Application of an in Vitro Assay to Identify Chemicals That Increase Estradiol and Progesterone Synthesis and Are Potential Breast Cancer Risk Factors

We identified E2- and P4-up chemicals using publicly available data of two HT-H295R steroidogenesis assay experiments conducted by the U.S. EPA and first reported by Karmaus et al. (2016) and Haggard et al. (2018). This HT-H295R assay was developed and conducted as part of the U.S. EPA’s ToxCast screening program. In these experiments, 2,012 chemicals were selected from ToxCast Phase I, II, and III and the endocrine 1000 (E1K)libraries, which include potential EDCs and other chemicals of regulatory interest, such as pesticides. The methodology for these experiments has been described in detail by Karmaus et al. (2016) and Haggard et al. (2018) and is briefly summarized below.

Human H295R adrenocortical carcinoma cells were prestimulated with forskolin for 48 h, followed by a 48-h chemical exposure at a single maximum tolerated concentration (MTC), usually $100\mathrm{\mu }\mathrm{M}$. The sample was diluted if the cell viability did not exceed 70% as tested using a [3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide] tetrazolium (MTT) cytotoxicity assay. Concentrations of 13 steroid hormones (Figure 1) were measured in the culture media using high performance liquid chromatography with tandem mass spectroscopy. The hormones dehydroepiandrosterone (DHEA) and pregnenolone were frequently detected below the lower limit of quantification (LOQ) and were excluded from further analyses. A treatment was considered to have a significant effect on a hormone if it produced a fold change $\ge 1.5$ that of the dimethyl sulfoxide (DMSO) solvent control. Treatments that significantly affected four or more hormones ($n=571$) were then tested again using a six-point concentration–response (CR) format with the MTC as $100\mathrm{\mu }\mathrm{M}$ or, if that dose was cytotoxic, the maximum concentration with $>70\%$ viability in an MTT cytotoxicity assay, and using half-log serial dilutions below that. An additional 85 chemicals that were not tested at a single MTC were also tested in the CR format (Haggard et al. 2018), for a total of 656 chemicals tested in the CR format. Across all the chemicals, administered concentration ranged from $0.041\text{\hspace{0.17em}nM}$ to $100\mathrm{\mu }\mathrm{M}$.

Karmaus et al. (2016) and Haggard et al. (2018) then used different approaches to determine the chemical–hormone pair significance (i.e., the hit-call) for each chemical run in the CR format. Karmaus et al. (2016) determined significance using the ToxCast automatic data processing pipeline (tcpl) (Filer et al. 2017), which is usually applied to all ToxCast data to standardize it with other HT data and to provide a preliminary look at the data, with the expectation that subsequent analyses would improve interpretation (Haggard et al. 2018). To be assigned a positive hit-call, the chemical data had to fit a hill or gain–loss model, and the top of the curve (the efficacy) had to exceed a cut off of six times the baseline (which was taken by normalizing the response values of the two lowest tested concentrations) (Filer et al. 2017; Karmaus et al. 2016). In contrast, Haggard et al. (2018) first determined significance for each chemical concentration–hormone pair using an analysis of variance (ANOVA) followed by post hoc Dunnett’s tests ($p\le 0.05$) on fold-change data using the DMSO solvent as the control (Haggard et al. 2018). A chemical was defined as hormonally active on the basis of logic applied in the OECD Test Guideline (TG) No. 456 (OECD 2011)—that is, either two consecutive concentrations had a significant effect or a significant result was observed at the maximum tested noncytotoxic concentration (Haggard et al. 2018).

We defined E2- and P4-up chemicals as those that significantly increased the concentrations of E2 or P4 in the CR experiments conducted by Karmaus et al. (2016) and Haggard et al. (2018), per the “Methods” section, as described above. These were identified from the subset of 656 chemicals tested in the CR format, of which 654 had adequate CR data for analysis (Haggard et al. 2018). We used the Haggard et al. (2018) hit-call data, which employed an ANOVA-based approach, to identify the E2- and P4-up chemicals because this approach was developed and validated for this assay by the OECD in TG 456 (OECD 2011) and the interlaboratory validation report for the OECD TG 456 (Hecker et al. 2011). The U.S. EPA’s ToxCast data, which uses the tcpl hit-call data from Karmaus et al. (2016), can also be used to identify E2- and P4-up chemicals, and we included the chemical concentration resulting in 50% of maximal hormone increase (AC50) and the chemical concentration resulting in 10% of maximal hormone increase (AC10) values from the tcpl data in Excel Tables S1 and S2. The ANOVA-based approach may be more sensitive to effects at the lower doses because tcpl uses the lower doses to establish a baseline or control response, whereas the ANOVA compares responses at each dose to the control and so is able to detect increases at the two lower doses. The hit-call data we used can be found in the Supplemental Data File 2 of Haggard et al. (2019), an extension of the Haggard et al. (2018) experiments and data, which reports the direction of steroidogenesis for each hormone measured after chemical treatment. If a chemical was run multiple times, we used only the data from the earliest block (as indicated by the lowest numbered plate). All data processing and analysis that followed was conducted using R (version 3.6.3; R Development Core Team; https://github.com/SilentSpringInstitute/Cardona-and-Rudel-2021/).

We used three measures to represent the potency and/or efficacy of each E2- and P4-up chemical: a) the maximal fold change (MFC) in hormone concentration; b) the lowest effective concentration (LEC) at which the tested chemical produced a significant hormone increase; and c) the adjusted maximal mean Mahalanobis distance (adj.maxmMd). The adj.maxmMd, described by Haggard et al. (2018), is a unitless statistical value reflecting a chemical’s magnitude of steroidogenic pathway disruption for 11 hormones across the tested chemical concentration range; it controls for correlation and covariance among the hormone analytes (Haggard et al. 2018).

To find the MFC, a proxy for efficacy, we used the raw hormone data found in the Haggard et al. (2018) Supplemental Data File 3. Using the raw hormone data, we first calculated the fold change relative to the average of the DMSO solvent control duplicates from the same plate at each of the tested concentrations. We then used Haggard et al. (2018) Supplemental Data Files 4 and 7 to identify the chemical concentrations with significant hormone increases. Following OECD hit-call logic, we identified the chemical concentrations with consecutive significant hormone increases or a significant hormone increase at the highest tested noncytotoxic chemical concentration. Of these chemical concentrations, the highest fold change was chosen as the MFC.

We also identified the LEC, a proxy for potency, using the raw hormone data from the Haggard et al. (2018) Supplemental Data File 3. As with the MFC, we used the Haggard et al. 2018 Supplemental Data Files 4 and 7 to identify the chemical concentrations with consecutive significant hormone increases, or significance at the highest tested noncytotoxic chemical concentration, and then selected the lowest chemical concentration from these as the LEC. Furthermore, we compared the LEC values with the potency values calculated using the tcpl: the AC50 and the AC10. We compiled the AC50 and AC10 values for each hormone-chemical pair using the ToxCast and Tox21 Summary Files for invitroDBv3.2 (U.S. EPA 2019a). Tcpl calculated AC50s only for chemicals it identified as active, and AC10s for chemicals that were able to fit a model even if their hit-call was not active. We conducted a Pearson’s correlation analysis to assess the relationship between the LEC and the AC50 and AC10 values.

Last, the adj.maxmMd calculated by Haggard et al. (2018) was used to take both potency and efficacy into account and identify chemicals with a clear response above assay noise considering data for 11 hormones (excluding DHEA and pregnenolone owing to their frequent detection below the lower LOQ). The adj.maxmMd for each chemical tested can be found in the Haggard et al. (2018) Supplemental Data File 11.

