Metal mixtures and DNA methylation measures of biological aging in American Indian populations

Abstract

Introduction:

Native American communities suffer disproportionately from elevated metal exposures and increased risk for cardiovascular diseases and diabetes. DNA methylation is a sensitive biomarker of aging-related processes and novel epigenetic-based “clocks” can be used to estimate accelerated biological aging that may underlie increased risk. Metals alter DNA methylation, yet little is known about their individual and combined impact on epigenetic age acceleration. Our objective was to investigate the associations of metals on several DNA methylation-based aging measures in the Strong Heart Study (SHS) cohort.

Methods:

Blood DNA methylation data from 2,301 SHS participants was used to calculate age acceleration of epigenetic clocks (PhenoAge, GrimAge, DunedinPACE, Hannum, Horvath). Urinary metals [arsenic (As), cadmium (Cd), tungsten (W), zinc (Zn), selenium (Se), molybdenum (Mo)] were creatinine-adjusted and categorized into quartiles. We examined associations of individual metals through linear regression models and used Bayesian Kernel Machine Regression (BKMR) for the impact of the total metal mixture on epigenetic age acceleration.

Results:

The mixture of nonessential metals (W, As, Cd) was associated with greater GrimAge acceleration and DunedinPACE, while the essential metal mixture (Se, Zn, Mo) was associated with lower epigenetic age acceleration. Cd was associated with increased epigenetic age acceleration across all clocks and BKMR analysis suggested nonlinear associations between Se and DunedinPACE, GrimAge, and PhenoAge acceleration. No interactions between individual metals were observed. The associations between Cd, Zn, and epigenetic age acceleration were greater in never smokers in comparison to current/former smokers.

Conclusion:

Nonessential metals were positively associated with greater epigenetic age acceleration, with strongest associations observed between Cd and DunedinPACE and GrimAge acceleration. In contrast, essential metals were associated with lower epigenetic aging. Examining the influence of metal mixtures on epigenetic age acceleration can provide insight into metals and aging-related diseases.

1. Introduction

Metals are essential and nonessential elements that are present in the environment. They can be found in water, air, soils, and in many consumer products, leading to ubiquitous exposures. An ever-growing body of experimental and human population research implicates metals in many deleterious health effects, such as cardiovascular disease (CVD) and neurodegeneration (Jaishankar et al., 2014). This burden is disproportionately shared by American Indian communities, who are often exposed to higher concentrations of environmental toxicants via drinking water from unregulated private wells, proximity to legacy mining sites, or traditional diets (Lewis et al., 2017; Nigra et al., 2020). In the Strong Heart Study (SHS), a population-based prospective cohort study of American Indian communities from Arizona (AZ), Oklahoma (OK), and North and South Dakota (ND/SD), metal exposure is a serious public health concern. This population is exposed to higher levels of arsenic (As), cadmium (Cd), and tungsten (W) compared to other US populations (Pang et al., 2016) and has higher levels of urinary zinc (Zn) – probably because hyperglycemia due to a high burden of type 2 diabetes increases urinary Zn excretion (Galvez-Fernandez et al., 2022). Furthermore, compared to the rest of the United States, the SHS and other Native American communities have higher rates of aging-related diseases, such as heart disease, diabetes, and stroke (Howard et al., 1996; Lee et al., 2002, 1990). Metal exposures have been shown to contribute to disease in this cohort, particularly As and Cd exposure with CVD, diabetes, and peripheral arterial disease (Kuo et al., 2015; Newman et al., 2016; Nigra et al., 2018; Tellez-Plaza et al., 2013b, 2013a). Although As and Cd are now well-established risk factors for CVD, less is known about the health impacts of other metals such as W, molybdenum (Mo), selenium (Se), and Zn (Nigra et al., 2018). Additionally, some of these metals, such as Zn, Se, and Mo, are essential for biological function and their levels in biological samples might reflect metabolic alterations. These metal exposures often occur in mixtures with nonessential metals and may interact to impact health, complicating our understanding of aging-related diseases (Nigra et al., 2018).

A challenge in investigating how metal exposures contribute to risk for aging-related disease is that pathologic levels of exposure may accumulate decades in advance of disease onset. A new class of measures that aim to quantify biological processes of aging (Rutledge et al., 2022) can help address this challenge. These measures can reveal the progress and pace of aging processes in individuals for whom disease processes may still be latent. The current state-of-the-art in methods to quantify biological aging are algorithms derived from analysis of blood DNA methylation, often referred to as epigenetic clocks. Several of these clocks are strongly associated with patterns of aging-related disease and mortality and can reveal differences in risk among individuals who are the same chronological age (Binder and Horvath, 2022). Epigenetic clocks have the potential to serve as biomarkers for diseases or as the basis of preventative measures. Previous studies have shown an association between metal exposures and DNA methylation (Navas et al., 2005; Sobel et al., 2021; Tellez-Plaza et al., 2013b), but less is known about the associations of essential and nonessential metals with epigenetic age (Bozack et al., n.d.; Domingo-Relloso et al., 2020; Tellez-Plaza et al., 2014). For instance, previous work found that Zn was negatively associated with epigenetic aging in China (Xiao et al., 2021). Interrogation of the impacts of environmental exposures on independent epigenetic clocks can provide insight to the impact of metal exposures on aging-related diseases.

