Metabolomic signatures of the short-term exposure to air pollution and temperature

Abstract

Background:

Short-term exposures to air pollution and temperature have been reported to be associated with inflammation and oxidative stress. However, mechanistic understanding of the affected metabolic pathways is still lacking and literature on the short-term exposure of air-pollution on the metabolome is limited.

Objectives:

We aimed to determine changes in the plasma metabolome and associated metabolic pathways related to short-term exposure to outdoor air pollution and temperature.

Methods:

We performed mass-spectrometry based untargeted metabolomic profiling of plasma samples from a large and well-characterized cohort of men (Normative Aging Study) to identify metabolic pathways associated with short-term exposure to PM_2.5, NO₂, O₃, and temperature (one, seven-, and thirty-day average of address-specific predicted estimates). We used multivariable linear mixed-effect regression and independent component analysis (ICA) while simultaneously adjusting for all exposures and correcting for multiple testing.

Results:

Overall, 456 white men provided 648 blood samples, in which 1,158 metabolites were quantified, between 2000 and 2016. Average age and body mass index were 75.0 years and 27.7 kg/m², respectively. Only 3% were current smokers. In the adjusted models, NO₂, and temperature showed statistically significant associations with several metabolites (19 metabolites for NO₂ and 5 metabolites for temperature). We identified six metabolic pathways (sphingolipid, butanoate, pyrimidine, glycolysis/gluconeogenesis, propanoate, and pyruvate metabolisms) perturbed with short-term exposure to air pollution and temperature. These pathways were involved in inflammation and oxidative stress, immunity, and nucleic acid damage and repair.

Conclusions:

This is the first study to report an untargeted metabolomic signature of temperature exposure, the largest to report an untargeted metabolomic signature of air pollution, and the first to use ICA. We identified several significant plasma metabolites and metabolic pathways associated with short-term exposure to air pollution and temperature; using an untargeted approach. Those pathways were involved in inflammation and oxidative stress, immunity, and nucleic acid damage and repair. These results need to be confirmed by future research.

Introduction

Short-term exposures to air pollution and temperature have been reported be associated with inflammation and oxidative stress^1,2; biological processes that are linked to many adverse health effects such as pulmonary³, cardiovascular⁴, and neurological diseases⁵, and mortality^6–10. The plasma metabolome represents a collection of biologically active chemicals derived from endogenous processes and exogenous exposures¹¹. A metabolic signature in the blood associated with ambient air pollution exposure is plausible because some ambient air pollutants (ultrafine particles and gaseous air pollutions) have been reported to enter the bloodstream directly from the lungs¹². In addition, larger particles unable to cross the lung epithelium as well as ambient temperature can induce inflammation in the lungs and trigger a systemic response observed in the peripheral blood^2,13. Although metabolomic profiling of plasma represents a powerful tool to increase mechanistic understanding, to date analytical and scientific uncertainties in metabolomics have limited its application for the measurement the response of exposure to air pollution.

Recently, several studies have examined the association between short-term exposure to air pollution and changes in the plasma metabolome and have reported perturbed pathways implicated in inflammation and oxidative stress^14–22. However, replication of these findings is still lacking and the existing literature on the short-term exposure to air-pollution on the metabolome is limited. To date, no human studies have examined the untargeted metabolomic signature of ambient temperature despite the strong association of ambient temperature with mortality and morbidity, and only a few studies have controlled for ambient temperature in models examining the association with air pollution. In addition, existing studies have tended to analyses each metabolite individually, which fails to capture the correlation patterns of metabolites that exist within co-regulated metabolic pathways and which could be important for elucidating mechanisms.

Therefore, in this study we aimed to determine changes in the plasma metabolome related to short-term exposure to outdoor air pollution and temperature, while using novel statistical methods e.g., independent component analysis (ICA) for highly dimension data. We performed mass-spectrometry based plasma untargeted metabolomic profiling of a large and well-characterized cohort of men (the Normative Aging Study). We hypothesized that short-term exposure to air pollution and temperature are associated with changes in the plasma metabolome consistent with perturbation of metabolic pathways that are related to the adverse health effects of air pollution and temperature exposure. We further tested if the metabolic signatures would be modified by metabolic conditions such as obesity and diabetes.

Materials and methods

Participants and study design

The Normative Aging Study (NAS) established in 1963 is a longitudinal study of aging among men based in Boston, Massachusetts. Men (N=2,280), 21–80 years old between 1963 and 1970 and free of known chronic diseases were enrolled in NAS and have been followed since²³. Participants periodically self-reported information about their medical history, dietary intake, and other health-related history throughout the follow-up. Men had physical examinations and laboratory tests every 3–5 years. At every visit, a 7ml fasting venous blood sample was collected from participants in a trace metal-free tube containing ethylenediaminetetraacetic acid. The samples were spun for 15 min at 3,000 revolutions per minute. Serum and plasma samples were placed in 1.8 ml Nunc tubes for long-term storage at −80°C. NAS was approved by the review boards of Harvard T.H. Chan School of Public Health and the Department of Veterans Affairs. All participants provided written informed consent.

For the current analysis, we considered 464 men who contributed 659 visits between 2000 and 2016 and had plasma metabolomics measured. Blood samples collected prior to 1999 were not suitable for profiling due to their storage conditions. Because only 8 men were non-white, we restricted our analysis to 456 white men who contributed 648 visits/blood samples.

