Fine particles, genetic pathways, and markers of inflammation and endothelial dysfunction: analysis on particulate species and sources

Abstract

Background

Studies have found associations between PM_2.5 and cardiovascular events. The role of different components of PM_2.5 is not well understood.

Methods

We used linear mixed-effects models with the adaptive LASSO penalty to select PM_2.5 species and source(s), separately, that may be associated with markers of inflammation and endothelial dysfunction, with adjustment for age, obesity, smoking, statin use, diabetes mellitus, temperature, and season as fixed effects, in a large longitudinal cohort of elderly men. We also analyzed those associations with source apportionment models and examined genetic pathway-air pollution interactions within three relevant pathways (oxidative stress, metal processing, and endothelial function).

Results

We found that independent of PM_2.5 mass vanadium (V) was associated with intercellular adhesion molecule-1 (ICAM-1) and vascular cell adhesion molecule-1 (VCAM-1). An IQR increase (3.2 ng/m³) in 2-day moving average V was associated with a 2.5% (95% CI: 1.2 – 3.8%) change in ICAM-1, and a 3.9% (95% CI: 2.2 – 5.7%) change in VCAM-1, respectively. Additionally, an oil combustion source rich in V was linked to these adhesion molecules. People with higher allelic risk profiles related to oxidative stress may have greater associations (p-value of interaction = 0.11).

Conclusions

Our findings suggest that particles derived from oil combustion may be associated with inflammation and endothelial dysfunction, and it is likely that oxidative stress plays a role in the associations.

Introduction

There is consistent evidence that fine particulate matter (PM_2.5, particles with an aerodynamic size of ≤ 2.5 µm) is associated with increased risk in cardiovascular morbidity and mortality.^{1, 2} Yet, it is not well understood which PM_2.5 species are of greater toxicity since PM_2.5 mass contains various species that originate from different sources.

Biological mechanisms whereby particles affect cardiovascular health are not fully described, but they include oxidative stress, inflammation, activation of C fibers in the lung, and endothelial changes.² In particular, it has been hypothesized that particles deposited in the lung can cause airway injury or activation of blood cells followed by a release of proinflammatory cytokines interleukins, such as interleukin 6; increased interleukins then activate mononuclear and endothelial cells, initiating the release of acute-phase reactants (e.g., C-reactive protein) and an upregulation of adhesion molecules.³

Metals have been hypothesized to be particularly toxic component of PM_2.5. For example Ghio et al.⁴ reported that metal extracts from particle filters were capable of generating reactive oxygen species. Tarantini et al.⁵ took blood samples of metal foundry workers after a weekend off from work, and again near the end of the work week; they found decreases in the methylation of the iNOS gene after 3 days of work exposure.

Therefore, to better understand the mechanisms as well as inform effective emission-control strategies, we examined associations of PM_2.5 species and sources with markers of inflammation and endothelial dysfunction. These include interleukin 6 (IL-6), C-reactive protein (CRP), intercellular adhesion molecule-1 (ICAM-1) and vascular cell adhesion molecule-1 (VCAM-1), in a large longitudinal cohort of elderly men.

Methods

Study population

Our study population consisted of Normative Aging Study (NAS) participants. This study was established in 1963 by the Veterans Administration.⁶ NAS is an ongoing longitudinal study of health and aging. Participants were free of known chronic medical conditions at enrollment and have undergone examinations every 3 to 5 years. We obtained written informed consent from all participants. The study was approved by the Institutional Review Boards of all participating institutions.

We excluded participants with incomplete information on any of the covariates of interest, and those who died or moved out of New England. There were 1,565, 1,566, 1,342, and 1,624 observations included in analyses of ICAM-1 (years 1999–2010), VCAM-1 (years 1999–2010), IL-6 (years 2000–2008), and CRP (years 1999–2010), corresponding to 735, 735, 712, 739 subjects, respectively.

Outcome assessment

Blood samples were collected at each clinical visit. We measured plasma ICAM-1 and VCAM-1 levels with an enzyme-linked immunoabsorbent assay (ELISA) method (R&D Systems, Minneapolis, MN). IL-6 was assayed from serum (MILLIPLEX Human Cytokine/Chemokine; EMD Millipore, Billerica, MA) and quantified with the Luminex^® 200™ System (Luminex Corp., Austin, TX). CRP levels were measured using an immunoturbidimetric assay on the Hitachi 917 analyzer (Roche Diagnostics, Indianapolis, IN).

