Phenol and Phthalate Effects on Thyroid Hormone Levels during Pregnancy: Relying on In Vitro Assays and Adverse Outcome Pathways to Inform an Epidemiological Analysis

The prospective SEPAGES [Suivi de l’Exposition à la Pollution Atmosphérique durant la Grossesse et Effets sur la Santé (Assessment of air pollution exposure during pregnancy and effect on health)] cohort recruited 484 pregnant women from eight obstetrical ultrasonography practices located in Grenoble area of France, between July 2014 and July 2017.¹⁸ Women were included based on the following criteria: $\ge 18\mathrm{y}$ of age, being pregnant for 19 gestational weeks or less, having a singleton pregnancy, residing in the study area, and planning to give birth in one of the four maternity clinics from the Grenoble area that were near the SEPAGES biobank.

Ethical agreements were obtained from the Comité de Protection des Personnes Sud – Est V (CPP) and the Comité Nationale de l’Informatique et des Libertés (CNIL), the French data privacy institution. All participating women gave written consent.

This analysis was restricted to the 437 pregnant women who did not report taking medication for any thyroid diseases (questionnaire completed during the first trimester that specifically asked about thyroid disorders) and had blood and urine samples collected during pregnancy (See flowchart in Supplemental Material, Figure S1).

Urine samples were collected over a week in the second trimester (median 17.7 gestational weeks (GW); 5th and 95th percentiles: 14.4, 20.0 wk, respectively), during which time women were requested to collect three spot urine samples per day (in the morning, midday, and evening). Samples were collected in $60\mathrm{mL}$ polypropylene tubes and stored in the participants’ freezer ($-20°\mathrm{C}$). At the end of the collection week, samples were transported on ice by a study fieldworker to the certified biobank of Grenoble University Hospital (bb-0033-00069). Samples were thawed overnight at 4°C, and for each subject a pool of the same volume of all the spots collected over the week were made following a previously validated protocol.¹⁹ Although assessments of biomarker concentrations did not formally account for urine dilution of individual samples, biomarker concentrations assessed in such equal volume pools have been shown to correlate well with those assessed in a pool of all urine volume collected over 24 h or a week.²⁰ For each woman, this pool was aliquoted and stored at $-80°\mathrm{C}$. Daily pools (equal volume pools of all samples collected over a day) were also made.

Pools of urine samples collected over a week were sent on dry ice with a temperature sensor to the Norwegian Institute of Public Health (NIPH), where measurements of phenol and phthalate metabolite concentrations were carried out (see Table 1 for a detailed list of biomarkers assessed). Phthalate and di(isononyl)cyclohexane-1,2-dicarboxylate (DINCH) biomarkers were analyzed and quantified using high-performance liquid chromatography coupled to mass spectrometry (HPLC-MS-MS).²¹ Phenols were analyzed and quantified using ultra high-performance liquid chromatography coupled to mass spectrometry (UPLC-MS-MS).²² The free and conjugated forms of phenol biomarkers were preliminarily measured in samples from 50 women. These preliminary measurements did not suggest external contamination,²³ so for the remaining participating women, we relied on analysis of the total form ($\text{free}+\text{conjugated}$). Bisphenols AF, B, F, and triclocarban were detected in $<\mathrm{5\%}$ of the pooled samples and were not considered in our analysis.

Nonfasting maternal blood was collected by trained SEPAGES fieldworkers during a study visit at the participants’ homes. For 85% of the women, blood was collected at the end of the urine collection week, and blood for the other 15% was collected several weeks after urine (median 9 wk, 5th and 95th percentiles: 6, 12, respectively). After collection, samples were transported on ice to the biobank of Grenoble University Hospital; there, blood was processed, and serum aliquots were stored at $-80°\mathrm{C}$.