Our objective was to investigate the relationship between metal exposures and epigenetic age acceleration in the SHS cohort. We used Bayesian Kernel Machine Regression (BKMR) to model the associations between metal mixtures and epigenetic age acceleration from different epigenetic clocks, which allowed us to explore flexible, non-linear relationships between different mixture components (Bobb et al., 2015). We hypothesized that the nonessential metals (As, Cd, W) would be associated with greater biological age acceleration and faster pace of aging. In parallel, we hypothesized that higher levels of essential metals (Zn, Se, Mo) would be associated with lower biological age acceleration and slower pace of aging.

2. Methods

2.1. Study population

Participant samples and data were from the Strong Heart Study (SHS), consisting of 4,549 men and women ages 45–74 from 13 tribes from three study centers in South Dakota, Oklahoma, and Arizona (Stoddart et al., 2000). Baseline visits were conducted between 1989 and 1991. One tribal nation declined to participate in 2016, leaving 3,517 participants. Among those, urine and blood samples were collected in 2,351 participants at baseline who were free of cardiovascular disease and diabetes and had sufficient blood and urine for analysis. A further 26 participants were excluded based on failed blood DNA methylation analyses as previously described (Navas-Acien et al., 2021b). Participants missing data from key confounders and exposures (i.e. sex, metals, education level, smoking status, study center, BMI, eGFR, and fasting plasma glucose), including fasting glucose (n = 6), As (n = 1), and W (n = 3) were excluded. This left 2,320 participants with the necessary information on covariates (See 2.6 Statistics for full list of covariates), DNA methylation, and metal measurements. To meet modeling assumptions and reduce the influence of outliers on our models, removed individuals with metal concentrations greater than five standard deviations away from the mean, resulting in a sample size of 2,301 (Supplemental Fig. S1).

2.2. Sociodemographic and lifestyle assessment

Sociodemographic information (age [years], sex [male, female], years of education, study center [OK, AZ, SD], smoking status [current, former, never]) and medical history (Diabetes: present or absent) were collected at the baseline visit through an interviewer-administered standardized questionnaire. Education was converted to a categorical variable with 3 levels: less than high school (<10 years), high school or equivalent (10–15 years), and more than high school (>15 years). Physical examinations collected participant height and weight to calculate body mass index (BMI in kg/m²) along with urine and blood samples. Blood samples were used for fasting glucose and DNA methylation measurements. Given a large number of missing values for pack-years, and inconsistencies between self-reported smoking status and an individual’s corresponding smoking-pack years value, we also used DNA methylation data to impute an epigenetic based measure of smoking history with the EpiSmokEr R package (See 2.4 Blood DNA methylation measurement) (Bollepalli et al., 2019). This value was included in regression models along with the categorical smoking variable collected from the questionnaire.

2.3. Urinary metals assessments

Urine samples collected at the baseline visit were used to measure metal levels. Spot urine samples were stored in metal-free polypropylene screw cap tubes, frozen within 1–2 h of collection and shipped to Medstar Health Research Institute in Hyattsville, MD and stored at −80 °C. Sample aliquots (1 mL) were shipped from Medstar to the Trace Element Laboratory of the University of Graz, Austria, on dry ice between 2009 and 2010 and then stored at −80 °C (Navas-Acien et al., 2021a). Total metals (Cd, W, Zn, Se, and Mo) were measured by inductively coupled plasma mass spectrometry (ICP-MS) using a previously published protocol (Agilent 7700x ICP-MS; Agilent Technologies) (Scheer et al., 2012). Arsenic speciation was performed to calculate total urinary As was calculated as previously described (Navas-Acien et al., 2009). Limits of detection (LOD) for the metals were as follows: 0.1μ g/L for As species, 0.015μ g/L for Cd, 0.1μ g/L for Mo, 2.0μ g/L for Se, 0.005μ g/L for W, and 10.0μ g/L for Zn. LOD was calculated from three times the standard deviation of blanks, accounting for dilution factors and averaged over multiple batches. For measurements below LOD, the values were replaced by LOD divided by $\sqrt{2}$. For Mo, there were 26 samples below LOD (SHS, 1989).

All metals were standardized for urinary creatinine (μ g/g creatinine) to account for urinary dilution. Urinary creatinine was measured at the National Institute of Diabetes and Digestive and Kidney Diseases Epidemiology and Clinical Research Branch laboratory (Phoenix, AZ, USA) using an automated alkaline picrate methodology on a Alpkem Rapid Flow Analyzer in duplicate or triplicate (Lee et al., 1990; SHS, 1989). Estimated glomerular filtration rate (eGFR) was calculated using the Chronic Kidney Disease - Epidemiology Collaboration formula (Levey et al., 2009). eGFR was included as a covariate in adjusted models to account for the role of kidney function on metals excretion (Jin et al., 2018).