Quantification of air pollution and temperature (exposure)

Because previous studies showed robust inflammatory responses after short-term exposures to air pollution and temperature variations^8–10, we focused in the current study on short-term air pollution and temperature exposures measured on the same day of the visit (extremely-short) and mean values at 7-days (medium-short) and 30-days (longer-short) before the visit for each blood draw. We chose those three windows pre-hoc considering the tradeoff by the multiple comparison issue and the subsequent correction needed. The air pollutants were particulate matter ≤2.5 μm (PM_2.5) and gaseous pollutants: nitrogen dioxide (NO₂) and ozone (O₃)^24–26. The exposures in the primary analysis came from prediction models we fit that provided daily estimates for a 1-km grid of the contiguous U.S ^27–29. The models used data from predictions of chemical transport models (GEOS-Chem, CMAQ, CAMS, and MERRA-2), meteorological data from the National Oceanic and Atmospheric Administration (NOAA), land-use terms from the national land cover dataset, road density data from the Census, traffic data from the environmental systems research institute (ESRI), and satellite-based measures of aerosol optical depth, NO₂, O₃, normalized difference vegetation index (NDVI; a measure of greenness), surface reflectance, and absorbing aerosol index. Using these variables, we trained three models: a neural network, a random forest, and a gradient boosting machine to US Environmental Protection Agency (EPA) Air Quality System and National Park Service monitoring data from the contiguous US to generate daily predictions of each of the pollutants (PM_2.5, NO₂, and O₃) on a 1×1-km grid. We combined the 3 predictions from the 3 models for each pollutant in a nonlinear geographically weighted regression to generate a single prediction / day / grid cell for each pollutant. The models performed well, with 10-fold cross-validation on held out monitoring sites yielding an out of sample R² =0.89, 0.86, and 0.84 respectively for annual average predictions of PM_2.5, O₃, and NO₂. The daily O₃ predictions were for the 8-hr daily maximum, the NO₂ predictions were for the 1-hr maximum, and PM_2.5 predictions were for the daily mean. The temperature and relative humidity data are from Gridmet (http://www.climatologylab.org/gridmet.html). The data are daily data with a 4-km resolution ³⁰.

As a sensitivity analysis, we repeated the analysis using all exposures as measured at a fixed monitoring site located at the Harvard University Countway Library near downtown Boston, MA and approximately 1 km from the examination center. Because study participants lived in greater Boston with a median distance of 20.8 km from the examination center ³¹, and since within that area, daily levels of each pollutant tend to go up and down together regardless of spatial heterogeneity in mean concentrations, we considered these measurements a reasonable surrogates of participants’ exposures. The advantage is we have real, not modeled measurements, the disadvantage is the lack of spatial resolution. Levels of PM_2.5 (μg/m³) were measured hourly using a tapered element oscillation microbalance (Model 1400A, Rupprecht and Pastashnick)³². Hourly NO₂ and O₃ levels [in parts per billion (ppb)] were measured by local state monitors in the greater Boston area and were averaged based on data from all available sites. We also used temperature data as well as relative humidity (as a covariate), which we obtained from the National Weather Service Station at Logan Airport (Boston, MA), located approximately 12 km from the examination site. Because study participants lived throughout the metropolitan area, we considered that the monitored temperature could serve as surrogates of their exposures.

Metabolomic Profiling (outcomes)

All samples were sent to the lab and analyzed at the same timepoint. Metabolomic profiling was conducted by Metabolon Inc. (Durham, NC, USA) using untargeted high-resolution Ultrahigh Performance Liquid Chromatography Coupled Tandem Mass Spectroscopy (UPLC-MS/MS) enabling the broadest coverage of the metabolome. The methods were previously described in detail ³³. In brief, the 4 platforms were: (1) UPLC-MS/MS under positive ionization for early eluting metabolites, (2) UPLC-MS/MS under positive ionization for late eluting metabolites, (3) UPLC-MS/MS under negative ionization, and (4) UPLC-MS/MS, polar platform (negative ionization). The samples were randomized for profiling and in-house standards and internal QCs were used ³³. Metabolites were identified by their mass-to-charge (m/z) ratio, retention time, and through a comparison to a library of purified known standards. Metabolites were quantified using area-under-the-curve of the peak and processed according to our in-house standard quality control (QC) pipeline³⁴. In brief missing values for a given metabolite were imputed with half the minimum observed value for that metabolite. Metabolite levels were then log transformed and pareto scaled. A total of 1,301 metabolites were profiled. Of these 143 had an interquartile range of zero and were therefore considered uninformative and excluded. This resulted in 1,158 metabolites available for the current analysis.

Statistical Analysis

We calculated descriptive statistics to summarize socio-demographics and lifestyle factors for the participants at the first visit and all visits. We examined whether the same-day, 7-, and 30-days prior to visit average air pollution exposures and temperature were associated with changes in the levels of the 1,158 metabolites. We first conducted generalized additive mixed models (GAMM) with a penalized spline for temperature to check if linearity would fit better. We then used time-varying linear mixed-effect regression models (LMEM) with random participant-specific intercepts, accounting for the correlation of repeated measures. We treated the outcomes and the exposures (including temperature) at every exposure window as continuous variables. Based on substance knowledge and the Directed Acyclic Graph (DAG) (Supplemental Figure 1), we adjusted for potential confounders and predictors for the outcomes including age (years), body mass index (kg/m²), cigarette pack-years, alcohol intake (< or ≥2 drinks per day), socioeconomic status (income and years of education), season (warm/cold), and relative humidity. We assumed that air pollution and ambient temperature have a confounding relationship with each other. Therefore, we simultaneously adjusted for PM_2.5, NO₂, O₃, and temperature (multi-pollutant models) for each metabolite as below. Multi – pollutant models:

$$E\left(Y_{ij}\right)={\beta }_0+{\mathrm{\mu }}_i+{\beta }_1{\mathit{PM}}_{\mathit{\text{2.5ij}}}+{\beta }_2{NO}_{2ij}+{\beta }_3{O}_{3ij}+{\beta }_4{\mathit{Temperature}}_{ij}+{\beta }_{5-n}{\mathit{Covariates}}_{ij}$$

, where Y_ij is the metabolome level of subject i at visit j, (β₀ is the fixed intercept, μ_iis the random intercept for subject i, and the air pollutants and temperature are the moving averages for subject i at visit j.

