Analysis of post-deployment serum samples identifies potential biomarkers of exposure to burn pits and other environmental hazards

Abstract

Objective.

The potential health risks of deployment to sites with open burn pits remains poorly understood, in part, because personal exposure monitoring was not performed. Here, we investigated whether post-deployment serum samples contain biomarkers associated with exposure to burn pits.

Methods.

237 biomarkers were measured in 800 serum samples from deployed and never-deployed subjects. We used a regression model and a supervised vector machine to identify serum biomarkers with significant associations with exposures and deployment.

Results.

We identified 101 serum biomarkers associated with polycyclic aromatic hydrocarbons, dioxins or furans, and 54 biomarkers associated with deployment. 26 of these biomarkers were shared in common by the exposure and deployment groups.

Conclusions.

We identify a potential signature of exposure to open burn pits, and provide a framework for using post-exposure sera to identify exposures when contemporaneous monitoring was inadequate.

Introduction

During Operation Iraqi Freedom (OIF) and Operation Enduring Freedom (OEF), the United States Department of Defense adopted a policy of destroying trash generated at forward operation bases through the use of open burn pits (1–3). These were literally open pits dug into the ground, into which various types of garbage were thrown, doused with jet fuel, and burned. At Joint Base Balad in Iraq, for example, it was estimated that the base generated 5 pounds of trash per person per day. With a population of about 50,000 service personnel at any given time, this amounted to 125 tons per day of trash (2). Although open burn pits have largely been replaced by controlled incinerators at most established bases, they are still sometimes used in forward areas and temporary operating areas.

Burn pits were justified as operationally necessary, but there was only limited environmental monitoring (2, 4–6). Burn pits produce a number of airborne toxicants of concern including polyaromatic hydrocarbons (PAHs), polychlorinated biphenols (PCBs), polychlorinated dibenzo-p-dioxins (PCDDs) and polychlorinated dibenzofurans (PCDFs) (5, 6). A pilot study of breathing zone samples from security personnel at Bagram Airfield documented the presence of 23 volatile organic compounds and PAHs, although only acrolein naphthalene exceeded military exposure guidelines, and only on some occasions (4, 7). Individual exposures varied from 10 to 100-fold from day to day and among different work locations. Because of the lack of widespread exposure monitoring, exposure to burn pits poses an unknown level of risk of long-term health impacts. Although there are reports of respiratory effects, as yet, there are no clearly confirmed long-term health consequences associated with burn pit exposure (8–11); thus, there is ongoing concern, and the Veterans Administration has established a Burn Pit Registry to collect self-reported exposure and health outcome records from service personnel (12, 13). There remains a critical need to be able to assess personal exposures in order to understand the biological impacts of burn pit exposure and the potential for future health risks.

The long-term goal of our overall study was to determine whether archived serum samples from service personnel deployed to areas with open burn pits could be used to identify personal exposures and biological responses to burn pits and other environmental hazards associated with deployment to Iraq and Afghanistan during OIF and OEF (3) (see the Introduction to this issue). More broadly, can post-exposure serum samples be used to gauge exposures to occupational hazards retrospectively, when contemporaneous environmental and personal monitoring is inadequate. We obtained pre- and post-deployment serum samples from 200 service personnel deployed to Iraq or Afghanistan, and an equal number of control samples from service personnel who were never deployed. We performed high-dimension, high-throughput analysis including 29 cytokines and chemokines related to inflammation and cardiopulmonary risk, 177 microRNAs, and a metabolomics analysis with over 10,000 features, of which several hundred have been identified with high confidence (14–18). To date, we have demonstrated that archival serum samples contain metabolic signatures of nutritional status and PAH exposure (19) and we identified a number of microRNAs that were correlated with deployment and with serum levels of PCDDs and PCDFs (14, 20). However, we did not detect a strong signature of either PAHs, PCDDs or PCDFs in post-deployment serum; the only PCDD that had a statistically significant association between deployment and serum level was 1,2,3,4,6,7,8-heptachloro-p-dioxin (15). Since lightweight PAHs are metabolized relatively quickly (21, 22), and PCDDs and PCDFs accumulate in fatty tissues (23), it is likely that serum is not the best place to search for direct evidence of exposure to these compounds. However, prior exposure may result in long-term changes to the metabolome or to expression of microRNAs, cytokines or chemokines, that could be detected in serum. Here, we apply multiple statistical and “big data” approaches to tentatively identify a “signature” group of biomarkers that can classify subjects based on their exposure to burn pits during deployment.

Materials and Methods

Ethics statement

Serum samples and demographic data were de-identified, and the study was approved by the Institutional Review Boards of the Uniformed Services University and The University of Rochester.

