We collected blood samples via venipuncture from all study participants and isolated DNA. We measured DNA methylation in all three studies utilizing the HumanMethylation450 BeadChips (Illumina HM450K arrays) and applied standard processing pipelines and quality controls as described previously (Phillips et al., 2019). We calculated all epigenetic aging biomarkers utilizing the Horvath’s new online calculator (http://dnamage.genetics.ucla.edu/). The outcomes of interest were EAA and age-adjusted DNAmTL (DNAmTLAdj), defined as the residuals of epigenetic aging and DNAmTL biomarkers linearly regressed on chronological age and pre-selected confounding covariates.

We calculated five different epigenetic clocks: Horvath’s Pan-Tissue Clock, Skin-Blood Clock, Hannum’s Clock, PhenoAge (Levine’s Clock) and GrimAge. Two additional measures of EAA were calculated that reflect different aspects of immunological aging: Extrinsic Epigenetic Age Acceleration (EEAA) is derived from Hannum’s Clock, which upweights the contributions of blood cell composition known to change with age (naïve cytotoxic T cells, cytotoxic T cells, and plasmablasts), and Intrinsic Epigenetic Age Acceleration (IEAA) is derived from the Pan-Tissue Clock, which is independent of age-related changes in leukocyte composition (Chen et al., 2016). Although Hannum’s EAA measure is known to be correlated with blood cell counts, its extrinsic EAA derived measure has been shown to predict all-cause mortality more strongly. Similarly, while the pan-tissue EAA is generalizable across tissues, correlations with leukocyte cell counts remain; however, its derived measure of intrinsic EAA has been shown to better reflect systemic cell-type aging, as it is preserved across tissues and cells collected from the same individual (Chen et al., 2016; Marioni et al., 2015). Therefore, these two measures provide a nuanced understanding of biological aging related to age related immune changes (EEAA) and biological aging across cells and tissues independent of immune changes (IEAA).

In addition, we tested associations with estimates of DNAmTL, adjusted for age and confounders. In this paper, we will generally refer to the adjusted DNAmTL biomarker as a measure of epigenetic age acceleration. However, this biomarker is a predictor for telomere length in units kb, and therefore, it is not necessarily a measure of telomere length itself but rather correlated with it.

On average, we expect the age acceleration (adjusted residuals) for the cohort of interest to be zero for all epigenetic clocks. However, a particular subgroup may on average have negative or positive values for the EAA measures in which case the subgroup may be viewed as being, on average, biologically younger or older than those not in the subgroup, even after adjusted for chronological age and confounders. When comparing the EAA of multiple subgroups, one should interpret the EAA values relative to the other subgroups (i.e., compare the differences in the EAA values between subgroups).

We modeled the residuals as a measure of biological age acceleration, as opposed to the difference in predicted and chronological age for each individual, as these measures become uncorrelated with predicted age and chronological age. Additionally, to calculate epigenetic residuals measures, we adjusted residuals models for BMI, smoking, recent alcohol consumption and self-reported recent infection. We did not adjust the EEAA and IEAA measures for confounders other than regressing chronological age. In sensitivity analyses, we tested age acceleration residual associations uncorrected for confounders. The results were similar but correcting for potential confounders generally improved the precision of estimates.

