A methodology to identify a DNA methylation-based signature of prenatal air pollution exposure in the ECHO Cohort

Abstract

Prenatal exposure to ambient air pollutants has been associated with adverse child health across different populations. In the absence of direct air pollution exposure measures, epigenetic biomarkers can serve as indicators of cumulative exposure. Previously, we developed a methodological pipeline to build a prediction model to distinguish infants highly exposed to air pollutants in utero based on DNA methylation in cord blood samples. Here, we test and refine this pipeline using data from multiple pregnancy cohort sites included in the Environmental influences on Child Health Outcomes (ECHO) Cohort. Levels of prenatal exposure to fine particulate matter (PM_2.5), nitrogen dioxide (NO₂), and 8-hour maximum ozone (O₃) were modeled using geocoded address data and machine learning. DNA methylation was measured using 450K or EPIC arrays. Training and validation sets were created and matched based on exposure level, cohort site, and array type to test the performance of the pipeline prediction model using area under the curve metrics. Among the three exposures, epigenetic-based predictions of PM_2.5 and NO₂ produced higher area under the curve values (0.64–0.68 and 0.62–0.95, respectively) than those for O₃ (0.50–0.75)_. We identified differences in the performance of the pipeline model across geographical regions, exposure windows, and thresholds used to distinguish “high” and “low” air pollutant exposure. Overall, this pipeline enabled the estimation of air pollution from DNA methylation data, though careful consideration of model inputs is required. Further studies will test the performance of the CpG sites identified by the pipeline to predict child health outcomes.

Introduction

Prenatal exposure to high levels of ambient air pollution has been associated with altered neurodevelopment [1–3], lung function [4–6], immune cell development [7,8], and other adverse health outcomes at birth and in childhood [9–12]. Awareness of prenatal air pollutant exposure may allow caregivers and medical providers to identify children at increased risk of disease, thereby supporting earlier interventions and improved disease management [13–15]. Additionally, exposure to environmental pollutants can be prevented through policies that reduce their generation and ensure necessary protections for highly exposed groups.

Studies of air pollution exposure at the population level often rely on area- or residential-level estimates of pollutants using geocoded addresses and data from stationary air pollution monitors managed by the Environmental Protection Agency (EPA), the National Aeronautics and Space Administration (NASA), and local organizations. However, such methods may not accurately quantify exposure. For example, people do not spend all their time in their residence, and access to air pollution mitigation strategies, such as in-home air purifiers, differs across populations. It is also difficult to quantify the effects of air pollution that are accumulated across pregnancy. Finally, differences in fetal biology, maternal compensatory mechanisms, and placental function may impact fetal resiliency to air pollution exposures [16]. This poses a challenge to identify children who are at true risk for adverse health outcomes associated with air pollutant exposure in utero. Here, we propose the use of a commonly collected, non-invasive biomarker—DNA methylation data measured in umbilical cord blood—to estimate the impact of prenatal air pollution exposure on fetal biology.

In previous work, we used cord blood DNA methylation data to develop a methodological pipeline that generated a biomarker that distinguished infants exposed to high levels of air pollution in utero using an estimated exposure variable [17]. DNA methylation measured in umbilical cord blood has proven to be sensitive to several environmental exposures, including air pollution, maternal diet, lead, and tobacco smoke [18–21]. DNA methylation is also relatively stable and reflects the accumulation of exposures over time [22,23]. In addition, cord blood serves as an accessible, non-invasive biosample from which to gather molecular information about an infant’s in utero environment. Cord blood DNA methylation at specific regions and CpG sites has been associated with childhood health outcomes, including those purported to be associated with prenatal air pollution exposure, suggesting that DNA methylation may serve as a biological mediator [24,25].

In our initial analysis, we were able to classify participants with high exposure to fine particulate matter (PM_2.5), nitrogen dioxide (NO₂), and polycyclic aromatic hydrocarbons (PAHs) in the validation set with modest or poor area under the curve (AUROC) values, depending on the pollutant (AUROCs = 0.46–0.51 for PM_2.5, AUROCs = 0.56–0.60 for NO₂, and AUROC = 0.44 for PAHs). However, the sample sizes in these initial cohorts were relatively small (training cohort n = 432, validation cohort n = 67). Importantly, the goal of our previous work was not to identify a set of CpGs that should be applied across all populations, but to offer a methodology to identify such a set of CpGs that could be applied within specific contexts. Herein, we apply our methodological pipeline in the Environmental influences on Child Health Outcomes (ECHO) Cohort, which provides high-quality geocoded address data, air pollution estimates at point locations from sophisticated ensemble-based models, and greater participant racial, ethnic, and geographic diversity. The main goals of this analysis were to 1) test the methodological pipeline in a larger cohort and 2) identify important factors that need to be considered when adapting the pipeline to novel populations, such as the choice of threshold value for high pollutant exposure, geographic region, and timing of the exposure relative to biosample collection.

We calculated AUROC metrics to evaluate how well the methodological pipeline performed among participants in the ECHO Cohort with prenatal exposure to PM_2.5, NO₂, and 8-hour maximum ozone (O₃). We then tested different pregnancy time windows and defined “high pollutant exposure” using three different thresholds: the ECHO Cohort median, the World Health Organization (WHO) air quality guidelines, and the EPA regulatory thresholds. Our methodological approach is shown in Fig. 1.

Figure 1.

Methodological approach to identify a DNA methylation-based signature of prenatal air pollution from six study sites enrolled in the ECHO Cohort.

Abbreviations: AUC, area under the curve.

Results

Descriptive characteristics and air pollutant concentrations

Table 1 shows the descriptive characteristics of the six ECHO cohort sites included in this analysis. Cohort sites are hereafter referred to as “sites” to distinguish them from the overall ECHO Cohort. Of the 1093 samples, 79% (N = 823) had cord blood DNA methylation measured on the 450K array, and the remainder were run on the EPIC v1 array. Most children (N = 661, 60.5%) were White (parent reported). The median gestational age at birth was 39.2 weeks, and site-specific median gestational ages ranged from 38.8 to 39.4 weeks. Four sites were located in the Northeastern region of the United States, and two sites were based in the Western United States. Participant recruitment across the six sites spanned from 1998 to 2023, although the current analysis only included births included in the air pollution model estimates (2000 to 2017).

