Metabolomic Evaluation of Air Pollution-related Bone Damage and Potential Mediation

Ambient air pollution has been associated with bone damage. However, no studies have evaluated the metabolomic response to air pollutants and its potential influence on bone health in postmenopausal women. We analyzed data from WHI participants with plasma samples. Whole-body, total hip, femoral neck, and spine BMD at enrollment and follow-up (Y1, Y3, Y6). Daily particulate matter NO, NO2, PM10 and SO2 were averaged over 1-, 3-, and 5-year periods before metabolomic assessments. Statistical analyses included multivariable-adjusted linear mixed models, pathways analyses, and mediation modeling. NO, NO2, and SO2, but not PM10, were associated with taurine, inosine, and C38:4 phosphatidylethanolamine (PE), at all averaging periods. We found a partial mediation of C38:4 PE in the association between 1-year average NO and lumbar spine BMD (p-value: 0.032). This is the first study suggesting that a PE may partially mediate air pollution-related bone damage in postmenopausal women.


Introduction
Osteoporosis, a condition that weakens bones, making them more susceptible to sudden and resulting in bone fractures, occurs especially during aging and remains as a leading cause of disability, particularly in women, reducing independence and quality of life, 1,2 and increasing morbidity and mortality in older individuals. 3,4 About 2.1 million bone fractures are reported each year in the US, resulting in up to $20.3 billion in annual direct health costs, which could be dramatically reduced through early detection and effective preventive strategies. 5 Risk factors for osteoporosis and bone fractures during aging include smoking, family history, low body mass index, long-term use of corticosteroids, anorexia/bulimia, heavy drinking, and long-term inactivity, 6 but osteoporosis impacts women more than men, with 80% of the estimated 10 million Americans with osteoporosis being women. 7 The postmenopausal period is a critical stage that enormously contributes to osteoporosis and fracture risk. Postmenopausal women have the highest fracture rates of any stage of lifespan. 8 The absolute risk of fractures is much higher among postmenopausal women than pre-and perimenopausal women. Also, one in two women over 50 will experience a bone fracture because of osteoporosis. 9 Therefore, studying risk factors for fracture postmenopausal women, such as factors that may have deleterious effects on bone density, can identify those at higher risk of bone fractures and allow us to target such individuals for early intervention strategies. evaluated, including 2-phosphoglycerate, alpha-glycerophosphate, C38:5 PE, alphaglycerophosphocholine, CMP, and UMP (Fig. 2, panel D). For the one-year average, metabolic pathway enrichment analysis with Metaboanalyst revealed several altered metabolic pathways (Fisher exact test p-value < 0.05), including purine, glycerophospholipid, as well as ascorbate and aldarate metabolism, among others (Fig. 2, panel E, left). For the 3-year average, metabolic pathway enrichment analysis revealed alterations in several pathways, including purine, glycerophospholipid, as well as ascorbate and aldarate metabolism, among others (Fig. 2, panel E, center). For the 5-year average, metabolic pathway enrichment analysis revealed alterations in purine, ascorbate and aldarate metabolism, and glycerophospholipid metabolism, among others (Fig. 2, panel E, right).
Metabolites and pathways associated with NO 2 After adjustment for multiple hypotheses (i.e., Bonferroni), we found that NO 2 exposure, at the three different averaging periods (1-, 3-, and 5-year average before plasma sampling), was associated with several metabolites, often showing a negative association after adjusting for multiple comparisons (Bonferroni < 0.05). At the 1-year average window, NO 2 was signi cantly associated with 27 metabolites, including 2-phosphoglycerate, UDP-galactose/UDP-glucose, GDP, ADP, GMP, UDP, taurine, lactose, and niacinamide, among others (Fig. 3, panel A). A complete list of metabolites, estimates, and p-values (raw and adjusted) for the 1-year time window is shown in Supplementary Table 5. At the 3-year average window, NO 2 was signi cantly associated with 31 metabolites, including 2-phosphoglycerate, GDP, UDPgalactose/UDP-glucose, ADP, niacinamide, GMP, UDP, hexose monophosphate, AMP, UMP, C38:4 PE, and taurine, among others (Fig. 3, panel B). A full list of metabolites, estimates, and p-values (raw and adjusted) for the 3-year time window is shown in Supplementary Table 6. At the 5-year average window, NO 2 was also signi cantly associated with 28 metabolites, including UDP, ADP, GDP, L-threo-Sphingosine, hexose monophosphate, UDP-galactose/UDP-glucose, C38:4 PE, and taurine, among others (Fig. 3, panel C). We found 26 shared metabolites between the three averaging periods evaluated, including 2phosphoglycerate, 6-phosphogluconate, adenosine, ADP, C38:4 PE, lactose, niacinamide, sucrose, taurine, and others (Fig. 3, panel D). A full list of metabolites, estimates, and p-values (raw and adjusted) for the 5-year time window is shown in Supplementary Table 7. Metabolic pathway enrichment analysis revealed similar alterations for the three averaging periods evaluated, including purine metabolism, glycerophospholipid metabolism, ascorbate and aldarate metabolism, pyrimidine metabolism, butanoate metabolism, starch and sucrose metabolism, pentose and glucuronate interconversions, pyruvate metabolism, and others, which were similar between the three averaging periods evaluated (Fig. 3, panel E).