We described each study with means and proportions for the variables analyzed and evaluated accuracy of epigenetic aging biomarkers via their empirical correlation with chronological age as well as scatterplots of all six epigenetic aging biomarkers across all three studies. Additionally, for the TCE study, raw measurements of the DNA telomere length were compared to the DNAmTL predictions via stratified scatter plots and correlations. To test for differences in epigenetic age acceleration (EAA) measures between exposure groups, we employed non-parametric statistical tests for differences in the median and distribution between the exposure groups. We defined the main test statistic of interest as the difference in the median value of age acceleration between exposure groups. We employed permutation-based tests to obtain exact p-values for the difference in the median age acceleration measure for each pair of exposure groups across all epigenetic aging biomarkers. The median test allows for exact inference and is preferable for its robustness against outliers, which were prevalent in the predictions of the epigenetic clocks. It should be noted that EAA residuals are obtained by performing a linear regression on all the data. To specifically test the null hypothesis of conditional independence, we derived and applied a conditional permutation test, whereby the exposure group labels are randomly permuted and the null permutation distribution of the median difference in residualized epigenetic age is computed and obtained quasi-exact p-values for the null hypothesis. The residualized epigenetic age is defined as the residuals of the linear regression of the unadjusted epigenetic age outcome on chronological age and the selected confounders (BMI, smoking, recent alcohol consumption and self-reported recent infection), omitting the exposure group assignment. By quasi-exact, we mean that the test has exact finite-sample type-1 error control against the null hypothesis of marginal independence and asymptotically exact type-1 error control against the null hypothesis of conditional independence. This and the permutation nature of the test suggest the method has better finite-sample control than asymptotic model-based methods. For robustness, we used conditional quantile regression for the EAA measures, adjusting for the pre-selected confounders of age, alcohol intake, sex, BMI, and smoking history in the models as sensitivity analyses for epigenetic clocks. To assess the continuous association between exposure and age acceleration, we performed Spearman’s test of correlation, which is a nonparametric test of monotonicity, and report p-values. We considered a p-value<0.05 as statistically significant monotonic association between exposure and age acceleration. We perform the trend test (p_trend) across all participants, pooling across the control and exposed participants.

We expected that both the median permutation test and quantile regression estimates would give similar results and control type I errors. We report 95% confidence intervals (95% CIs) for differences in medians using exposure-group stratified bootstraps along with exact p-values. We did not adjust for multiple testing but provide estimates and measures of uncertainty for interpretability. All analyses were performed with the R statistical software (https://www.R-project.org/).

A total of 98 workers were analyzed in the original benzene study: 48 were controls and 50 participants were exposed to benzene with a mean exposure of 6.02 ppm (SD=12.91). In this study, the mean chronological age was 31.2 (SD=8.3) years. In the TCE study, a total of 128 workers were analyzed. Of these, 68 were controls exposed to <0.0047 ppm TCE, 25 were exposed to levels between >0.005 to <10 ppm with an average of 4.22 ppm (SD=2.56) and 35 were exposed to levels ≥10 ppm, with an average of 38.7 ppm (SD=47.07). The mean age of study participants was 26.9 (SD=6.9 years). Participant characteristics for the benzene and TCE studies are described in Table 1. In the formaldehyde study, 39 control workers were exposed on average to 0.02 ppm (SD=0.007), while 31 exposed workers had a mean exposure of 1.25 ppm (SD=0.62). In this current study, the mean age was 30.5 (6.8) years. Participant characteristics for the formaldehyde study are shown in Supplementary Table S1.

In the benzene study, correlations between epigenetic aging biomarkers and chronological age were high, ranging from r=0.93 for the Skin-Blood and Grim Age Clocks to r=−0.64 for DNAmTL estimates. For the TCE study, epigenetic clock predictions for chronological age were also high, ranging from r=0.92 for the Skin-Blood Clock to r=−0.73 for DNAmTL. A slightly lower performance was observed in the formaldehyde study, with the highest correlation of r=0.88 observed in the Skin-Blood Clock to r=−0.68 for DNAmTL (Figure 1). Across all studies, the Skin-Blood Clock had the highest correlation, while the DNAmTL showed the lowest absolute correlations. For the TCE study, qPCR measurements of the DNA telomere length were available and compared to the DNAmTL estimated using Pearson’s test of correlation. Across all TCE study participants, the correlation was moderate between DNAmTL and qPCR measured leukocyte telomere length (r=0.31; p=0.0019) (Supplementary Figure S1-A), but similar in magnitude to the original DNAmTL publication. We found that the control and lower exposure group’s DNA telomere length measurements respectively had a moderate correlation of r=0.47 (p=0.0026) and r=0.51 (p=0.0087) with DNAmTL, while the higher exposure group had a nonsignificant correlation of r=−0.0051 (p=0.98) (Supplementary Figure S1-B). We additionally compared the difference in correlation between the exposure groups using the Fischer-transform test of difference in correlation. We found that the correlation of the control and lower exposure groups were not significantly different from each other (p=0.59), but both the correlations comparing the higher exposure group to the control (p=0.044) and lower (p=0.02) exposure groups were significantly different (Supplementary Table S2).