TSH was quantified in maternal sera by LOCI Chemiluminescence on Dimension Vista analyzer (Siemens).²⁴ Serum concentrations of protein-bound and free T4 and 3, 5, 3′-triiodothyronine (T3) were quantified by RIA-Gnost (CisBio Bioassays). Maternal total T3 and T4 were obtained by summing the free and protein-bound concentrations. The ratio of total T3 to T4, an indicator of T4 deiodination into the bioactive T3 form, was then calculated. Selenium, an essential micronutrient required for biosynthesis of selenoproteins involved in the peripheral conversion of free T4 to free T3, as well as being vital antioxidants in the thyroid gland, was also measured in maternal sera using inductively coupled plasma mass spectrometry (ICP-MS).²⁵

Iodine, an essential element in the synthesis of thyroid hormones, was measured in daily pooled samples of maternal urine using inductively coupled plasma mass spectrometry (ICP-MS).²⁶

The research effort herein relied on existing AOP networks that map the causal and putative MIEs in the thyroid pathway that have been demonstrated or hypothesized to be the chemical targets. To this end, there have been a number of efforts in the development and evolution of the thyroid AOP network, including an effort recently by Noyes et al.⁷ Several ToxCast/Tox21 in vitro HTS assays were evaluated for chemical interactions with MIEs in the thyroid pathway.¹⁷ MIEs targeted are highlighted in the Supplemental Material, Figure S2 (thick, green border in left-hand column) and include those involved in thyroid hormone biosynthesis in the thyroid gland (${\mathrm{Na}}^{+}{/\mathrm{I}}^{-}$ symporter [NIS,^27,28 thyroperoxidase (TPO²⁹)]; receptor-based interactions [thyroid hormone (TR) receptors ($\text{TR}\mathrm{\alpha }$, $\text{TR}\mathrm{\beta }$³⁰) thyroid stimulating hormone (TSH) receptor,³¹ and thyrotropin releasing hormone (TRH) receptor]; thyroid hormone peripheral tissue metabolism [iodothyronine deiodinases (DIO1, DIO2, DIO3^32,33)], and activation of hepatic T4 catabolism [e.g., constitutive androstane receptor (CAR), pregnane X receptor (PXR), uridine diphosphate glucuronosyl transferases (UDPGTs)]. The MIEs in the thyroid pathway have been described in detail elsewhere and readers are referred to reviews for additional background.^6,7,34

ToxCast/Tox21 database outputs are typically presented as positive ($\text{hitcall}=1$) or negative ($\text{hitcall}=0$) with associated half-maximal activity concentration (${\text{AC}}_{50}$) values and efficacy values (cutoff and maximum responses) for bioactive substances. To derive point estimates, raw chemical screening data in assay tests are processed through the ToxCast data analysis pipeline involving several steps in data normalization and dose–response modeling. ToxCast applies three dose–response models in its data evaluations: a) constant model with a constant value of zero response and only one parameter, the scale term; b) Hill model that is a constrained three parameter model; and c) Gain-Loss model that is a constrained five-parameter model. For assay results to be considered bioactive, the modeled concentration–response curves must meet three criteria: a) Hill or Gain-Loss curve fit models must be the winning models; b) the modeled curve fit top must exceed the efficacy cutoff for at least one dose; and c) the median response must exceed the efficacy cutoff. Automated flags identify potentially anomalous outputs. To further limit potential false positives, we also considered only ToxCast positive hitcalls that matched the following criteria: a) concentration–response curves had no more than three flags; b) response curve fits displayed a sigmoidal shape; c) more than one data point was above the efficacy cutoff; and d) data for low concentrations were present (i.e., ${\text{AC}}_{50}$ should not have been extrapolated^35,36). Concentration–response curves and associated data were extracted from the ToxCast chemistry dashboard for each chemical.³⁷ Those not meeting these criteria and excluded from the analysis are displayed in Supplemental Material, Figure S3.