Exposure assessment

We obtained daily concentrations of PM_2.5 and its components (K, S, Se, Al, Si, Fe, Ni, V, Cu, Zn, and Na) from the stationary ambient monitoring site at the Harvard University Countway Library. Details about the monitoring site can be found elsewhere.⁷ We focused on these components as their concentrations are mostly above detection limits and they are representative of different PM sources. Also, they were under studied for the outcomes of interest of our study. PM_2.5 was measured using the Tapered Element Oscillating Microbalance (TEOM; Model 1400A, Rupprecht & Patashnick Co., Albany, NY), whereas, components were measured by the Energy Dispersive X-ray Fluorescence Spectrometer (Epsilon 5, PANalytical, Almelo, The Netherlands). We applied the Specific Rotation Factor Analysis (SRFA) to identify and apportion sources of PM_2.5.⁸ Based on analysis of the covariance matrix, SRFA increases the statistical weight for the tracer elements and decrease the weight for the elements that present relatively high analytical errors. Source contributions are estimated by regressing elemental and mass concentrations on the factor loadings in SRFA.

Statistical analysis

Main analysis

The adaptive LASSO selection procedure was used to select PM_2.5 components and sources that may be associated with outcomes of interest. More detailed descriptions can be found in our previous study, in which we found that Ni was selected as an important predictor that may contribute to the PM-related effects on increased blood pressure.⁹ Briefly, the LASSO (Least Absolute Shrinkage and Selection Operator) puts a ℓ₁ penalty on the regression coefficients in a procedure that minimizes the sum of squared errors subject to the sum of the absolute values of the coefficients being less than a given value.¹⁰ The adaptive LASSO is a refinement of the LASSO that uses weights for penalizing different coefficients in the ℓ₁ penalty to achieve asymptotical normality and consistent selection.¹¹ Its estimates are given by,

$$\beta =\text{arg min }{\Vert y-\mathrm{\Sigma }X\beta \Vert }^2+\lambda \mathrm{\Sigma }w|\beta |,$$ [1]

where, λ is the penalty parameter and the adaptive weight w is calculated as the inverse of coefficient from the ordinary linear mixed-effects model (w = 1/β_lme). It has been shown that the adaptive LASSO consistently selects predictors with non-zero coefficients and provides estimates with asymptotic normality (known as oracle properties).¹¹

To identify and select PM_2.5 components associated with the health outcomes, we applied such penalty to all PM_2.5 components in the models, but not to covariates of interest (a detailed description of covariates can be found below). The penalty parameter, λ, determines how strongly one penalizes, or constrains, the magnitude of the PM_2.5 components regression coefficients. When λ is small, components are weakly penalized and estimates are close to those from the standard linear mixed-effects model, whereas when λ is large, shrinkage is heavy so that all component coefficients equal to zero, yielding a model that includes fixed covariates only. If λ takes some value between the extremes, we get a penalized model where some coefficients are zero and some are non-zero. We considered the components with non-zero coefficients. We ran the models across the full range (i.e., from the model with no penalty to the model without any components) of λ’s and chose the λ having the smallest Bayesian Information Criterion (BIC).¹² Finally, we excluded unselected components to fit a mixed-effects model and estimate the effects and corresponding 95% confidence intervals. Additionally, we ran the same selection procedure with PM_2.5 source contributions in order to investigate which sources of PM_2.5 were most predictive of the outcomes of interest.

Previous studies have shown that particulate pollution has acute effects on ICAM-1, VCAM-1 and IL-6,^{13, 14} so we calculated 2-day moving average concentrations of PM_2.5 and components for the analyses of blood marker outcomes. For CRP, we used 28-day moving average concentrations instead since earlier studies have suggested particles averaging over that time period were associated with CRP.^{15, 16} Values of all outcomes were log-transformed so that the normality assumption for the residuals was satisfied.