In Supplementary Table S3, we report the Spearman correlation matrix between the age acceleration measures of all clocks (residuals of predicted age based on the adjusted linear model). All age acceleration measures were uncorrelated with age. We found that IEAA and the Horvath Pan-Tissue Clock are highly correlated (r=0.94), and the EEAA and the Hannum EAA were also highly correlated (r=0.95). This is expected since the IEAA is derived from the Horvath Pan-Tissue Clock, and similarly, the EEAA measure is derived from the Hannum-Blood Clock. As expected, the residuals of DNAmTL were negatively correlated with all EAA measures.

When comparing unexposed controls (<0.035 ppm, n=48) to benzene exposed workers (n=50), we observed that the benzene exposed group had a median of 1.75 years greater EAA for the Skin-Blood Clock (95% CI: 0.27, 3.28; p=0.012) and 0.09 kb shorter DNAmTLAdj (95% CI: −0.18, −0.001; p=0.023) relative to controls (Table 2). No significant results were found for the median EAA of the other clocks. Results for all epigenetic aging biomarkers and their association with benzene exposure are summarized in Table 2. The results for difference in mean EAA between the control and exposed groups using adjusted quantile regression can be found in Supplementary Table S4A, for controls vs. exposed, and Supplementary Table S4B, for controls vs. lower vs. higher exposure groups. Similar to the median permutation test results, unexposed controls were observed to have a median EAA adjusted difference for the Skin-Blood Clock that is 1.74 years greater than the exposed group (p=0.008), and median DNAm-age adjusted telomere length 0.10 kb shorter than the exposed group (p=0.037). No significant differences were observed for the remaining mean EAA measures.

To assess the association of higher benzene exposure and age acceleration, we further classified the exposure groups into controls (<0.035 ppm, n=48), a lower exposure group (<10 ppm, mean 1.35 ppm, n=41) and higher exposure group (≥10 ppm, mean 27.3 ppm, n=9). When comparing the lower exposure to the control group, EAA for the Skin-Blood Clock remained increased by 1.69 years (95% CI: −0.09, 3.17; p=0.034). The higher exposure group had a median increase of 3.01 years in EAA for the Skin-Blood Clock (95% CI: 0.01, 4.98; p=0.019) relative to controls. Median adjusted DNAmTL residual was lower in the higher exposure group by −0.25 kb (95% CI: −0.47, 0.06; p=0.0038) and by −0.09 kb (95% CI: −0.15, 0.02; p=0.06) for the lower exposure group relative to controls. Benzene results across the three exposure groups are summarized in Table 3. When testing for trends across all participants, we found a significant increasing trend in EAA for the Skin-Blood Clock (p=0.0021) and a significant decreasing trend in DNAmTLAdj (p=0.0213). Associations for EAA for each clock across exposure groups are shown in Figure 2. No further significant trends were reported.

In the TCE study, no significant differences were observed when comparing controls to all TCE exposed workers (Supplementary Table S5) and no consistent exposure-response patterns were present. When comparing controls, lower and higher exposure groups certain exposed groups were significantly different from controls. Namely, control workers had a median Skin-Blood Clock EAA of −0.54 years (n=71) compared to 1.63 years among workers exposed to <10 ppm of TCE (n=27; mean TCE = 4.22 ppm; p=0.035) (Table 4). The results for the adjusted median difference in epigenetic aging between exposure groups estimated with quantile regression can be found in Supplementary Table S6, with consistent estimate of age acceleration observed between workers exposed to <10 ppm and controls of 2.13 years (95% CI: 0.37, 4.84).

No statistically significant differences were observed between formaldehyde exposed workers relative to controls by permutation testing or estimates of quantile regression. The results are shown in Supplemental Table S7.