Exposure biomarker concentrations below the limit of detection (LOD) and between the limit of detection and the limit of quantification (LOQ) were singly imputed by values randomly selected between 0 and LOD and between LOD and LOQ, respectively, based on the estimated underlying distribution.^38,39 To limit the impact of between-subject variations in conditions related to urine processing (i.e., sample transport time from participant’s home to the biobank, time during which the individual samples were thawed at 4°C during the pooling procedure) and assay (analytical batches), we standardized the measured phenol and phthalate metabolite concentrations when needed. We first estimated the associations between each biomarker concentration assessed in pools (natural log-transformed) and the factors above-mentioned using adjusted linear regression. If processing/assay conditions were identified as associated with the biomarker urine concentrations ($p<0.2$), we then used the measured biomarker concentrations and the estimated effects of the processing/assay conditions to predict standardized concentrations (i.e., concentrations that would have been observed if all samples had been processed under the same conditions and assayed in the same analytical batch).^40,41 We used these standardized concentrations in our statistical analyses.

To limit the impact of extreme values, thyroid hormone concentrations, as well as the T3/T4 ratio, were ln-transformed. Phenol and phthalate metabolite concentrations were considered as continuous (ln-transformed) variables, as well as categorized into tertiles, except for compounds detected in $<\mathrm{70\%}$ of the samples that were dichotomized (detected/undetected). We computed the molar sum of metabolites from the same parent [e.g., di(2-ethylhexyl) phthalate (DEHP), diisononyl phthalate (DiNP) and DINCH parents, Table 1]. We used adjusted linear regression to assess the associations between each selected phenol and phthalate biomarker and thyroid hormone (TSH, T4, T3) or the T3/T4 ratio.

Adjustment factors for our statistical analyses were selected a priori and included variables likely to be common causes of both the exposures and the thyroid hormones without being likely consequences thereof and factors that were possible predictors of the thyroid hormones only,^42,43 such as maternal age (quadratic terms), body mass index (BMI, kilograms per square meter) before pregnancy (continuous), education level (three categories: $\le 2\mathrm{y}$ after high school, 3–4 y after high school, and postgraduate or $\ge 5\mathrm{y}$ after high school), maternal smoking during the first trimester of pregnancy (Yes/No), parity (nulliparous and parous), gestational age at serum collection (continuous, weeks), hour of serum collection (categorized), maternal urinary iodine concentrations (ln-transformed, micrograms per liter), and selenium concentrations in sera during pregnancy (tertiles, micromoles per liter). Models were also adjusted for analytical batch for all hormones except TSH, for which no batch effect was detected. A directed acyclic graph representing these relationships is displayed in Supplemental Material, Figure S4. Missing values for covariates were handled using multiple imputation (20 imputed data sets). Effect estimates were reported as percentage change in the hormone concentrations for a doubling of urinary biomarker concentrations. For bisphenol S (BPS) and butylparaben, which were dichotomized, effect estimates represent percentage change in hormone concentrations between those with detected and undetected concentrations.

Betas were expressed as percentage change in outcomes for each doubling of urinary biomarker concentrations except for butylparaben and BPS, for which effect estimates represent percentage change in outcomes between those with undetected and detected urinary concentrations. These percentage changes were obtained from the original betas using the following formulae: $(2^{\mathrm{\beta }}-1)\times 100$ for continuous exposures and $\left(e^{\mathrm{\beta }}-1\right)\times 100$ for categorial exposures (butylparaben and BPS). We report in the “Results” section all associations with $p<0.05$. Associations with $p$--values between 0.05 and 0.10 were also reported as a trend when the biomarker was associated ($p<0.05$) with another hormone.

Iodine status might modify the associations between chemical exposure and thyroid hormone levels.⁴⁴ For this reason, in sensitivity analyses, we explored modification by adding an interaction term between iodine levels and urinary phenol and phthalate metabolite concentrations. When an effect modification was suggested (p_interaction $<0.10$), stratified analyses on iodine status were performed. For this analysis, we dichotomized iodine levels according to the World Health Organization (WHO) threshold for iodine deficiency in pregnant women ($150\mathrm{\mu }\mathrm{g}/\mathrm{L}$⁴⁵).

Although serum samples were collected at the end of the urine collection week for most women, serum samples were collected for 15% of participants several weeks after the urine samples (median 9 wk, 5th and 95th percentiles: 6, 12, respectively). For this reason, we ran an additional analysis restricted to the 373 women for whom serum samples were collected at the end of the urine collection week.