To account for multiple testing while also accounting for the high correlation between metabolites that are closely connected through interlinked biological pathways, we utilized the “number of effective tests approach”^35,36. This method determines the number of principal components required to explain a given percentage of the variance in the data (i.e., the number of effective/independent tests (ENT)). We calculated the adjusted p-value threshold as α/m where α denotes the nominal p-value threshold of 0.05, and m denotes the number of principal components. We used thresholds of 95% (ENT95%) and 99% (ENT99%) variance explained for these analyses and further divided by 3 (moving averages) and then by 4 (investigated exposures).

In order to explore metabolomic profiles rather than individual metabolites, we next used an independent component analysis (ICA)³⁷ as an unsupervised technique to reduce the highly-correlated metabolites into five independent factors. Principal component analysis (PCA) is a tool for dividing multiple correlated variables into uncorrelated factors. If the variables are Gaussian distributed then uncorrelated factors are independent, however that is not true for non-Gaussian distributed variables such as metabolites’ levels. ICA is a computational method for separating multivariate signal into additive factors that are maximally independent that assumes non-Gaussian signals³⁷. For each factor, there is an attached weight for each metabolite to determine its individual contribution. We then conducted LMEM where those five independent factors were the outcomes while exposures and covariates remained as above. We corrected for multiple testing in the ICA models by applying Benjamini-Hochberg false discovery rate (FDR) and set the false positive threshold as 5%³⁸ and further divided by 3 (moving averages: same-day, 7-, and 30-days prior to visit) and then by 4 (investigated exposures: PM_2.5, NO₂, O₃, and temperature).

As sensitivity analyses, we repeated the analysis using the central site exposures. Analyses were conducted using R version 3.6.0 and SAS version 9.4 (SAS Institute Inc., Cary, NC, USA).

Only for the pollutants that showed any significant associations after correcting for multiple testing at level of ENT95%, we then further conducted pathway analysis for the significant metabolites (at p-value <0.01) from the multi-pollutant model and the 100 metabolites that had the greatest weighting on the significant ICA factor. To do that, we used the ‘Pathway Analysis’ functionality in MetaboAnalyst 4.0, that accounted for both over-representation analyses (i.e., how many significant metabolites are within a pathway) and pathway topology (i.e., how influential those metabolites are to that given pathway)³⁹ and used Human Metabolome Database (HMDB), and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases. We considered statistical significance for the pathways at level of p-value ≤0.1 and considered additional noteworthy pathways if the impact score was ≥0.5 while p-value <0.3⁴⁰.

Results

Overall, 456 men provided 648 blood samples between 2000 and 2016 were included in the analysis; 1,158 distinct metabolites were quantified and passed our in-house QC pipeline in these samples. We had access to all metabolites as quantified by Metabolon in these samples. Approximately 64% of the participants provided one blood sample, 31% provided two samples, and 6% provided three samples. The highest percentage of the measured metabolites were lipids (39%), followed by metabolites of un-identified origin (19%), amino acids (17%), xenobiotics (12%), nucleotides (3%), cofactor and vitamins (3%), carbohydrates, peptides, partially characterized molecules (2% each), and energy metabolites (1%) (Supplemental Figure 2). At the initial visit, on average, men’s age was ~75.0 years, body mass index was 27.7 kg/m², mean annual income was $8, 610 in 1965 $US, and they had received 15.1 years of education (Table 1). Approximately 69% of the participants were former smokers, and 3% were current smokers. Most of the participants (79%) consumed <2 alcoholic drinks/day. Around 60% of the visits were in the warm season (April to September) (Supplemental Tables 1). Average values of air pollution and temperature over the follow-up period remained stable across the 30-day exposure windows (Table 2)

Table 1.

Characteristics of the study population in the Normative Aging Study (2000 to 2016)

Demographic Characteristics	At first visit (2000 to 2013) (N=456)
Age, Mean (SD)	75.0 (6.70)
Body mass index (kg/m2), Mean (SD)	27.7 (4.19)
Years of education, Mean (SD)	15.1 (3.00)
Baseline annual income, thousands of 1965 $US, Mean (SD)	8.61 (3.77)
Smoking
Never, N (%)	128 (28)
Current, N (%)	14 (3)
Former, N (%)	314 (69)
Pack-year smoked (years), Mean (SD)	21.4 (25.2)
Season
Cold: October - March, N (%)	189 (41)
Warm: April - September, N (%)	267 (59)
Alcohol consumption (≥2 drinks per day), N (%)	95 (21)

Abbreviations: N, number of participants; SD, standard deviation; Kg, Kilogram; m, meters.

Table 2.

Moving averages of the concentrations of the ambient air pollution and temperature in the Normative Aging Study (2000 to 2013) at the first visit, N=456

Moving averages of the concentrations	Predicted models, mean (SD)	Central site, mean (SD)
Average of 24 hours of PM2.5 (μg/m3)	10.5 (7.11)	10.4 (5.92)
Average of 7 days of PM2.5 (μg/m3)	9.83 (4.09)	10.1 (3.71)
Average of 30 days of PM2.5 (μg/m3)	9.67 (2.96)	10.2 (2.73)
Average of 24 hours of NO2 (ppb)	24.6 (12.34)	20.0 (6.02)
Average of 7 days of NO2 (ppb)	22.8 (10.47)	18.3 (3.95)
Average of 30 days of NO2 (ppb)	22.6 (9.71)	18.4 (3.30)
Average of 24 hours of O3 (ppb)	36.4 (16.13)	24.3 (12.5)
Average of 7 days of O3 (ppb)	36.5 (11.83)	24.8 (8.69)
Average of 30 days of O3 (ppb)	36.3 (10.54)	24.9 (7.50)
Average of 24 hours of Temperature (°C)	12.2 (9.89)	13.6 (8.68)
Average of 7 days of Temperature (°C)	12.0(9.47)	13.0 (7.85)
Average of 30 days of Temperature (°C)	12.1 (8.81)	12.9 (7.56)

Abbreviations: N, number of participants; SD, standard deviation; ppb, is part per billion; C, Celsius.