We controlled for the following covariates a priori: age, obesity (defined as body mass index (BMI) ≥ 30), smoking status (current, former, and never smokers), pack years, statin use, physician-diagnosed diabetes mellitus, temperature, quadratic temperature, and season (defined as spring: March-May, summer: June-August, fall: September-November, winter: December-February). As Mostofsky et al.¹⁷ illustrated, it is likely one would observe effects of PM_2.5 components well correlated with mass merely because they are highly correlated with mass and not because they are themselves particularly toxic. Therefore, we also controlled for PM_2.5 mass which could confound the effects of components. All these covariates were forced to stay in the models without penalty. Our model is shown below:

$$Y_i=({\alpha }_1{X}_{i1}+\dots +{\alpha }_P{X}_{iP})+({\beta }_1{Z}_1+\dots +{\beta }_M{Z}_M)+{\alpha }_0+{\mu }_i+{\epsilon }_i,$$ [2]

where, Y_i is the blood marker level of subject i, μ_i is the random intercept for subject i, and ε_i is the residual error term; α₁,…, α_P are estimates of the fixed effects of PM_2.5 mass and other covariates X_i1, …, X_iP, and β₁, …, β_M are estimated effects of PM_2.5 components Z₁, …, Z_M that are determined by Equation 1.

Analysis of effect modification

We further explored genetic pathway-air pollution interactions by including interaction terms between species’ concentrations and genetic scores within relevant pathways: oxidative stress, metal processing, and endothelial function, respectively.¹⁸ The genetic score, developed by Bind et al., is a novel approach to investigate interactions between pathways and environment. The authors related genes to one of the three pathways based on their biological functionality provided by GeneCards,¹⁹ and considered independent outcomes representative of each pathway: 8-hydroxydeoxyguanosine for oxidative stress, augmentation index for endothelial function, and patella lead for metal processing. They used the LASSO method to select the most appropriate gene variants for the above outcomes. Gene variants that were selected (please see Tables S1 to S3 in the supplement for a full list of selected gene variants.) had non-zero coefficients and were summed up to construct the scores for all study subjects. Finally, a binary score was created by dichotomizing using the median of each score’s distribution. In the current study, we examined potential effect modification by including interactions between the binary genetic scores and PM_2.5 species in our models, in addition to a main effect for the binary score.

The statistical model that examined interactions was,

$$Y_i=({\gamma }_1{X}_{i1}+\dots +{\gamma }_P{X}_{iP})+{\gamma }_{P+\mathit{1}}Z+{\gamma }_{P+\mathit{2}}Z\times \mathit{\text{Score}}+{\gamma }_{P+\mathit{3}}\mathit{\text{Score}}+{\gamma }_0+{\mu }_i+{\epsilon }_i,$$ [3]

where, Y_i is the blood marker level of subject i, γ₁, …, γ_P are estimates of the fixed effects of PM_2.5 mass and other covariates X_i1, …, X_iP, μ_i is the random intercept for subject i, and ε_i is the residual error term; Z indicates PM_2.5 species, and Score presents the genetic score.

Results

Descriptive results

Characteristics of study populations are summarized in Table 1. There were 1,565, 1,566, 1,342, and 1,624 observations for ICAM-1, VCAM-1, IL-6, and CRP, respectively. Across all visits, the mean ICAM-1 level was 285.0 (SD = 1.3) ng/mL, while mean VCAM-1 was 986.6 (SD = 1.4) ng/mL; IL-6 had a mean of 33.6 (SD = 8.0) pg/mL, and CRP had a mean of 1.6 (SD = 3.0) mg/L.

Table 1.

Characteristics of subjects in analyses of different health outcomes in the study.

	ICAM-1	VCAM-1	IL-6	CRP
Variable	n=1,565	n=1,566	n=1,342	n=1,624
	(1999–2010)	(1999–2010)	(2000–2008)	(1999–2010)
Mean ± SD
Health outcomea	285.0 ± 1.3	986.6 ± 1.4	33.6 ± 8.0	1.6 ± 3.0
Age (years)	74.9 ± 6.7	74.9 ± 6.7	74.3 ± 6.7	74.9 ± 6.7
Pack years	20.0 ± 24.0	19.9 ± 24.0	20.4 ± 24.4	20.0 ± 24.0
Number (%)
Subjects with 1 visit	241 (33%)	241 (33%)	263 (37%)	227 (31%)
Subjects with 2 visits	206 (28%)	205 (28%)	272 (38%)	197 (27%)
Subjects with ≥ 3 visits	288 (39%)	289 (39%)	177 (25%)	315 (43%)
Current smokers	53 (3.4%)	53 (3.4%)	49 (3.7%)	54 (3.3%)
Former smokers	1,045 (66.8%)	1,046 (66.8%)	891 (66.4%)	1,082 (66.6%)
Use of statin	781 (49.9%)	782 (49.9%)	636 (47.4%)	813 (50.1%)
Diabetes	246 (15.7%)	247 (15.8%)	203 (15.1%)	257 (15.8%)
Obesity	414 (26.5%)	414 (26.4%)	355 (26.5%)	429 (26.4%)

^aValues of geometric mean and geometric standard deviation were reported for these health outcomes. Units of measurement: ICAM-1 (ng/mL), VCAM-1 (ng/mL), IL-6 (pg/mL), CRP (mg/L).