We defined E2- and P4-up chemicals with higher potency and efficacy as active on the basis of the following criteria: $\text{MFC}\ge 1.5$, $\text{LEC}\le 33\mathrm{\mu }\mathrm{M}$, and adj.maxmMd $>0$. These active chemicals were the focus of subsequent analysis because they are less likely to be false-positives (based on an $\text{MFC}\ge 1.5$, as suggested by the OECD interlaboratory analysis) and more likely to be active at environmentally relevant concentrations (by limiting to more potent chemicals). An adj.maxmMd $>0$ is also less likely to result in a false-positive for steroidogenic pathway disruption, per Haggard et al. (2018). We defined the excluded chemicals as borderline active. We also removed the following synthetic or endogenous steroid hormones because they can interfere with steroidogenesis by acting as substrates: 17-methyltestosterone, $17\mathrm{\alpha }\text{-estradiol}$, $17\mathrm{\alpha }\text{-ethinylestradiol}$, $17\mathrm{\alpha }\text{-hydroxyprogesterone}$, $17\mathrm{\beta }\text{-estradiol}$, 4-androstene-3,17-dione, $5\mathrm{\alpha }\text{-dihydrotestosterone}$, androsterone, dehydroepiandrosterone, equilin, estrone, P4, and testosterone propionate.

We classified the active chemicals into three groups corresponding to their potency and efficacy by ranking them on the basis of the average of their MFC percentile rank and LEC percentile rank. We considered those chemicals in the top 25% of the rank as having higher (compared with the other chemicals in the ranking) efficacy and potency, chemicals in the middle 25–75% as having intermediate potency and efficacy, and chemicals in the bottom 25% as having lower potency and efficacy.

Although chemicals that increase E2 or P4 will not necessarily also activate the ER directly, we wanted to identify the E2- and P4-up chemicals that may also act on the ER to provide insight on whether the two activities appear to be independent vs. dependent. We gathered the ER area under the curve (AUC), as calculated by Judson et al. (2015), which is a score between 0 and 1 describing the probability of a chemical to be active at the ER on the basis of its potency and efficacy across 18 of ToxCast’s in vitro assays that measure ER-associated pathways such as binding, dimerization, and ER-dependent cell proliferation (Judson et al. 2015). We considered chemicals positive for ER interaction if they had an AUC score $\ge 0.1$, positive per Judson et al. (2015). Chemicals with an $\text{AUC}\ge 0.01$ but $<0.1$ were considered as having ambiguous activity, and chemicals with an AUC $<0.01$ were considered likely in vivo inactive (Judson et al. 2015). The ER AUC for agonism and antagonism, describing the probability of a chemical to be either an ER agonist or an ER antagonist, was also calculated by Judson et al. (2015), and we considered both probabilities, identifying the activity with the higher probability as being the more likely.

We conducted three analyses to evaluate whether in vivo effects consistent with increased E2 or P4 are observed for chemicals that were active in vitro. First, we looked for any evidence of carcinogenicity or reproductive or developmental toxicity in authoritative databases and mammary gland–focused review articles. Second, we evaluated how many of the chemicals that have been listed in review articles as causing mammary gland tumors or other mammary gland effects also show in vitro E2- and P4-up activity. Finally, because the authoritative carcinogenicity or reproduction and developmental toxicity assessments are general and may not specifically reflect effects due to increased E2 or P4, we conducted a more focused, but preliminary, primary literature review for a subset of the most potent and efficacious E2- and P4-up chemicals to identify in vivo effects that may be consistent with increased E2 and P4, focusing on mammary gland effects. Of course, the end points we looked for in rodent studies are imperfect proxies for in vivo effects in humans, but relevant data on humans is generally not available. The details of these analysis follow.

To collect information on in vivo reproductive toxicity, developmental toxicity and carcinogenicity, we used the U.S. EPA’s Toxicity Value Database (ToxValDb) (Judson 2019; Williams et al. 2017). ToxValDb summarizes in vivo studies by various parameters, such as type of study, species tested, and toxicity values [e.g., no observed effect level (NOEL) or lowest observed effect level (LOEL)].

We gathered measures of exposure potential (i.e., the likelihood of general population exposure) and exposure sources for the active E2- and P4-up chemicals. We used data from the U.S. EPA’s HT exposure prediction meta-model, which estimates the population intake rate of chemicals ($n=\mathrm{479,926}$) by using structural and physiochemical properties, production information, and predicted exposure pathway(s) leading from source to consumer (e.g., pesticides, industrial processes, consumer products and diet) (Ring et al. 2019). Using these predictors, the model estimates a median intake rate for the general U.S. population (units of milligrams per kilogram of BW per day) with a 95% credible interval (CI; quantiles for 0.025 and 0.975) for each chemical. This CI holds that the true median has a 95% probability of falling within the interval. Because of the tremendous uncertainty in these exposure estimates and the lack of any estimate of high-end population exposures, we used the upper quantile of the CI; if an upper 95% CI was not calculated, we used the predicted median instead.

To identify potential sources of exposure, we used the U.S. EPA CPDat (Dionisio et al. 2018), which compiles data on chemical composition of products and product/chemical usage as reported in material data safety sheets, ingredient lists or online retail sites. Each chemical is assigned a cassette or a product use category composed of terms describing the product/chemical usage. We used these categories to assign the E2- and P4-up chemicals into one or more of the following exposure sources: consumer (e.g., in furniture, toys or apparel), industrial (e.g., product manufacturing or processing), pesticide, diet (e.g., food flavoring/preservatives/colorants or food/water contaminants), and pharmaceutical; Excel Table S6 provides a descriptor for each. We also used data source information from CPDat (Dionisio et al. 2018) to assign chemicals into the exposure source categories; for example, every chemical/cassette pair originating from the Retail Product Categories was categorized as consumer use. Some sources were not exposure-source specific so we manually categorized these cassettes and product use categories based on their terms (Excel Table S7). Some chemical/cassette pairs were classified into multiple exposure sources, as is the case for pesticides whose data originated from an industry-specific source and a consumer use-specific source, and were thus categized under the pesticide, consumer use, and industrial categories.

To provide additional detail, we identified several exposure source categories that describe the functional uses of a chemical or the type of product it may be found in. These included antimicrobials, cigarettes, drinking water contaminants, flame retardants, food additives, food contact, food residue, fragrance, hair-dyes, human metabolites, personal care products, plastics, or textiles—descriptions of these can be found in Excel Table S8. Some of these categories were chosen because they have previously been reported as associated with breast cancer or as containing endocrine disruptors (e.g., personal care products, textiles, fragrance, antimicrobials), or because they have wide exposure potential (e.g., water and food-related contamination). To assign chemicals into these exposure sources, we used text search to match the terms in each chemical’s corresponding cassette or product use category to our categories of interest. A list of cassettes and/or product use categories and the associated categories can be found in Excel Table S9.

If a chemical was not assigned to an exposure source using CPDat (Dionisio et al. 2018) data, we used the exposure pathway assumed by Ring et al. (2019) to classify the chemical. Ring et al. (2019) used machine learning to assign a likely exposure pathway on the basis of a chemical’s structure and physicochemical properties.

We also identified the chemicals that are found in currently used pesticide products, pharmaceuticals, or consumer products; or if they are biomonitored in the Centers for Disease Control and Prevention (CDC) National Health and Nutrition Examination Survey (NHANES) because there is suggested exposure in the U.S. population. To identify current U.S. EPA-registered pesticides, we used the Pesticide Product Information System (PPIS) (U.S. EPA 2021), which lists pesticide products registered for use in the United States and their active ingredients. To identify pharmaceuticals in current U.S. Food and Drug Administration (FDA)-approved products, we used the Orange Book: Approved Drug Products with Therapeutic Equivalence Evaluations publication (FDA 2020), which lists drugs approved for use by the FDA and their active ingredients. Although there is no comprehensive list of chemicals in consumer goods, we noted if our chemicals were listed in CPDat (Dionisio et al. 2018) as consumer use products, which is compiled from material data safety sheets, ingredient lists and online retail sites. Finally, we used NHANES list of biomonitored chemicals, to identify if any of the E2-up or P4-up chemicals are biomonitored; this was downloaded from the CompTox Chemical Dashboard’s downloadable list of NHANES chemicals (Williams et al. 2017).

Based on an HT CR screening of 654 chemicals (of an initial 2,012) in the H295R steroidogenesis assay (Haggard et al. 2018, 2019), 418 chemicals significantly increased E2 (274) or P4 (283), with 139 increasing both (Excel Tables S1 and S2). To be included in the 418 chemicals, the chemical also had to show a significant effect at two consecutive concentrations or at the highest noncytotoxic concentration tested (i.e., the MTC), consistent with Haggard et al. (2018) criteria.

