Individual and joint effects of phthalate metabolites on biomarkers of oxidative stress among pregnant women in Puerto Rico

Abstract

Exposures to phthalate compounds have been linked to adverse birth outcomes, potentially through oxidative stress mechanisms. We explored associations between mixtures of biomarkers of phthalate and phthalate replacement metabolites and oxidative stress using lipid peroxidation biomarker 8-iso-prostaglandin-F2α (8-iso-PGF2α). As 8-iso-PGF2α can be generated via both chemical (nonenzymatic) and enzymatic lipid peroxidation pathways, we calculated the ratio of 8-iso-PGF2α/prostaglandin F2α in an attempt to distinguish the potential contributions of the two pathways. Urinary biomarker measurements were taken from 775 pregnant women in the Puerto Rico Testsite for Exploring Contamination Threats (PROTECT) longitudinal birth cohort at up to three time points during gestation (16–20, 20–24, and 24–28 weeks gestation). Adaptive elastic net with pairwise linear interaction terms (adENET-I) was used to determine individual phthalate metabolites and phthalate interactions that were predictive of lipid oxidative stress biomarkers, and to subsequently create environmental risk scores (ERS) to represent weighted sums of phthalate exposure for each individual at each study visit. Repeated ERS were then used in linear mixed effects models to test for associations between biomarkers of phthalate mixtures and biomarkers of oxidative stress. We also used Bayesian kernel machine regression (BKMR) to explore nonlinearity and interactions between phthalate metabolites within the mixture. An increase from the first to fourth quartile of phthalate ERS derived from adENET-I was associated with a 96.7% increase (95% CI: 74.0, 122) in the hypothesized chemical fraction of 8-iso-PGF2α and a 268% increase (95% CI: 139, 465) in the hypothesized enzymatic fraction of 8-iso-PGF2α. BKMR analyses also suggested strong linear associations between the phthalate mixture and biomarkers of lipid oxidative stress. Various phthalates displayed nonlinear relationships with both chemical and enzymatic fractions of 8-iso-PGF2α, and we observed some evidence of interactions between metabolites in the mixture. In conclusion, exposure to phthalate mixtures was strongly associated with linear increases in biomarkers of lipid oxidative stress, and differences observed between hypothesized chemical and enzymatic lipid peroxidation outcomes highlight the need to critically assess pathways of 8-iso-PGF2α generation in relation to environmental exposures.

1. Introduction

Puerto Rico has historically been subject to extensive levels of environmental contamination. The karst aquifer system in the northern coastal region of Puerto Rico, which allows contaminated drinking water to travel freely over long distances (Padilla et al. 2011), creates additional concern for human exposures to contaminants. The elevated risk of human exposure to environmental toxicants on the island, coupled with the observation of higher incidence of adverse health outcomes such a preterm birth, have lead researchers to hypothesize that increased exposure to environmental chemicals is contributing to greater risk of adverse health outcomes (Cantonwine et al. 2014). One group of environmental toxicants of special interest is phthalates – a class of synthetic plasticizers used in manufacturing of consumer products such a plastic food packaging and personal care products (Hauser and Calafat 2005). Past research has provided compelling evidence that phthalate exposure can lead to deleterious physiological outcomes, particularly when exposure occurs during pregnancy. For example, numerous studies have observed associations between gestational phthalate exposure and increased risk of preterm delivery (Ferguson et al., 2019; Gao et al., 2019; Meeker et al., 2009; Watkins et al., 2016; Weinberger et al., 2014), however, the mechanisms by which this occurs are poorly understood.

Phthalates could exert their effects on pregnancy via changes in oxidative stress. Biomarkers of oxidative stress have been shown to be associated with reduced gestational age at birth (Ferguson et al., 2015a; Longini et al., 2007; Peter Stein et al., 2008) and other adverse pregnancy outcomes (Ferguson et al., 2017b; Kim et al., 2005). Preliminary work in the PROTECT cohort showed positive associations between single phthalate metabolites and other biomarkers of oxidative stress among a smaller sample of PROTECT women. Specifically, metabolites of phthalates including DEHP, DBP, and DiBP were associated with increases in 8-isoprostane and 8-hydroxydeoxyguanosine (8-OHdG) biomarkers (Ferguson et al. 2014). The lipid peroxidation product, urinary 8-isoprostane-prostaglandin-F2α (8-iso-PGF2α), is widely regarded as one of the best biomarkers for measuring oxidative stress because of its reliability and stability during pregnancy. In contrast to 8-OHdG, a commonly used biomarker of oxidative stress which occurs via DNA damage, 8-Iso-PGF2α is a stable oxidative stress measure and is specific to arachidonic acid peroxidation by reactive oxygen species (Roberts and Morrow 2000). It’s concentrations are also not influenced by dietary lipid intake (Richelle et al. 1999) and have been shown to be a consistent indication of oxidative damage relating to chemical exposures (Kadiiska et al., 2013,1998). However, 8-iso-PGF2α may not solely be a biomarker of lipid oxidative stress – it is not only generated by chemical (nonenzymatic) lipid peroxidation, but it is also produced by prostaglandin-endoperoxide synthase (PGHS)-mediated enzymatic lipid peroxidation, which is significantly induced in inflammation. Van’t Erve and colleagues have developed a computational method for determining the estimated fractions of 8-iso-PGF2α deriving from chemical lipid peroxidation and enzymatic lipid peroxidation (Van’t Erve et al. 2016). It is important to be able to disentangle the pathways being activated when interrogating the mechanisms of action of environmental toxicants such as phthalates.