Metabolites and pathways associated with SO 2
After adjustment for multiple hypotheses (i.e., Bonferroni), at the 1-year average window, SO 2 was not signi cantly associated with any of the metabolites evaluated ( Supplementary Fig. 1, panel A). The full list of metabolites, estimates, and p-values (raw and adjusted) for the 1-year time window is shown in Supplementary Table 8. However, at the 3-year average window and after adjustment for multiple hypotheses (i.e., Bonferroni), SO 2 was signi cantly associated with eight metabolites, including GDP, 6phosphogluconate, ADP, GMP, adenosine, hexose diphosphate, UDP-galactose/UDP-glucose, and AMP ( Supplementary Fig. 1, panel B). A full list of metabolites, estimates, and p-values (raw and adjusted) for the 3-year time window is shown in Supplementary Table 9. At the 5-year average window, SO 2 was signi cantly associated with only three metabolites, including taurine, ADP, and GMP ( Supplementary  Fig. 1, panel C). We found three shared metabolites between the 3-and 5-averaging periods, including ADP, GMP, and taurine ( Supplementary Fig. 1, panel D). A full list of metabolites, estimates, and p-values (raw and adjusted) for the 3-year time window is shown in the Supplementary Table 10. Metabolic pathway enrichment analysis revealed ve altered metabolic pathways (Fisher exact test p-value < 0.05), including purine metabolism, ascorbate and aldarate metabolism, pentose and glucuronate interconversions, amino sugar and nucleotide sugar metabolism, and taurine and hypotaurine metabolism. Similar pathways were found associated with the 5-year average averaging periods ( Supplementary Fig. 1, panel E).