Table 1.

Descriptive statistics of the six ECHO sites included in the current analysis.

Study Site	Site Aa	Site B	Site C	Site D	Site E	Site F	Full
N for this analysis (max)	229	209	130	61	393	71	1093
Years of recruitment	1998–2023	1999–2002	2009–2017	2016–2023	2009–2014	2013–2105	1998–2023
Inclusion criteria	▪ 18–35 years of age▪ ≤20 weeks gestation▪ Singleton pregnancy▪ Dominican or African American▪ Living in Manhattan or the South Bronx▪ Non-smoking▪ No illicit drug use▪ No preexisting hypertension, HIV, diabetes	▪ <22 weeks gestation▪ Singleton pregnancy▪ English speaking▪ Receive prenatal care selected study practices▪ Delivering at study hospitals	▪ 18–45 years of age▪ Singleton pregnancy▪ English speaking▪ Use of a private drinking-water system at home	▪ ≥18 years of age▪ ≤18 weeks gestation▪ Planned to deliver at a study hospital site	▪ ≥15 years of age▪ ≤23 weeks gestation▪ Singleton pregnancy▪ No history of stillbirths▪ No preexisting serious chronic disease (cancer, psychiatric diseases, steroid-dependent asthma, diabetes mellitus)	▪ ≥15 years of age▪ ≥ 23 weeks gestation▪ Singleton pregnancy▪ English speaking▪ Current cigarette smoker (≥1 cigarette/wk)	N/A
Geographic location	Northeast	Northeast	Northeast	Northeast	West	West	Northeast & West
Urbanicity	Urban	Urban & Suburban	Suburban	Urban	Urban & Suburban	Suburban	Urban & Suburban
Methylation platform	450K/EPIC	450K	450K	EPIC	450K	EPIC	450K/EPIC
Race (parent report)
American Indian or Alaska Native	<5 (<2.2%)	<5 (<2.4%)	<5 (<0%)	<5 (<8.2%)	<5 (<1.3%)	<5 (<7.0%)	<5 (<0.1%)
Asian	<5 (2.2%)	6 (2.9%)	<5 (<3.8%)	7 (11.5%)	8 (2.0%)	<5 (<7.0%)	24 (2.2%)
Black	69 (30.1%)	28 (13.4%)	<5 (<3.8%)	<5 (<8.2%)	40 (72.0%)	<5 (<7.0%)	141 (12.9%)
Multiple	<5 (<2.2%)	16 (7.7%)	<5 (<3.8%)	7 (11.5%)	26 (6.6%)	15 (21.1%)	66 (6.0%)
Other	<5 (<2.2%)	<5 (<2.4%)	<5 (<3.8%)	8 (13.1%)	<5 (1.3%)	<5 (<7.0%)	43 (3.9%)
White	<5 (2.2%)	158 (75.6%)	128 (98.5%)	34 (55.7%)	283 (72.0%)	55 (77.5%)	661 (60.5%)
Inconsistent	18 (7.9%)	<5 (<2.4%)	<5 (<3.8%)	<5 (<8.2%)	<5 (<1.3%)	<5 (<7.0%)	18 (1.6%)
Don’t Know	134 (58.5%)	<5 (2.4%)	<5 (<3.8%)	<5 (<8.2%)	<5 (<1.3%)	<5 (<7.0%)	136 (3.9%)
Self-reported ethnicity (% Hispanic)	167 (72.9%)	22 (10.5%)	<5 (<3.80%)	20 (32.8%)	97 (24.7%)	11 (15.5%)	317 (29.0%)
Child sex (% female)	116 (50.7%)	119 (56.9%)	66 (50.8%)	32 (52.5%)	188 (47.8%)	36 (50.7%)	557 (51.0%)
Gestational age (mean, SD)	39.3 (1.62)	39.4 (1.60)	39.4 (1.46)	38.8 (1.47)	39.1 (1.29)	38.9 (1.24)	39.2 (1.46)

^aSite A includes three studies that recruited using the same geographic area and inclusion criteria.

Abbreviations: ECHO, Environmental influences on Child Health Outcomes; SD, standard deviation.

The distributions of the trimester-averaged concentrations of PM_2.5, NO₂, and 8-hour maximum O₃ for all six sites are shown in Fig. 2. Table 2 shows the exposure thresholds published by the EPA and WHO, along with references. The ECHO Cohort median concentrations of PM_2.5 (first trimester = 8.2 ug/m³, second trimester = 8.1 ug/m³, third trimester = 8.3 ug/m³) fell in between the 2012 EPA threshold (12 ug/m³) and the WHO threshold (5 ug/m³), with all of the site-specific median concentrations of PM_2.5 above the WHO threshold. Exposure distributions by site differed the most for NO₂. Site-specific median concentrations of NO₂ were all below the EPA primary threshold (53 ppb) but above the WHO threshold (5 ppb), with the study-wide median concentration falling in the middle (first trimester = 35.9 ppb, second trimester = 36.4 ppb, third trimester = 35.6 ppb). For O₃, site-specific medians were most consistent, and all observations of 8-hour maximum O₃ were below the EPA primary threshold (70 ppb), with the cohort median of O₃ closer to the WHO threshold for O₃ (30 ppb) (first trimester = 37.5 ppb, second trimester = 37.9 ppb, third trimester = 40.0 ppb). We also observed that the two sites with the majority of participants located in the Western United States (Sites E and F) had large differences in distributions of NO₂ levels. This may be due to the urbanicity of the recruitment area, with Site E participants coming primarily from the Denver, Colorado area and Site F participants coming from Oregon and Washington. Additionally, variation in factors that influence air pollutant formation between these locations, such as climate, topography, and altitude, would likely contribute to the differences in pollutant levels, particularly for NO₂ and O₃. Among the sites located in the Northeast, Site C had much lower median NO₂ exposure levels than the other sites. Differences in exposure levels and variability across sites supported our approach to balance training and validation groups based on cohort site and geographic region in addition to methylation array type.