For the main models, we plotted residuals to visually identify influential individuals. If any, we run a sensitivity analysis without these individuals to assess result’s robustness.

Finally, we estimated the joint effects of the selected phenols and phthalates on each hormone using adjusted Bayesian Kernel Machine Regressions (BKMR; R package: bkmr). Such modeling allows estimation of mixture effect and identification of the important components of the mixture. It also accommodates for nonlinear relationships and correlated exposures.^46,47 Categorical exposures (BPS and butylparaben) were not considered in this analysis. Each exposure biomarker concentration was standardized [i.e., divided by their standard deviation (SD)]. For each model, we ran 50,000 iterations, dropped the first 25,000, and kept every fifth iteration among the last 25,000 for inference. The overall mixture effect was given by a figure showing the expected change in hormone concentration with concomitant increase quantiles of all exposure biomarkers, relative to when they are fixed at their 25th percentile. When this graph was suggestive of an effect of the mixture, we provided the posterior inclusion probability (PIP) for each biomarker and plotted the estimated effect of an increase from the 25th to 75th percentile in a single biomarker concentration when all other biomarker concentrations were fixed at either their 25th, 50th, or 75th percentiles.

Analyses were carried out with Stata/SE (version 15.1; StataCorp LLC) and R (version 4.0.4; R Development Core Team).

Median of maternal age at recruitment was 32.5 y old (Table 2). The majority (55%) already had a child and were highly educated (56% had pursued education for $\ge 5\mathrm{y}$ after high school). Most (74%) had a BMI in normal range ($18.5–25{\mathrm{kg}/\mathrm{m}}^2$), and 6% reported smoking during their first trimester of the pregnancy (after they knew they were pregnant). Median of gestational age at birth was 40 wk, and 54% of the infants were males. Median (percentiles: 5th and 95th) number of samples in each pool was 21 (17 and 21).

Phenol and phthalate metabolite concentrations have been described elsewhere.^23,48 Briefly, bisphenols AF, B, F, and triclocarban were detected in $<\mathrm{5\%}$ of the pooled samples and were not considered in our analysis. Except for butylparaben and BPS detected in 25% of the samples, frequencies of detection for the other compounds were above 83% (Table 3).

For TSH, 95% of the women ($n=414$) were within the reference range defined for pregnant women by the French Health Authority (0.358 to $2.500\text{\hspace{0.17em}mUI}/\mathrm{L}$ for samples collected during the first trimester and $0.358\text{\hspace{0.17em}to\hspace{0.17em}}3.000\text{\hspace{0.17em}mUI}/\mathrm{L}$ for samples collected later in pregnancy). Median iodine urinary concentration assessed in urine daily pools was relatively low ($89\mathrm{\mu }\mathrm{g}/\mathrm{L}$; Table 4), and 80% of the women had a iodine urinary concentration lower than the WHO guideline for pregnant women.⁴⁵

Three compounds (methylparaben, ethylparaben, and diethyl phthalate) were not bioactive at any of the thyroid-related MIEs evaluated and so were not included in the statistical analysis (Table 5; Supplemental Material, Excel Table S1). Among the phenols and phthalates assessed in SEPAGES, one phthalate, [bis(2-propylheptyl) phthalate (DPHP)] was not screened for thyroid activity in ToxCast. We nevertheless retained its metabolite in the tier-2 statistical analysis, based on thyroid toxicity (i.e., thyroid gland hypertrophy/hyperplasia) reported in adult male rats in a two-generation reproductive toxicity study.⁴⁹

This selection step allowed us to restrict our set of 16 (13 individual compounds and 3 molar sums) initially considered chemicals to 13, leading to a reduction in the number of tests performed in our main analysis (Table 6) by approximatively 19%.