After adjusting for covariates in the multi-pollutant models, NO₂, and temperature showed a statistically significant associations at the 95% level of the number of effective tests (ENT95%) (19 metabolites for NO₂, and 5 metabolites for temperature out of 1,158 metabolites) (Figure 1, Tables 3 and 4). However, PM_2.5 and O₃ did not have any significantly associations at the ENT95% level (Figure 1). In Figure 1, x-axis represents the beta coefficients from the adjusted LMEM and Y-axis represents the -logarithm (base 10) of the corresponding p-values. Each point on the graph represents a metabolite of the 1,158 examined metabolites. The transverse lines represent different statistical significance level of p-value: 0.05, 0.01, ENT95%, and ENT99%, from lowest to highest line in the figure. Similarly, in the ICA analysis, only NO₂ was consistently and significantly associated with the second factor from the independent component analysis (ICA-factor 2) only and not the other four ICA-factors (Table 5). Many of the most contributing metabolites to ICA-factor 2 were also individually significant with exposures of interest in the multi-pollutant models (Tables 3 and 4).

Figure 1. Volcano Plots presenting the adjusted associations between short-term exposures to air pollutants (predicted models) and temperature with metabolomics (multi-pollutant models).

These models were linear mixed-effect regression models (LMEM) with random participant-specific intercepts and simultaneously adjusted for PM_2.5, NO₂, O₃, and temperature (multi-pollutant models) for the same exposure window.

All models were adjusted for age (years), body mass index (kg/m2), cigarette pack-years, alcohol intake (< or ≥2 drinks per day), socioeconomic status (income payment and years of education), season (warm/cold), and relative humidity.

The transverse dashed lines represent different statistical significance levels of p-values (from lower to upper): 0.05, 0.01, ENT95%, and ENT99%.

Note the different scale of the X axes.

Abbreviations: ENT: Effective/independent number of tests; log10, logarithmic base 10; PM_2.5, particulate matter ≤2.5 μm; NO₂, nitrogen dioxide; and O₃, ozone.

Table 3.

The significant adjusted associations between short-term exposure to NO₂ (predicted models) with metabolomics (multi-pollutant models) at ENT95% significance level

Biochemical	Super-pathway	Sub-pathway	1 day window, Beta (SE) for individual metabolites	1 day window, p-value for individual metabolites	7 day window, Beta (SE) for individual metabolites	7 day window, p-value for individual metabolites	30 day window, Beta (SE) for individual metabolites	30 day window, p-value for individual metabolites	ICA_factor 2 Weight	ICA_factor 2 Rank
Hypotaurine	Amino Acid	Methionine, Cysteine, SAM and Taurine Metabolism	−0.013(0.003)	4.62E-07	NS	NS	NS	NS	0.517134	20
Maltotriose	Carbohydrate	Glycogen Metabolism	−0.017(0.004)	7.11E-06	−0.021(0.004)	1.54E-06	−0.022(0.005)	6.54E-06	0.615801	9
3-phosphoglycerate	Carbohydrate	Glycolysis, Gluconeogenesis, and Pyruvate Metabolism	−0.016(0.003)	5.82E-06	NS	NS	NS	NS	0.856077	2
Nicotinamide	Cofactors and Vitamins	Nicotinate and Nicotinamide Metabolism	−0.012(0.002)	6.3E-07	NS	NS	NS	NS	0.514755	22
Dihomolinoleoylcarnitine (C20:2) *	Lipid	Fatty Acid Metabolism (Acyl Carnitine, Polyunsaturated)	−0.008(0.002)	1.16E-05	NS	NS	NS	NS	0.354472	66
Palmitoloelycholine	Lipid	Fatty Acid Metabolism (Acyl Choline)	0.014(0.003)	5.83E-06	NS	NS	NS	NS	−0.44944	37
Palmitoylcholine	Lipid	Fatty Acid Metabolism (Acyl Choline)	0.013(0.003)	9.48E-06	NS	NS	NS	NS	−0.47882	30
Oleoylcholine	Lipid	Fatty Acid Metabolism (Acyl Choline)	0.014(0.003)	3.33E-06	NS	NS	NS	NS	−0.48131	28
Stearoylcholine *	Lipid	Fatty Acid Metabolism (Acyl Choline)	0.015(0.003)	1.4E-05	NS	NS	NS	NS	−0.5512	15
8-hydroxyoctanoate	Lipid	Fatty Acid, Monohydroxy	0.013(0.003)	1.02E-05	NS	NS	NS	NS	−0.30251	95
Linolenate [alpha or gamma; (18:3n3 or 6)]	Lipid	Long Chain Polyunsaturated Fatty Acid (n3 and n6)	NS	NS	−0.01(0.002)	8.52E-06	NS	NS	0.055594	500
Palmitate (16:0)	Lipid	Long Chain Saturated Fatty Acid	NS	NS	−0.008(0.002)	7.3E-06	NS	NS	0.059332	473
1-stearoyl-2-oleoyl-GPS (18:0/18:1)	Lipid	Phosphatidylserine (PS)	−0.013(0.003)	6.87E-06	NS	NS	NS	NS	0.640158	8
Phosphoethanolamine	Lipid	Phospholipid Metabolism	−0.012(0.002)	5.48E-07	NS	NS	NS	NS	0.551825	14
Choline phosphate	Lipid	Phospholipid Metabolism	−0.009(0.002)	2.84E-06	NS	NS	NS	NS	0.422478	46
Sphingadienine	Lipid	Sphingolipid Synthesis	−0.015(0.003)	1.81E-06	NS	NS	NS	NS	0.696453	6
Sphingosine	Lipid	Sphingosines	−0.011(0.002)	8.05E-06	NS	NS	NS	NS	0.56189	13
Adenosine 5’-monophosphate (AMP)	Nucleotide	Purine Metabolism, Adenine containing	−0.015(0.003)	4.08E-06	NS	NS	NS	NS	0.796578	4
X - 18914	Xenobiotics	Chemical	NS	NS	NS	NS	−0.01(0.002)	3.99E-06	0.099545	274

These models were linear mixed-effect regression models (LMEM) with random participant-specific intercepts and simultaneously adjusted for PM_2.5, NO₂, O₃, and temperature (multi-pollutant models) for the same exposure window.