Metabolites and pathways associated with PM 10
After adjustment for multiple hypotheses (i.e., Bonferroni), we found that PM 10 at three different averaging periods (1-, 3-, and 5-year average before plasma sampling) was not signi cantly associated with any of the available metabolites ( Supplementary Fig. 2). A full list of metabolites, estimates, and pvalues (raw and adjusted) for the 1-, 3-, and 5-year averaging periods are shown in Supplementary Air pollution-related metabolites also linked with BMD We tested whether air pollution-related metabolites were also linked with absolute bone mineral density (in g/cm 2 ) and T-scores. In total, 36 metabolites were explored for this association. After adjustment for multiple hypotheses (i.e., Bonferroni), we found that C38:4PE, C38:5PE, taurine, and lactose were signi cantly associated with lumbar spine BMD (Fig. 4, panel A). For total hip BMD, only inosine and CMP were signi cantly associated at a Bonferroni-adjusted threshold (Fig. 4, panel B). For femoral neck BMD, only inosine was signi cantly associated (Fig. 4, panel C). For total body BMD, no air pollution-related metabolites were signi cantly associated (Fig. 4, panel D). Metabolic pathway enrichment analysis showed four altered metabolic pathways, including taurine and hypotaurine metabolism, glycosyphosphatidylinositol (GPI)-anchor biosynthesis, glycerophospholipid metabolism, and primary bile acid biosynthesis (Fig. 4, panel E). A full list of metabolites, estimates, standard errors, and p-values for the association with BMD are shown in Supplementary Table 14. Similar associations were observed for T-scores. A full list of metabolites, estimates, standard errors, and p-values for the association with Tscores are shown in Supplementary Table 15. Some changes were observed when sensitivity models explored the effects of smoking, alcohol consumption, physical activity, dietary modi cation or hormone therapy trials (Supp. Figure 3-7) on the association between air pollutants and BMD.

Mediation modeling
We ran mediation modeling focusing on pollutants (i.e., nitrogen oxide), metabolites (i.e., inosine, taurine, C38:4 PE, C38:5 PE), and anatomical sites (i.e., lumbar spine BMD) with most substantial potential for mediation. Table 1 shows the mediation of the NO-lumbar spine BMD association by C38:4 PE. Following Baron and Kenny's steps for mediation models, 25 we determined the role of air pollution-related metabolites in lumbar spine BMD (b path). Although values of beta for the b path differ from the null for C38:4 PE, taurine, and lactose, no signi cant indirect effect was found for taurine or lactose. We found a partial mediation of the association between NO and lumbar spine BMD by C38:4 PE (Table 1) for the 1year average time window, where C38:4 PE signi cantly accounted for 31% of the association (p-value: 0.032, raw p-value). No other metabolites showed statistical signi cance for mediation for any pollutant and averaging periods.

