by prenatal exposure to scriptomal signature profiles: Evidence from studies of xenoestrogens. Proc Natl Acad the obesogen tributyltin in mice. Environ Health Perspect 121(3):359–366. Sci USA 103(32):12033–12038. 7. Zoeller TR (2010) Environmental chemicals targeting thyroid. Hormones (Athens) 9(1): 23. Wood SN (2006) Generalized Additive Models: An Introduction with R (Chapman and 28–40. Hall/CRC, Boca Raton, FL). 8. Vandenberg LN, et al. (2012) Hormones and endocrine-disrupting chemicals: Low- 24. Freudenberg JM, Sivaganesan S, Phatak M, Shinde K, Medvedovic M (2011) Gener- dose effects and nonmonotonic dose responses. Endocr Rev 33(3):378–455. alized random set framework for functional enrichment analysis using primary ge- 9. Taylor JA, et al. (2012) Dose-related estrogen effects on gene expression in fetal nomics datasets. 27(1):70–77. mouse prostate mesenchymal cells. PLoS ONE 7(10):e48311. 25. Birnbaum LS (2012) Environmental chemicals: Evaluating low-dose effects. Environ 10. Do RP, Stahlhut RW, Ponzi D, Vom Saal FS, Taylor JA (2012) Non-monotonic dose Health Perspect 120(4):A143–A144. effects of in utero exposure to di(2-ethylhexyl) phthalate (DEHP) on testicular and 26. Won Jeong K, Chodankar R, Purcell DJ, Bittencourt D, Stallcup MR (2012) Gene-spe- serum testosterone and anogenital distance in male mouse fetuses. Reprod Toxicol cific patterns of coregulator requirements by estrogen receptor-α in breast cancer 34(4):614–621. cells. Mol Endocrinol 26(6):955–966. 11. Hill AB (1965) The environment and disease: Association or causation? Proc R Soc Med 27. Klinge CM, Jernigan SC, Mattingly KA, Risinger KE, Zhang J (2004) Estrogen response 58:295–300. element-dependent regulation of transcriptional activation of estrogen receptors 12. vom Saal FS, et al. (2010) Flawed experimental design reveals the need for guidelines alpha and beta by coactivators and corepressors. J Mol Endocrinol 33(2):387–410. requiring appropriate positive controls in endocrine disruption research. Toxicol Sci 28. Carroll JS, et al. (2005) Chromosome-wide mapping of estrogen receptor binding 115(2):612–613; author reply 614–620. reveals long-range regulation requiring the forkhead FoxA1. Cell 122(1): 13. Myers JP, Zoeller RT, vom Saal FS (2009) A clash of old and new scientific concepts in 33–43. toxicity, with important implications for public health. Environ Health Perspect 29. DuSell CD, McDonnell DP (2008) 27-Hydroxycholesterol: A potential endogenous 117(11):1652–1655. regulator of estrogen receptor signaling. Trends Pharmacol Sci 29(10):510–514. 14. Vandenberg LN, et al. (2013) Regulatory decisions on endocrine disrupting chemicals 30. Evans RM, Scholze M, Kortenkamp A (2012) Additive mixture effects of estrogenic should be based on the principles of endocrinology. Reprod Toxicol 38:1–15. chemicals in human cell-based assays can be influenced by inclusion of chemicals with 15. Almstrup K, et al. (2002) Dual effects of phytoestrogens result in u-shaped dose- differing effect profiles. PLoS ONE 7(8):e43606. response curves. Environ Health Perspect 110(8):743–748. 31. Lindberg K, Helguero LA, Omoto Y, Gustafsson JA, Haldosén LA (2011) Estrogen re- 16. Li L, Andersen ME, Heber S, Zhang Q (2007) Non-monotonic dose-response re- ceptor β represses Akt signaling in breast cancer cells via downregulation of HER2/ lationship in steroid hormone receptor-mediated gene expression. J Mol Endocrinol HER3 and upregulation of PTEN: Implications for tamoxifen sensitivity. Breast 38(5):569–585. Cancer Res 13(2):R43.

6of6 | www.pnas.org/cgi/doi/10.1073/pnas.1315929110 Shioda et al. Downloaded by guest on September 30, 2021