Propylparaben was negatively associated with free T3 and the T3/T4 ratio. A similar pattern was observed for butylparaben in our analysis restricted to the 373 women for whom serum samples were collected at the end of the urine collection week. Propylparaben was also identified as a major contributor of the negative association observed between the mixture and T3/T4 ratio. Among the few studies that have explored associations between parabens and thyroid hormone concentrations during pregnancy,^50,51 only one assessed the T3/T4 ratio and did not report association for propylparaben and butylparaben.⁵⁰ To our knowledge, no study assessed free T3, limiting comparison with our results. Although studies in animal are also scarce, exposure to butylparaben has been shown to increase TPO activity and reduce DIO activity.⁵² Regarding other potential mechanisms by which parabens may affect thyroid hormone homeostasis, the in vitro HTS data indicated activation of xenobiotic nuclear receptors (e.g., CAR, PXR) regulating expression of genes encoding metabolizing enzymes, which could in turn enhance thyroid hormone (Table 5).

When categorized in tertiles, BPA was negatively associated with TSH. This result was in line with those of two previous studies,^53,54 whereas for one the association was seen only among women with a prepregnancy $\text{BMI}>23{\mathrm{kg}/\mathrm{m}}^2$.⁵⁴ Five other human studies reported null associations with TSH.^50,55–58 We did not observe any association between BPA and the other hormones assessed nor with the T3/T4 ratio, whereas a few previous studies did.^50,55,57 Changes in thyroid hormone concentrations have also been reported in pregnant females (and/or their offspring) following low-dose exposures in experimental animal models.^59–61 These results, along with ToxCast data indicating BPA bioactivity on several relevant MIEs in the thyroid pathway, strengthen the overall body of evidence of its potential to perturb thyroid hormone signaling during sensitive developmental periods.

We did not observe any association for BPS. One study, with a bigger sample size and a higher frequency of detection than ours, reported a positive association between BPS and total T4.⁵⁶

We observed a non-monotonic decrease in TSH concentrations with increased triclosan urinary concentrations. Wang et al. also reported a U-shape association with TSH,⁶² whereas other studies report no association with TSH^50,51,57; however, only one has explored nonmonotonic associations.⁵⁷ Mechanisms by which triclosan may affect thyroid hormone homeostasis include inhibiting DIO, TPO, and NIS, binding ${\text{TR}}_{\mathrm{\beta }}$, and activating xenobiotic nuclear receptors (Table 5), which generally aligns with other mechanistic evidence.^63–68 Additionally, an in vitro screening assay indicated some capacity for triclosan to inhibit iodotyrosine deiodinase involved in iodide recycling in the thyroid gland.⁶⁹

MBzP, a metabolite of BBzP, was positively associated with total T4 and negatively with the T3/T4 ratio. MBzP was also one of the major contributors of the negative association observed between the mixture and T3/T4 ratio. To the best of our knowledge, associations with the T3/T4 ratio has been evaluated in only two studies.^10,70 Consistent with our results, one reported a negative association,⁷⁰ whereas the other did not.¹⁰ Our results for total T4 are not aligned with previous studies that reported a negative⁹ or no association with T4.^10,43,70 BBzP can up-regulate the transcriptional activity of NIS.⁷¹ Consistent with results herein, it is possible up-regulation of NIS may prompt increased iodine intake into the thyroid and increased thyroid hormone production. The fact that we observed a positive association only with total T4 and not total T3 might partly be related to difference in half-lives across hormones with T3 (1 d) having shorter half-life than T4 (5 to 7 d).

Effect modification by iodine concentrations were observed for two phthalates, $\Sigma \text{DEHP}$ and MnBP. Only Villanger et al. examined interactions with iodine concentrations.⁴⁴ That study relied on a factor analysis and reported interactions between iodine status and the factor containing MnBP along with MiBP and MBzP. However, associations were observed for TSH and total and free T4, whereas we observed associations with total and free T3. Villanger et al. did not report interactions with the factor containing $\Sigma \text{DEHP}$.