2.4. Blood DNA methylation measurement

Blood specimens were collected during physical exams at the initial visit and stored at < −70 °C. DNA from blood cells was extracted and stored under a strict quality control system at the Penn Medical Laboratory, MedStar Health Research Institute. Genomic DNA was bisulfite-converted using the EZ DNA Methylation kit (Zymo Research, Irvine, CA), eluted in buffer, and DNA methylation was measured at a single nucleotide resolution of > 850,000 CpG sites using the Infinium MethylationEPIC BeadChip (850 K) (Illumina, San Diego, CA). The R package, minfi (version 1.18.4) function, preprocessNoob, was used for single sample normalization to account dye-bias and background correction (Fortin et al., 2017; Triche et al., 2013). Quality checks, preprocessing, and beta-value calculations were done through minfi functions as previously described (Domingo-Relloso et al., 2020; Fortin et al., 2017).

Batch effects for plate and by row were corrected using the ComBat function in sva R packages (version 3.36.0). Cell type proportions (CD4T, CD8T, NK, Monocytes, B cells, granulocytes) were estimated using the Houseman projection method (Houseman et al., 2012). Additionally, cross-hybridizing probes, and probes located on sex chromosomes or those on SNPs with a minor allele frequency > 0.05, were removed from analysis (McCartney et al., 2016). EpiSmokEr, a DNA methylation-based predictor of smoking behavior, was used to more precisely define a continuous value of smoking history in the absence of complete information on pack years (Bollepalli et al., 2019) (Supplemental Fig. S2). The EpiSmokEr R package calculates the probability of an individual being classified as a current smoker, which was used directly as a covariate in regression analyses in addition to the questionnaire information on smoking status.

2.5. Calculation of epigenetic clocks

We computed DNA methylation age using algorithms developed by Hannum et al. (Hannum et al., 2013), Horvath et al. (Horvath, 2013), Levine et al. (PhenoAge) (Levine et al., 2018), Lu et al. (GrimAge) (Lu et al., 2019), and Belsky et al. (Belsky et al., 2022) [Dunedin Pace of Aging (DunedinPACE)]. For each clock, except for the DunedinPACE, which is already calculated as a rate, age acceleration was calculated by extracting the residuals from the regression of chronological age on each epigenetic age measure, unadjusted for cell types. We selected these five clocks as they each provide a unique perspective on biological aging (Li et al., 2022). However, for simplicity of interpretation, we present the results from the “second-generation” clocks (i.e. GrimAge, PhenoAge and DunedinPACE), trained on health and mortality as primary analyses, and the “first-generation” clocks (i.e. Hannum and Horvath), trained on chronological age, as secondary analyses. Each clock was developed using penalized elastic net regression methods (Zou and Hastie, 2005). For full information on each clock, please see the Supplemental Materials. Horvath’s online calculator (https://dnamage.genetics.ucla.edu) was used to calculate epigenetic age using Horvath, GrimAge, and PhenoAge methods. We calculated age acceleration by extracting the residuals from the regression of chronological age on each epigenetic age measure, unadjusted for cell types. These residuals were included as the independent variable in regression analyses. DunedinPACE was calculated using the software available on GitHub (https://github.com/danbelsky/DunedinPACE).

2.6. Statistical analysis

We estimated Spearman correlation coefficients between the measures of epigenetic age, the calculations of epigenetic age acceleration, and urinary metals (As, Cd, W, Zn, Se, Mo).

All models were run with epigenetic age acceleration calculated as the residuals from models of epigenetic age and chronological age instead of raw epigenetic age, to better estimate the impact of metals on biological aging. To account for complex relationships between metals and health, metals were divided into quartiles based on the creatinine-adjusted values for regression models. Linear regression models were estimated to calculate the association between each measure of epigenetic age acceleration by individual metal quartiles, with quartile 1 as the reference category:

$$ageaccelerationresiduals={\beta }_0+{\beta }_1(quartile2)+{\beta }_2(quartile3)+{\beta }_3(quartile4)$$

We performed three progressively-adjusted models to examine the associations between each individual metal and an epigenetic age acceleration measure. Model 1 was adjusted for sex, chronological age, estimated cell type proportions (CD4T, CD8T, NK, Monocytes, and B cells), and five genetic principal components to account for genetic ancestry. Model 2 included the factors in Model 1 and additionally included education level (<high school, high school, >high school), smoking status (current, former, never), EpiSmokEr probability values (Bollepalli et al., 2019), study center (AZ, OK, and ND/SD), BMI, and eGFR (Levey et al., 2009). eGFR was included to account for variability in metal excretion determined by kidney function. Model 3 included all variables from Model 2 as well as fasting plasma glucose due to the strong impact of hyperglycemia on Zn excretion in urine (Farooq et al., 2020). The SHS population has a high burden of uncontrolled diabetes, including participants with official diagnoses and or biomarkers that meet the criteria for diabetes (Howard et al., 1996; Wang et al., 2011, 2002). Effect estimates from linear models can be interpreted as the mean difference in epigenetic age acceleration between quartile 1 and the indicated quartile for each metal.

2.6.1. Bayesian Kernel Machine Regression

BKMR (Bayesian Kernel Machine Regression) is a powerful mixture method that can identify important mixture components, account for multicollinearity between exposures, and estimate non-linear and non-additive interactions (Bobb et al., 2018). Metals were included in the models as continuous, log2-transformed, and scaled values. We used 50,000 iterations of the Markov chain Monte Carlo to provide posterior distributions for variables. We inspected trace plots for convergence of each component in each model. We included hierarchical variable selection for pre-specified groups of essential and nonessential metals (As, Cd, W & Zn, Se, Mo), which provides posterior inclusion probabilities (PIPs) for each group and component within the group. PIPs indicate the relative contribution of a mixture component on the outcome.