Discussion
This study provides the rst evidence of a metabolomic response mediating air pollution-related bone damage, particularly in a highly susceptible population: postmenopausal women. We identi ed an intense metabolomic response to two of the four pollutants evaluated, NO and NO 2 , but not SO 2 or PM 10 .
We also found that some of these metabolites were associated with BMD, particularly at the lumbar spine, total hip, and femoral neck. Remarkably, mediation models showed that C38:4 PE partially mediated the effect of NO on lumbar spine BMD. Our results suggest, for the rst time, the involvement of critical bioactive phospholipids in air pollution-related bone damage.
In our study, we evaluated a large number of metabolites determined using an untargeted approach, including water-soluble metabolites (HILIC-pos and HILIC-neg columns), free fatty acids and bile acids (C18-neg columns), and polar and nonpolar plasma lipids (C18-pos). Initially, we observed extensive negative associations between NO exposure and multiple metabolites, indicating a potential vulnerability to long-term ambient air pollution. Metabolites commonly identi ed as associated with air pollutants included ADP, CMP, and taurine. Pathway analyses identi ed 16 metabolic pathways perturbed with longterm exposure to NO, including purine, glycerophospholipid, ascorbate and aldarate, and butanoate metabolism, among others. We also identi ed 11 metabolic pathways perturbed by long-term exposure to NO 2 , many shared with NO x , including purine, glycerophospholipid, ascorbate and aldarate, and butanoate metabolism, but not others, such as pyruvate metabolism, glycolysis/gluconeogenesis, and linoleic acid metabolism. SO 2 exposure also showed an association with purine and ascorbate and aldarate metabolism, but also with amino sugar and nucleotide metabolism, as well as taurine and hypotaurine metabolism. Interestingly, no metabolomic response was observed in response to PM 10 .
These data suggest that air pollutants could trigger similar metabolomic pathways such as purine metabolism. Studies from Weichtal and colleagues have suggested that postmenopausal women are particularly susceptible to air pollution (i.e., NO), increasing their health risk of, for example, breast cancer. 26 Ascorbate and aldarate metabolism has been reported as important for antioxidant activity in the human body, potentially acting as the rst line of defense against inhaled pollutants. 27 Our team has previously reported results from mixture analyses using Bayesian kernel machine regression models suggesting that bones in postmenopausal women may be particularly susceptible to air pollutants, especially NO. 18 The current study offers the initial clues on the potential mechanisms underlying air pollution-related bone damage. To our knowledge, our results are the rst to suggest strong and speci c metabolomic signatures of long-term air pollution exposure in this susceptible population.
Arginine, one of the most versatile amino acids, is metabolically interconvertible with the amino acids proline and glutamate. 28 It serves as a precursor for the synthesis of protein, nitric oxide, creatine, polyamines, agmatine, and urea. 28 Similar to our ndings for 1-and 3-year average NO exposure, a shortterm (48 hour) exposure study by Liang and colleagues, identi ed arginine metabolism (as well as histidine, γ-linolenic acid, and hypoxanthine metabolism) as signi cantly associated with tra c indicators, including black carbon, carbon monoxide, nitrogen oxides, and ne particulate matter. In ammatory and oxidative stress-related pathways, such as leukotriene and vitamin E metabolism, were also associated in that study, but not ours (except by aldarate metabolism that was not found in Liang's study). 29 Also, in a randomized study of clean air interventions, Li and colleagues found that arginine metabolism also was associated with PM 2.5 exposure. 30 These results suggest that arginine is consistently associated with air pollution and could be used as an exposure biomarker for biological effects of air pollutants.
On the other hand, Li and colleagues reported that purine metabolism was also linked to PM 2.5 exposure, a report consistent with our nding of NO, NO 2 , and SO 2 associations with purine metabolism at all averaging periods evaluated (except for 1-year average SO 2 ). 30 Purine metabolism plays an essential role in nucleic acid synthesis. Hu and colleagues showed that polycyclic aromatic hydrocarbon (PAH) and other air pollutant (e.g., PM 10 , NO 2 , and SO 2 ) exposures were also linked with purine metabolism. 31 Other studies on the effect of indoor air pollutants in older individuals with chronic obstructive pulmonary disease have also revealed a potential effect on purine metabolism. 32 Other studies on the metabolomic response to air pollutants 33,34 did not show metabolites to those observed by us, although the methodological design limits comparability between studies.