Figure 2.

Trimester-averaged distributions of exposures across ECHO sites. The dashed blue (middle) line represents the cohort median (PM_2.5 = 8.3 ug/m³, NO₂ = 35.6 ppb, O₃ = 40.0 ppb). The dashed orange (top) line represents the United States Environmental Protection Agency/National Ambient Air Quality Standards primary threshold for each pollutant (PM_2.5 = 12 ug/m³ as of 2012, NO₂ = 53 ppb as of 2010, O₃ = 70 ppb as of 2015). The dashed magenta (bottom) line represents the World Health Organization air quality guidelines threshold for each pollutant (PM_2.5 = 5 ug/m³, NO₂ = 5 ppb, O₃ = 30 ppb).

Abbreviations: ppb, parts per billion; ug/m³, micrograms per cubic meter.

Table 2.

References to thresholds published by the EPA and WHO for PM_2.5, NO₂, and 8-hour max O_3.

Pollutant	EPA threshold	EPA reference	WHO threshold	WHO reference
PM2.5	12 µg/m3	2020 NAAQS primary annual mean, averaged over 3 years standard (https://www.govinfo.gov/content/pkg/FR-2020-12-18/pdf/2020-27125.pdf)	5 µg/m3	2020 long-term/annual AQG (https://www.who.int/publications/i/item/9789240034228)
NO2	53 ppb	2018 NAAQS primary 1-year annual mean standard (https://www.federalregister.gov/documents/2018/04/18/2018-07741/review-of-the-primary-national-ambient-air-quality-standards-for-oxides-of-nitrogen)	5 ppb	2020 long-term/annual AQG of 10 µg/m3, which corresponds to approximately 5.07 ppb. Same reference as above.
O3	70 ppb	2020 NAAQS primary annual fourth-highest daily maximum 8-hour concentration, averaged over 3 years (https://www.govinfo.gov/content/pkg/FR-2015-10-26/pdf/2015-26594.pdf)	30 ppb	2020 long-term, peak season AQG of 60 µg/m3 (which corresponds to approx. 30 ppb). Same reference as above.

Abbreviations: AQG, Air Quality Guidelines; EPA, Environmental Protection Agency; NAAQS, National Ambient Air Quality Standards; WHO, World Health Organization.

Performance of known predictors and negative control

Fig. 1 depicts our approach to determining CpG inclusion, ranking, and performance assessment by AUROC for the pipeline. We first examined whether cord blood DNA methylation could accurately predict variables such as child sex and preterm birth. The highest mean AUROCs across all CpG sets and validation iterations for these variables were 1.00 and 0.91, respectively. We also demonstrated that the highest mean AUROC for a randomly generated binary variable, our negative control, was 0.54. The AUROCs for child sex remained at 1.00 for each leave-one-out sensitivity analysis, while AUROCs ranged from 0.80 to 1.00 for preterm birth in the main analysis (0.80 to 1.00 in the sensitivity analyses) and ranged from 0.50 to 0.60 for the randomly generated variable, depending on the site that was left out of the training set. When the data were subset by region (six Northeastern sites, two Western sites), all AUROCs for sex remained at 1.00, whereas AUROCs for preterm birth and the randomly generated variable were 0.89 and 0.56 for the Northeastern data, respectively, and 0.63 and 0.48 for the Western data, respectively. Detailed information, including AUROCs for all main and sensitivity analyses and sample sizes of training and validation sets, are provided in the Supplementary Data.

Comparing pipeline performance under different thresholds for high pollutant exposure

Table 3a shows the AUROCs from models when prenatal PM_2.5, NO₂, and O₃ exposures were computed by averaging daily measures from different days before birth and either the cohort median, EPA threshold, or WHO threshold was used as the cutoff for high versus low exposure. Given the imbalance in the sizes of the high- and low-exposure when using the EPA and WHO thresholds, we also show the area under the precision-recall curve (AUPRC) on top of the AUROC, as well as the prevalence of high exposure for each pollutant and threshold in Table 3b. For PM_2.5, the EPA threshold and ECHO Cohort median yielded similar AUROCs (within ± 3%) across the tested prenatal windows, whereas the WHO threshold yielded lower AUROCs. Trimester-averaged values of PM_2.5 exposure yielded higher AUROCs than shorter prenatal windows across all thresholds; this trend was most evident using the lower WHO threshold. The highest AUROC for NO₂ occurred using the EPA threshold when NO₂ concentration was averaged over 60 days prior to birth (AUROC = 0.95). Conversely, the AUROCs for NO₂ exposure were higher when averaging shorter exposure time periods (two and seven days before birth) using the WHO threshold. AUROCs for O₃ were generally lower across all tested thresholds compared with AUROCs for PM_2.5 and NO₂, and we were only able to calculate AUROCs for averages from two days before birth using the EPA threshold due to insufficient numbers of participants above the EPA threshold in training or validation sets (fewer than five participants in the high exposure group, Supplementary Data). AUROCs were slightly higher using the WHO threshold compared to using ECHO Cohort median, and both increased as the time over which O₃ was averaged increased. Table 3b shows the AUPRCs corresponding to the AUROCs in Table 3a, along with prevalence of the high-exposure group based on the tested threshold and exposure window. AUPRCs for all pollutant exposure models were greater than the prevalence of high exposure based on any threshold, suggesting the predictability of these models even with imbalanced (EPA and WHO thresholds) data. WHO thresholds gave the most imbalanced groupings compared to the ECHO cohort median and EPA thresholds and correspondingly, these models tended to have the lowest AUROCs. The models using the WHO thresholds also had the smallest difference between the prevalence of the high-exposure group to the AUPRC

Table 3a.

Area under the receiver operating characteristic curve (AUROC) values of models when using all available data averaged over different prenatal windows and using the cohort median, EPA threshold, and WHO threshold for PM_2.5, NO₂, and O_3.

Color scale: Green indicates a higher area under the curve (closer to 1.00), while yellow indicates a lower area under the curve (closer to 0.50).

T-Inv indicates that the training set does not have enough people in each level to run t-test.

Abbreviations: EPA, Environmental Protection Agency; WHO, World Health Organization.