We present our findings as the cumulative effect of the metal mixture on each measure of age acceleration, the dose–response relationship for a metal and age acceleration with others fixed at their median, and metal–metal interactions. For each age acceleration measure, the cumulative effect is interpreted as the change in mean outcome (i.e. epigenetic age acceleration) in comparison to the 50th percentile when all metals are increased in 5 percentile increments from the 25th to 75th percentile (Bobb et al., 2018). Univariate exposure–response plots estimate the shape of the dose–response relationship between individual metals and each outcome, with all other metals fixed at their median. Interactions between two metals are represented by bivariate response plots, which plot the response function of one metal while the second is held at the 10th, 50th and 90th percentile. In addition to a BKMR model including all metals, we also ran separate models for essential and nonessential elements. For each of these models, the metals belonging to the other group were added to the covariate matrix so that the essential elements BKMR was accounted for the nonessential metals and vice versa. All analyses were performed using R (version 4.1.2), and BKMR specifically used the bkmr package (Bobb et al., 2018).

2.6.2. Effect modification

Based on previous research, we hypothesized a priori that participant sex, study center (SD, OK, AZ), and smoking status (never, current/former) may modify the impact of metals on epigenetic age acceleration (Kuo et al., 2022; Lieberman-Cribbin et al., 2022; Tellez-Plaza et al., 2013a). Hence, we examined effect modification by these variables by including an interaction term with each metal quartile. To calculate strata specific estimates, stratified models, without the interaction term, were created for each category of the potential effect modifier. Models were adjusted for covariates from Model 3 above.

2.6.3. Sensitivity analysis

We conducted several sensitivity analyses to determine the robustness of our findings. First, we removed the cell type adjustments from our models to determine the impacts of metals on DNA methylation aging that may proceed through alterations in cell type proportions. Next, we removed EpiSmokEr values from our regression to determine the role of this smoking estimate on our results. Additionally, as BMI and eGFR may be mediators of the association between metals and DNAm age, we repeated Model 2 removing these factors. Finally, the Horvath Skin and Blood Clock (“SkinBlood” herein) is an updated measure of chronological DNAm age specifically optimized for forensic applications in skin and blood and previously shown to be a better estimate of chronological age than the original Horvath clock (Horvath et al., 2018). To determine if our results were similar using this updated clock, we used the methylclock R package to calculate the SkinBlood age in our participants (Pelegí-Sisó et al., 2021), then regressed the values on chronological age to obtain the residuals as above. We repeated our adjusted linear models as above using the SkinBlood age acceleration as the outcome.

3. Results

3.1. Population characteristics

The demographic and clinical characteristics of study participants are included in Table 1. About 42% of participants were male, and 71% of participants identified as current or former smokers. The mean ± standard deviation (SD) chronological age of participants was 51.2 ± 8.09 years and for BMI it was 30.3 ± 6.1 kg/m². Women had higher concentrations of all metals compared to men (Table 1). Spearman correlations showed that metal concentrations were moderately correlated, the lowest correlations observed between Cd and W (Spearman’s rho = 0.25) and the highest correlations observed between Zn and Se (Spearman’s rho = 0.66) (Supplemental Fig. S3A). The mean ± SD epigenetic age acceleration was −0.03 ± 6.56 years for PhenoAge, 1.4E-3 ± 4.54 years for GrimAge, −0.02 ± 4.89 years for Hannum, and −0.03 ± 5.03 years for Horvath (Table 2). The DunedinPACE mean ± SD was 1.12 ± 0.13. DunedinPACE is denominated in Pace-of-Aging units, i.e. years of aging-related physiological deterioration per chronological year. The raw DNAm clock measures (Hannum, Horvath, PhenoAge, and GrimAge) were highly correlated (Spearman’s rho range: 0.67 – 0.83) (Supplemental Fig. S3B). However, the correlations between clocks calculated as age acceleration were weaker (Spearman’s rho range: −0.04 – 0.61) (Supplemental Fig. S3C). DunedinPACE was most strongly correlated with GrimAge acceleration (Spearman’s rho = 0.55) (Supplemental Fig. S3C).

Table 1.

Population characteristics of participants included in analyses of metals and epigenetic aging.