Beta (SE) and p-values presented are from individual metabolites analysis.

All models were adjusted for age (years), body mass index (kg/m2), cigarette pack-years, alcohol intake (< or ≥2 drinks per day), socioeconomic status (income payment and years of education), season (warm/cold), and relative humidity.

Significant metabolites at ENT95% significance level only are presented.

ICA_factor2 rank represents the rank of the corresponding metabolite contributing to factor 2 from the independent component analysis (ICA). Higher rank (and higher weights) means higher contribution to factor 2 of the ICA. We show alongside the weights and ranks of each metabolite that contributes to factor 2 from the ICA only because it was the only significant factor.

Range of facto2-ICA rank is from 1 to 1158 i.e., number of examined metabolites. Range of facto2-ICA weight is from −1 to 1.

The metabolites with an X-XXXX format are unknown but reproducible.

A single asterisk after a metabolite (*) indicates a compound that has not been confirmed based on a standard, but Metabolon is confident in its identity based on a subset of these analytical parameters.

Abbreviations: ENT: Effective/independent number of tests; PM_2.5, particulate matter ≤2.5 μm; NO₂, nitrogen dioxide; and O₃, ozone; ICA, independent component analysis; SE, standard error.

Table 4.

The significant adjusted associations between short-term exposure to temperature (predicted models) with metabolomics (multi-pollutant models) at ENT95% significance level

Biochemical	Super-pathway	Sub-pathway	1 day window, Beta (SE) for individual metabolites	1 day window, p-value for individual metabolites	7 day window, Beta (SE) for individual metabolites	7 day window, p-value for individual metabolites	30 day window, Beta (SE) for individual metabolites	30 day window, p-value for individual metabolites	ICA_factor2 Weight	ICA_factor2 Rank
guanosine	Nucleotide	Purine Metabolism, Guanine containing	NS	NS	NS	NS	0.032(0.006)	1.1E-06	0.33	78
2’-deoxyuridine	Nucleotide	Pyrimidine Metabolism, Uracil containing	NS	NS	NS	NS	0.027(0.005)	2.29E-06	−0.15	206
X - 25271	Un-identified	-	NS	NS	0.025(0.005)	5.58E-06	NS	NS	0.02	848
X - 24306	Un-identified	-	NS	NS	NS	NS	0.022(0.004)	9.08E-07	−0.17	184
X - 24307	Un-identified	-	NS	NS	NS	NS	0.027(0.006)	4.74E-06	−0.34	74

These models were linear mixed-effect regression models (LMEM) with random participant-specific intercepts and simultaneously adjusted for PM_2.5, NO₂, O₃, and temperature (multi-pollutant models) for the same exposure window.

Beta (SE) and p-values presented are from individual metabolites analysis.

All models were adjusted for age (years), body mass index (kg/m2), cigarette pack-years, alcohol intake (< or ≥2 drinks per day), socioeconomic status (income payment and years of education), season (warm/cold), and relative humidity.

Significant metabolites at ENT95% significance level only are presented.

ICA_factor2 rank represents the rank of the corresponding metabolite contributing to factor 2 from the independent component analysis (ICA). Higher rank (and higher weights) means higher contribution to factor 2 of the ICA. We show alongside the weights and ranks of each metabolite that contributes to factor 2 from the ICA only because it was the only significant factor.

Range of facto2-ICA rank is from 1 to 1158 i.e., number of examined metabolites. Range of facto2-ICA weight is from −1 to 1.

The metabolites with an X-XXXX format are unknown but reproducible.

A single asterisk after a metabolite (*) indicates a compound that has not been confirmed based on a standard, but Metabolon is confident in its identity based on a subset of these analytical parameters.

Abbreviations: ENT: Effective/independent number of tests; PM_2.5, particulate matter ≤2.5 μm; NO₂, nitrogen dioxide; and O₃, ozone; ICA, independent component analysis; SE, standard error; NS, not significant at the ENT95% level.

Table 5.

The adjusted associations between short-term exposure to air pollutants (prediction models) and temperature with the 5 factors of ICA (multi-pollutant models).