Human exposure to phthalates occurs in complex mixtures. Various parent phthalate compounds may be used in a single product, individuals may consume numerous products containing different phthalates, parent compounds may be metabolized into various metabolites once inside the body, and multiple phthalates may impact the same adverse health outcomes and relevant pathways. Consequently, robust studies of phthalate exposure effects on health endpoints must consider phthalate exposures within a mixtures framework. Therefore, the goal of this study was to utilize the Puerto Rico Testsite for Exploring Contamination Threats (PROTECT) pregnancy cohort to interrogate the associations between a mixture of 16 metabolite biomarkers of phthalate exposure and 8-iso-PGF2α originating from chemical and enzymatic lipid peroxidation pathways. Further, we explored the potential for phthalate metabolites within the mixture to interact with one another in both linear and nonlinear ways on their associations with lipid oxidative stress and inflammation.

2. Methods

2.1. Study population

Women in the present study were part of the PROTECT birth cohort which aims to uncover potential environmental exposures contributing to incidence of preterm birth and other adverse birth outcomes on the island. Details on study methods have been previously described (Cantonwine et al. 2014). Briefly, pregnant women living in the Northern karst region of Puerto Rico were recruited from 2012 to 2017 from seven hospitals and prenatal clinics at 14 ± 2 weeks gestation. Eligible participants were 18–40 years old, had their first clinic visit before 20 weeks gestation, did not use oral contraceptives within 3 months of getting pregnant, did not use in vitro fertilization to get pregnant, and did not have any known medical or obstetric conditions. Participating women provided spot urine samples for analysis at three time points during pregnancy (median 18, 22, and 26 weeks gestation). Demographic information was collected from all participants at the first study visit. The present analysis included 775 women who had complete data on at least 1 exposure-outcome pair for at least one of the three study visits. This study was approved by the research and ethics committees of the University of Michigan School of Public Health, University of Puerto Rico, Northeastern University, and participating hospitals and clinics. All study participants provided full informed consent prior to participation.

2.2. Urinary biomarker measurements

All urine samples were analyzed for 15 phthalate metabolites: mono-2-ethylhexyl phthalate (MEHP), mono-2-ethyl-5-hydroxyhexyl phthalate (MEHHP), mono-2-ethyl-5-oxohexyl phthalate (MEOHP), mono-2-ethyl-5-carboxypentyl phthalate (MECPP), monoethyl phthalate (MEP), mono-n-butyl phthalate (MBP), monobenzyl phthalate (MBzP), mono-isobutyl phthalate (MiBP), mono-hydroxyisobutyl phthalate (MHiBP), mono-3-carboxypropyl phthalate (MCPP), mono carbox-yisononyl phthalate (MCNP), mono carboxyisooctyl phthalate (MCOP), mono-hydroxybutyl phthalate (MHBP), mono isononyl phthalate (MNP), and mono oxononyl phthalate (MONP). Four additional phthalate replacement metabolites were later added to the analytical panel: cyclohexane-1,2-dicarboxylic acid monohydroxy isononyl ester (MHiNCH), cyclohexane-1,2-dicarboxylic acid monocarboxy isooctyl ester (MCOCH), mono-2-ethyl-5-carboxypentyl terephthalate (MECPTP), and mono-2-ethyl-5-hydrohexyl terephthalate (MEHHTP). Samples were frozen at −80 °C and shipped over night on dry ice to the CDC for all analyses. Analyses were performed using solid phase extraction high-performance liquid chromatography-isotope dilution tandem mass spectrometry. Each analytical batch included 40 unknown samples, five reagent blanks, and two high- and two low-concentration quality control materials. All quality control materials were characterized by 60 repeated measurements in a 3 week period to define control limits for each phthalate metabolite. Further details on sample analysis and quality control are published elsewhere (Kato et al., 2005; Silva et al., 2007, 2019).

Stable isotope dilution gas chromatography-negative ion chemical ionization-mass spectrometry was used to measure free 8-isoprostane-prostaglandin-F_2α (8-iso-PGF2α), a widely used biomarker for lipid oxidative stress, and Prostaglandin-F_2α (PGF2α) at the Eicosanoid Core Laboratory at Vanderbilt University Medical Center (Nashville, TN). The method utilized an internal standard and was highly precise and accurate, giving 6% precision and 96% accuracy (Milne et al. 2007). The main metabolite of 8-iso-PGF2α, 2,3-dinor-5,6-dihydro-15- F_2t-iso-prostane, was also quantified utilizing the same method due to its potential as a superior biomarker of oxidative stress (Dorjgochoo et al. 2012). We also utilized the quantitative method proposed by Van’t Erve and colleagues to distinguish the absolute concentration of 8-iso-PGF2α hypothesized to be derived from chemical lipid peroxidation (absolute chemical lipid peroxidation; aCLP) versus prostaglandin-endoperoxide synthases (absolute prostaglandin h-synthase; aPGHS). Briefly, the method utilizes the ratio of 8-Iso-PGF2α to PGF2α produced, which markedly differs between chemical and enzymatic lipid peroxidation pathways, to quantitatively determine the contribution of each pathway to the total amount of measured 8-Iso-PGF2α. We have hypothesized these values to reflect the amount of the association with phthalates attributable to chemical lipid peroxidation and inflammation induced enzymatic lipid peroxidation, respectively (Van’t Erve et al. 2016).