Characteristics	Total (N = 2301)	Female (N = 1343)	Male (N = 958)
Age (years) [mean ± SD] Centera [N (%)]	56.2 ± 8.09	56.5 ± 81.5	55.6 ± 8.00
SD	1022 (44.4)	579 (43.1)	443 (46.2)
OK	973 (42.3)	567 (42.2)	406 (42.2)
AZ	306 (13.3)	197 (14.7)	109 (11.4)
Education [N (%)]
< HSb	544 (23.6)	326 (24.3)	218 (22.8)
HS or equivalent	1462 (63.5)	858 (63.9)	604 (63.0)
>HS	295 (12.8)	159 (11.8)	136 (14.2)
Smoking Status [N (%)]
Current	889 (38.6)	474 (35.3)	415 (43.3)
Former	738 (32.1)	375 (27.9)	363 (37.9)
Never	674 (29.3)	494 (36.8)	180 (18.8)
BMI (mean ± SD)	30.3 ± 6.10	30.8 ± 6.28	29.6 ± 5.76
Diabetes [N (% Yes)]c	959 (41.7)	596 (44.4)	185 (19.3)
eGFRd (mean ± SD)	97.4 ± 16.8	96.3 ± 17.6	98.84 ± 15.5
Fasting Glucosee (mean ± SD)	139 ± 68.2	143 ± 72.8	133.2 ± 60.7
Metals (mg/ g creatinine) [Median (IQR)]
As	11.6 (6.97, 19.1)	12.5 (7.6, 20.2)	10.6 (6.44, 17.6)
Cd	0.97 (0.62, 1.5)	1.17 (0.77, 1.78)	0.71 (0.47, 1.10)
W	0.11 (0.06, 0.22)	0.12 (0.07, 0.24)	0.10 (0.05, 0.19)
Zn	556.3 (399, 797.3)	608 (438, 877)	505 (360, 709)
Se	49.1 (36.7, 67.3)	51.9 (39.6, 71.7)	45.1 (34.1, 61.2)
Mo	29.7 (20.6, 41.4)	31.7 (22.4, 43.9)	26.8 (18.0, 38.1)

^aSouth/North Dakota (SD); Oklahoma (OK); Arizona (AZ).

^bHigh School

^cIncident T2DM was defined as fasting plasma glucose ≥ 126 mg/dL or a 2-h postload plasma glucose level ≥ 200 mg/dL or HbA1c ≥ 6.5% or self-reported use of insulin or ora l diabetes treatment.

^dFasting glucose determined by enzymatic methods.

^eEstimated glomerular filtration rate (eGFR).

Table 2.

Epigenetic clock values for participants included in the analysis (N = 2301).

Clock	DNAm Age (mean ± SD)	DNAm Age (Median, IQR)	Acceleration* (mean ± SD)	Acceleration* (Median, IQR)
Hannum (years)	50.5 ± 8.08	49.7 (44.4, 56.2)	−0.02 ± 4.89	−0.23 (−3.22, 2.87)
Horvath (years)	58.3 ± 7.97	57.9 (52.5, 63.8)	−0.03 ± 5.03	−0.06 (−3.13, 3.17)
PhenoAge (years)	50.1 ± 9.51	49.5 (43.3, 56.7)	−0.03 ± 6.56	−0.13 (−4.45, 4.31)
GrimAge (years)	40.7 ± 7.41	40.0 (35.1, 445.7)	1.4E-3 ± 4.54	−0.5 (−3.41, 2.80)
DunedinPACE (physiological change / year)			1.12 ± 0.13	1.12 (1.04, 1.21)

^*Acceleration is estimated as a rate for DunedinPACE and as a difference between the actual biological age vs the expected biological age based on birthdate using a residual based approach.

3.2. Associations of metals with DNAm age acceleration in linear models

For ease of interpretation, we present the age acceleration results from “second-generation” clocks (i.e. GrimAge, PhenoAge and DunedinPACE), trained on health and mortality, as primary analyses, and the “first-generation” clocks (i.e. Hannum and Horvath), trained on chronological age, as secondary analyses. In our primary models adjusting for all covariates (Model 3), we observed robust associations between Cd and DunedinPACE [mean difference in rate between 1st and 4th quartile Cd (95% CI) = 0.02 (0.01, 0.04)] and GrimAge acceleration [mean difference in years of acceleration between 1st and 4th quartile Cd (95% CI) = 1.05 (0.7, 1.41)] across all quantiles (Fig. 1, Supplemental Table S1). The 3rd quartile of Zn was also associated with GrimAge acceleration [mean difference in years between 1st and 3rd quartile Zn = 0.45 (0.14, 0.75)] and DunedinPACE [mean difference in rate between 1st and 3rd quartile Zn = 0.01 (0, 0.03)]. No robust associations were observed between epigenetic age acceleration and Se, Mo, As, and W or with PhenoAge acceleration.

Fig. 1. Forest plots for Model 3 associations between metal quartiles and measures of epigenetic age acceleration.

Effect estimates with 95% confidence intervals are the mean difference in age acceleration in years for each metal quartile in comparison to the first quartile, except for DunedinPACE, which is unitless. Models are adjusted for sex, estimated cell type proportions (CD4T, CD8T, NK, Monocytes, and B cells), genetic principal components, education level, smoking status, EpiSmokEr probability values, study center, BMI, eGFR, and fasting plasma glucose.

In minimally adjusted models (Model 1), Cd and Zn were positively associated with epigenetic age acceleration for all clocks, with the strongest associations observed for GrimAge acceleration and DunedinPACE (Supplemental Fig. S4, Supplemental Table S1). The mean difference (95% CI) for the DunedinPACE rate comparing Cd quartiles 2–4 vs. the lowest quartile was 0.03 (0.01, 0.04), 0.04 (0.03, 0.06), 0.05 (0.04, 0.07); the corresponding mean difference (95% CI) in years for GrimAge acceleration was 1.41 (0.97, 1.85), 2.83 (2.38, 3.28), 3.68 (3.15, 4.10). Additionally, Mo was negatively associated with GrimAge acceleration. The mean difference (95% CI) for GrimAge acceleration (in years) comparing quartiles 2–4 vs. the lowest quartile of Mo was −0.47 (−0.93, −0.01), −0.59 (−1.05, −0.13), and −0.55 (−1.01, −0.08). When adjusting for BMI, eGFR, smoking status, education, EpiSmokEr, and study center in Model 2 the association of Mo with GrimAge acceleration was no longer significant and associations with Cd and Zn were attenuated (Supplemental Fig. S5, Supplemental Table S1). No consistent associations were observed between As, Se, or W and the epigenetic age acceleration. Results for first-generation clocks can be found in Supplemental Figs. S4–S6.