	24 hour- window, PM2.5			24 hour- window, NO2			24 hour- window, O3			24 hour- window, Temperature
ICA Factors	Beta (SE)	P-value	FDR P-value	Beta (SE)	P-value	FDR P-value	Beta (SE)	P-value	FDR P-value	Beta (SE)	P-value	FDR P-value
ICA_factor1	0.003(0.006)	0.61	0.99	−0.002(0.003)	0.48	0.99	0.001(0.003)	0.78	0.99	0.001(0.005)	0.80	0.99
ICA_factor2	−0.001(0.007)	0.92	0.99	−0.019(0.004)	9.30E-07	5.58E-05	−0.001(0.003)	0.77	0.99	0.006(0.006)	0.27	0.99
ICA_factor3	0.015(0.006)	0.02	0.99	0.004(0.003)	0.21	0.99	−0.004(0.003)	0.20	0.99	0.004(0.005)	0.39	0.99
ICA_factor4	−0.001(0.006)	0.88	0.99	0.004(0.003)	0.27	0.99	−0.002(0.003)	0.49	0.99	−0.001(0.005)	0.81	0.99
ICA_factor5	0.002(0.005)	0.73	0.99	−0.002(0.003)	0.47	0.99	0.002(0.002)	0.52	0.99	−0.004(0.004)	0.33	0.99
	7-day window, PM2.5			7-day window, NO2			7-day window, O3			7-day window, Temperature
ICA Factors	Beta (SE)	P-value	FDR P-value	Beta (SE)	P-value	FDR P-value	Beta (SE)	P-value	FDR P-value	Beta (SE)	P-value	FDR P-value
ICA_factor1	0.005(0.008)	0.53	0.99	−0.001(0.004)	0.87	0.99	−0.002(0.004)	0.65	0.99	0.004(0.005)	0.41	0.99
ICA_factor2	−0.002(0.011)	0.83	0.99	−0.019(0.004)	3.09E-05	0.002	−0.005(0.005)	0.33	0.99	0.009(0.007)	0.16	0.99
ICA_factor3	0.018(0.01)	0.06	0.99	0.011(0.004)	0.008	0.25	−0.008(0.004)	0.05	0.99	0.009(0.006)	0.12	0.99
ICA_factor4	−0.001(0.01)	0.89	0.99	0.008(0.004)	0.06	0.99	−0.0002(0.004)	0.95	0.99	−0.002(0.006)	0.68	0.99
ICA_factor5	0.0001(0.008)	0.99	0.99	−0.0001(0.004)	0.99	0.99	−0.001(0.003)	0.77	0.99	−0.0002(0.005)	0.96	0.99
	30-day window, PM2.5			30-day window, NO2			30-day window, O3			30-day window, Temperature
ICA Factors	Beta (SE)	P-value	FDR P-value	Beta (SE)	P-value	FDR P-value	Beta (SE)	P-value	FDR P-value	Beta (SE)	P-value	FDR P-value
ICA_factor1	−0.0001(0.011)	0.99	0.99	−0.0001(0.004)	0.99	0.99	0.001(0.005)	0.76	0.99	0.003(0.005)	0.61	0.99
ICA_factor2	−0.003(0.014)	0.85	0.99	−0.02(0.005)	7.92E-05	0.005	−0.005(0.006)	0.44	0.99	0.009(0.007)	0.22	0.99
ICA_factor3	0.025(0.013)	0.05	0.99	0.008(0.005)	0.11	0.99	−0.013(0.005)	0.01	0.70	0.006(0.006)	0.32	0.99
ICA_factor4	−0.003(0.013)	0.82	0.99	0.009(0.005)	0.05	0.99	0.002(0.005)	0.67	0.99	−0.003(0.006)	0.66	0.99
ICA_factor5	−0.007(0.01)	0.51	0.99	−0.002(0.004)	0.68	0.99	−0.0004(0.004)	0.93	0.99	−0.003(0.005)	0.57	0.99

These models were linear mixed-effect regression models (LMEM) with random participant-specific intercepts and simultaneously adjusted for PM_2.5, NO₂, O₃, and temperature (multi-pollutant models) for the same exposure window.

All models were adjusted for age (years), body mass index (kg/m²), cigarette pack-years, alcohol intake (< or ≥2 drinks per day), socioeconomic status (income payment and years of education), season (warm/cold), and relative humidity.

ICA_factor2 rank represents the rank of the corresponding metabolite contributing to factor 2 from the independent component analysis (ICA). Higher rank (and higher weights) means higher contribution to factor 2 of the ICA. We show along slide the weights and ranks of each metabolite that contributes to factor 2 from the ICA only because it was the only significant factor.

Range of facto2- ICA rank is from 1 to 1158 i.e., number of examined metabolites. Range of facto2- ICA weight is from −1 to 1.

Abbreviations: ENT: Effective/independent number of tests; PM_2.5, particulate matter ≤2.5 micrometers; NO₂, nitrogen dioxide; and O₃, ozone; ICA, independent component analysis; FDR, Benjamini-Hochberg false discovery rate; SE, standard error.

In the sensitivity analysis where exposures were based on the central site, in the adjusted multi-pollutant models, PM_2.5, NO₂, and temperature showed multiple statistically significant associations at the 95% level of the number of effective tests (ENT95%) (40 metabolites for PM_2.5 mostly with the 30-day average, 100 metabolites for NO₂, and 14 metabolites for temperature). However, O₃ was significantly associated (beta coefficient of −14.95 (p-value= 1.26e-05)) with only cysteine which is a component of methionine, cysteine, SAM and taurine metabolism (Supplemental Figure 3, Supplemental Tables 2-5). Consistent with this analysis, in the ICA analysis, PM_2.5, NO₂, and temperature but not O₃ were consistently and significantly associated with the second factor from the independent component analysis (ICA-factor 2) only and not the other four ICA-factors (Supplemental Table 6). Many of the most contributing metabolites to ICA-factor 2 were also individually significant with exposures of interest in the multi-pollutant models (Supplemental Tables 2-5).

In the pathway analysis, based on the significant metabolites (p-value<0.01) from the regression models, short-term exposure to NO₂ was associated with sphingolipid metabolism (p-value=0.01) and butanoate metabolism (p-value=0.10) (Figure 2 and Supplemental Table 7). Short-term exposure to temperature was associated with pyrimidine (p-value=0.01), glycolysis/gluconeogenesis (p-value=0.04), propanoate and pyruvate metabolisms (p-value=0.11). The pathways that were associated with the 100 highest contributing metabolites to ICA-factor 2 were purine (p-value=0.02), sphingolipid (p-value=0.03), beta-alanine metabolism (p-value=0.06), and glycerophospholipid (p-value=0.1) metabolisms.

Figure 2. Metabolic Pathways identified as enriched among the metabolites significantly associated with predicted exposures: (A) Short-term exposure to NO₂; (B) Short-term exposure to temperature; and (C) top 100 metabolites loading on to ICA-Factor 2.

Pathway analysis conducted using Metaboanalyst³⁹ is based on both over enrichment analysis, i.e., how many significant metabolites fall within a given pathway, and pathway topology analysis, i.e., how important those metabolites are to that pathway.

The figure shows the −log (natural logarithm) of the enrichment p-value on the x-axis, and the impact score on the y-axis which is based on the cumulative importance of all the significant metabolites within the pathway.

The size of each bullet represents the impact value.

The color of each bullet represents the significance of the enrichment.

Abbreviations: PM_2.5, particulate matter ≤2.5 μm; NO₂, nitrogen dioxide; and O₃, ozone; ICA, independent component analysis; FDR, Benjamini-Hochberg false discovery rate.