All urinary biomarkers measured showed log-normal distributions and were natural-log transformed for all statistical analyses. Any metabolite measured below its limit of detection (LOD) was replaced by the LOD/√2 (Hornung and Reed 1990). Specific gravity was measured using a digital handheld refractometer (AtagoCo., Ltd., Tokyo, Japan) and utilized to adjust for differences in urinary dilution between samples. Specific gravity is likely a better indicator of urinary dilution during pregnancy than creatinine concentration because creatinine clearance increases significantly during pregnancy and is not consistent between women. Specific gravity correction was employed in preliminary bivariate analyses and used the formula P_C = P [(SG_m – 1) / (SG_i – 1)], where P_c is the specific gravity-corrected biomarker concentration (ng/mL), P is the measured biomarker concentration, SG_m is the median specific gravity value of the study population (1.019), and SG_i is the specific gravity value for each individual (Meeker et al. 2009).

2.3. Statistical analyses

Descriptive statistics were used to describe the demographics of the study population. One-way ANOVA tests were used to test for differences in means between visits for all biomarkers measured. Pearson correlation coefficients were calculated on natural log-transformed phthalate metabolite concentrations to create a heat map of correlation coefficients between metabolites across the study duration. Single-pollutant models were conducted using linear mixed effect models with random intercepts for repeated correlated outcomes. Phthalate exposures were modeled both as continuous variables and as quartiles. All statistical models were adjusted for categorical forms of maternal age, maternal education, and maternal pre-pregnancy BMI, as well as specific gravity. Covariates were selected based on a priori knowledge and impact on the main effect estimate (≥10%).

2.3.1. Adaptive elastic net and construction of ERS

We utilized adaptive elastic net with pairwise linear interactions between exposure variables (adENET-I) to construct overall environmental risk scores (ERS), which represent a weighted sum of an individual’s total exposure to a mixture of phthalates (Park et al. 2017). Similar to LASSO (least absolute shrinkage and selection operator), elastic net (ENET) penalizes unimportant predictors and shrinks them to zero, but ENET can additionally accommodate highly correlated predictors, whereas LASSO tends to randomly select one predictor from a group of correlated predictors (Tibshirani 2011). Ridge regression can handle multiple correlated predictors, but it will never shrink unimportant predictors all the way to zero. Thus, ENET is a hybrid of the LASSO and ridge regression methods which can simultaneously shrink unimportant predictors to zero and also accommodate highly correlated predictors. Adaptive ENET further provides standard errors and p-values and allows us to conduct hypothesis testing on the selected predictors. Optimal tuning parameters (λ₁ and λ₂) were chosen based on 5-fold cross-validation to minimize prediction error. P-values and 95% confidence intervals were calculated from standard errors on selected non-zero coefficients. We did not penalize coefficients of covariates. Coefficients derived from adENET-I were used as weights to reflect the relative importance for prediction of the outcome, with positive and negative weights denoting positive and inverse associations, respectively, between the predictor and the outcome. ERS was then computed as a weighted sum of all non-zero coefficients from adENET-I. adENET-I and subsequent calculation of ERS were conducted on data stratified by study visit to investigate potential windows of susceptibility to phthalate exposure. Repeated ERS were then used as predictor variables in linear mixed models to test for associations between exposure to phthalate metabolite mixtures and each outcome. ERS were modeled as continuous variables and as quartiles. The R package gcdnet (R version 3.5.1) was used to implement adaptive elastic net. The Benjamini and Hochberg method was used to determine the false discovery rate on these models (Benjamini and Hochberg 1995).

2.3.2. Bayesian kernel Machine regression

Bayesian Kernel Machine Regression (BKMR) utilizes a kernel function to flexibly model the relationship between a large number of predictors and a response variable (Bobb et al. 2015). It is a useful tool for extracting important predictors and for visualizing complex (non-linear) interactions between predictors in the mixture. We utilized 2000 iterations to ensure convergence of the models and calculated posterior inclusion probabilities (PIP) to determine relative importance of different predictors.

Our BKMR workflow utilized a stepwise process. First, a hierarchical model which included all 16 phthalate metabolites (grouped based on correlation coefficients roughly above 0.5) was run to establish a subset of predictors that BKMR deemed important. Then, a second BKMR model was run which included only important predictors from the first model (group and conditional PIPs roughly above 0.5). A hierarchical model was used if multiple predictors within the same group were selected from the first model. If some PIPs from the second model were below 0.5, a third model was run which eliminated those less important predictors. BKMR was implemented in R using the package bkmr (R version 3.5.1).