3.3. Metal mixture analysis with BKMR

The group-level posterior inclusion probabilities (PIPs), which indicate the relative contribution of a mixture component on the outcome, for nonessential metals were greater than the PIPs for essential metals in GrimAge acceleration, and DunedinPACE while the opposite was true for PhenoAge acceleration (Supplemental Fig. S7A). Individual metal PIPs were greatest for Cd and Se across models, which passed our threshold of 0.5 for DunedinPACE, GrimAge acceleration, and PhenoAge acceleration, indicating that these elements were included in the models >50% of the time (Supplemental Fig. S7B) (Barbieri and Berger, 2004). The PIP was also > 0.5 for W with GrimAge acceleration.

Estimated increases in the concentrations of total metals from their 20th to 50th percentile in the population was associated with an increase in GrimAge acceleration (Fig. 2). As PIPs suggested varying importance of essential and nonessential metals to the models by clock, we conducted additional BKMR analyses stratified by essential (Zn, Se, Mo) and nonessential (As, Cd, W) metals and examined their overall effects. Estimated GrimAge acceleration and DunedinPACE decreased at higher levels of the essential metal mixture, yet increased with greater levels of nonessential metals (Fig. 2). There were no meaningful changes in PhenoAge acceleration for either metal group.

Fig. 2.

Plots of the overall metal mixture (circles) on each epigenetic clock (as age acceleration or rate) from BKMR models, as well as separated by essential (triangle) and nonessential (square) metals (bottom panels). Plots represent the change in epigenetic age acceleration when all metals are at their median by either increasing or decreasing all metals in decile increments.

To investigate the linearity of associations between metals and epigenetic age acceleration, we generated exposure–response curves for individual metals when all other metals are fixed at their medians (Fig. 3). The findings for Cd were largely robust for all clocks. Se was non-linearly negatively associated with age acceleration for GrimAge, PhenoAge, and DunedinPACE (Fig. 3). Results for first-generation clocks can be found in Supplemental Figs. S8 and S9.

Fig. 3.

Individual exposure–response plots for each metal and epigenetic age acceleration measure with all other metals fixed at their median. Shaded areas represent 95% credible intervals and the dashed line represents an effect estimate of zero.

To investigate interactions between metal:metal pairs, we generated bivariate interaction plots for each clock (Supplemental Fig. S10 – S14). No interactions between metal pairs were observed, except for an indication of an interaction between Cd and Zn/Se on Horvath acceleration (Supplemental Fig. S13).

3.4. Effect modification by sex, study site and smoking status

We next examined effect modification by sex, study site, and smoking status in fully adjusted models (Model 3), focusing our description of results on those that were detected in our main analyses and followed clear patterns across measure of epigenetic age and metal quartiles. Smokers were categorized into a binary variable of Never (n = 674) vs. Current/Former (n = 1627) for ease of interpretation. Effect modification by smoking status was observed for all clocks except DunedinPACE, with associations for Cd and Zn greater among never smokers than current/former smokers (Supplemental Table S2). For instance, in comparison to the first quartile, the fourth quartile of Cd in never smokers was associated with a mean difference of 1.56 (0.07, 3.04) years of PhenoAge acceleration and a mean difference of −0.02 (−0.98, 0.93) years of PhenoAge acceleration in current/former smokers (interaction p-value = 0.01). Furthermore, increasing quartiles of Zn were positively associated with GrimAge acceleration in smokers but not never-smokers with a mean difference of 0.91 (0.31, 1.51) years in never-smokers and 0.26 (−0.11, 0.62) in current/former smokers (interaction p-value = 0.03). Most associations did not differ by sex; however, the association between Cd and DunedinPACE was stronger in men than in women (interaction p-value = 0.004) (Supplemental Table S3). Finally, although most associations did not differ by study site, marked differences were observed in associations in those living in SD in comparison to OK and AZ for GrimAge acceleration (Supplemental Table S4).

3.5. Sensitivity analyses

Removing immune cell types from the model also did not meaningfully impact our conclusions; however, some effect estimates were slightly stronger suggesting that some of the impact of metals on epigenetic age acceleration may proceed via alterations in immune cell function (Supplemental Table S5). We reasoned that BMI and eGFR may be on the causal pathway for the relationship between metals and epigenetic age acceleration, so we repeated Model 2 without these adjustments (Supplemental Table S6). No meaningful differences in our effect estimates were observed.