Air pollution-related metabolites, including taurine, lactose, C38:4 PE and C38:5 PE (two phosphatidylethanolamines), inosine, and CMP, were associated with bone mineral density, particularly at the lumbar spine, but also on femoral neck and total hip. Zhang and colleagues have previously identi ed 27 metabolites associated with femoral neck BMD. 35 The glycine, serine, and threonine metabolism pathway (including four identi ed metabolites: creatine, dimethylglycine, glycine, and serine) were associated with BMD and improved the prediction and the classi cation of osteoporotic fracture risk beyond conventional risk factors. 35 Early studies from You and colleagues using proton nuclear magnetic resonance spectroscopy suggested that, among postmenopausal women, elevated glutamine was signi cantly associated with low BMD and that elevated lactate, acetone, lipids, and very low-density lipoprotein were associated with high BMD. 36 Miyamoto and colleagues, in unadjusted analyses, showed that postmenopausal women might show differential levels of metabolites such as pyruvate, lactate, succinate, urea, and creatine, among others in low-estrogen and low-BMD. Coincident with our results, they reported taurine as reduced in postmenopausal women with low BMD. 37 Absence of common metabolites or pathways between these results and ours, except for taurine, may suggest that air pollution response is not related to age-or hormone-related bone damage. Taurine (2-aminoethane-1sulfonic acid) is a sulfur-containing amino acid with a β-amino group and an acidic sulfonic group (R-SO3H) separated by two methylene (CH 2 ) moieties. In humans, taurine plays a functional role in vital organs, such as the brain, eyes, kidneys, and heart. It performs several primary physiological functions, including osmotic regulation, and has antioxidant, antiapoptotic, and anti-in ammatory effects. 38 Although taurine is not a structural component of proteins, it is metabolically involved in many processes that in uence bone development and promote osteoblastogenesis. Studies in postmenopausal Brazilian women with osteopenia or osteoporosis have shown reduced taurine levels in plasma compared to healthy volunteers. 39 Similar results were observed in White pre-menopausal women in the USA, Japanese women with low estradiol and BMD levels, and older Chinese adults with low BMD. 40 Metabolomic signatures associated with age-related BMD reductions go further than our intended research. Further research about the potential implications of taurine in bone metabolism for air pollutionrelated bone mineral reductions is guaranteed.
Our results from mediation modeling suggest a critical role of C38:4 PE, a phosphatidylethanolamine, in the association between air pollution (i.e., NO) and bone damage. For many years, long-term exposures to air pollution have been linked to oxidative stress. 41 Reactive oxygen species (ROS) include unstable shortlived molecules that contain oxygen (O 2 ., H 2 O 2 , and OH − ) and are highly reactive in cells. 42 Superoxide anion may give rise to a variety of reactive carbonyl species called reactive aldehydes (RA). RAs can be more destructive than ROS, as their lives are longer (minutes to hours), and their structure allows them to migrate long distances and induce damage to cellular components, such as phospholipids, including phosphatidylethanolamines (PEs). PEs are considered the second most abundant phospholipid in mammalian cells and comprises about 15-25% of the total lipid in mammalian cells, after phosphatidylcholine. 43 PE is enriched in the inner lea et of membranes, and it is especially abundant in the inner mitochondrial membrane. 43 Due to its conical shape, PE modulates membrane curvature and lateral pressure 44,45 and thus supports membrane fusion 46-48 and the function of several membrane proteins. 49,50 Oxidative damage by RAs on PE may induce PE adducts (e.g., HNE-Michael adducts, HNE-Schiff adducts, and ONE-Schiff adducts), 42 potentially affecting cell membrane domains stability (lipid rafts), including bone cells. PE-derived adducts have been proposed as mediators of RA effects on membrane proteins. 42 In fact, evidence suggests that PE added to metal implants increases mesenchymal stem cell osteoblastogenesis. 51 Osteoclasts, one essential component of bone resorption, require PE for osteoclastogenesis. 52 Studies from Irie and colleagues have found that immobilization of the cell surface PE blocked osteoclast fusion, suggesting a critical role of PE abundance and distribution for osteoclast generation. 52 In this context, we hypothesize that air pollution-induced oxidative damage could be inducing RAs that interact with bone cells PE, particularly mononuclear pre-osteoclasts, affecting osteoclastogenesis and, therefore, bone resorption and bone physiology in general. A summary of this potential physiopathological mechanism is shown in Supplementary Fig. 8.
Like any other study, ours has some limitations. First, as this study was performed only in the US and included only menopausal women, these results may not be generalized to other populations. However, our population included White, Black, Hispanic, and women from other races/ethnicities, contributing to the potential generalizability of the results. Also, our and other studies have suggested the effect of air pollutants on men. Therefore, it is necessary to con rm whether this metabolomic response is sexspeci c or if there is a common metabolomic response and mediation between men and women. Another limitation of our study was our inability to include PM 2.5 , which may also trigger bone damage.
Unfortunately, most air pollution predictions were made in the 1990s, and the US EPA Federal Reference Method network for PM 2.5 was established in 1999, 53 this fraction was not included in the analysis. Our results could also be in uenced by self-correlation (e.g., NO and NO 2 are pollutants with similar sources).
However, these self-correlated pollutants exposure occur in similar contexts. On the other hand, our study has several strengths, including a well-validated metabolomics platform, detailed covariate information with a prospective design (baseline and year 1 of follow-up), and a robust methodology.