Table 2 summarizes concentrations of PM_2.5 and its species. We used 2-day moving average PM_2.5 in analyses of ICAM-1, VCAM-1, and IL-6. The means were 10.8, 10.8, and 11.2 µg/m³, respectively. 28-day moving average PM_2.5 was used in the CRP analysis, with a mean of 9.3 µg/m³.

Table 2.

Summary of PM_2.5 and components in analyses for different health outcomes across all study visits.

Pollutant	ICAM-1	VCAM-1	IL-6	CRP
Meana ± SD
PM2.5 (µg/m3)	10.8 ± 6.0	10.8 ± 6.0	11.2 ± 6.2	9.3 ± 8.5
Components (ng/m3)
Fe	76.8 ± 37.4	76.8 ± 37.4	78.9 ± 37.9	67.8 ± 17.5
K	37.8 ± 22.2	37.8 ± 22.2	37.9 ± 21.7	42.1 ± 23.9
S	1,127.0 ± 848.6	1,126.6 ± 848.4	1,194.6 ± 875.1	1063.6 ± 404.5
Al	54.6 ± 33.5	54.6 ± 33.5	56.3 ± 33.8	53.1 ± 23.3
Si	82.3 ± 58.3	82.3 ± 58.3	84.2 ± 58.5	77.0 ± 41.7
Ni	3.0 ± 2.9	3.0 ± 2.9	3.4 ± 3.0	3.1 ± 2.6
V	3.4 ± 2.6	3.4 ± 2.6	3.8 ± 2.6	3.5 ± 2.0
Cu	3.7 ± 2.4	3.7 ± 2.4	3.8 ± 2.4	3.6 ± 1.5
Zn	13.4 ± 11.4	13.4 ± 11.4	14.2 ± 11.4	11.5 ± 4.8
Se	0.1 ± 0.4	0.1 ± 0.4	0.1 ± 0.4	0.2 ± 0.3
Na	195.6 ± 114.0	195.6 ± 114.0	202.5 ± 114.7	198.7 ± 64.3

^a2-day moving averages of concentrations were calculated in analyses of ICAM-1, VCAM-1, and IL-6; 28-day moving averages were calculated for CRP analysis.

Main analysis

LASSO coefficient paths are illustrated in Figure 1. Every PM_2.5 species has a coefficient path, which shows how its coefficient started at a non-zero value and moved towards zero at a certain rate. The vertical dashed line indicates the λ with minimum BIC. For both ICAM-1 and VCAM-1, V was selected using the value of λ with the smallest BIC value; however, no species was selected for either IL-6 or CRP (data not shown). In both plots, V remained the longest as it was the last component that hit zero. Interestingly, for VCAM-1, Ni behaved similarly to V and was excluded in the model at a value of λ just below that for V.

Figure 1.

LASSO coefficient paths: plot of coefficient profiles for PM_2.5 species as a function of log(λ).

The Specific Rotation Factor Analysis indicated 6 factors as source contributions to PM_2.5: (1) road dust (Al, Si, etc.), (2) wood burning and local pollution (K, Cu, etc.), (3) oil combustion (V, Ni), (4) traffic (Fe, Zn, etc.), (5) sea salt (Na, Cl), and (6) regional pollution (S). Analysis of the association of ICAM-1 with PM_2.5 sources selected factor 3, which represented oil combustion, using the value of λ that minimized the BIC. As for VCAM-1, there were two equivalent minima of BIC and factor 3 was consistently selected (data not shown). In addition, the results suggested factor 2, in which wood burning and local pollution contributed the largest fraction, might also be associated with levels of VCAM-1 as it was selected at one of the BIC minima.

In our final models, an IQR increase (3.2 ng/m³) in 2-day moving average V was associated with a 2.5% (95% CI: 1.2 – 3.8%) change in ICAM-1, and a 3.9% (95% CI: 2.2 – 5.7%) change in VCAM-1. In the same models, each 10 µg/m³ in PM_2.5 was associated with a 2.6% (95% CI: 0.3 – 4.9%) change in ICAM-1 and a 2.8% (95% CI: −0.2 – 5.9%) change in VCAM-1.