3. Results

3.1. Demographics and biomarker distributions

Demographic characteristics of the study population are shown in Table 1. Women were generally under the age of 30 (67.9%), had attained at least a bachelor’s degree (44%), were currently employed (61.6%), lived in a home earning less than $30,000 per year (63.3%), reported never smoking or being exposed to environmental tobacco smoke (85–90%), did not consume alcohol (50.2%), had<2 total pregnancies (56.9%), and had a BMI under 30 kg/m² (75.8%).

Table 1.

Demographic characteristics for 775 women in PROTECT.

Maternal Age	N	%
18–24	279	36.0%
25–29	247	31.9%
30–34	164	21.2%
35–41	85	11.0%
Maternal Education
GED or less	164	21.3%
Some college	267	34.7%
Bachelors or higher	339	44.0%
Currently Employed
Yes	474	61.6%
No	295	38.4%
Annual Household Income
<10k	223	32.3%
10k - <30k	214	31.0%
30k - <50k	155	22.5%
≥50k	98	14.2%
Smoking Status
Never	660	85.5%
Ever	98	12.7%
Current	14	1.8%
ETS
Never	651	89.7%
1 h	27	3.7%
> 1 h	48	6.6%
Alcohol Use
Never	386	50.2%
Yes, before pregnancy	333	43.3%
Yes, currently	50	6.5%
Number of Pregnancies
1	326	56.9%
2	77	13.4%
3–8	170	29.7%
BMI
(0, 25]	329	45.1%
(25, 29.9]	224	30.7%
(29.9, 51]	177	24.2%

Distributions of phthalate biomarker concentrations, over the study duration and at each visit, are shown in Supplementary Table 1. All phthalate biomarkers showed log-normal distributions and most did not show significant changes between study visits (exceptions being MBzP, MCOP, and MCPP, all of which showed their highest concentrations at the second study visit). A heat map of spearman correlation coefficients between phthalate metabolite concentrations, corrected for specific gravity, over the study period is shown in Fig. 1. As expected, metabolites from the same parent compound showed strong correlations (R = 0.80 – 0.98), and metabolites from distinct parent compounds also showed moderate correlations (R = 0.09 – 0.66), supporting the exploration of mixtures methods to investigate phthalate effects on adverse health outcomes.

Fig. 1.

Heat map of spearman correlation coefficients between phthalate metabolite concentrations, adjusted for specific gravity, measured over the study duration.

Distributions of lipid oxidative stress biomarkers, over the study duration and at each visit, are shown in Table 2. All biomarkers showed log-normal distributions. Significant differences in concentrations between study visits were observed with aCLP. Calculated fractions of differential pathway contributions to total 8-Iso-PGF2α indicated that approximately 78% was derived from chemical lipid peroxidation and 22% was derived from enzymatic lipid peroxidation.

Table 2.

Distribution of specific gravity-adjusted oxidative stress biomarker concentrations.

		Measured Biomarkers
		n	min	25th	50th	75th	90th	95th	Max	GM	GSD	ICC (95% CI)
8-Iso-PGF2α	Total	1420	0.127	1.26	1.87	2.64	3.90	4.78	28.4	1.82	1.89	0.31 (0.27, 0.34)
	V1	510	0.333	1.36	1.87	2.49	3.48	4.31	11.7	1.87	1.63
	V2	495	0.127	1.21	1.95	2.73	4.23	5.23	11.0	1.80	2.03
	V3	415	0.178	1.19	1.84	2.64	3.86	5.11	28.4	1.77	2.00
8-Iso-PGF2α metab.	Total	1402	0.04	0.67	0.97	1.44	2.04	2.74	19.5	0.98	1.85	0.33 (0.30, 0.37)
	V1	505	0.13	0.71	0.97	1.38	1.74	2.13	6.82	0.98	1.63
	V2	490	0.13	0.65	0.97	1.49	2.29	2.91	19.5	0.98	1.95
	V3	407	0.04	0.61	0.94	1.46	2.15	3.01	10.1	0.96	1.97
PGF2α	Total	1420	0.034	1.83	2.80	4.55	6.98	9.26	527	2.82	2.16	0.28 (0.24, 0.32)
	V1	510	0.311	1.95	2.70	3.99	6.00	7.84	29.8	2.77	1.87
	V2	495	0.143	1.75	2.91	4.75	7.03	10.7	24.6	2.79	2.31
	V3	415	0.034	1.72	2.97	4.83	7.57	9.32	527	2.91	2.32
		Calculated Values
		n	min	25th	50th	75th	90th	95th	Max	Mean	SD	ICC (95% CI)
aPGHS	Total	1420	0.0001	0.08	0.33	0.64	1.04	1.38	5.69	0.14	9.7	0.40 (0.36, 0.44)
	V1	510	0.0005	0.08	0.32	0.55	0.90	1.11	2.53	0.12	9.3
	V2	495	0.0001	0.08	0.34	0.68	1.07	1.58	4.17	0.14	10.4
	V3	415	0.0004	0.09	0.34	0.70	1.17	1.44	5.69	0.16	9.45
aCLP	Total	1420	0.058	0.91	1.41	2.18	3.24	4.16	28.4	1.37	2.09	0.36 (0.32, 0.40)
	V1	510	0.073	1.00	1.44	2.16	3.13	3.68	11.7	1.44	1.87
	V2	495	0.058	0.92	1.44	2.29	3.43	4.29	10.9	1.36	2.18
	V3	415	0.081	0.83	1.37	2.11	3.16	4.44	28.4	1.29	2.22
fPGHS	Total	1626	0.001	0.05	0.21	0.35	0.47	0.55	0.99	0.22	0.18	0.41 (0.37, 0.44)
	V1	540	0.001	0.04	0.17	0.32	0.44	0.53	0.93	0.20	0.18
	V2	631	0.001	0.05	0.21	0.36	0.46	0.55	0.90	0.22	0.18
	V3	455	0.001	0.07	0.23	0.37	0.49	0.57	0.99	0.24	0.19
fCLP	Total	1626	0.014	0.65	0.79	0.95	1.00	1.00	1.00	0.78	0.18	0.37 (0.33, 0.40)
	V1	540	0.074	0.68	0.83	0.96	1.00	1.00	1.00	0.80	0.18
	V2	631	0.105	0.64	0.79	0.95	1.00	1.00	1.00	0.78	0.18
	V3	455	0.014	0.63	0.77	0.93	1.00	1.00	1.00	0.76	0.19