To determine the ability of EpiSmokEr values to control for confounding due to smoking, we repeated our models not adjusting for EpiSmoker (Supplemental Fig. S15). We find that without adjustment for EpiSmokEr, effect estimates for associations between Cd and all clocks are increased. In contrast, associations with other metals are unchanged, suggesting that EpiSmokEr provides some level of control of confounding due to smoking given the high levels of Cd in tobacco smoke.

The mean SkinBlood age and SkinBlood age acceleration values were 62 ± 8.66 years and 0 ± 0.4 years, respectively. SkinBlood age acceleration was moderately correlated with Horvath and Hannum acceleration (Pearson’s r = 0.67 and 0.61, respectively). Linear regression analyses of metals and SkinBlood acceleration in fully adjusted model (Model 3) were not meaningfully different from associations of metals with Horvath acceleration and Hannum acceleration (Supplemental Table S7).

4. Discussion

Here we present a comprehensive analysis of the associations of metal mixtures with epigenetic aging in a population of American Indians. Epigenetic clocks are in wide use in aging-related research. These clocks were developed with different approaches and there are varying levels of evidence for their validity as biomarkers of risk for aging-related health decline. PhenoAge, GrimAge, and DunedinPACE have stronger evidence of association with morbidity and mortality and with risk factors for shorter and less healthy lives (Belsky et al., 2022). In parallel, these clocks showed stronger associations with metals exposure in the present study, likely because the CpG sites included in these clocks are more indicative of biological processes both underlying aging and impacted by metal exposures. We find that increasing concentrations of nonessential metals were associated with increased GrimAge acceleration, while essential metals were associated with decreased acceleration. This is consistent with previous work comparing the utility of the different epigenetic clocks at predicting disease and mortality (Belsky et al., 2022; Lu et al., 2019; Sugden et al., 2022; Wang et al., 2021).

To date, few studies have examined the associations between epigenetic age acceleration and metal mixtures. Nwanjai-Enwerem et al. studied metal exposures and age acceleration within a smaller cohort of elderly US Veterans from Normative Aging Study (n = 48), and observed a significant association between PhenoAge acceleration and manganese (Mn) exposure (Nwanaji-Enwerem et al., 2020). In another study of elderly Chinese participants (n = 288) that examined metal mixtures and epigenetic age acceleration, increasing blood Zn levels were associated with decreased Horvath, GrimAge, and Hannum acceleration (Xiao et al., 2021).

In the present study, we observed robust positive associations between urinary Cd and DunedinPACE and GrimAge acceleration. Cd is a toxic, nonessential element that has previously been associated with the development of several diseases, including several cancers and CVD (Lamas et al., 2021). Cd toxicity is thought to occur via oxidative stress, DNA damage, and cell death (Đukić-Ćosić et al., 2020); however, experimental studies also suggest that Cd can modify DNA methyltransferase activity to alter DNA methylation (Domingo-Relloso et al., 2020; Poirier and Vlasova, 2002; Takiguchi et al., 2003; Yuan et al., 2013). In fact, previous work in the SHS identified six CpGs associated with Cd exposure in an epigenome-wide associations study, with many sites overlapping with tobacco smoke exposure (Domingo-Relloso et al., 2020). Tobacco smoke is a significant source of Cd exposure (Genchi et al., 2020). Pack-years and other smoking biomarkers were included in the development of the GrimAge calculator (Lu et al., 2019). In addition to the numerous health effects linked to smoking and secondhand smoke exposure, GrimAge acceleration is elevated among those who smoked in childhood and had parents who were smokers (Klopack et al., 2022). Smoking is a significant risk factor for many aging-related diseases, such as cancers and lung disease. However, in stratified analyses, associations between Cd and epigenetic aging biomarkers appeared to be stronger in never smokers than current/former smokers, suggesting that smoking status was not driving our findings.

Zn is an micronutrient, with roles in over 3000 proteins that provide a variety of essential functions (Maret, 2013). Urinary Zn does not reflect Zn intake, but rather increased urinary loss of Zn related to hyperglycemia (Galvez-Fernandez et al., 2022). We observed positive associations between the third quartile of Zn and DunedinPACE, PhenoAge, and GrimAge acceleration. As Zn is an essential metal, both Zn deficiency and elevated exposures can lead to adverse health impacts, suggesting that nonlinear associations between Zn and epigenetic clocks may be related to complex Zn regulation. Zn is a key co-factor in the one-carbon metabolism pathways required for DNA methylation (Azimi et al., 2022). A previous epigenome-wide study in a lung cancer cohort from China found 28 CpGs enriched in biological pathways for antioxidant stress, protein transmembrane transport, and mitosis, associated with plasma Zn levels (Meng et al., 2022). Interestingly, research shows direct Cd and Zn interactions in CVD and hypertension (Lin et al., 2014; Ponteva et al., 1979; Thind and Fischer, 1976) and suggests that Zn may be protective against Cd toxicity (Jemai et al., 2007; Yu et al., 2021). Studies on DNA methylation of imprinted genes suggest that Zn can modify the associations of Cd (Vidal et al., 2015). We observed an interaction between Zn and Cd on Horvath acceleration, but not on other clock measures. These associations require greater exploration in future studies.