Table 3b.

Area under the precision-recall curve (AUPRC) values of models when using all available data averaged over different prenatal windows and using the cohort median, EPA threshold, and WHO threshold for PM_2.5, NO₂, and O_3. Prevalence of the high exposure group is provided in parentheses below the AUPRC value for comparison.

Color scale: Green indicates a higher area under the curve (closer to 1.00), while yellow indicates a lower area under the curve (closer to 0.50).

T-Inv indicates that the training set does not have enough people in each level to run t-test.

Abbreviations: EPA, Environmental Protection Agency; WHO, World Health Organization.

Comparing performance across different geographical regions

The AUROCs of the geographical region-specific models, using the trimester- and pollutant-specific ECHO Cohort medians, are shown in Table 4. For the models to predict PM_2.5 exposure using four sites from the Northeast, AUROCs remained generally consistent with the main analysis (AUROCs ranging from 0.70 to 0.85 across prenatal windows). However, the prediction model using participants from the Western sites performed poorly, with AUROCs ranging from 0.51 to 0.62. For NO₂ exposure, the prediction model using the Northeastern data alone performed better than the model using the entire study population (“full” model), whereas the West-specific model performed either on par with or worse than the full model. The Northeast-specific model for O₃ exposure performed slightly worse than the full model across prenatal windows. The West-only model tended to perform better than the full model for most prenatal windows considered and performed fairly for the prenatal window of 60 days prior to birth (AUROC = 0.77) and over the third trimester (AUROC = 0.81). For all air pollutants, predictions tended to improve as the prenatal window time span over which the pollutants were averaged increased. The exception to this was for the models predicting O₃ exposure using the Northeastern sites, where the model predicting first trimester-averaged O₃ (AUROC = 0.74) performed the best.

Table 4.

Area under the receiver operating characteristic curve (AUROC) values of models when using all available data averaged over different prenatal windows using the cohort median, compared across geographical regions.

Color scale: Green indicates a higher area under the curve (closer to 1.00), while yellow indicates a lower area under the curve (closer to 0.50).

N values refer to number of cohorts included in each geographical region-specific analysis

Comparing performance in the leave-one-out sensitivity analysis

AUROCs from the sensitivity analyses using the leave-one-out approach with ECHO Cohort medians as cutoffs for high/low exposure are shown in Tables S1–S3. The “all but” column shows the AUROCs of all of the sites except for the left-out site, while the “left out” column shows the AUROCs for the left-out site only. AUROCs for the leave-one-out models were generally similar for all PM_2.5 sensitivity analyses, although they dropped somewhat when Site B was left out. AUROCs from leave-one-out models for NO₂ were generally better or on par with those from the full model. AUROCs remained fairly stable across all sensitivity analyses for O₃ exposure. However, for all pollutants, the models performed worse for the left-out site compared with the full and “all but” analyses. For NO₂ exposure, time windows > 2 days could not be calculated for Sites C, D, and F alone due to invariant NO₂ exposure in these sites. This also occurred for O₃ exposure in Site D from exposure-averaged windows of seven to 30 days.

Comparing CpG probes from different models

To better understand potential consistencies and differences in the biological underpinnings of these models across different pollutants, regions, and exposure windows, we extracted the probes that contributed to each model that used the cohort median cutoff for high/low exposure. Each prediction model could contain from 10 to 500 probes, and this range was reflected in the results. Table S4 shows the number of probes for the best performing models (which correspond to the AUROC values in Table 4). The model for PM_2.5 averaged over 30-days before delivery used only 10 probes, whereas four models (first trimester PM_2.5 in the West, 14-day averaged O₃ overall, 14-day averaged O₃ in the Northeast, and 30-day averaged O₃ in the West) used the maximum allotted 500 probes. The most common number of probes used in models was 70, which applied to seven of the models. Among models for PM_2.5 exposure for the entire region, cg00777627 was the only probe that contributed to models for all eight tested exposure windows. Several other probes contributed to six (cg02786368 and cg11164639) or seven (cg02912316 and cg10601100) of the exposure windows. The majority of probes (n = 409) contributed to only one of the eight tested exposure windows. There was less consistency in probes among models for NO₂ exposure. No probe contributed to all eight tested windows. Only two probes (cg1152206 and cg14286941) contributed to all three trimesters, suggested that the impact of NO₂ exposure on the cord blood methylome differs throughout gestation. Models for O₃ exposure showed similar inconsistency, with no probes contributing to all tested time points or even all three trimesters. Models for O₃ also tended to use larger numbers of probes for predictions, compared with probe sets for PM_2.5 and NO₂.

When we compared models for the entire study cohort, there were few overlapping probes across models for different pollutants. PM_2.5 and O₃ models shared 25 probes, many of which were also shared across windows of exposure. Models for the first, second, and third trimester exposure windows of PM_2.5 and O₃ shared 23, 24, and 21 probes, respectively. Conversely, NO₂ models did not share any probes with PM_2.5 models or O₃ models. We also compared probes from the same pollutant models in different regions (Northeast versus West). PM_2.5 models shared a total of seven probes between regions, two of which were present in models for the Northeast and West for all three trimesters (cg05434655 and cg27007358). NO₂ models shared six probes between regions, three of which contributed to all three trimesters in both regions (cg00884529, cg08393828, and cg25018755). All six probes contributed to the third trimester models in the Northeast and West. Finally, there were 10 shared probes across Northeast and West models for O₃ exposure. Six of these probes (cg03576854, cg05215748, cg05500774, cg16140548, cg16246962, and cg17697633) contributed to models for all three trimesters and all 10 probes contributed to the third trimester O₃ exposure model in both regions.

Discussion

We applied a pipeline to a national cohort that was previously used in a smaller, New York City-based cohort to identify a set of differentially methylated CpG sites in cord blood that could reflect in utero pollutant exposure [17]. In this analysis, we assessed the pipeline’s applicability over different prenatal windows of exposure, using different thresholds for classifying “high” exposure and using data from populations living in differing regions of the country. Our results show that the performance of the pipeline was variable based on these parameters. Performance also varied by pollutant, with models for PM_2.5 and NO₂ generally performing better than those for O₃. This could arise from lower variability of O₃ levels across the study population compared to the other pollutants. It may also be due to true differences in biological response based on the pollutant. This may be supported by the generally larger probe sets observed for the best O₃ prediction models, compared with those for PM_2.5 and O₃.