Conclusions
This is the rst study showing a strong metabolomic response to criteria air pollutants, with potential effects on bone mineral density, suggesting that PEs are critical mediators of air pollution-related damage in postmenopausal women. Studies that discover the potential implications of these metabolites and pathways on clinically relevant outcomes (i.e., bone fractures) are needed. Experimental intervention studies, increasing the levels of those metabolites negatively affected by air pollution and directly linked with bone health, are also encouraged by our results.

Population
The population included in this study was drawn from the Women's Health Initiative -Observational Study (WHI-OS) and Clinical Trials (WHI-CT) of postmenopausal United States women who enrolled between 1993 and 1998. 54 All participants provided written informed consent. In the original metabolomic study, 55 784 participants who developed CHD after the baseline examination (cases) were selected, and 787 participants who did not develop CHD were matched on 5-year age, race/ethnicity, hysterectomy status, and 2-year enrollment groups. From them, 140 participants with CHD and 138 controls were included in the current study based on the availability of air pollution and BMD data.
Air pollution exposure data WHI participant addresses from study initiation to date have been geocoded. 56,57 Geocoded participant address-speci c daily mean concentrations of PM 10 (µg/m 3 ) )between 1993 and 2012 have been spatially estimated using available US Environmental Protection Agency Air Quality System (AQS) data and national-scale, log-normal, ordinary kriging. 58-60 Analogous participant address-speci c daily mean concentrations of gaseous pollutants (nitrogen oxides [NO, [parts per million, ppm]; nitrogen dioxide [NO 2 , ppm]; and sulfur dioxide [SO 2 , ppm]) also have been estimated using the same methods. In addition, monthly mean geocoded participant-address speci c concentrations of PM 10 have been spatiotemporally estimated using generalized additive mixed models and geographic information system-based predictors. The pollutant-, duration-and model-speci c estimates were averaged over one, three, and ve years before (and ending on) dates of plasma metabolomic assessment. As most of the air pollution predictions were made in the 1990s and the US Environmental Protection Agency (EPA) Federal Reference Method network for particulate matter < 2.5 µm (PM 2.5 ) was established in 1999, 53 this particle fraction was not included in the analysis.

Metabolomic data
Plasma samples were collected using EDTA tubes in the WHI at baseline and Y1 and processed immediately. All WHI specimens were stored in a − 70°C freezer within 2 hours of collection or stored at − 20°C for up to 2 days, then shipped on dry ice and stored at − 70°C until processing. The majority of the WHI samples had been thawed once before our study. Metabolomic measurements were made using four complementary liquid chromatography-tandem mass spectroscopy (LC-MS) methods resulting in 371 metabolites. For each liquid chromatography (LC) method ([HILIC]-positive, C8-positive, C18-negative), pooled plasma reference samples were included after every 20 samples, and results were standardized using the ratio of the value of the sample to the value of the nearest pooled reference multiplied by the median of all reference values for the metabolite. Other details about the metabolomic assessment of these samples are described elsewhere. 55 Bone mineral density assessments The WHI Bone Density Substudy, in which BMD (g/cm 2 ) was measured at enrollment, year 1, year 3, and year 6 clinic visits, included all participants at three clinical centers (Birmingham, AL; Pittsburgh, PA; and Tucson, AZ) chosen to maximize racial diversity; and a satellite clinic (Phoenix, AZ) (N = 11,020). Those participants without air pollution estimations were excluded. Therefore, for the current study, we analyzed data from the WHI -Clinical trial (CT) and WHI -Observational Study (OS) participants (n = 9,041 participants: n = 4,202 from the CT and n = 4,839 from the OS). Participants of this WHI BMD cohort underwent dual-energy x-ray absorptiometry (DXA) measurement using Hologic machines (QDR2000, 2000+, or 4500). Quality assurance methods included cross-clinic calibration phantoms and a review of a random sample of scans. When the Hologic QDR 2000 machines were upgraded to QDR 4500 machines, in vivo cross-calibration procedures were performed, and results were adjusted for these correction factors and for longitudinal changes in scanner performance. T-score was also evaluated as an outcome. The reference standard from which the T-score is calculated is the female, White, age 20-29 years, NHANES III database. 61 We used measurements of BMD at the total hip, lumbar spine, and total body in g/cm 2 from DXA scans performed on each participant at enrollment and the years 3 and 6 clinic visits.