Of these 418 E2- and/or P4-up chemicals, we classified 296 as active (based on our previously defined criteria for potency and efficacy and excluding hormones) and these were the focus of subsequent analyses receiving an efficacy/potency score of higher, intermediate, or lower (Figure 2; Excel Tables S1 and S2) based on the average of their respective efficacy (i.e., the MFC) and potency (i.e., the LEC) percentile ranks (Excel Tables S1 and S2) as described in the “Methods” section. This set includes 182 E2-up and 185 P4-up chemicals, with 71 increasing synthesis of both E2 and P4. We classified the 122 chemicals that did not meet the criteria for active as borderline actives (indicating exclusion based on our criteria for potency and efficacy) or hormone substrates (indicating a hormone that may act on the steroidogenesis pathway as a substrate); these are included in Excel Tables S1 and S2 and can be identified in the efficacy/potency column. Figure 2 shows the 296 active chemicals that increased E2 and P4 arranged by their potency (i.e., the LEC) and efficacy (i.e., the MFC) (Excel Tables S1 and S2).

For the 182 active E2-up chemicals, the LECs ranged from 0.0041 to $33\mathrm{\mu }\mathrm{M}$ (first and third quartile: 1.2, 20), and the MFCs ranged from 1.5 to 82.1 (first and third quartile: 1.8, 2.6). The E2-up chemicals with the highest efficacy/potency on the basis of the average of the LEC and MFC percentile ranks (indicating highest potency and efficacy) were three pesticides (hexythiazox, oxyfluorfen, pirimiphos-methyl) and two failed drug candidates (CP-61237 and PharmaGSID47263) (Excel Table S1).

For the 185 active P4-up chemicals, the LECs ranged from 0.037 to $33\mathrm{\mu }\mathrm{M}$ (first and third quartile: 1.2, 11), and the MFCs ranged from 1.5 to 39.3 (first and third quartile: 1.9, 4.7). The P4-up chemicals with the highest efficacy/potency on the basis of the average of the LEC and MFC percentile ranks (indicating highest potency and efficacy) were imazalil, prochloraz, and triflumizole (three pesticides), mifepristone (a drug), and 3,3′-dimethylbenzidine (used in producing dyes and pigments) (Excel Table S2).

We compared the LEC measurements that were calculated using data of Haggard et al. (2018) with the AC50 and AC10 potency estimates calculated using the tcpl (Filer et al. 2017) (Figure S1). Only 25% (46) of the 182 E2-up chemicals were considered active using tcpl compared with 60% (111) of the 185 P4-up chemicals; these tcpl actives also had an AC50 available. An additional 86 E2-up and 42 P4-up chemicals had an AC10 calculated by tcpl but were not considered active in the ToxCast hit-calls. In general, for E2-up chemicals the AC50 approximated the LEC most closely [E2-up: $r\left(44\right)\mathrm{=}0.92$, $p<{20.2}^{-16}$; P4-up: $r\left(109\right)=0.74$, $p<{2.2}^{-16}$] and these values were more highly correlated compared with the AC10s (Figure S1). For both E2- and P4-up chemicals, the correlations between LEC and AC10s limited to the tcpl actives were generally lower than with the AC50s but still highly correlated [E2-up: $r\left(44\right)=0.81$, $p={1.3}^{-11}$; P4-up: $r\left(109\right)=0.69$, $p<{2.2}^{-16}$], and for the E2-up chemicals, correlations were further reduced when combining the chemicals that the tcpl considered to be active or inactive [E2-up: $r\left(130\right)=0.66$, $p<{2.2}^{-16}$; P4-up: $r\left(151\right)=0.68$, $p<{2.2}^{-16}$].

Some of the chemicals that increased E2 or P4 synthesis also showed ER agonism, but most did not (Excel Tables S1 and S2). Ten of the 182 E2-up chemicals and 15 of the 185 P4-up chemicals had an AUC $\ge 0.1$ and were thus considered active at the ER; all 10 of the E2-up ER-active chemicals were agonists whereas 14/15 P4-up chemicals were. An additional 34 E2-up chemicals and 33 P4-up chemicals had an $\text{AUC}\ge 0.01$ but $<0.1$ and so were considered to have ambiguous ER activity. Twenty-four of the E2-up chemicals and 32 of the P4-up chemicals had not been tested for ER activity. A majority of the chemicals were not active at the ER (114 E2-up, 105 P4-up).

To investigate whether chemicals that increased E2 or P4 in vitro may be reproductive toxicants, developmental toxicants, or carcinogens in vivo, we compiled data from the U.S. EPA’s ToxValDb, from California Prop65 listings, and from review articles of chemicals observed to produce rodent mammary gland tumors or alter mammary gland development. We gathered data for the 182 E2-up chemicals and 185 P4-up chemicals that had been prioritized by potency and efficacy. As described in the “Methods” section, we conducted three analyses to evaluate whether in vivo effects consistent with increased E2 or P4 are observed for the chemicals that are active in vitro.

We found that 33% (60) of the 182 E2-up chemicals were likely repro/dev toxicants and 30% (54) were likely carcinogenic; 12% (21) were positive for both end points (Figure 3A). Meanwhile, 6% (11) of chemicals were unlikely repro/dev, and 13% (24) were designated unlikely carcinogenic (Figure 3A). Of the 185 P4-up chemicals, 33% (61) were likely repro/dev toxicants and 28% (50) were likely carcinogens; 9% (16) were positive for both of these end points (Figure 3B). In contrast, 5% (8) of chemicals were classified as unlikely repro/dev and 11% (20) as unlikely carcinogenic (Figure 3B). The majority of the chemicals did not have enough information to determine repro/dev toxicity [61% (111) of E2-up, 63% (116) of P4-up] or carcinogenicity [57% (104) of E2-up, 62% (115) of P4-up]; 43% (78) of E2-up chemicals and 44% (82) of P4-up chemicals had insufficient data to evaluate either end point (Figure 3).

In addition, this research team previously compiled $\sim 250$ chemicals with reported mammary gland tumors or other mammary gland effects, such as altered mammary gland development (Cardona and Rudel 2020; Rudel et al. 2007, 2011). For some of these chemicals, although mammary effects were reported, they were also dismissed as not treatment related (Cardona and Rudel 2020). Forty-five chemicals from the lists of chemicals with mammary gland effects (excluding endogenous or synthetic steroidal hormones) were tested in the H295R assay, and 29 of the 45 increased E2 ($n=21$) or P4 ($n=23$). Six of those had been removed from our detailed analysis because of their lower potency ($\text{LEC}>33\mathrm{\mu }\mathrm{M}$) or efficacy ($\text{MFC}<1.5$); these six are 2,4-diaminotoluene, etridiazole, captafol, biochanin A, 1,2-diphenylhydrazine, and propazine (Table 1). It is notable that these six compounds that we classified as borderline active were associated with in vivo mammary gland effects including tumors. The remaining twenty-three chemicals with mammary effects that we had classified as active for E2-up or P4-up included 19 that induced mammary tumors (Table 1). Notable in our list of E2/P4-up chemicals is 7,12-dimethylbenz(a)anthracene (DMBA), a research chemical that is commonly used to induce mammary gland tumors in experimental animal models (Abba et al. 2016; Currier et al. 2005; Liu et al. 2015; Welsch 1985). Also on the list are chemicals, or chemical classes, that consistently produce mammary tumors (Rudel et al. 2007, 2014) including aromatic amines (such as benzidine compounds, anilines, and diaminotoluenes), nitro polycyclic aromatic hydrocarbons (PAHs), and triazines (Table 1).

We conducted a preliminary literature review for the five most potent and five most efficacious chemicals that increased E2 or P4 in vitro to identify reports of in vivo effects that are consistent with increased E2 and/or P4 levels, as described in the “Methods” section. We limited to chemicals in the higher efficacy/potency classification. These findings are described in Table 2.