aPGHS and aCLP are absolute concentrations of 8-so-PGF2α from enzymatic and chemical oxidation pathways, respectively, while fPGHS and fCLP are fractions. ICC: Intraclass correlation coefficient; measured biomarkers were corrected for specific gravity to calculate ICCs.

3.2. Mixtures analyses

Before exploring the results of our mixtures analyses, we first provide some context for interpretation of these results. First, selections from adENET analyses are given as weights. These weights reflect the β estimate assigned by adENET, which is a coefficient indicating the importance of each predictor for determining the outcome. Thus, a positive weight that is large in magnitude indicates that predictor has a strong positive association with the outcome. Second, BKMR results are shown in univariate and bivariate plots. The univariate plots show the shape of the association between one phthalate metabolite (x-axis) and the percent change in the outcome (y axis), while holding the rest of the metabolite in the mixture at their 50th percentile. These plots are an easy way to depict nonlinear associations. Lastly, bivariate plots show the shape of the association between phthalate 1 (x axis) and the percent change in the outcome (y axis), conditional on varying concentrations of phthalate 2 (given by different colored lines), while holding the rest of the phthalates in the mixture at their 50th percentiles (given by the notation h(phthalate 1 ∣ phthalate 2_percentile, Others₅₀). These plots show interactions between phthalate 1 and phthalate 2 within the mixture when the different colored lines are not parallel to each other.

3.2.1. Total 8-Iso-PGF2α metabolite

Two main effects and two interaction terms were selected by adENET-I as being important predictors for the total 8-Iso-PGF2α metabolite (Supplementary Table 2). MBP had a significant positive weight (visit 2, β: 0.191, p = 6.77e⁻⁹), while MECPP also had a positive weight but was not statistically significant (visit 2, β: 0.066, p = 0.141). Interactions between MBP and MECPP (visit 3, β: 0.042, p = 3.79e⁻⁷), and between MCOP and MECPP (visit 1, β: 0.030, p = 0.001) were observed, both with significant positive weights for 8-Iso-PGF2α. In accordance with adENET-I, BKMR models also selected MBP (PIP = 1) and MECPP (PIP = 0.992) as important predictors of 8-Iso-PGF2α, and additionally selected MCNP (PIP = 1). BKMR provided evidence of inverted U-shaped associations between 8 and Iso-PGF2α and MEHHTP and MCNP while fixing the other metabolites in the mixture at their 50th percentiles (Fig. 2a). Fig. 2b shows evidence of interactions between some metabolites in the mixture, given by bivariate relationships (i.e. different colored lines) that are not parallel to one another. The association between MCNP and 8-Iso-PGF2α changed slightly in a linear fashion across exposure levels of MBP and MECPP. The general linear relationships with MECPP and MBP increased in magnitude with increasing concentrations of MCNP. Other depictions of results from BKMR, including single-predictor and overall risks, are shown in Supplementary Fig. 1. Finally, an increase in ERS from the first quartile to the second, third, and fourth quartiles resulted in a 23.1% increase (95% CI: 10.9, 36.7), 32.0% increase (95% CI: 19.1, 46.2), and 54.3% increase (95% CI: 39.4, 70.9) in the concentration of the 8-Iso-PGF2α metabolite, respectively (Fig. 3).

Fig. 2.

Univariate and bivariate predictor-response functions for the effects of the phthalate metabolite mixture on the 8-Iso-PGF2a metabolite estimated by BKMR. A) univariate exposure-response functions and 95% confidence intervals for each phthalate metabolite while fixing the rest of the metabolites in the mixture at their 50th percentiles. B) bivariate predictor-response functions of one phthalate metabolite, conditional on four different quantiles of a second phthalate metabolite, while fixing the rest of the metabolites in the mixture at their 50th percentiles.