In our sensitivity analysis, we showed that the pipeline did not apply well to left-out cohorts. However, leaving out a site did not majorly affect the performance in the remainder of the sites combined. As discussed below, these findings have implications for how to apply such a pipeline in the future and highlight considerations for DNA methylation-based indicators of environmental exposures.

In our previous work using cohorts based in Northern Manhattan and the South Bronx, we tested the number of CpG sites that was sufficient, as determined by positive predictive values, to screen exposure measures that are potentially predictable by DNA methylation, where the positive predictive value was defined as the percentage of CpGs that can be defined as “true positives.” Specifically, using the training set, we applied a two-sample t-test at each CpG site examining methylation measures between high- and low-exposure groups and selected top-ranked CpGs based on P-values. We similarly applied a two-sample t-test using the validation set to the selected CpG sites. CpGs with P-values < 0.05 and with the same directional change in methylation measures in both the training and validation sets were considered “true positives.” [17] We used this methodology to test inclusion of 50 to 1 500 CpG sites, increasing by 50, and found that inclusion of 500 CpG sites was sufficient for screening. Therefore, in this analysis we applied the same logic that 500 CpG sites would be sufficient to screen and predict DNA methylation associated with PM_2.5, NO₂, and O₃ exposure.

In our previous study, the AUROCs from the third trimester-averaged exposure models using the study sample median were 0.54 for PM_2.5 and 0.61 for NO₂ [17]. This can be compared to values of 0.88 and 0.82 for PM_2.5 and NO₂, respectively, in the current analysis. These differences may likely be due to increased sample size (n = 432 versus n = 969). It is important to note that prediction for specific cohorts in the leave-one-out sensitivity analysis performed less well, even when the distribution of air pollutants in the training set reflected that of the validation set. This suggests that there are variables beyond the distribution of air pollutants that influence the pipeline and supports usage of the pipeline with larger sample sizes.

Several patterns stood out in this study. In most cases, models that used larger windows of exposure resulted in higher AUROCs. This is most likely due to an increasing number of exposure measurements as the time frame increases, making the individual more likely to be accurately assigned to high or low exposure in the training set. Therefore, the length of exposure time and number of exposure measurements should be considered when choosing pipeline parameters. To align our pipeline with current nationwide and global standards, we also assessed performance using the WHO air pollution guidelines and EPA regulatory thresholds. As these guidelines are based on existing studies of adverse effects from PM_2.5, NO₂, and O₃, we expected that they would apply to the pipeline. However, we observed that the established thresholds from the EPA and WHO often did not apply well to the study population, particularly when certain sites were left out of the training phase. This often resulted in too few participants above the threshold, as was the case for the EPA standards, or below the threshold, as was the case for the WHO standards. Thus, although these thresholds are grounded in what is theoretically meaningful for health, they may not be feasible for prediction purposes.

Our ability to predict differences based on the cohort median suggests that there are true biological effects of air pollution. This is supported by other studies on prenatal air pollution exposure and cord blood DNA methylation [26–28]. While we did consider geographic region-specific medians, one alternative option that was not explored in the current study was the use of regionally relevant thresholds for pollutants, such as the top 50^th percentile of a pollutant based on historical regional norms. Additionally, we used the cohort median to characterize high and low pollution. However, such characterization may not appropriately capture the complexities of air pollution exposure. Continuous and multi-level modelling may improve predictive performance and should be considered in further applications of the pipeline.

One of the most interesting findings from this analysis was the differences in model performance when only study participants living in the Northeast or West were considered in the model. This could be due to changes in the variability of pollutant exposure levels after sub-setting the sample. Less variability in exposure levels may make it more difficult to identify CpG sites that differentiate between high and low exposure in the training set. There may also be underlying study characteristics that affect DNA methylation. One of the sites located in the West, Site F, exclusively recruited pregnant participants who reported smoking during pregnancy [29]. The exposure to cigarette smoke, which has been shown to affect DNA methylation in cord blood, may have modified the association between cord blood DNA methylation and air pollution exposure, thus altering model performance [21]. However, when this site was left out of the training set, the model performance did not markedly change in a consistent manner for any of the pollutants. This suggests that the pipeline accurately identifies high prenatal exposure to PM_2.5, NO₂, and O₃ without major interference from other exposures that are known to affect DNA methylation in utero. Other potential confounders, including race, ethnicity, and maternal education, were not included in the pipeline, as they were not strongly associated with both air pollution and DNA methylation in preliminary analyses. This could change in different contexts and study populations. Potential confounders should be explored prior to pipeline implementation. As t-tests cannot adjust for covariates, we suggest matching the training and validation sets based on confounders, as was done in this study to control for region of residence. In the same vein, we are unable to truly determine whether changes to prediction performance in the sensitivity analyses were due to cohort site characteristics or the recruitment period for the cohort. We see that sites that began recruitment earlier in time (Sites A and B) had higher levels of PM_2.5 and NO₂, likely due to changes in environmental regulations over time. However, it is difficult with the data available at this time to fully untangle the effect of differing pollutant levels versus other cohort site characteristics on the prediction performance.

Another consideration that we did not explore in the current study was differences between array type, as all but one of the cohort sites only used either the EPIC or 450K array to measure DNA methylation. This barred us from formally investigating the influence of array type on model performance, as we would be unable to discern whether differences arose from cohort site or array type. The major difference between the platforms is the number of measured CpG sites, with 485 000 CpG sites on the 450K array and 850 000 CpG sites on the EPIC v1 array. Here we used only probes present on both arrays (357 156 CpG sites, after quality control). However, there may be other differences between the arrays that may result in differences in model performance. The 450K array accommodates 12 samples in a single batch/slide, while the EPIC v1 can only accommodate eight samples per batch. Another consideration here is the probe chemistry used on the arrays. The 450K array contains only Type I probes, which use two separate probes sequences to measure a single CpG site. The Type II probes use a single sequence to measure a CpG site and tend to perform better than Type I probes [30]. The EPIC array contains a mix of both probe chemistry types, and even those probes present on both arrays may be Type I on the 450K and Type II on the EPIC. In the current analysis, we split our training and validation sets equally based on methylation array type to overcome these differences and produce models that were applicable across arrays. However, if one were to apply the pipeline using only one array type, model performance and top contributing CpG probes would likely change.