In the sensitivity pathway analysis using central site exposures, short-term exposure to PM_2.5 was associated with sphingolipid (p-value=0.007), beta-alanine (p-value=0.03), pyrimidine (p-value=0.07), glutathione (p-value=0.07), and glycerophospholipid (p-value=0.10) metabolism (Supplemental Figure 4). Short-term exposure to NO₂ was associated with glutathione (p-value=0.006), butanoate (p-value=0.03), glycerophospholipid (p-value=0.04), propanoate (p-value=0.1), and sphingolipid (p-value=0.1) metabolism. Short-term exposure to temperature was associated with glycerophospholipid (p-value=0.007), glutathione (p-value=0.08), and beta-alanine (p-value=0.1) metabolism. Similar to the significant super-pathway (methionine, cysteine, SAM and taurine metabolism) perturbed by O₃ exposure, taurine and hypotaurine metabolism was also noteworthy due to high impact score of 0.99 for temperature exposure and of 0.71 for the ICA-factor2.

Discussion

This is the first study, to date, to report the untargeted metabolomic signature of temperature exposure and the largest study to report untargeted metabolomic signature of air pollution. The observed results for ambient temperature exposure could help explain the previously reported excess mortality with both high and low short-term temperature exposure⁴³ and increased hospital admissions with high short-term temperature exposure ^44–46. The latter studies reported linear associations consistent with the linear associations with temperature that we found best fit our data. These results warrant intervention given the global climate change. We observed several significant associations between address-specific short-term air pollution as measured from prediction models (mainly NO₂) and temperature, with several blood metabolites. We identified two unique metabolic pathways (sphingolipid and butanoate metabolisms) associated with short-term exposure to air pollution and four pathways (pyrimidine, glycolysis/gluconeogenesis, propanoate, and pyruvate metabolisms) that were associated with temperature exposure. There was an overlap between the pathways when we used different statistical methods. This is the first study to use ICA to examine metabolic signature of air pollution. Previously most studies^47,48 have used PCA which produces uncorrelated factors, a necessary but not sufficient condition for independence. While ICA is a computational method for separating multivariate signal into additive factors that are maximally independent and assumes non-Gaussian signals³⁷.

With the central site exposures, we observed more significant associations between short-term air pollutants (mainly PM_2.5 and NO₂ and less with O₃) and temperature, while simultaneously examined, with several blood metabolites. We identified eight unique metabolic pathways (glycerophospholipid, sphingolipid, glutathione, beta-alanine, pyrimidine, butanoate, propanoate, and methionine, cysteine, SAM and taurine metabolisms) associated with short-term exposure to air pollution including four pathways that were also associated with temperature exposure. There was an even bigger overlap between the pathways when we used ICA. Significant metabolites and metabolic pathways were similar after we weighted the regression models by the inverse of the distance between the residential address and the central sites (Results not shown). The fact that the central site exposures had more significant hits than the address-specific predicted exposures suggests that the former captures purely temporal variation, or that it captures regional patterns more than the address specific exposures which could be more important with short-term exposures. The prediction models capture both spatial differences and temporal patterns, whereas the central site just captures temporal differences. The main source of variability in the short-term exposure is likely to be driven by the day to day variability as pollution generally goes up and down together over the region i.e., greater temporal variability than spatial variability. Therefore, the temporal variability might be more relevant than the spatial variability. Nevertheless, the results overlapped, and we present both results that will need to be replicated in future research.

The identified pathways were mainly related to fat metabolism (glycerophospholipid and sphingolipid metabolisms), inflammation (glycerophospholipid metabolism), oxidative stress (sphingolipid, glutathione, propanoate, glycolysis/gluconeogenesis, pyruvate, and taurine and hypotaurine metabolisms), immunity (glycerophospholipid and butanoate metabolisms), and nucleic acid damage and repair (purine, pyrimidine, and beta-alanine metabolisms). Short-term air pollution and temperature has been previously linked to inflammation^1,2 and oxidative stress^1,2. We observed perturbations in the glycerophospholipid metabolism with PM_2.5, NO₂, and temperature exposure. Glycerophospholipid is the main component of biological membranes. We observed perturbations in sphingolipid metabolism that is involved in inflammation and immunity^49,50 consistent with previous finding with short-term air pollution⁵¹. Sphingolipid has a role in membrane biology⁵² and has been linked previously to diabetes, Alzheimer’s disease, and hepatocellular carcinoma^53,54, heart failure⁵⁵, and cancer⁵⁰. Air pollution generates free radicals or acts directly as free radicals that induce oxidative stress⁵⁶. The cell membrane is one of the primary targets of reactive oxygen species derived from air pollutants. Oxidative stress can induce the activation of phospholipase A2 that hydrolyze phospholipid from the cell membrane to generate polyunsaturated free fatty acid and lyso-phospholipid⁵⁷. This metabolism has been reported in association to short-term traffic related air pollution in pregnant women^51,58 and to exposure to cigarette smoke in mice, another source of particulate exposure⁵⁹. We also observed perturbations in glutathione metabolism which is involved in xenobiotic-mediated oxidative stress^60,61 and has been reported before to be associated with traffic-related air pollution⁵⁸. We observed perturbations in methionine, cysteine, SAM & taurine and hypotaurine metabolisms with short-term O₃ and temperature exposures. Those are sulfur-containing amino acids that are readily oxidized⁶². Previous studies reported oxidation of methionine to be associated with ambient air-pollution exposure in pregnant women⁵¹ and in mice⁶³.We observed perturbations in propanoate metabolism that is involved in branched-chain amino acids metabolism and the short-chain fatty acid metabolism⁶⁴. Propanoate metabolism is also involved in colon microbiome and gluconeogenesis in the liver and acetate achieving the highest systemic levels in blood⁶⁵. We observed perturbations in purine, pyrimidine, and beta-alanine metabolisms; all related to nucleic acid metabolism, damage and repair⁶⁶ consistent with reported association with air pollution^19,67 and temperature⁶⁸. We observed perturbations in butanoate involved in immune regulation⁶⁹ and this metabolite has been reported previously to be associated with long-term air pollution⁷⁰.