Fig. 3.

Associations and 95% confidence intervals between each outcome and an increase from the first quartile to the second, third, and fourth quartiles of ERS derived from adENET-I.

3.2.2. Chemical lipid peroxidation (aCLP)

Five predictors were selected from adENET-I as having an effect on the chemical fraction of 8-Iso-PGF2α (aCLP), which is hypothesized to be a measure of chemical lipid peroxidation (Supplementary Table 2). Four of these five were main effects and had positive weights [MCOP (visit 1, β: 0.072, p = 0.044; visit 2, β: 0.066, p = 0.052), MECPP (visit 1, β: 0.263, p = 6.38e⁻⁷), MEHP (visit 2, β: 0.214, p = 4.15e⁻⁶; visit 3, β: 0.299, p = 3.31e⁻¹¹), and MiBP (visit 1, β: 0.211, p = 2.60e⁻⁵; visit 2, β: 0.130, p = 4.94e⁻⁴)]. Conversely, the interaction between MEHHP and MHiBP had a negative weight (visit 1, β: −0.096, p = 1.19e⁻⁵). Only 3 predictors selected by adENET-I were also selected by BKMR: MECPP (PIP = 1), MiBP (Group PIP = 0.971, Cond. PIP = 0.416), and MHiBP (Group PIP = 0.971, Cond. PIP = 0.583). BKMR also selected MCNP (PIP = 1), MEP (PIP = 1), and MEHHTP (PIP = 1). BKMR provided strong evidence of nonlinear relationships between aCLP and most phthalate metabolites, while fixing the rest of the phthalates in the mixture at their 50th percentiles (Fig. 4a). There was also strong evidence of a nonlinear relationship between aCLP and MEHHTP which was modified when MEP was at its 25th percentile, but only when MEHHTP concentrations were high. The direction and the shape of the association between MHiBP and aCLP was dependent on varying quantiles of exposure to both MEHHTP and MECPP (Fig. 4b). Other depictions of results from BKMR, including single-predictor and overall risks, are shown in Supplementary Fig. 2. Finally, an increase in ERS from the first quartile to the second, third, and fourth quartiles were associated with a 47.8% increase (95% CI: 31.0, 66.7), 67.0% increase (95% CI: 48.3, 88.2), and 96.7% increase (95% CI: 74.0, 122) in aCLP, respectively (Fig. 3).

Fig. 4.

Univariate and bivariate predictor-response functions for the effects of the phthalate metabolite mixture on oxidative stress (aCLP) estimated by BKMR. A) univariate exposure-response functions and 95% confidence intervals for each phthalate metabolite while fixing the rest of the metabolite in the mixture at their 50th percentiles. B) bivariate predictor-response functions of one phthalate metabolite, conditional on four different quantiles of a second phthalate metabolite, while fixing the rest of the metabolites in the mixture at their 50th percentiles.

3.2.3. Enzymatic lipid peroxidation (aPGHS)

Three main effects and 10 interactions were selected as being important for the enzymatic fraction of the 8-Iso-PGF2α (aPGHS) that is hypothesized to be a measure of inflammation induced enzymatic lipid peroxidation (Supplementary Table 2). MBP (visit 1, β: 0.012, p = 0.906; visit 3, β: 0.022, p = 0.856), MEOHP (visit 1, β: 0.012, p = 0.921), and MHiBP (visit 1, β: 0.017, p = 0.874), all had positive weights but were not statistically significant. Interactions between MBP and MECPTP (visit 2, β: 0.086, p = 3.85e⁻⁴), and between MHBP and MiBP (visit 3, β: 0.165, p = 0.002) had positive weights and were the only interactions that were statistically significant. BKMR selected MECPP (PIP = 0.992) and MHBP (PIP = 1), which were parts of selected interaction terms from adENET-I, and uniquely selected MCNP (PIP = 0.978) and MEP (PIP = 0.783) as important predictors of the calculated marker of inflammation. BKMR provided evidence of nonlinear associations between aPGHS and MECPP and MCNP (Fig. 5a). Fig. 5b shows interactions between various metabolites on their associations with aPGHS. Low concentrations of MCNP show a strong positive association with aPGHS, which increases in magnitude with greater concentrations of MHBP. Similarly, a positive linear association which increased in magnitude with greater MEP was observed between aPGHS and MECPP. However, higher concentrations of MECPP were inversely associated with aPGHS only when MEP concentrations were below the 90th percentile. The association between aPGHS and MECPP became inverse at even lower MECPP concentrations with increasing concentrations of MHBP. Finally, MEP displayed associations that were similarly dependent on concentrations of MCNP and MECPP; MEP showed a strong positive association with aPGHS when MCNP and MECPP were at or above the 75th percentile, a largely null association with aPGHS when MCNP and MECPP were at their median, and an inverse association with aPGHS when MCNP and MECPP were at the 25th percentile. Other depictions of results from BKMR, including single-predictor and overall risks, are shown in Supplementary Fig. 3. Finally, an increase in ERS from the first quartile to the second, third, and fourth quartiles was associated with a 54.8% increase (95% CI: 1.08, 137), 191% increase (95% CI: 89.1, 348), and 268% increase (95% CI: 139, 465) in aPGHS, respectively (Fig. 3).