There are several limitations to this analysis, which could affect implementation of the pipeline beyond the current study. First, our air pollution estimates were not derived from personal air pollution monitors, which could affect the accuracy of the measures for each participant. To enhance accuracy to the best of our ability, we used sophisticated ensemble models to estimate air pollutant exposures that incorporated physical monitoring, satellite data, meteorological variables, land use variables, elevation, and chemical transport model predictions [31–33]. Nevertheless, we were unable to account for time spent inside versus outside the home, and we only made air pollution estimates based on home address. Such imprecision may lead to misclassification of the exposure, although we expect that such misclassification would be non-differential. Second, our study sample consisted mainly of participants who lived in the Northeast or West of the United States during their pregnancy. This limited our ability to compare performance in the Midwest, Southeast, and Southwest. Finally, we did not explore the effects of potentially important variables such as population density, climate, altitude, and seasonal variation. While we expect that these variables would be closely associated with study site, we did not explore them in depth here.

This study draws strength from the use of a demographically diverse cohort of United States-based pregnancy sites, providing large training and validation sets and allowing for model testing with different parameters. This major strength provides valuable insight into biomarker development and raises considerations for applying existing DNA methylation-based biomarkers to other cohorts. The current study also serves as an important step towards validating the proposed pipeline to identify newborns who may be at higher risk for adverse health outcomes commonly associated with prenatal air pollutant exposure. We show that the pipeline can be implemented in varying contexts, but choice of the threshold and window of exposure should be considered. Additionally, our results stress the importance of the exposure distribution in the training sample reflecting that of the validation sample for biomarkers of exposure.

This study contributes to the growing body of research aimed at elucidating molecular mechanisms underlying associations between prenatal air pollution and child health outcomes. In further studies, we will examine whether the CpG sites identified by the pipeline are associated with childhood health outcomes, such as neurodevelopment, obesity, and asthma, and whether these CpG sites mediate associations between exposure and outcome. In doing so, we can understand whether the DNA methylation signatures identified here predict childhood health outcomes better than measurements of ambient air pollutants. Such work will also inform how inter-individual differences alter the ability of the fetus to mitigate air pollution. Importantly, these studies could help to improve risk stratification and identify the most vulnerable populations during pregnancy. The current study serves as an important step in this goal. Given our results, we will focus these analyses in the ECHO Cohort on PM_2.5 and NO₂ exposure for populations living in the Northeastern region of the United States during the time period studied in this analysis (2000 to 2017). Other air pollutants, regions, and calendar times require further investigation in larger sample sizes to validate our findings.

Materials and methods

Study participants

The ECHO Cohort is an National Institutes of Health-funded study initiated in 2016 to study and improve children’s health in the United States [34]. The cohort consists of 69 existing study sites from across the United States and includes harmonized extant data and newly collected data using a standardized protocol. Study consent, enrollment, and data harmonization have been described in detail elsewhere [3,35]. In this study, we included eight ECHO sites with 1) umbilical cord blood DNA methylation measured in cord blood using either the Infinium 450K array or the EPIC array and 2) air pollution estimates (PM_2.5, NO₂, and O₃) derived from residential address throughout pregnancy, accounting for mobility, as of the March 2023 data lock (Fig. S1). Three of the sites that recruited using the same geographic area and inclusion/exclusion criteria were combined into a single site (Site A) in the current analysis due to overlap in recruitment population, participants, and study design. The names and details of the sites included in the analysis are provided in Table 1. Across all sites, study participants were recruited during pregnancy and followed through delivery. The study protocol was approved by the local or single ECHO institutional review board (IRB), and written informed consent was obtained from the mother for participation in the specific site protocols and, for some, participation in the ECHO-wide Cohort Data Collection Protocol. The study protocol was approved by the single ECHO institutional review board, WCG IRB. Written informed consent or a parent’s/guardian’s permission was obtained along with child assent, as appropriate, for the ECHO Cohort Data and Biospecimen Collection Protocol participation and for participation in specific study sites.

Point estimates of PM_2.5, NO₂, and O₃

Participants provided address data over the course of their pregnancy, which were systematically processed and then manually checked, where needed, by ECHO’s Data Analysis Center. Processed address data were geocoded using ArcGIS StreetMap Premium (Esri, Redlands, CA, USA). To balance sample size and air pollutant estimate accuracy, we included participants whose addresses could be matched to a street, but not necessarily a specific building number. Geocoded prenatal residential addresses were used to estimate air pollutant exposure with a sophisticated exposure model developed by Di and colleagues [31–33]. Briefly, these models integrates multiple machine learning algorithms, including neural network, random forest, and gradient boosting, in ensemble models and generalized additive models to predict daily PM_2.5, NO₂, and 8-hour O₃ levels at ground level at 1 km² grid cells across the United States. Model input data include monitored daily pollutant (PM_2.5, NO₂, and O₃) levels from extant stationary monitors, land use terms, chemical transport model predictors, meteorological variables, and satellite measurements. The cross-validated R² values from the estimation models were 0.86 for PM_2.5 [31], 0.79 for NO [2,33] and 0.76 for O₃ [32]. A detailed description of air pollutant model development has been published elsewhere [31–33]. The models estimated daily air pollutant levels between 2000 and 2016. Across all prenatal addresses, and accounting for time spent at the address, daily prenatal air pollution exposure estimates were assigned based on the nearest 1 × 1 km grid cell. Daily estimates were then averaged across the gestational period to construct exposure metrics representing average exposure during the first, second, and third trimesters as well as during specific acute prenatal exposure windows (i.e. lagged exposures from 2–7 days before birth as well as 14, 30, and 60 days before birth). Averages were not calculated for exposure windows when the participant was missing data from more than 25% of the possible days in that window (i.e. if the participant did not have at least 45 daily estimates from the period 60 days before birth, then that average was not calculated).