Of the chemicals that increased E2, the most efficacious chemicals (i.e., the highest MFC) were the herbicide precursor 2,4-dichlorophenol (2,4-DCP), three pesticides (oxyfluorfen, cyfluthrin, and coumaphos), and the supplement forskolin (Figure 2). The most potent chemicals (i.e., the lowest LEC) were four pesticides (hexythiazox, difenoconazole, pirimiphos-methyl, and dimethomorph) and the methoxychlor metabolite 2,2-bis($p$-hydroxyphenyl)-1,1,1-trichloroethane (HPTE) (Figure 2). These represent 10 unique chemicals with LECs ranging from 0.04 to $11\mathrm{\mu }\mathrm{M}$ and MFCs ranging from 1.9 to 82.1 (Table 2). Of these 10, 3 were classified likely and 3 unlikely to be carcinogenic, 4 had inadequate evidence; 5 were classified likely repro/dev toxicants, and 5 inadequate evidence (Table 2).

We found reports of mammary tumors or other mammary gland effects for four of these 10 E2-up chemicals: 2,4-DCP, cyfluthrin, HPTE (as methoxychlor), and hexythiazox (Table 2). In addition, possible effects on lactation are suggested for four chemicals where lower pup weight during lactation was reported: hexythiazox, difenoconazole, cyfluthrin, and oxyfluorfen. For most (7) of the top 10 E2-up chemicals, we found reports of some effect potentially related to increased E2 synthesis. In some cases, potentially relevant effects were dismissed as not treatment related—owing to a lack of statistical significance (i.e., cyfluthrin) or not following a monotonic increasing dose–response pattern (oxyfluorfen)—or as unimportant (i.e., 2,4-DCP) (Table 2). For two chemicals (coumaphos and pirimiphos-methyl) the U.S. EPA reported no E2-relevant effects in summary documents, but the underlying studies or data evaluations records were not available for review (Table 2). Where we did not find mammary gland effects, it may reflect limited assessment and reporting.

Of the chemicals that increased P4, the ones with the highest MFCs included a pharmaceutical (mifepristone), two pesticides (prochloraz and imazalil), a dye intermediate (3,3′-dimethylbenzidine), and the industrial chemical 4-(2-phenylpropan-2-yl)-N-[4-(2-phenylpropan-2-yl)phenyl]aniline (Figure 2). The chemicals with the lowest LECs were four pesticides (imazalil, triflumizole, prochloraz, and fenbuconazole) and the industrial chemical 2,3-dinitrotoluene (2,3-DNT) (Table 2). These represent eight unique chemicals with LECs ranging from 0.037 to $3.7\mathrm{\mu }\mathrm{M}$ and MFCs ranging from 2.7 to 39 (Table 2). Of these eight, four were classified as likely and one as unlikely to be carcinogenic, three had inadequate evidence; three were classified as likely repro/dev toxicants and five had inadequate evidence (Table 2).

Of the eight P4-up chemicals, mammary tumors are reported for two (3,3′-dimethylbenzidine and a mixture that contains 2,3-DNT), whereas possible lactation-related effects are reported for three (mifepristone, triflumizole, and imazalil) (Table 2). Effects on hormone levels or other reproductive effects are reported for five chemicals (triflumizole, fenbuconazole, mifepristone, prochloraz, and imazalil); prochloraz increased serum concentrations of P4 in vivo, and prochloraz, imazalil, and triflumizole increased gestational length.

We used data from the U.S. EPA’s ExpoCast HT exposure modeling (Ring et al. 2019) and a chemical uses database (Dionisio et al. 2018) to estimate general population exposure levels and classify the chemicals by type of use, for example, diet, consumer products, pesticide, industrial, pharmaceutical. Of the 182 E2-up and 185 P4-up chemicals, 169 and 165 chemicals, respectively, had an identified exposure source or exposure prediction by the U.S. EPA (271 unique chemicals). Of the 169 E2-up chemicals, 93 are in consumer products, 121 in dietary sources, 103 in industrial sources, 127 in pesticide products, and 36 in pharmaceuticals (Figure 4A). Of the 165 P4-up chemicals, 84 are in consumer products, 110 in dietary sources, 110 in industrial sources, 108 in pesticide products, and 39 in pharmaceuticals (Figure 4A). Many of the chemicals that were identified in pesticide products have dietary exposures: 102 E2-up, and 87 P4-up (Excel Tables S1 and S2). Based on more specific data about exposure sources, 119/271 E2 and/or P4-up chemicals are used as food additives, 110 are found as food residue, 50 have food contact, and 77 are considered drinking water contaminants (Figure 4B). Fifty-eight are found in personal care products, with 17 of these found in hair dye (Figure 4B). Twenty-two are reported in textile-related products, 36 in plastic-related products, and 14 in flame retardants (Figure 4B).

The U.S. EPA’s exposure modeling (Ring et al. 2019) generated median population exposure estimates for 181/182 E2-up chemicals and 183/185 P4-up chemicals. As discussed in the “Methods” section, because of uncertainties in these estimates, we used upper CI estimates of the predicted medians (UCI median). For the E2-up chemicals, UCI median exposure ranged from $<0.0001$ to $8.74\mathrm{mg}/\mathrm{kg}\text{\hspace{0.17em}BW}/\text{day}$, and 33 chemicals had a UCI median $>0.01\mathrm{mg}/\mathrm{kg}\text{\hspace{0.17em}BW}/\text{day}$ (Figure 5). For the P4-up chemicals, UCI medians ranged from $<0.0001$ to $52\mathrm{mg}/\mathrm{kg}\text{\hspace{0.17em}BW}/\text{day}$, and 28 chemicals had a predicted exposure $>0.01\mathrm{mg}/\mathrm{kg}\text{\hspace{0.17em}BW}/\text{day}$ (Figure 5). Some of the highest exposure E2-up chemicals were nitrilotriacetic acid (a consumer-use product used for chelating), atraric acid (a natural compound often used as a drug), 10-undecenoic acid (used in drugs and consumer products), dimethyl isophthalate (found in food additives and adhesives), and $p$-cresol (used as a disinfectant and antiseptic); for P4-up, they included 3-isopropylphenol, $p$-cresidine (used in dyes and pigments), 2,4,6-tribromophenol (a fungicide and intermediate in flame retardant production), 2,5-dichlorophenol (an intermediate in pesticides), and 3-tert-butylphenol (Figure 5).

We identified 119/182 E2-up chemicals and 115/185 P4-up chemicals that have a higher likelihood of exposure because they are found in current U.S. EPA-registered pesticide products (55 E2-up and 37 P4-up), are active ingredients in current FDA-approved drug products (5 E2-up and 8 P4-up), are reported in consumer use products (76 E2-up and 77 P4-up), or are currently biomonitored in NHANES (30 E2-up and 28 P4-up) (Excel Tables S1 and S2). These include many common pesticides (e.g., diazinon, atrazine, cyfluthrin, malathion, permethrin, imazalil) and commercial chemicals [e.g., di(2-ethylhexyl) phthalate, 2,4-DCP, tetrabromobisphenol A, methyl paraben, nitriloacetic acid].

Interestingly, 53 of the 104 active E2-up chemicals that have not been evaluated for carcinogenicity are among the 119 chemicals currently used in consumer goods, pesticides, or drugs (Figure 3A; Excel Tables S1 and S2). The two chemicals in this group with a higher potency/efficacy and the highest predicted exposure rates are 2,4-DCP (a precursor and metabolite of the herbicide 2,4-D) and 2,4-dimethylphenol (used as a fungicide and disinfectant, also known as 2,4-xylenol). Sixty-two of the 111 E2-up chemicals with inadequate evidence to assess reproduction and development are also currently used (Figure 3A; Excel Tables S1 and S2); the two chemicals with a higher potency/efficacy as well as highest predicted exposure are nitrilotriacetic acid (used in washing and cleaning products) and 2,4-dimethylphenol. For P4-up chemicals, 59 of 115 chemicals with inadequate evidence to assess carcinogenicity have current known uses (Figure 3B; Excel Tables S1 and S2). Among these chemicals, the higher potency/efficacy P4-up chemicals with the highest predicted exposure rates are 3,4-dichloroaniline (a precursor of the commonly used herbicide propanil and an azo dye for polyester fabrics), 2,3-DNT (an isomer of dinitrotoluene that is used in explosives), and 4,4′-sulfonylbis[2-(prop-2-en-1-yl)phenol] (a $p,p$′-bisphenolic compound also known as TGSA). Of the P4-up chemicals with the inadequate evidence to assess reproduction and development, 66 of 116 chemicals have current known uses (Figure 3B; Excel Tables S1 and S2).