In comparison with an agnostic approach, the hypothesis-driven selection of phenols and phthalates using the thyroid AOP network and in vitro HTS assays allowed for the reduction of the number of statistical tests performed and limited FWER. However, the complexity of the thyroid system that includes tightly controlled compensatory signaling, and differences in pharmacokinetic parameters in vivo and in vitro, make predictions with in vitro screening data a challenge. Adaptive responses at higher levels of biological organization also may impart some protection to chemicals, with in vitro assays typically predicting effects at lower dose levels (more sensitive). This complexity may explain why the direction of associations observed in our study are not always aligned with those described in AOPs for the targeted MIEs. In addition, although several of the phthalates and phenols evaluated in this study activated xenobiotic nuclear receptors (and the mechanistic literature suggests a role for up-regulated T4 catabolism pathways), the performance of these and other assays in predicting effects at putative targets in the thyroid hormone conjugation and excretion pathway is an area of ongoing study. In addition, the ToxCast in vitro HTS assays rely on testing individual chemicals, so the potential role of chemical mixtures in affecting thyroid regulation was not considered in chemical selection. Finally, it is possible that a direct chemical effect on another biological target (e.g., immune system/inflammation⁷²) may prompt secondary effects on thyroid regulation that in turn elicit effects elsewhere in the body, depending on the life stage, timing, severity, and duration of the hormonal disruption. By filtering for compounds with bioactivity at thyroid MIEs, we may have excluded chemicals that elicit thyroid disruption by secondary pathways. Nonetheless, although beyond the scope of this study, integration of biological pathway models such as the AOP network approaches used herein provide a framework from which to begin capturing some of these more complex modes of action and effects of thyroid hormone reductions in pregnant women exposed to environmental chemicals. The approach herein allowed for the selection of chemicals with increased biological plausibility for interacting with MIEs in the thyroid pathway and thus provided a useful application in mixture and exposome studies that test larger sets of chemicals.

We relied on within-subject pools of repeated urine samples collected over a week to assess exposure. This approach is of importance for compounds with high intra-individual variability, such as BPA and DEHP metabolites (intraclass correlation coefficients of about 0.2–0.3^11–13). For such compounds, reliance on multiple urine samples during relevant time windows should lead to decreased measurement error and, for a given sample size, increased power.¹⁵ SEPAGES correlation coefficients between two non-consecutive weeks of pregnancy was still relatively low for some compounds,^23,48 highlighting the need to collect repeated biospecimens in sensitive time windows. For most women (85%), urine samples were collected the week preceding blood draws, allowing for evaluation of short-term effects of phenols and phthalates. In addition to thyroid hormone concentrations, we also explored associations with the T3/T4 ratio. Although clinically relevant, this indicator, which was negatively associated with MBzP, propylparaben, and butylparaben, as well as the mixture, has not been extensively studied in association with prenatal exposure to phenols and phthalates. In our study population, 80% of the women were below the WHO threshold for iodine deficiency. Such low iodine concentrations have been previously described among pregnant women from France^73,74 and other European countries.⁷⁵ Further studies should consider this essential element, given results here and among other researchers⁴⁴ that effects of some phthalates and phenols on the thyroid may be modulated by iodine status.

Although we accounted for many potential confounders, residual confounding cannot be ruled out. There are indeed other synthetic chemicals known to disrupt thyroid functioning that were not assessed in our study, and that may have confounded the observed associations (e.g., halogenated BPA compounds such as tetrabromobisphenol A). In addition, we did not assess thyroid antibodies, which might be important predictors of thyroid hormone concentrations. We also did not recruit women early enough to assess thyroid hormones during the first trimester of pregnancy, a key period during which the fetal thyroid gland is immature and the fetus is dependent on maternal sources of thyroid hormone. Despite our chemicals’ a priori selection, the number of associations tested was still high. We did not apply any formal correction for multiple comparisons. This and the fact that in vitro data may not always be predictive of in vivo biological events suggest that part of the associations we observed may still have resulted from chance findings and thus should be interpreted cautiously.

Relying on pools of multiple urine samples and on a novel hypothesis-driven method to reduce false positives and chance findings, we observed negative associations between several phenols and phthalates and TSH, free T3, and the T3/T4 ratio. Given widespread exposure to these compounds in the general population and the crucial role of thyroid hormones in development, the impact on fetal and child health might be substantial.