Cord blood DNA methylation pre-processing

Cord blood was collected at birth by trained study personnel following individual study site protocols. DNA methylation was measured using the Infinium HumanMethylation450 BeadChip (450K) array or the Infinium MethylationEPIC v1 BeadChip (EPIC) array (Illumina, Inc., San Diego, CA, USA). Data from both platforms were preprocessed in parallel by the ECHO Data Analysis Center using the R package minfi 1.36.0 [36]. Samples were removed if they showed poor bisulfite conversion efficiency, low overall array intensity, low call rate, or mismatch between reported and predicted sex or if they were replicate samples. Probes were removed if they showed > 1% of samples with a detection P-value > 0.05. Cross-reactive probes and probes mapping to single nucleotide polymorphisms at the single base pair extension or CpG site were also removed [37]. Background signal and dye bias were corrected for using the normal-exponential using out-of-band probes method (noob) [38]. Samples were further removed if they showed predicted epigenetic age not corresponding to a sample from cord blood. We then applied a beta-mixture quantile probe type normalization and removed probes mapping to sex chromosomes [39]. DNA methylation values were calculated as M-values, which represent the log2 ratios of the intensity of the methylated versus the unmethylated probe. After merging with phenotypic data, samples were removed if they had an incomplete dataset. This resulted in 1197 samples from the 450K array (431 455 probes) and 451 samples from the EPIC array (736 881 probes). After combining the methylation and air pollution data, there were 1093 samples and 357 156 overlapping probes (Fig. S1).

Methodological approach

Our methodological approach is outlined in Fig. 1. For the overall ECHO Cohort, we calculated time-frame-specific (i.e. different sizes of prenatal window to average daily measures of PM_2.5, NO₂, and O₃) medians to establish ECHO Cohort-wide thresholds for “high” and “low” pollutant exposure. Medians were chosen as the threshold to reduce the effect of outliers. We examined the confounding potential for several important covariates, including race, ethnicity, maternal education, region where the mother spent the majority of the pregnancy, and DNA methylation platform. Only region and DNA methylation platform were associated with both air pollutant concentrations and DNA methylation variability, based on a principal component analysis (PCA). Site also showed some association with DNA methylation values based on the PCA. Data were, therefore, split evenly into training and validation sets with respect to cohort site, array platform, region of residence, and the exposure of interest to ensure similarity across the training and validation sets. Cell type composition and gestational age were not included in the prediction model although they are known to influence DNA methylation measures, as the goal of the analysis was prediction rather than inferring biological pathways.

In our previous work, we tested the positive predictive value of using 50 to 1 500 CpGs, increasing by units of 50. True positives were defined as CpGs with a P-value < 0.05 (between high and low exposure groups, split based on the pollutant median concentrations) and with the same direction of change in methylation in the training and test sets. Results from this pre-screening step supported the use of 500 CpGs, as increasing the number of CpGs from 500 did not provide additional information or improve predictive power [17]. We carried this determination over into the current analysis, where we applied two-sample t-tests to each CpG to first identify the top 500 CpGs associated with high versus low pollutant exposure by P-value in the training set. We did not adjust for multiple comparisons given that the goal of these tests was to inform modelling decisions. We ranked these 500 selected CpGs by Gini importance using a random forest model. We used the randomForest package in R with the default settings, which utilizes 500 decision trees, and the option for importance set to True.

Data in the validation set were further split into a validation training set (60%) and a validation test set (40%), balancing on exposure of interest, array platform, region, and site. Subsets of the selected 500 CpGs were entered into another random forest using the validation training data, with the CpGs added to this random forest in increments of 10 based on their importance ranking from the training model, resulting in 50 different random forest models using the validation training set for each pollutant. After running each random forest on the validation training set, we assessed the predictive performance of the model on the validation test set by calculating the AUROC. This process, starting from splitting the validation set into training and testing, was repeated 200 times, and AUROCs were averaged across all 200 iterations per CpG set. AUROCs reported throughout this paper correspond to the highest mean AUROC across all CpG sets.

We repeated this pipeline using the averaged pollutant concentrations for each trimester as well as average exposure in the two, seven, 14, 30, and 60 days before birth. We also tested the pipeline using two known predictors of DNA methylation levels, child sex, and preterm birth (defined as < 37 weeks gestation). We used a randomly generated binary variable as a negative control. This provided a range in which we would expect AUROCs to fall for the pollutant prediction models. We expected the pollutant model to perform better than a randomly generated binary variable, but perhaps not as well as more definitive characteristics like sex and preterm birth.

We re-ran the pipeline with a few variations. First, we tested categorization of “high” and “low” pollutant levels based on thresholds specified by the WHO and the United States EPA. Given the imbalanced groupings using these two thresholds, we additionally calculated the AUPRC for these prediction models. Second, because geographic and climatic regions may differ in the concentration and composition of air pollution, we ran the pipeline separately for participants whose delivery addresses were located in the 1) Northeastern region and 2) Western region of the United States. Although some participants resided in the Midwest and Southern regions at delivery, there were too few observations to make comparisons for these regions.

Sensitivity analyses

To assess the robustness of the pipeline across different sites, we conducted a sensitivity analysis using a leave-one-out approach. Here we repeated the pipeline leaving out participants from one of the sites during the training and validation phases. This process was repeated for each site.

CpG probe comparison across models

To provide further evaluation of the final models, we compared the probes that contributed to each model that used the cohort median cutoff for high/low exposure. This resulted in 24 sets of probes per pollutant (eight exposure windows, three regions), for a total of 72 sets of probes. The number of probes was determined by the number that produced the best mean AUROC. The probe set for a model could, therefore, range from 10 to 500 probes. The CpGs in each probe set and their corresponding models can be provided upon request.