Study Design

The overall study population has been described in detail elsewhere (1, 3, 14, 19). The Case group consisted of 200 service personnel who were deployed to Joint Base Balad in Iraq between 2005–2007 (Operation Iraqi Freedom) or Bagram Airbase in Afghanistan from 2011–2012 (Operation Enduring Freedom) who had been exposed to open burn pits, and for whom pre- and post-deployment serum samples were available at the Department of Defense Serum Repository (DoDSR) in sufficient quantity to complete the planned analyses. The Control group consisted of 200 personnel who were never deployed, who had consecutive serum samples available in the DoDSR, and who were matched to the Case group for time in service at the time of the first blood draw and matched for the length of time between first and second blood draw (1, 3). This strategy was adopted to control for biomarker changes associated with military service but not deployment. Demographic information was drawn from the Defense Medical Surveillance System (DMSS) and linked to the serum samples by anonymized code numbers. Upon examining the demographic database, we identified some minor discrepancies, such as post-deployment blood samples recorded as being drawn before the start of deployment. Rather than assume that the database was incorrect and impute a more logical value, we developed a strategy to exclude subjects with inconsistent dates (Case group, N=8) (see Figure S1, Supplemental Digital Content). Since we wanted to assess the effects of deployment-related environmental exposures on serum biomarkers, we also excluded a small number of subjects (Case group, N=5) with deployment duration <150 days. Since the Control group was never deployed and the dates of “pre” and “post” samples in the Control group were arbitrarily assigned, we only excluded Control subjects with duration between serum samples of <150 days (N=14). The remaining population of 187 Case and 186 Control subjects is described in Table S1, Supplemental Digital Content. Briefly, and as previously described (1), the Control group contained more female personnel, due to DoD policies regarding deployment of female personnel that were in place during the relevant time frames. The Case group also skewed slightly higher in rank, with a higher proportion of E5–E9 personnel. The Case group was made up of Army and Air Force personnel due to the study design (1) while the Control group was drawn randomly from the DMSS database. There were no significant differences in Age, Race, or Smoking Status.

Biomarker analysis

The serum sample was divided into several aliquots and processed as previously reported (see Khan et al., this issue). A panel of 21 cytokines and chemokines associated with inflammation, and a panel of 8 cardiovascular disease markers, was measured by multiplex analysis (FlexMAP 3D, Luminex Corp., Austin, TX) using Milliplex MAP kits (EMD Millipore, Billerica, MA) according to the manufacturer’s instructions. Serum cotinine was measured by commercial EIA (Calbiotech, Spring Valley, CA). Serum IgE was measured by commercial ELISA (Bethyl Laboratories, Montgomery, TX). Benzo(a)pyrene diol epoxide (BPDE)-protein adducts were measured by ELISA (OxiSelect™ BPDE Protein Adduct ELISA Kit, Cell Biolabs Inc., San Diego, CA) (20). RNA was extracted and used to determine levels of a panel of 144 human miRNAs commonly found in serum, as reported previously (14). A panel of 59 metabolites related to nutritional status and pesticide exposure was measured by mass spectrometry and reported previously (19). Serum levels of 17 polyaromatic hydrocarbons and 25 polychlorinated dibenzo-p-dioxins and dibenzofurans were measured by us using a novel small volume extraction method, as previously described (15). The final analysis included 279 biomarkers, as listed in Table S2, Supplemental Digital Content. Since the dioxins and furans were sparse and zero inflated, we performed principal component analysis on the dioxin and furan measurements and retained the first principal component (PC1) to use in subsequent analyses (see Khan et al., this issue). Thus, the regression and machine learning models used 236 variables.

Statistical Analysis

We developed an ordinary least squares regression model for 236 biomarker values using age, sex and tobacco use as covariates. Complete details of the regression analysis are reported in Khan et al., this issue.

Machine learning model for classification

We used a supervised linear support vector machine (SVM) (24) model to classify subjects as either Case (deployed) or Control (not deployed). Briefly, the SVM uses a subset of the data (a training set) to build a model that identifies which biomarker features are most useful in correctly classifying the training set as Case or Control. Then, the model is tested with a validation set to determine how accurately that classification is reproduced. In this analysis, we used 10-fold cross-validation, which generalizes the model and minimizes the chances of overfitting the data. The data set was divided into 10 subgroups, and for each run, 9 subgroups were used as the training set and the 10^th was used for validation. This process was repeated 10 times, with each subsample used once as the validation set. The results of the 10 validation runs were averaged to produce a single estimate of the overall accuracy of the SVM model. To further improve the accuracy of the model and for feature selection, we used recursive feature elimination (25). The SVM was first run using all 236 biomarker features. After each run, the least-weighted biomarker was removed and the SVM re-run on the reduced data set. Each iteration was assessed by 10-fold cross-validation. The SVM was developed both with and without adjustment for age, sex and tobacco use. Both models assigned similar weights to most features, but at each step in the RFE, the unadjusted model required fewer features to achieve the same accuracy and cross-validation, so only the SVM using the unadjusted data is reported here. Complete details of the SVM model are reported in Khan et al., this issue.