Finally, we observed nonlinear associations of urinary Se concentrations with DunedinPACE, PhenoAge, and GrimAge acceleration in our BKMR models. Se is also an essential element with a key role as a cofactor in several enzymes involved in oxidative stress and that has been non-linearly associated with cardiovascular disease risk (Bleys et al., 2008; Zhao et al., 2022). Indeed, previous research in vitro suggests that elevated Se has the potential to inhibit DNA methyltransferase activity, whereas Se deficiency may interfere with one-carbon metabolism (Speckmann and Grune, 2015).

We also investigated the association of the overall metal mixture with epigenetic aging biomarkers and observed an increase in GrimAge acceleration with increasing total metals levels. Previous studies found increases in the mixture of As, Cd, Mn, Pb, and Hg, to be associated with greater PhenoAge acceleration, but not GrimAge (Nwanaji-Enwerem et al., 2020). Furthermore, when models were split into essential and nonessential metals, increasing concentrations of nonessential metals were associated with increased epigenetic aging, while essential metals were associated with decreased epigenetic aging. This is likely due to the known antioxidant and essential role of essential metals, in contrast to nonessential metals that have no known function in the body. Essential metals and nonessential metals often undergo complex interactions, with many sharing metal transporters and opposing antioxidant and oxidant functions (Goyer, 1997). For instance, Zn and Cd compete for binding to metallothionein, which plays a key role in essential metal homeostasis (Brzóska and Moniuszko-Jakoniuk, 2001; Goyer, 1997), and Zn has been suggested as a counter measure for Cd toxicity (Yu et al., 2021). Nonetheless, our results indicate that even when essential and nonessential metals are combined, the adverse impact of metals on epigenetic age acceleration persists, suggesting that essential metals do not fully counteract nonessential metals. Future studies should extend metals measures to include key essential and nonessential metals not measured in this analysis. Few studies have compared the impacts of essential vs nonessential metals on epigenetic age acceleration. Previous work examining essential [magnesium (Mg), calcium (Ca), vanadium (V), Mn, iron (Fe), cobalt (Co), nickel (Ni), copper (Cu), Zn, Se, and Mo] and nonessential [titanium (Ti), arsenic (As), rubidium (Rb), strontium (Sr), Cd, barium (Ba), and lead (Pb)] metals observed similar results with Horvath age acceleration. While the observed trends are similar, the discrepant findings between clocks may be due to the metals included in the analysis, the use of urinary vs blood-based metal biomarkers, and the population studied.

The urinary metal levels measured in this study population are generally higher compared to other populations in the US. Compared to the SWAN cohort, composed of women, the participants in the current analysis had higher median levels of urinary Cd, W, and Zn (Wang et al., 2022, Wang et al., 2020). Furthermore, the current sample from the SHS cohort had higher median levels of urinary As, Cd, and W; and lower median levels of urinary Mo for all cycles of NHANES from 2011 to 2018. Metals levels in urine can represent different exposure durations depending on the toxicokinetics and half-life of each metal. Urine is the preferred biomarker for Cd and generally reflects body burden exposure (Vorkamp et al., 2021). Arsenic in urine has a half-life of approximately 2–4 days (Bügel et al., 2008), whereas Se is excreted on the order of days to months depending on the species absorbed (Alexander, 2015). Zn and Mo can be rapidly excreted in urine when levels exceed the bodily requirements (Nordberg et al., 2014).

This study has several notable strengths, including the use of a community-engaged, understudied population of American Indians from across the US, a large sample size, and the inclusion of advanced mixture methods. Furthermore, epigenetic clocks provide several key advantages over studying the direct phenotypic impacts of metals on aging. First, they encompass multiple components of the biological aging process that span across aging-related diseases rather than examining each disease individually. Second, they are predictive of future disease before the aging phenotype manifests, which allows us to capture changes in the biological aging process in advance of the disease.

We also acknowledge several limitations. Our analysis was cross-sectional, limiting our ability to draw inference on the direction of the observed associations. The mechanisms behind these associations remain unknown, as epigenetic aging biomarkers were designed to have more predictive than mechanistic value. Furthermore, since this work began, more refined reference libraries for cell type composition estimates from DNA methylation data have been established (Salas et al., 2022). Future work should evaluate these expanded cell type estimates to determine the full impact of cell type composition on epigenetic aging. Furthermore, the clocks used in this study were developed in cohorts of predominantly White participants, and thus might not accurately predict epigenetic aging in our population of American Indians (Philibert et al., 2020). To account for genetic variability, we adjusted for five principal components in our models. However, we are unable to compare the effectiveness of these clocks between different races with our current study. Research in this area is necessary to develop inclusive measures of biological aging. However, American Indians are at particularly high risk of aging related diseases, such as diabetes and cardiovascular disease, and at greater risk of metals exposure (Lewis et al., 2017).

5. Conclusions

In conclusion, Cd, Zn, and Se were associated with epigenetic biomarkers of biological aging, with stronger associations observed in “second-generation” clocks. These findings suggest that metals exposures, in particular higher levels of Cd and Zn in the urine, may be related to biological processes that underly important aging-related diseases. Given that American Indian populations suffer disproportionately from diseases such as CVD and diabetes in comparison to other race/ethnic groups within the US, these observations would benefit from future research examining longitudinal exposures and biological aging outcomes and a research focus on untangling the epigenetic mechanisms of environmental disease.

Data availability

The authors do not have permission to share data.