Effect modification

Figure 2 shows the percent changes in ICAM-1 and VCAM-1 per IQR increase in V according to the three types of genetic scores (high versus low). Our results were suggestive that people with a high genetic score within the oxidative stress pathway had stronger associations between V and ICAM-1 levels (p-value of interaction = 0.11).

Figure 2.

Percent change in ICAM-1 and VCAM-1 levels per IQR increase in 2-day moving average V by genetic scores in oxidative stress, metal processing, and endothelial dysfunction pathways.

Discussion

In this study, we estimated the differential effects of PM_2.5 species and sources on blood markers of inflammation and endothelial dysfunction in a large longitudinal cohort. By using the adaptive LASSO method, we found that among the species we examined (K, S, Se, Al, Si, Fe, Ni, V, Cu, Zn, and Na), V was associated with increased levels in both ICAM-1 and VCAM-1. Our analyses on the genetic pathway-air pollution interaction suggested that such associations in individuals with higher allelic risk profiles related to oxidative stress might be stronger. In addition, we observed that pollution contributions from oil combustion were associated with each of these two adhesion molecules. This supported our species-specific results, as V and Ni are predominantly from fuel oil combustion.

Adhesion molecules

Cellular adhesion molecules (CAMs) stabilize the interactions between cells and mediate the process of leukocyte adhesion and trans-endothelial migration. ICAM-1 and VCAM-1 are two important members of the immunoglobulin superfamily of CAMs. Increased levels of ICAM-1 and/or VCAM-1 have been associated with development of cardiovascular diseases and cardiovascular death.^20–22

Because of the important role of adhesion molecules in inflammation and endothelial function, researchers have conducted epidemiological studies on associations between these molecules and particulate air pollution. A prospective panel study on coronary heart disease patients has found that the odds of observing high ICAM-1 levels (i.e., above the 90^th percentile) were associated with ambient particles.³ In a cross-sectional study on individuals with type-2 diabetes, the authors documented that VCAM-1 was associated with PM_2.5 exposures.²³ More recently, Bind et al.²⁴ reported short- and intermediate-term effects of PM_2.5 on ICAM-1 and VCAM-1 levels in the NAS cohort. There have also been controlled-exposure studies on the toxicity of particles on adhesion molecules. For example, Salvi et al.²⁵ exposed healthy individuals to diesel exhaust (DE) and found significant increases in endothelial ICAM-1 and VCAM-1 expression. Likewise, Stenfors et al.²⁶ observed DE-induced upregulation of VCAM-1 in healthy humans.

Effects of PM_2.5 species and source contributions

Although a large number of studies have examined the association between PM mass and adhesion molecules, there is very limited research on the effects of PM_2.5 species. Moreover, most of the existing studies examined only a few components. Black carbon (BC), a marker of traffic, has attracted the most attention to date, and has been found to be adversely associated with ICAM-1 and/or VCAM-1 in several studies.^{13, 23, 24, 27} O’Neill et al.²³ did not find associations between sulfate on ICAM-1 and VCAM-1 levels among people with type-2 diabetes, whereas Bind et al.²⁴ observed significant associations in the NAS cohort. In addition, elemental carbon (EC) and organic carbon (OC) were examined by Rückerl et al.³ and were found to be associated with high VCAM-1 levels.

Our findings that fine particles were associated with markers of inflammation and endothelial dysfunction are in agreement with those from previous studies. Furthermore, we investigated a broad range of PM components emitted from different sources, particularly, metals. Toxicity of PM-related metals on health has been documented in the literature;^28–30 however, to the best of our knowledge, effects of PM-related metals on adhesion molecules have never been reported in previous epidemiological studies. On the other hand, toxicological research has observed such toxicity. For example, endothelial cells exposed to vanadium pentoxide (V₂O₅) resulted in an increase in reactive oxygen species (ROS) and nitric oxide production, and a decreased proliferation.³¹ Ni-induced changes in heart rate and heart rate variability were observed in animal studies.^{30, 32} The current study provides epidemiological evidence on associations between PM_2.5 metals and adhesion molecules for the first time.

Both V and Ni are tracers of oil combustion. In the analysis of VCAM-1, Ni behaved similarly to V in terms of when it was shrunk to zero in the adaptive LASSO. Specifically V was the last component shrunken to zero, with Ni next to last. These results suggest oil combustion is associated with adhesion molecules. We further examined this hypothesis by selecting individual source daily contributions, in lieu of daily species concentrations, and found evidence of an association between oil combustion contributions and both ICAM-1 and VCAM.