We found that chemicals that increased E2 and/or P4 were more likely to be carcinogens and/or repro/dev toxicants than to not cause those types of effects. Many of the chemicals were identified as likely repro/dev toxicants [33% (60) E2-up and 33% (61) P4-up] or potentially carcinogenic [30% (54) E2-up and 28% (50) P4-up] by authoritative sources or in comprehensive reviews (Figure 3). Generally, fewer than 6% of the chemicals appeared to be unlikely reproductive or developmental toxicants ($\text{NOEL}\ge 100\mathrm{mg}/\text{kg-day}$) (Figure 3). Fewer than 13% were classified as unlikely to be carcinogenic (Figure 3); however, this number could be smaller if mammary tumors were observed but dismissed, as we have previously reported for pesticides (Cardona and Rudel 2020) and as we found to be the case for some of the E2- and P4-up chemicals we identified (Table 2). In addition, well over half of the chemicals with observed mammary gland effects and that were tested in the HT-H295R assay increased E2 or P4 (Table 1), so this mechanism appears to be common in chemicals that cause mammary gland tumors and other effects. Of course, the effects we identified are not necessarily a consequence of the E2-up activity of the chemical, and there are several examples of chemicals with multiple activities related to estrogen pathways, for example, HPTE (methoxychlor).

Through our preliminary literature review of the 18 most potent and efficacious E2- and P4-up chemicals, we found in vivo evidence of effects that may be due to increased hormone synthesis, including increased hormone concentrations, mammary gland effects including tumors, and other reproductive and developmental toxicity (Table 2). Of concern, we found cases where effects that are plausibly related to the mechanism of increasing E2 or P4 were dismissed in regulatory evaluations, and cited as not statistically significant or not biologically relevant. For example, a multigeneration reproduction and development study of 2,4-DCP (a substrate and degradant of the widely used pesticide 2,4-D)—which produced the highest increase in E2 production in the H295R assay—observed alterations in mammary gland histopathology at all doses in all generations (Aoyama et al. 2005). Despite these histopathological changes, a no observed adverse effect level (NOAEL) was set at the mid dose, ignoring these and other estrogenic effects. Surprisingly, these effects were not mentioned in the abstract or conclusions of the article or as a keyword, making it impossible to find this article by searching “mammary,” our primary search term for mammary effects, in PubMed. We also found that cyfluthrin, one of the more potent E2-up chemicals, produced mammary adenocarcinomas in male and female rats in the cancer bioassay, yet the tumors in females were dismissed by the U.S. EPA based on statistical significance and the tumors in males were not discussed (U.S. EPA 2001c). The adenocarcinomas in the males and females exceeded rates in historical controls (Haseman et al. 1998), and mammary hyperplasia, inflammation, and fibroadenomas were also reported in some low-dose males and females (Table 2) (U.S. EPA 2001c). It is possible that severe BW decreases observed at the mid and high doses (U.S. EPA 2001c) masked tumors in the mammary gland as it has been previously shown that weight decreases can lead to reduced mammary tumors in rodents (Haseman et al. 1998). The acaricide hexythiazox was also found to produce mammary tumors, although a lack of mutagenicity was cited as a reason to develop a reference dose over a cancer slope factor (U.S. EPA 2011, 2012a). Although not mutagenic, hexythiazox may act as an EDC and increase risk of mammary gland tumors and lactation effects through the E2 synthesis pathway, given that there was also an observed decrease in pup weight during lactation (U.S. EPA 2011, 2012a).

Because testing of pesticides is required before registration (including tests for developmental and reproductive toxicity and cancer), whereas there are no similar testing requirements for many other common chemicals, finding that many of the chemicals with in vivo evidence are pesticides is not surprising. However, as we have reported previously (Cardona and Rudel 2020), we have found many examples where mammary tumors were dismissed for pesticides, so pesticide registrations for chemicals that cause mammary tumors and also increase E2 or P4 or have other endocrine activity should be reevaluated. Pesticides that increased E2 or P4 in the HT-H295R assay and have reported mammary gland tumors but were classified as unlikely carcinogenic by the U.S. EPA’s pesticide office include atrazine (E2-/P4-up), simazine (E2-/P4-up) and propazine (E2-up). In addition, of the chemicals with inadequate evidence for carcinogenicity, two pesticides—ametryn (E2- and P4-up) and terbutylazine (E2-up)—have reported mammary tumors in vivo, but these were also dismissed by the U.S. EPA (Cardona and Rudel 2020). There are likely to be additional pesticides where mammary tumors were dismissed that were not identified in Cardona and Rudel 2020, given that that report only included pesticides where mammary tumors were reported in the Reregistration Eligibility Decisions, and the cancer bioassays or their summaries were not reviewed. For example, in this article we also identified cyfluthrin as one of the most effective/potent E2-up chemicals, and in searching for in vivo data, we discovered that mammary tumors had been reported and then dismissed in studies submitted to the U.S. EPA’s Office of Pesticide Programs (U.S. EPA 2001c).

To investigate potentially relevant in vivo effects for the 20 most potent and efficacious E2- and P4-up chemicals, we read original studies (or data evaluation records for pesticides) and because mammary effects were not consistently reported, we had to review study findings extremely closely to identify relevant effects. A more systematic approach would be a useful next step but is challenging because mammary gland assessment methods are not standardized or routine (Makris 2011; Rudel et al. 2011), hormone measurements may not be made with sufficiently sensitive methods (OECD 2018a; Schwarzman et al. 2015), and underlying study data for many pesticides are not available and so cannot be reviewed (Cardona and Rudel 2020; https://iaspub.epa.gov/apex/pesticides/f?p=CHEMICALSEARCH:1:0::NO:1). Taken together, these limitations in the conduct and reporting of studies suggest that in vivo effects that result from chemicals that increase E2 and P4 levels may have been missed in many cases.

Although we did not conduct a primary literature review to identify relevant in vivo findings for the remaining E2- and/or P4-up chemicals, a few chemicals and chemical groups stood out because of their well-established connections with breast cancer. These include DMBA [a chemical commonly used to induce mammary tumors as an experimental model of breast cancer (Welsch 1985)], aromatic amines (such as benzidine compounds, anilines, and diaminotoluenes), and nitro PAHs, all of which consistently induce mammary tumors in rodents (Rudel et al. 2014). Rodent cancer bioassays are the most common approach to predicting if a chemical may cause cancer in humans (IARC 2019) and they are typically used to identify carcinogens and establish exposure limits. Russo and Russo (1996) documented many aspects of mammary gland hormonal regulation, development, and carcinogenesis that are similar in rats and humans, although some important differences have also been highlighted (Thayer and Foster 2007). It is also interesting to note that chemically induced mammary tumors are more often seen in rats compared with mice (Dunnick et al. 1995; Thayer and Foster 2007), and the observation that aromatase promoters are not present in mammary adipose in mice (Zhao et al. 2012, 2016) might explain this difference in part.