Fig. 5.

Univariate and bivariate predictor-response functions for the effects of the phthalate metabolite mixture on inflammation (aPGHS) estimated by BKMR. A) univariate exposure–response functions and 95% confidence intervals for each phthalate metabolite while fixing the rest of the metabolite in the mixture at their 50th percentiles. B) bivariate predictor-response functions of one phthalate metabolite, conditional on four different quantiles of a second phthalate metabolite, while fixing the rest of the metabolites in the mixture at their 50th percentiles.

4. Discussion

We examined the associations between repeated measures of phthalate metabolites, as well as mixtures of phthalate metabolites, and biomarkers of lipid oxidative stress, including 8-iso-PGF2α, its main metabolite, and fractions derived from the 8-iso-PGF2α/PGF2α ratio described previously (Van’t Erve et al. 2016). Exposure to phthalates may lead to increases in oxidative stress and inflammation by inducing the prevalence of reactive oxygen species and leukocytes, respectively (Ferguson et al. 2014). Overall, our results suggested increased concentrations of phthalate metabolites were significantly associated with higher levels of lipid oxidative stress biomarkers. Notably, the direction and shape of associations among certain individual phthalate metabolites, such as MHiBP and oxidative stress measures, were dependent on varying quantiles of exposure to other phthalates, including MEHHTP and MECPP. Meanwhile, associations between oxidative stress biomarkers and both MECPP and MCNP remained unaffected by increasing quantiles of exposure to the rest of the mixture. These findings suggest that complex interactions between individual phthalate metabolites exist within phthalate mixtures. Further study of these interactions could lead to identification of potential toxicity drivers with respect to lipid oxidative stress.

According to a recent review assessing phthalate biomonitoring studies around the world, some observed phthalate metabolite concentrations in the PROTECT cohort were lower than those observed in other populations, while other metabolites were higher. Factors including the year the study was conducted, location of the study, and differential product use patterns likely contribute to this variability (Wang et al., 2019).

4.1. 8-Iso-PGF2α

Previous studies have indicated that multiple individual phthalate metabolites have been shown to have strong associations with higher 8-iso-PGF2α levels in both pregnant and nonpregnant women. For example, a positive association between MBP and 8-Iso-PGF2α was reported among a preliminary subpopulation of PROTECT participants (Ferguson et al. 2014) and another pregnancy cohort with similar exposure levels and detection rates described previously (Van’t Erve et al. 2019). The findings from the present study are also similar to those reported in an earlier cohort of couples seeking fertility treatment (Wu et al. 2017), and the general U.S. population (Ferguson et al. 2012). Interestingly, a case-control study of pregnant women found the strongest evidence of mediation by 8-Iso-PGF2α on the association between MBP and preterm birth, relative to other phthalate metabolites in a study of pregnant women from the Boston area (Ferguson et al. 2017a). In our study, results from adENET-I and BKMR both identified MBP as a significant predictor of 8-Iso-PGF2α. These findings substantiate previous work suggesting MBP as a significant predictor of 8-Iso-PGF2α levels during pregnancy.

Previous studies have predominantly focused on linear associations between phthalate metabolites and 8-Iso-PGF2α. In Salinas Valley, California, a longitudinal birth cohort comprised of Mexican-American farmworkers and known as the Center for Health Assessment of Mothers and Children of Salinas (CHAMACOS) Study, Holland et al reported linear associations among 8-Iso-PGF2α and several phthalate metabolites (MBP, MECPP and MCNP) among pregnant women. Summations of DEHP metabolites and additional high molecular weight metabolites were also assessed and followed similar trends (Holland et al., 2016). These findings are similar to those in the present study. Based on the modeling of categorical ERS, a strong linear relationship was detected among the overall phthalate mixture and 8-Iso-PGF2α. BKMR analysis also indicated a linear association between the overall phthalate mixture and 8-Iso-PGF2α. Taken together, these results suggest that the combined effect of the overall phthalate mixture should be considered to better understand the effects of exposure to phthalates on concentrations of 8-Iso-PGF2α.