High-Resolution Mass Spectrometry and Data Pre-processing

The metabolome was processed as described in Go et al. in this issue. In brief, a 65 μL serum aliquot was added to 130 μL of acetonitrile containing a mixture of stable isotope-labeled internal standards and prepared for metabolomics analysis using our previously established method (26–28). Samples were analyzed in triplicate by liquid chromatography with Fourier transform mass spectrometry (Dionex Ultimate 3000, Q-Exactive, Thermo Fisher) with C₁₈ chromatography/positive electrospray ionization (ESI) mode and resolution of 70,000 (29). Spectral m/z features were acquired in scan range 85–1,250 mass-to-charge ratio (m/z). Raw data files were extracted using apLCMSv6.3.3 (30) with xMSanalyzerv2.0.7 (31), followed by batch correction with ComBat (32). Uniquely detected ions consisted of m/z, retention time and ion abundance, referred to as m/z features.

Metabolic Feature Selection

Prior to data analysis, triplicate injections were averaged and only m/z features with at least 80% non-missing values in either of the groups and more than 50% non-missing values across all samples were retained. After filtering based on missing values, data were log2 transformed and quantile normalized (33).

Mapping the interactions of the metabolome with miRNAs using xMWAS

miRNA and HRM data from the Case sample group were integrated by using xMWAS (34). We used sparse partial least-squares regression, a variable selection and dimensionality reduction method with visualization performed by the package mixOmics, to conduct pairwise correlation analysis between the miRNA (364 samples × 14 microRNA’s) and the metabolome (364 samples × 3011 metabolic features) which had been quantile normalized and log-transformed). Ten samples were removed from the Case group before xMWAS analysis due to the absence of microRNA detection in either the pre or the post-deployment samples leaving a total of 364 samples in the Case group. Thresholds for determining significant associations must have met the correlation threshold criteria (|r|>0.3)) and p < 0.05 as determined by Student’s T-Test.

Metabolite annotation:

Metabolic features were annotated using xMSannotator (35); confidence scores for annotation by xMSannotator are derived from a multistage clustering algorithm. Identities of selected metabolites were confirmed by co-elution relative to authentic standards and ion dissociation mass spectrometry (Level 1 identification by criteria of Schymanski et al. (36)). Supplemental annotations were made based on high or medium confidence (≥ 2) with M+H adducts detected in the positive mode. Lower confidence annotations were made using either the KEGG, (Kyoto Encyclopedia of Genes and Genomes) (37) or the HMDB (Human Metabolome Database) (38) at a 5ppm tolerance.

Results

Micro-RNAs associated with serum levels of polychlorinated dibenzo-p-dioxins and dibenzofurans

We first investigated whether there were associations between serum miRNAs and serum levels of polychlorinated dibenzo-p-dioxins (PCDDs) and polychlorinated dibenzofurans (PCDFs). miRNAs were chosen for the initial examination since they are persistent in serum and might reflect epigenetic changes in gene expression due to prior exposures, and because our earlier analysis had reported only miRNA associations (14). This analysis used both Case and Control serum values and was agnostic as to the source of the exposure; we were looking for associations between miRNA and PCDDs/PCDFs regardless of source. Using an uncorrected regression model, we identified 56 miRNAs that were significantly associated with one or more species of PCDD or PCDF (14). Here, we repeated the analysis using our new regression model that is corrected for age, sex and smoking status. We identified 50 miRNAs with statistically significant associations serum levels of PCDDs and PCDFs, of which 39 miRNAs were in common between both models (Figure 1A and Tables S3 and S4, Supplemental Digital Content).

Figure 1.

A. Comparison of miRNAs that were significantly associated with polychlorinated dibenzo-p-dioxins and dibenzofurans in an uncorrected model (Woeller et al., J Occup Environ Med. 2016 Aug;58(8 Suppl 1):S89–96.) and a corrected model (this work). (a), see Table S3, Supplemental Digital Content; (b), see Table S4, Supplemental Digital Content. B. Comparisons of all biomarkers significantly associated with polychlorinated dibenzo-p-dioxins and dibenzofurans in the corrected model and biomarkers significantly associated with BDPE-protein adducts in serum. (c) see Table 1. All raw data for Figure 1B is found in Table S5, Supplemental Digital Content.