Notably, eight triazine pesticides (terbutylazine, deisopropylatrazine, deethylatrazine, phenothiazine, anilazine, simazine, atrazine, and 2,4,6-tris(allyloxy)-1,3,5-triazine) increased E2, and the latter four also increased P4 levels, suggesting that effects on steroidogenesis may be a mechanism of action for the observed increases in mammary tumors in rats after administration of several triazines (Cardona and Rudel 2020; Rudel et al. 2007). Those tumors were dismissed as not being relevant to humans because they were attributed to a persistent estrous via attenuation of luteinizing hormone surge (U.S. EPA 2018a). However, the observed increases in E2 and P4 by the triazines are consistent with previous in vitro reports of increased E2 production (Hecker et al. 2011; Higley et al. 2010; Tinfo et al. 2011), and P4 production (Pogrmic-Majkic et al. 2014; Tinfo et al. 2011), including in ovarian granulosa cells and H295R cells. Increased E2 has been reported in vivo in male rats (Stoker et al. 2000) and in ovariectomized female rats (Cooper et al. 2007). Triazines have also been shown to induce aromatase activity in vitro (Caron-Beaudoin et al. 2016; Heneweer et al. 2004; Holloway et al. 2008; Sanderson et al. 2000, 2001; Tinfo et al. 2011), suggesting aromatase induction as a possible mechanism for increasing E2 production. These effects represent an important mechanism of action that may be operative in humans by increasing local or systemic hormone levels and breast cancer risk. We propose that additional research to better characterize the significance of these effects should be a priority given that U.S. regulatory agencies have dismissed the mammary tumors but these herbicides are widely used and are common drinking water contaminants in agricultural areas. There is precedent for reviewing previously published toxicity studies and modifying the conclusions in response to new information about mechanism; such changes may be appropriate for the triazines and other chemicals discussed here, such as 2,4-DCP and cyfluthrin. For example, a multigeneration study of dibutyl phthalate was revised after low-incidence effects on the testis, epididymis, and penis of F1 male offspring, which had originally been dismissed as not treatment related, were reviewed in light of knowledge that phthalates inhibit fetal testosterone synthesis (Makris et al. 2013).

In addition, some chemicals have previously been shown to have potential E2/P4-related repro/dev effects in epidemiologic studies, including phenols and parabens. For instance, peripubertal exposure to methyl paraben (E2-up, P4-up, weak ER agonist) from personal care products was associated with an earlier age of breast development, pubic hair development and menarche in girls; peripubertal propyl paraben (P4-up, weak ER agonist) urinary concentrations were associated with earlier pubic hair development in girls and earlier genital development in boys; and peripubertal exposure to 2,5-DCP (E2-up) was associated with later-onset pubic hair development in girls (Harley et al. 2019). A possible explanation for some of these results is that children going through early puberty are more likely to use personal care products that contain these chemicals (Harley et al. 2019), so additional studies are required to establish a clear relationship. Harley et al. (2019) also reported that prenatal exposure to 2,4-DCP (E2-up) was associated with earlier age at menarche. A study in pregnant women reported that butyl and propyl paraben (both P4-up, weak ER agonists) concentrations were associated with increased odds of preterm birth, decreased gestational age at birth, decreased birth weight and decreased body length (Rattan et al. 2017). Butyl paraben levels in young adults were also associated with shortened menstrual cycles (reviewed by Rattan, Zhou et al. 2017). Another study of pregnant women reported decreased serum E2 and a decreased E2/P4 ratio associated with urinary butylparaben (P4-up, weak ER agonist) (reviewed by Rattan, Zhou et al. 2017). In addition, a case report of gynecomastia among male refugees exposed to lice treatments in a detention center investigated estrogenic (via the ER) and anti-androgenic effects of the lice treatments, which contained phenothrin, permethrin, and piperonyl butoxide. Although these investigators identified phenothrin as having some antiandrogen activity (Brody 2003), it is interesting to note that all three of those pesticides increased E2 in the HT-H295R assay, and permethrin also increased P4, so it is possible that effects on steroidogenesis might be involved in this outbreak. Genetic polymorphisms in aromatase and other CYP enzymes may influence individual responses to these chemicals in humans. In addition, there is supporting evidence of paraben effects in rodents, given that early life oral exposure to methyl paraben (E2- and P4-up, weak ER agonist) led to histological abnormalities in mammary glands and increased pup mortality that may be due to effects on lactation (Manservisi et al. 2015). We suggest that it should be a priority to evaluate whether exposure to chemicals that increased E2 or P4 are associated with breast cancer incidence or progression, breast density, breast development, or effects on lactation. Some methodological challenges must be addressed to allow measuring human exposures in a meaningful way and for measuring some of these outcomes.

Taken together, our analysis provides preliminary and compelling evidence that chemicals that increased E2 and P4 in the HT-H295R assay are of toxicological concern because many of these chemicals affected the mammary gland (e.g., causing tumors) and/or induced other repro/dev effects in vivo. However, almost half of the chemicals have not been tested in vivo or do not have adequate in vivo data to determine repro/dev toxicity or carcinogenicity (Figure 3; Excel Tables S1 and S2), and so these should be priorities for further study. Prioritization of these for further evaluation could consider exposure potential as well as potency and efficacy. For example, we identified 53 active E2-up and 59 P4-up chemicals that have not been evaluated for carcinogenicity and are currently used in consumer goods, pesticides, drugs, or are biomonitored in NHANES (Figure 3; Excel Tables S1 and S2) and so these should be considered for carcinogenicity assessment, carefully considering hormonal mechanisms. Similarly, there are 62 active E2-up and 66 P4-up chemicals with current uses that have inadequate evidence to assess repro/dev toxicity (Figure 3), so they should be prioritized for in vivo study.

Many of the E2- and P4-up chemicals represent common exposures. For example, 119/182 E2-up and 115/185 P4-up chemicals are reported to be currently used in pesticides, pharmaceuticals, and/or consumer products, or are found in NHANES biomonitoring samples (Figure 3; Excel Tables S1 and S2). Many of the chemicals in our E2- and P4-up chemical lists are pesticides (127 and 108, respectively), which is concerning, considering there are many ways that people can be exposed to pesticides, including aerial spraying, food residues, drinking water contamination, and use in the home. Indeed, many pesticides have been previously detected in environmental samples such as soil, water, indoor air and dust, food residues, and drinking water (Rudel et al. 2003; Zota et al. 2017). Consumer products are another major exposure source for these chemicals, including personal care products, such as hair dye, and materials used in buildings and furnishings, such as chemical flame retardants. This finding aligns with previous studies that have found various EDCs, such as phthalates, parabens and phenols, in commonly used products such as feminine hygiene products (Gao and Kannan 2020), hair products (Helm et al. 2018), and other personal care products (Dodson et al. 2012). Chemicals found in personal care products have also been found in in home air and dust (Rudel et al. 2003; Zota et al. 2017). Some of the consumer products also include pesticides that are used in or near homes and lawns, on children for lice control, or on pets, scenarios that present higher exposure potential. Imazalil, one of the most potent and effective P4-up chemicals, is a commonly used fungicide applied post-harvest to citrus fruits, which may result in elevated exposures through contact with and ingestion of citrus peels (Vass et al. 2015). A barrier to reducing exposure to these chemicals is that, with the exception of pesticides, ingredient disclosure is not available for most chemicals in most categories of products (Egeghy et al. 2012). The exposure sources we gathered from the U.S. EPA’s CPDat (Dionisio et al. 2018) are far from comprehensive, as there are no laws or regulations requiring full disclosure of consumer product ingredients.

The U.S. EPA’s estimates of median population exposures identified 33 E2-up and 28 P4-up chemicals with exposure rates $>0.01\mathrm{mg}/\mathrm{kg}\text{\hspace{0.17em}BW}/\text{day}$, and include butylphenols, parabens, BPA, benzophenone, nitrophenols, and fragrance chemicals (Figure 5; Excel Tables S1 and S2). Although there is substantial uncertainty in these estimates because information about chemical use is often not disclosed by manufacturers, the estimates can indicate which chemicals are expected to have ubiquitous population-level exposures. Moreover, it is well known that population distributions of environmental chemical exposure are lognormal and highly skewed, which means typically exposures among the higher exposed people can be several orders of magnitude above median exposures (Jia et al. 2008; Su et al. 2012). Thus, these median exposure estimates are of limited utility for identifying chemicals that pose a higher risk to subpopulations of frequent users, given that the median is low if most of the population does not use the chemical. For example, people using chemicals in the workplace, higher intensity users, children, and some racial/ethnic minorities are expected to have much higher exposures than those at the population median (Arcury et al. 2018; McKelvey et al. 2011; Nguyen et al. 2019, 2020; Trowbridge et al. 2020).

Based on our review of the HT-H295R assay data for E2 and P4 as well as analyses published by Haggard et al. (2018, 2019) and Karmaus et al. (2016), the HT-H295R assay appears to be a sensitive and robust HT assay that can measure chemical effects throughout the steroidogenic pathway using cells that are able to express the enzymes necessary to synthesize four major classes of hormones, including progestogens, corticosteroids, androgens, and estrogens. However, given that in vitro testing approaches can be vulnerable to interferences and other artifacts, it would be useful to supplement the HT-H295R assay with additional approaches for detecting effects on steroidogenesis.