Relationship between air pollutants and circulating metabolites
We performed the necessary data preprocessing and quality control steps, such as checking for normality and the presence of outliers. In the case of non-normal distributions of continuous variables, data transformation (e.g., log transformation) was used. In the case of missing data, multiple imputation protocols were established to maintain statistical power. For the effect of air pollutants on the metabolome, we examined whether 1-, 3-, and 5-year mean pollutant (PM 10 , SO 2 , NO, and NO 2 ) concentrations prior to the examination cycle (metabolomic assessment) were longitudinally associated with the relative abundance of the metabolites. We t linear mixed-effects models for continuous outcomes and adjusted for potential confounders (i. . For the linear mixedeffect models, we assumed the following model: Y ij = α 0 + β 1 Z 1i + β 2 X 2ij + b 0i + e ij , where Y ij and X ij represent the metabolomic levels and the potential confounders at the j th measurement on the i th subject, and Z 1i is the pollutant concentration for the i-th individual at time point j. b 0i and e ij correspond to random effects and random error effects, respectively. We used Bonferroni correction for multiple testing

Relationship between circulating metabolites and bone mineral density
Only air pollution-related metabolites were evaluated for associations with BMD using multivariableadjusted linear mixed models. Models were adjusted for potential confounders (i.e., age, body mass index, ethnicity, education, and coronary heart disease [yes; no] while controlling for the repeated evaluation [baseline and visit 1]).

Mediation analysis
We ran mediation models following Baron and Kenny's steps. 25 We assessed whether the effect of key air pollutants on target BMD anatomical sites was mediated by those metabolites associated with air pollutants and BMD using multiple mediator models (Supp. Figure 9). Coe cients were obtained from a linear regression analysis. Indirect effects were calculated using the product-of-coe cients method (a*b). 62 Standard errors and con dence intervals for mediation analyses were calculated by bootstrapping (5000 samples). 63 Outcome and mediating variables were adjusted for age (years), race/ethnicity (White vs. non-White), body mass index (kg/m 2 ), and education level (primary vs. higher education). The direct effect (c' path) did not have to be reduced to zero because an incomplete mediation of the effect was expected. Also, for signi cant mediating variables, the proportion mediated was calculated as an effect size measure ((a*b)/c). 64 Because of the low number of metabolites included in the mediation models, no correction for multiple testing was included. We performed the analyses using R software (R Project for Statistical Computing, CRAN, The Comprehensive R Archive Network, Vienna) and the command mediate() in the 'mediation' package for bootstrapping. A Bonferroni-adjusted p-value < 0.05 identi ed results as statistically signi cant.

Pathway Identi cation
To better understand the biological signi cance of the metabolites, we conducted pathway analysis using metabolites signi cantly associated (at p-value < 0.05) with air pollution and the top metabolites that have the greatest weighting on the signi cant independent component analysis-factor(s). We used the "Pathway Analysis" functionality in MetaboAnalyst 5.0, which accounts for both over-representation (i.e., how many signi cant metabolites fall within a given pathway) and pathway topology (i.e., how important those metabolites are to that pathway) 65 using the Human Metabolome Database for metabolite IDs and KEGG databases for pathway mapping. We considered statistical signi cance for pathways at a nominal p-value ≤ 0.1 and considered additional noteworthy pathways if the impact score was ≥ 0.5 while the nominal p-value < 0.3.

Ethics Statement
We con rm that the use of human tissue samples (i.e., plasma) was performed in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants at Table 1 Supplementary Tables S1-s15 And Figures S1-s8 Supplementary Tables S1-S15 and Figures S1-S8 are not available with this version