4.2. Chemical lipid peroxidation (aCLP)

Based on the computational methods described previously, the fraction of 8-iso-PGF2α hypothesized to be attributable to chemical lipid peroxidation (aCLP) may serve as a more accurate biomarker for lipid oxidative stress. Oxidative stress has been identified as a likely mechanism of toxicity for DEHP based on results from in vitro analyses (Mankidy et al. 2013). Other studies have reported similar findings. For example, in a CD-1 mouse model, isolated ovarian antral follicles exposed to DEHP showed increases in reactive oxygen species (ROS) as measured by the presence of free radicals and a reduction in antioxidant levels compared to controls (Wang et al. 2012). Similar results were reported in another animal model using cells derived from brain vesicles in mice (Wu et al. 2019). These in vitro examples are consistent with results from the present study. Some DEHP metabolites (MECPP and MEHP) were selected by adENET-I as important predictors of chemical fraction of 8-iso-PGF2α (aCLP). BKMR also suggested a nonlinear association between MECPP (a DEHP metabolite) and aCLP, independent of varying levels of the rest of the mixture. DEHP exposure has previously been shown to significantly impact lipid peroxidation in fathead minnow embryos (Mankidy et al. 2013). Mankidy et al. (2013) reported that a 1 mg DEHP/L dose caused a two-fold increase in lipid peroxidation, quantified based on the absorbance of lipid peroxide at 500 nm, and further determined DEHP as having the highest biologic effect score among phthalates (derived from a risk assessment grading scheme). The associations seen between DEHP metabolites and chemically-derived 8-iso-PGF2α were also consistent with findings in the Infant Development and Environment Study (TIDES) pregnancy cohort (Van’t Erve et al. 2019), which showed significant increases in aCLP with an IQR increase in all measured DEHP metabolites (MECPP, MEHHP, MEOHP, and MEHP) among pregnant women. The lack of significant associations between DEHP metabolites and enzymatically-derived 8-iso-PGF2α (aPGHS) in our study stresses the importance of using the ratio approach to better understand mechanistic pathways related to oxidative stress, as well as inflammation. Increases in oxidative stress may later induce inflammatory response, thereby producing more free radicals and leading to further oxidative stress. Therefore, understanding these highly intercorrelated processes and their respective relationships with pollutants is critical in determining their effect on other health endpoints.

4.3. Enzymatic lipid peroxidation (aPGHS)

In the present study, the fraction of 8-iso-PGF2α attributable to enzymatic lipid peroxidation (aPGHS) serves as a hypothesized biomarker of inflammation. Our study did not identify any individual metabolites as significant predictors of inflammation induced enzymatic lipid peroxidation. Earlier studies revealed similar null findings (Ferguson et al. 2015b). However, our study did identify significant interactions among individual metabolites that were predictive of aPGHS (MBP and MECPTP; MHBP and MiBP), as well as a significant positive association between the overall phthalate mixture and aPGHS. BKMR analyses further noted that the association between MHBP and aPGHS also did not change considerably with differing levels of other phthalate metabolites, with the exception of moderate increases in magnitude at higher concentrations of MCNP. Taken together, these results suggest that although individual phthalate metabolites may not have significant effects on the enzymatic pathway, when taken together they may result in elevated levels of enzymatic lipid peroxidation. Future studies assessing overall phthalate exposure are needed to better characterize the effects of phthalates, particularly phthalate mixtures, on inflammation.

There are limitations to the analyses in the present study. First, our results may not be generalizable to individuals outside of our study population (e.g., children, men, non-pregnant women, or pregnant women not participating in this cohort). Potential effect modification by unmeasured covariates was not a focus of this study. For example, control for antioxidant levels was not considered. Consumption of potential antioxidants such as fish oil has been shown previously to modify associations among phthalate exposure and 8-iso-PGF2α (Van’t Erve et al. 2019). Although, it is worth noting that this level of modification would have likely biased our effect estimates towards the null given the antioxidant properties present in fish oil. Similarly, interactions with other co-pollutants beyond phthalates is a possibility but was not a focus of this study. The ratio method used in our study to depict hypothesized levels of chemical and enzymatic lipid peroxidation may not accurately depict the true differences between oxidative stress and inflammation with respect to their impact on 8-Iso-PGF2α concentrations. Regarding our study design, measurement error may also be present given that this is a mathematically derived proxy and not a direct measurement of reactive oxygen species and leukocyte levels indicative of oxidative stress and inflammation, respectively. Also, a number of statistical comparisons were made in our study that may have led to chance findings. Lastly, despite the use of repeated measures of exposure levels, there remains a considerable amount of within-person variability and thus the potential for measurement error in our study.

In spite of these limitations, the present study had a number of strengths. In addition to the robust sample size, multiple biomarker measurements were taken per participant to assess exposure to phthalates, which provides a more robust measure of phthalate exposure over time compared to only one measurement. This study also applied previously established methods (e.g., adENET-I and BKMR) in a novel manner to assess the effects of exposure to phthalate mixtures and associations with lipid oxidative stress during pregnancy. These methods may more accurately reflect real-world exposure scenarios. Lastly, the ratio approach used to differentiate between chemical and enzymatic lipid peroxidation increased our ability to differentiate phthalate-induced toxicity pathways contributing to the two pathways induced by oxidative stress and inflammation.

In conclusion, we found several phthalate metabolites, and their mixture, to be associated with increased levels of oxidative stress. We further identified non-linear associations among phthalate metabolites and biomarkers of oxidative stress and inflammation. A linear association between exposure to the overall phthalate mixture and oxidative stress was also observed. Lastly, different interactions were seen between metabolites when assessing associations with chemically and enzymatically derived lipid peroxidation biomarkers. Taken together, these findings underscore the complexity of phthalate mixtures and highlight the need to assess associations with the two pathways separately. Future work is also needed to understand how environmental exposures may contribute to adverse pregnancy outcomes via oxidative stress and inflammatory pathways. Further study in other populations should also be conducted to substantiate these results.