Within the in vitro chemical screening programs Tox21 and ToxCast, two other assays—the Tox21 aromatase inhibition assay and the NovaScreen hCYP19A1 activation assay—relate to aromatase activity, which is one of the steroid biosynthesis mechanisms captured within the H295R assay and a mechanism that can influence E2 synthesis (Figure 1). However, neither of these assays provide insight into chemicals that increase aromatase activity, which can increase E2 synthesis. The Tox21 aromatase inhibition assay (Chen et al. 2015) only detects inhibition, so will not detect chemicals that increase E2. The cell-free NovaScreen hCYP19A1 activation assay only produced 3 positive results for aromatase activation, vs. 72 positives for aromatase inhibition, so it may not be sensitive to detecting activation (Williams et al. 2017).

We considered whether activity in steroidogenesis pathways might be accompanied by other endocrine activity, for example, activation at the ER. However, these activities appear to be fairly independent as only 10 of the 182 E2-up chemicals and 15 of the 185 P4-up chemicals were considered active at the ER (Judson et al. 2015).

Although the low-throughput OECD H295R assay also measures effects on steroidogenesis, it is not as efficient in screening many chemicals at once and only identifies chemicals that alter E2 and testosterone, ignoring effects on other hormones. However, a comparison of the HT-H295R assay results with the OECD reference chemical interlaboratory results for the low-throughout H295R steroidogenesis assay, which is run without forskolin prestimulation, showed good reproducibility in testosterone and E2 responses, with the HT-H295R assay showing a sensitivity of 0.75, a specificity of 0.85 and an accuracy of 0.81 for increased E2 (Haggard et al. 2018). We also noted some reports of increased aromatase and increased E2 associated with neonicitinoid pesticides in another study that used the low-throughput H295R assay (Caron-Beaudoin et al. 2016), but this result was not clearly replicated in the HT-H295R assay.

Despite its strengths identifying chemicals that increased E2 and P4 in vitro, there are some important limitations in using the HT-H295R assay specifically and in vitro assays in general. For example, a requirement of the HT-H295R assay, and most of the ToxCast assays, is that chemicals are soluble in DMSO (Karmaus et al. 2016), so chemicals that do not meet this requirement cannot be tested. In addition, the assays are not set up to test volatile chemicals (Thomas et al. 2019). As a result, it is not known whether steroidogenesis is a pathway affected by common solvents that increase mammary gland tumor incidence but may have been considered too volatile to test with the HT-H295R assay, such as methylene chloride, dichloroethane, di-chloropropane, carbon tetrachloride, and benzene (Rudel et al. 2014). In addition, because of prestimulation of steroidogenesis with forskolin, the HT-H295R assay appears to be less sensitive to E2 increases, and this may result in false-negatives (Haggard et al. 2018). Consistent with this finding, we observed much more robust MFC responses for P4 compared with E2 (Figure 2). We also noted that compared with the ANOVA-based data analysis presented by Haggard et al. (2018), the tcpl returned a much smaller fraction of positive hit-calls for E2-up chemicals compared with P4-up chemicals: Only 25% of E2-up chemicals were also active in tcpl, compared with 60% of P4-up. This is possibly due to the low responsiveness for E2 increases, so data analysis choices are influential. In fact, the differences between hit-calls for the HT-H295R assay using tcpl (as in ToxCast) vs. ANOVA (as used by Haggard et al. 2018, 2019) raises concerns about insensitivity of the tcpl approach. For example, two of the most potent/effective E2-up chemicals based on the ANOVA in the reports by Haggard et al. (2018, 2019), cyfluthrin and hexythiazox, were not designated active for E2-up based on the tcpl analysis, yet we found that both of these have studies that reported mammary tumors (U.S. EPA 2011, 2012a). Of particular concern is the tcpl practice of using the two lowest doses, rather than the control, to establish a baseline and looking for a response above that at higher doses.

There are also limitations to using data from in vitro assays such as the HT-H295R assay to predict in vivo outcomes (e.g., an increase in E2 or P4 in a live animal or a downstream effect). For instance, although in vitro assays can identify activity of a parent chemical, in vivo this activity can be modified by the effects of absorption, distribution, metabolism, and excretion, which can substantially increase or decrease activity. Without metabolic capability, an important limitation is the potential for false-negatives because the H295R cells cannot produce an active metabolite (Hecker et al. 2011). The HT-H295R assay also cannot demonstrate how a chemical may affect steroidogenesis in other endocrine tissues such as the hypothalamus or the pituitary gland, including effects on the release of follicle stimulating hormone or luteinizing hormone (Hecker et al. 2011). Chemicals may also have multiple biological effects in vivo that may overshadow the effects of increased hormone levels or which make it hard to separate effects due to steroidogenesis vs. receptor agonism. For example, in our literature review of chemicals with high potency/efficacy, mifepristone both increased P4 and is considered a strong PR antagonist (Chwalisz 1994; Ho et al. 2002; Kim et al. 2020), thus the effects of increased P4 levels on the PR may be blocked. HPTE increased E2, but is considered a strong ER agonist (Hewitt and Korach 2011), so it was difficult to tell which of the relevant outcomes observed were related to ER agonism vs. steroidogenesis.

Another uncertainty in generalizing from the H295R assay to steroidogenesis effects in other tissues or organisms is that cellular responses to chemicals will vary. For example, the aromatase promoter profile in H295R cells differs from that of breast adipose tissue or gonadal tissue (Bulun et al. 2004; Caron-Beaudoin et al. 2016), so it is not known whether mammary tissue responses to aromatase inducers will be the same as those observed in the H295R assay, which uses adrenocortical carcinoma cells. However, there is evidence to support the idea that some of the same promoters—PII and I.3—that activate aromatase transcript genes and activity in HR29R cells (Caron-Beaudoin et al. 2016; Watanabe and Nakajin 2004) are also active in adipose tissue (although used minimally) and are responsible for upregulating aromatase in breast adipose tissue of women with breast cancer (Agarwal et al. 1996; To et al. 2015; Zhao et al. 2016). Because aromatase is the enzyme that converts androgens to estrogens, higher levels of aromatase are predicted to lead to more E2 synthesis (Simpson 2003; Zhao et al. 2016).

Finally, in vitro assays are typically run in a cell line or primary cell culture derived from a single individual, so responses do not reflect genetic susceptibility or variation that exists in populations, which may make some individuals more prone or resistant to a given outcome. For example, women with a polymorphism that increased CYP19A1 gene expression (which encodes for aromatase) had a worse prognosis after breast cancer (Friesenhengst et al. 2018).

Despite the limitations of in vitro testing, these data can inform weight of evidence discussions about carcinogenicity or endocrine disruption by suggesting plausible mechanisms of action, and can identify new chemicals that alter steroidogenesis and so may pose a hazard. Although it would be useful to observe the effects of these chemicals on hormone synthesis in human breast tissue or gonads, this is difficult to do in practice. A crucial gap in screening programs for endocrine disruptors is that there are no adequate in vivo tests sensitive to local or systemic increases of E2 or P4. The rodent uterotrophic assay has in some cases mistakenly been assumed to be a useful screen for E2 activity in vivo, but it does not capture effects on estrogen synthesis if it is run in ovariectomized animals (OECD 2018a). For example, the U.S. EPA concluded that $o$-phenylphenol (an E2-up chemical) did not have sufficient evidence for an interaction with the estrogen pathway based on a negative result in a uterotrophic assay (U.S. EPA 2015).

Overall, the HT-H295R assay, with the ANOVA-based approach for assigning a hit-call, currently provides the best available in vitro prediction of effects on E2 and P4 synthesis in a HT format. New coculture models may provide even better breast cancer–relevant models for testing EDCs, for example, by increasing the sensitivity of E2-up responses, integrating measures of breast cell proliferation and aromatase activation, and by incorporating more relevant aromatase promoters (Yancu et al. 2020).