Genetic pathway-air pollution interaction

A gene-environment interaction has been observed in previous studies. GSTM1 (Glutathione S-Transferase Mu 1; a gene encoding enzymes that function in antioxidation) deletion modified the association between BC and VCAM-1, suggesting that oxidative stress may mediate the effects of BC on cardiovascular health.¹³ In addition, Zeka et al.¹⁶ found stronger effects of particles on inflammatory markers in GSTM1-null subjects than in GSTM1-present subjects. A study on acute exposure-induced gene expression profiles reported that V induced alterations in genes associated with passive and active transport of solutes across the membrane, and activated kinases involved in signal transduction pathways.³³ Bind et al.¹⁸ observed a genetic pathway-air pollution interaction in relation to CRP and ICAM-1, and reported greater effects of particle number on CRP and ICAM-1 for participants with higher metal processing genetic scores. We found stronger associations between V and ICAM-1 among subjects with a higher genetic score within the oxidative stress pathway, adding evidence to this growing body of literature.

IL-6 and CRP

IL-6 is a cytokine that plays a key role in the inflammatory response.³⁴ Studies have shown that IL-6 was related to particles. For example, a prospective longitudinal study of myocardial infarction survivors has found increases in IL-6 when particle number concentrations (PNC) were elevated; another study observed elevated circulating levels of IL-6 among individuals exposed to high levels of PM₁₀.^{14, 35} CRP is another marker of inflammation that is associated with particulate air pollution.³⁶ Pope et al. reported a relationship between PM_2.5 and CRP in a panel of elderly subjects.³⁷ It has been documented that particles were associated with the odds of observing CRP concentrations above the 90^th percentile.³ However, none of the species we examined was selected for either IL-6 or CRP. The reason might be that other components not included in this study, such as EC and OC, are more responsible for the effects of PM_2.5 mass. For instance, studies have found significant associations between IL-6 and OC.^{38, 39}

Implications

With the adaptive LASSO procedure, our analysis on species found that V may be the most toxic species for markers of inflammation and endothelial dysfunction, while analysis on source contributions showed that oil combustion contributed the most to PM-related effects on those markers. Our study adds to the evidence of V toxicity and is of importance to explain the health effects of residual oil fly ash (ROFA), which significantly contributes to ambient particles.⁴⁰ Researchers have postulated that V accounts for a remarkable portion of the biologic activity of ROFA: transition metals in ROFA, especially V, are involved in specific chemical reactions to produce reactive oxygen species that lead to phosphorylation-dependent cell signaling, transcription factor activation, induction of inflammatory mediator expression, and eventually inflammatory lung injury.⁴⁰ The results of the current study provide support to the hypothesis.

In a previous study, we applied the same methods and identified that Ni was associated with increased blood pressure in the NAS cohort.⁹ Together with those findings, we believe that oil combustion particles may be linked with cardiovascular health, perhaps by inducing inflammation and endothelial dysfunction. The U.S. National Research Council has made it a research priority to examine the differential toxicity of particulate matter species and regulate the most toxic species.⁴¹ Our study can be helpful to such emission-control strategies.

Strengths and limitations

The adaptive LASSO approach we used to select PM_2.5 species and sources posing higher risk is one of the main features of our study. This approach has the advantage to select the most appropriate variables asymptotically. In contrast, conventional methods, e.g., stepwise selection, do not guarantee consistent selection. To the best of our knowledge, this is the first longitudinal study on associations of markers of inflammation and endothelial dysfunction with daily PM_2.5 species concentrations, especially the metals, as well as source contributions. Our findings that PM_2.5 species and sources were differentially associated with these markers can be critical to the development of cost-effective mitigation strategies focusing on the most toxic PM_2.5 species or sources.

The main limitations in this study are the following. First, we lacked personal exposure information as we used PM_2.5 measurements from a stationary monitor. This may induce non-differential exposure misclassification that can attenuate our estimates. Second, we studied a population consisting of primarily white elderly men. Therefore we cannot be sure that our results are generalizable to women or younger and more diverse populations.

To summarize, we conducted mixed-effects models with the adaptive LASSO selection algorithms to identity PM_2.5 species and source contributions that may be related to biomarkers of inflammation and endothelial dysfunction in a large and stable longitudinal cohort. Our results suggested that particles derived from oil combustion may be associated with inflammation and endothelial dysfunction, and it is likely that oxidative stress plays a role in the associations.