By capturing those perturbations of multiple metabolic pathways, metabolomics will enable functionally relevant, holistic explorations of the impacts of short-term exposures to air pollution and temperature. Genomic, epigenetic, transcriptomic, and proteomic variability leads to alterations in metabolite levels in biological fluids and tissues⁷¹. Therefore metabolomics, as the downstream ‘ome’ most closely related to phenotype, provides an integrated profile of biological status and reflecting the ‘net results of these preceding omes’ and, most importantly, their environmental interactions. Furthermore, those metabolic signatures could be promising biomarkers for short-term exposure to air pollution and temperature. These results warrant more research in the promising field of human metabolomics to identify the specific pathways and metabolites perturbed by air pollution and temperature exposure and thus promising ways for prevention and therapy.

While yet limited, there is an increase in literature that examined the short-term effect of traffic-related air pollution on metabolomics^14–22. One study to date has examined the metabolic signature of temperature used targeted approach, limiting their ability to identify novel findings⁶⁸. None of the previous studies have used ICA in the analysis which is a new analytical approach for highly dimensional and correlated data and only a few co-adjusted for ambient temperature. Nevertheless, our results were consistent with the previous literature. A study conducted targeted metabolomics for 138 metabolites in older adults’ serum samples in Germany and reported association between short-term air pollution exposures, mainly NO₂, and lysophosphatidylcholines¹⁴. A randomized, crossover trial that enrolled 55 healthy college students in China reported that short-term exposure to PM_2.5 caused metabolic changes related to stress hormones, insulin resistance, and oxidative stress and inflammation markers¹⁵. Other experimental studies using a crossover design identified several metabolic features related to the short-term exposure to air pollution in the UK, Netherlands, and the US^16,17,21,22 as well as in urinary metabolomics and identified perturbation in energy metabolism, oxidative stress, and inflammation in a Chinese and US study^58,72. Another study used a multi-platform approach including gas chromatography–mass spectrometry, liquid chromatography-mass spectrometry, and nuclear magnetic resonance to identify metabolites in bronchial wash and bronchoalveolar lavage fluid samples from 15 healthy participants in Sweden exposed to biodiesel exhaust¹⁸. A panel study of 54 healthy college students in Georgia, USA characterized the metabolic profile of plasma and saliva samples associated with traffic-related air pollution exposure and identified metabolites and perturbed pathways e.g., oxidative stress, inflammation, and nucleic acid damage and repair¹⁹. A study that examined 160 pregnant women in California, USA, reported associations between maternal exposure during pregnancy to higher air pollution and metabolic pathways consistent with oxidative stress and inflammation pathways⁵¹.

Our study had a few limitations including that at present no profiling platform is capable of covering the entire metabolome, and there are likely a number of important metabolites, particularly metabolites of exogenous origin that we were unable to measure here, and which may be of importance to our exposure of interest. Further work should focus on leveraging other platforms targeted to these metabolites of exogenous origin. Nevertheless, Metabolon provides some of the broadest coverage of the metabolome to date using an untargeted approach, maximizing the potential for novel discoveries. Our study population were predominantly elderly white men who were residents in a low-polluted environment that may limit the generalizability of the results. Although, air pollution and temperature exposure continue to affect everyone, minority populations of certain racial and ethnic groups tend to be disproportionately more exposed than white men. Our results though were largely consistent with studies in pregnant women⁵¹. Nonetheless, future studies on more diverse populations (including younger participants, different racial and ethnic backgrounds, and women) and in moderately or highly-pollutant areas will be helpful. We did not have another population to replicate results. These results will need to be confirmed by future research among different population. We acknowledge the lack of data on recent diet and physical activity, however the limited data we had was on a different time scale (usual diet and physical activity habits and not short-term data). Furthermore, those variables are likely mediators as they are unlikely to predict exposures. Finally, given the observational nature of the study, we cannot exclude possible residual confounding. However, we adjusted for several covariates. Our study had several strengths including that it is the largest study to date that examined the untargeted metabolomic signatures of short-term exposures to air pollution and temperature simultaneously in the well-characterized longitudinal NAS. Furthermore, the NAS population were geographically stable and well followed-up since they were enrolled in 1963 with >80% response rates. In addition, we presented both address-specific and central site exposure and their associations with metabolomics to be further studied in eth future and we used several rigorous statistical methods as well as pathway analysis. Finally, we had 1,158 untargeted measured metabolites with repeat samples for a subset of men. Compared to targeted approaches, untargeted metabolomics has the potential to increase the coverage of the metabolome (increasing from several hundred markers to several thousands) and to reduce bias towards identifying well-studied metabolites¹⁶.

Conclusions

Metabolomics is a powerful tool for exploring the short-term effects of xenobiotic agents. In this study, we identified several significant metabolites and metabolic pathways associated with short-term exposure to air pollution and temperature; using a global and untargeted approach that both confirmed previous findings and revealed new ones. Those metabolic pathways were involved in inflammation and oxidative stress, immunity, and nucleic acid damage and repair. Cautions need to be taken while interpreting these exploratory results and they need to be confirmed by future research. These findings help us to better understand the biology of health effects of air pollution and temperature and may also support the development of biomarkers of exposure or mediators for the adverse health effects that could be targeted for prevention and therapeutic purposes. These findings demonstrate that exploiting the potential of metabolomics provides novel ways to help address the substantial global burden of air pollution and especially temperature exposure given global climate change.

Highlights.

1. Using untargeted metabolomics profiling, we identified significant metabolites associated with short-term exposure to air pollution and temperature.

2. We identified 6 metabolic pathways associated with short-term exposure to air pollution and temperature.

3. These pathways were involved in inflammation and oxidative stress, immunity, and nucleic acid damage and repair.