We identified 296 chemicals of particular concern for increasing breast cancer risk on the basis of their ability to increase the synthesis of E2 (182) and/or P4 (185). These chemicals were more likely to be carcinogens or repro/dev toxicants than to not cause those types of effects (Figure 3). In vivo effects of the 18 most potent and efficacious E2- and P4-up chemicals included increased hormone concentrations, mammary gland effects (including tumors), and other repro/dev toxicity. These examples of in vivo effects may be consistent with local or systemic increases in E2 and P4 levels and add confidence to the continued use of this assay as a screening tool for identifying chemicals that may increase the risk of breast cancer. For example, of 45 chemicals previously reported to cause mammary tumors or other mammary effects and that were also tested in the HT-H295R steroidogenesis assay, 29 increased E2 or P4, including the well-known mammary gland carcinogen DMBA (Table 1). Among these are 6 chemicals that we classified as borderline active because of their relatively low potency or efficacy in the H295R assay, which suggests that even these may have important in vivo effects. E2- and P4-up chemicals included pesticides, consumer product ingredients, food additives, and drinking water contaminants.

Many of the E2- and P4-up chemicals have not been evaluated in vivo or have inadequate data to evaluate for carcinogenicity or repro/dev potential, so we submit that follow-up studies of these chemicals is a priority, especially because they can be found in common exposure sources. Fifty-three active E2-up chemicals and 59 active P4-up chemicals that are currently used in consumer products, food, pesticides, or drugs have inadequate information to assess carcinogenic potential, so these are important priorities for additional study and exposure reduction (Figure 3; Excel Tables S1 and S2). Future studies on steroidogenic effects should also prioritize the chemicals that increased E2 or P4 at the MTC but which were not tested in the CR format because they did not affect four or more hormones, as well as the chemicals that we considered to be borderline active because they did not meet our predefined potency and efficacy criteria (Excel Tables S1 and S2).

Where in vivo assessments exist already, we propose that another priority for future assessment is to carefully review existing toxicity studies and associated risk assessments for the E2- or P4-up chemicals to note whether there are related effects that are reported, reported and dismissed, or were not ascertained because the appropriate end points were not measured. Based on the in vivo studies we reviewed (Table 2), we found incomplete reporting of outcomes related to steroidogenesis, especially for mammary gland effects. For example, mammary gland histopathology was altered by 2,4-DCP at all doses in all generations in a multigenerational rodent reproduction and development study (Aoyama et al. 2005) but despite these changes, an NOAEL was set at the mid dose, ignoring these and other estrogenic effects. In other cases, in vivo effects plausibly related to the mechanism of increased E2 or P4 were also dismissed in regulatory evaluations and were regarded as not statistically significant or not biologically relevant (Table 2) (Pesticide residues in food 2018; U.S. EPA 2001b, 2001c, 2019a). Studies often do not include end points that would be sensitive to increases in E2 or P4 synthesis, including mammary gland assessments (Makris 2011; Rudel et al. 2011).

There are opportunities to increase the ability to detect mammary gland effects in toxicology studies, and previous research by ourselves and others provides direction (Davis and Fenton 2013; Makris 2011; Rudel et al. 2011). Specifically, classical transverse or cross-sectioning of the mammary gland that includes the skin provides insufficient mammary tissue to detect hyperplasia or inflammation, and so frontal or longitudinal sectioning is recommended instead to increase by 8–10 times the mammary epithelium present for detection of morphological changes, hyperplasia, and inflammation that might reflect locally increased hormone action (Davis and Fenton 2013). These methods can be used in repeated-dose toxicity studies to detect mammary gland effects (Tucker et al. 2018). Developmental studies rarely evaluate the mammary gland, and it is not a required end point in test guideline studies, despite the fact that mammary gland development—as revealed by mammary gland whole mounts—has been shown to be altered following developmental exposure to some EDCs (reviewed by Rudel et al. 2011). Similarly, it is possible that effects on lactation could be detected with greater sensitivity than in current study designs, where reduced pup BW may indicate problems with lactation (Makris 2011). In addition, guideline toxicity studies rarely use sensitive measures of circulating E2 or P4 concentrations and never measure localized concentrations or aromatase activity in mammary tissue (OECD 2018b; Schwarzman et al. 2015), so effects of chemicals that increase these hormones are likely missed. Thus, in guideline studies currently, the ability to detect effects on mammary gland is limited. As a result of these limitations in toxicology study design and interpretation, there may be underreporting of relevant in vivo effects, including in authoritative chemical risk evaluations and databases, such as the U.S. EPA pesticide registrations and ToxValDB (Judson 2019; Williams et al. 2017), which are commonly used to identify effects of chemicals. The review of existing toxicity studies would be facilitated if the many studies submitted to the U.S. EPA for pesticide registrations were readily accessible and searchable—currently Freedom of Information Act requests are required to access many documents (https://iaspub.epa.gov/apex/pesticides/f?p=CHEMICALSEARCH:1:0::NO:1).

Additional research is needed to better understand dose–response relationships between these chemicals and effects such as serum or tissue hormone concentrations and altered mammary gland structure, development, histopathology, lactation, and cancer. The dose–response implications of co-exposures to chemicals that activate multiple important cancer pathways should also be investigated given that we expect strong interactions between agents that induce genomic damage and those that increase cell proliferation (Helm and Rudel 2020), such as these E2- and P4-up chemicals. Previous research has shown similar interactions between inflammation and DNA damage, for example, with a much higher number of mutated cells when the two stimuli are given together (Kiraly et al. 2015). It is also important that these chemicals be tested during different windows of susceptibility (WOS)—such as the prenatal, pubertal, pregnancy, and menopausal transitions—because exposures to these hormones during WOS may be especially consequential and important for later mammary carcinogenesis or lactation effects. For example, studies have shown that exposure of rodents to EDCs—including E2, P4, or both—in critical time periods of mammary gland development leads to changes in later development, structure, and function; these alterations may in turn increase susceptibility to breast cancer (Gore et al. 2015; Rudel et al. 2011).

The data presented here also support reconsidering the uses of these chemicals in light of the many common exposure sources, including pesticides and food and consumer products, and the potential for cumulative exposures, given that people are usually exposed to chemicals in mixtures. It is possible that E2- and P4-up chemicals may act additively with each other and with other chemicals active at the ER or in other biological pathways because there is ample evidence that EDCs can act cumulatively, even by varied mechanisms (Bois et al. 2017; National Research Council 2008). In our opinion, risk assessment for combined exposures is a priority, although it is rarely performed. In one notable mixtures risk assessment model, Bois et al. (2017) considered combined exposures to environmental chemicals that inhibit aromatase to predict the chance of effects on ovulation and fertility (Bois et al. 2017). This study found that although predicted exposures to individual chemicals were not expected to be high enough to cause effects on ovulation, simulations of exposures to mixtures of these chemicals predicted there would be effects on ovulation in a portion of the general population (Bois et al. 2017). Our findings in this study suggest that it is a priority to extend these risk assessment approaches to mixtures of chemicals that increase E2 and P4, especially in connection with their potential effects on breast cancer. In addition, existing risk assessments likely need to be modified to better consider cumulative exposures. Additional work is also needed to develop biomonitoring methods to guide exposure reduction and to use in epidemiologic studies. We propose that advances in analytical chemistry and data analysis for the epidemiologic study of mixtures as well as efforts to develop functional assays for integrated measurements of exposure, and to better understand the biological pathways, will contribute to this effort.

This analysis presents several hundred endocrine active chemicals that should be considered as potential risk factors for breast cancer on the basis of their ability to increase E2 or P4 synthesis and demonstrates use of the HT-H295R in vitro steroidogenesis screening assay for chemical screening. We demonstrate that many of these chemicals are likely or possible carcinogens and/or repro/dev toxicants and that about half of mammary carcinogens that were tested increased E2 or P4 synthesis. In addition, we found that toxicity testing is missing or incomplete for many chemicals that increased E2 and P4 steroidogenesis, and assessment of mammary gland effects is especially limited. Exposure to many of these chemicals is likely ubiquitous, based on exposure prediction models. We conclude that these EDCs are priorities for biomonitoring and exposure reduction as well as for additional study to better understand potential effects on breast cancer and other reproductive and developmental effects.