Author contributions

Abigail Gaylord (Visualization [lead], Writing—original draft [lead], Writing—review & editing [lead]), Julie White (Formal Analysis [lead], Methodology [equal], Writing—original draft [supporting], Writing—review & editing [supporting]), Tessa Bloomquist (Writing—review & editing [supporting]), Carrie Breton (Investigation [equal], Writing—review & editing [supporting]), Carlos A Camargo (Investigation [equal], Writing—review & editing [supporting]), Andres Cardenas (Investigation [equal], Methodology [supporting], Writing—review & editing [supporting]), Elena Colicino (Investigation [equal], Writing—review & editing [supporting]), Dana Dabelea (Investigation [equal], Writing—review & editing [supporting]), Suman Ranjan Das (Investigation [equal], Writing—review & editing [supporting]), Rebecca C Fry (Investigation [equal], Writing—review & editing [supporting]), Jia Guo (Investigation [equal], Writing—review & editing [supporting]), Luhang Han (Investigation [equal], Writing—review & editing [supporting]), Christine W. Hockett (Investigation [equal], Writing—review & editing [supporting]), Margaret R. Karagas (Investigation [equal], Writing—review & editing [supporting]), Janine LaSalle (Investigation [equal], Writing—review & editing [supporting]), Tengfei Ma (Investigation [equal], Writing—review & editing [supporting]), Amy Margolis (Investigation [equal], Writing—review & editing [supporting]), Grier P Page (Investigation [equal], Writing—review & editing [supporting]), Frederica Perera (Conceptualization [supporting], Funding acquisition [supporting], Investigation [equal], Writing—review & editing [supporting]), Alicia K. Peterson (Investigation [equal], Writing—review & editing [supporting]), Sima Ramratnam (Investigation [equal], Writing—review & editing [supporting]), Annemarie Stroustrup (Investigation [equal], Writing—review & editing [supporting]), Rebecca Schmidt (Investigation [equal], Writing—review & editing [supporting]), Jennifer K.Straughen (Investigation [equal], Writing—review & editing [supporting]), Kristina Walker Whitworth (Investigation [equal], Writing—review & editing [supporting]), Shuang Wang (Conceptualization [lead], Methodology [lead], Supervision [lead], Writing—review & editing [supporting]), Julie B. Herbstman (Conceptualization [lead], Funding Acquisition [lead], Investigation [lead], Supervision [lead], Writing—review & editing [supporting]).

Conflicts of interest

The authors have no conflicts of interest to disclose.

Funding

Research reported in this publication was supported by the Environmental influences on Child Health Outcomes (ECHO) Program, Office of the Director, National Institutes of Health, under Award Numbers U2COD023375 (Coordinating Center), U24OD023382 (Data Analysis Center), U24OD023319 with co-funding from the Office of Behavioral and Social Science Research (Measurement Core), U24OD035523 (Lab Core), ES0266542 (HHEAR), U24ES026539 (HHEAR Barbara O’Brien), U2CES026533 (HHEAR Lisa Peterson), U2CES026542 (HHEAR Patrick Parsons, Kannan Kurunthacalam), U2CES030859 (HHEAR Manish Arora), U2CES030857 (HHEAR Timothy R. Fennell, Susan J. Sumner, Xiuxia Du), U2CES026555 (HHEAR Susan L. Teitelbaum), U2CES026561 (HHEAR Robert O. Wright), U2CES030851 (HHEAR Heather M. Stapleton, P. Lee Ferguson), UG3/UH3OD023251 (Akram Alshawabkeh), UH3OD023320 and UG3OD035546 (Judy Aschner), UH3OD023332 (Clancy Blair, Leonardo Trasande), UG3/UH3OD023253 (Carlos Camargo), UG3/UH3OD023248 and UG3OD035526 (Dana Dabelea), UG3/UH3OD023313 (Daphne Koinis Mitchell), UH3OD023328 (Cristiane Duarte), UH3OD023318 (Anne Dunlop), UG3/UH3OD023279 (Amy Elliott), UG3/UH3OD023289 (Assiamira Ferrara), UG3/UH3OD023282 (James Gern), UH3OD023287 (Carrie Breton), UG3/UH3OD023365 (Irva Hertz-Picciotto), UG3/UH3OD023244 (Alison Hipwell), UG3/UH3OD023275 (Margaret Karagas), UH3OD023271 and UG3OD035528 (Catherine Karr), UH3OD023347 (Barry Lester), UG3/UH3OD023389 (Leslie Leve), UG3/UH3OD023344 (Debra MacKenzie), UH3OD023268 (Scott Weiss), UG3/UH3OD023288 (Cynthia McEvoy), UG3/UH3OD023342 (Kristen Lyall), UG3/UH3OD023349 (Thomas O’Connor), UH3OD023286 and UG3OD035533 (Emily Oken), UG3/UH3OD023348 (Mike O’Shea), UG3/UH3OD023285 (Jean Kerver), UG3/UH3OD023290 (Julie Herbstman), UG3/UH3OD023272 (Susan Schantz), UG3/UH3OD023249 (Joseph Stanford), UG3/UH3OD023305 (Leonardo Trasande), UG3/UH3OD023337 (Rosalind Wright), UG3OD035508 (Sheela Sathyanarayana), UG3OD035509 (Anne Marie Singh), UG3OD035513 and UG3OD035532 (Annemarie Stroustrup), UG3OD035516 and UG3OD035517 (Tina Hartert), UG3OD035518 (Jennifer Straughen), UG3OD035519 (Qi Zhao), UG3OD035521 (Katherine Rivera-Spoljaric), UG3OD035527 (Emily S Barrett), UG3OD035540 (Monique Marie Hedderson), UG3OD035543 (Kelly J Hunt), UG3OD035537 (Sunni L Mumford), UG3OD035529 (Hong-Ngoc Nguyen), UG3OD035542 (Hudson Santos), UG3OD035550 (Rebecca Schmidt), UG3OD035536 (Jonathan Slaughter), UG3OD035544 (Kristina Whitworth). Abigail Gaylord also received funding from the National Heart, Lung, and Blood Institute (T32HL007343).

Data availability

Select de-identified data from the ECHO Program are available through NICHD’s Data and Specimen Hub (DASH). Information on study data not available on DASH, such as some Indigenous datasets, can be found on the ECHO study DASH webpage.