We next investigated associations of all biomarkers with PCDDs and PCDFs and with benzo[a]pyrene diol epoxide (BDPE)-protein adducts. Benzo[a]pyrene is a polycyclic aromatic hydrocarbon (PAH) that is more likely associated with transportation sources (aerodrome, motor pool) and diesel electric generators than with burn pits (6), but its metabolite BDPE is highly reactive and a potent mutagen, and serum levels of BDPE-protein adducts reflect prior exposure to PAHs (39, 40). We reasoned that biomarkers common to serum levels of PCDDs, PCDFs and BDPE-protein adducts would likely be important indicators of past exposure to airborne hazards from combustion sources such as found at military bases with transportation hubs and burn pits. (Here, we are using total biomarkers including miRNA, cytokines, chemokines and metabolites, so the number of associated biomarkers is higher than in the previous analysis of miRNA only.) 84 biomarkers had significant associations with PCDDs and PCDFs, while 51 biomarkers had significant associations with BDPE-protein adducts (Figure 1B and Table S5, Supplemental Digital Content). Of these, 34 were common to both (Table 1). Of particular interest are oxoproline, two members of the tryptophan pathway, kyneurinine and indoleacrylic acid, miR-144–3p and two members of the let-7 family of miRNAs. These will be discussed in more detail below.

Table 1.

Biomarkers significantly associated with both polychlorinated dibenzo-p-dioxins and dibenzofurans, and BDPE-protein adducts.

Species	p-value (BDPE-protein adducts)	p-value (PCDD/PCDF PC1)
Acetyl carnitine	0.0165	2.37E-05
Adipsin	1.33E-04	1.46E-06
Aminobutyrate	8.85E-04	0.016
Asparagine	2.30E-03	6.45E-05
Creatinine	1.62E-07	0.016
Dibutyl phthalate	5.99E-04	7.67E-03
Fetuin A	3.70E-03	5.58E-03
Glucose	3.53E-11	1.44E-19
Glutamate	0.0438	5.83E-06
Glutamine	3.57E-06	2.84E-06
Hypoxanthine	2.19E-05	1.18E-04
Indoleacrylic acid	2.38E-03	5.09E-03
Kyneurinine	0.0704	9.48E-03
let-7b-5p	0.0204	8.59E-03
let-7i-5p	4.05E-03	2.22E-06
Lysine	9.91E-09	4.46E-06
LysoPC 20:4	4.23E-03	9.86E-07
miR-106a-5p	6.06E-04	1.02E-05
miR-107	0.0293	3.98E-03
miR-130b-3p	1.42E-04	0.021
miR-140–3p	4.70E-08	4.29E-12
miR-142–3p	7.19E-03	7.56E-15
miR-142–5p	7.19E-03	6.50E-03
miR-144–3p	0.0293	5.51E-03
miR-16–5p	0.0190	2.86E-08
miR-191–5p	5.88E-05	2.02E-09
miR-19b-3p	1.23E-03	1.02E-03
miR-22–3p	2.62E-04	8.55E-03
miR-320b	0.0114	0.016
Oxoproline	2.08E-07	2.07E-08
Proline	0.0197	3.60E-03
Serine	4.97E-06	0.025
sn-glycero-3-Phosphocholine	6.63E-07	9.60E-04
Sphingosine	0.0198	0.020
Tryptophan	7.92E-05	0.033

Biomarkers associated with deployment to Iraq or Afghanistan

We used the regression model corrected for age, sex and smoking status, to identify biomarkers that were significantly associated with Case Post-Deployment, compared to Case Pre-Deployment, Control Pre-Deployment, and Control Post-Deployment (Table 2). 28 biomarkers were significantly associated with deployment, including biomarkers previously shown to be associated with PCDDs, PCDFs and BPDE-adducts, including oxoproline, indoleacrylic acid, miR-144–3p and the let-7 family. Seven additional biomarkers that were significant between Case Post-Deployment and Case Pre-Deployment were also significant between Control Pre-deployment and Control Post-deployment, suggesting that they are likely related to age or time in service rather than deployment status.

Table 2.

Significant features in the Case Post-Deployment group using the corrected regression model.

Feature	P-value
Acetylphosphate	0.035
Alpha2 macroglobulin	0.011
Eotaxin	0.048
Indoleacrylic_acid	0.043
Kynurenine	5.3E-04
let-7d-5p	0.027
LysoPC 18:0	0.030
Methionine	1.2E-04
miR-101–3p	0.014
miR-144–3p	0.026
miR-145–5p	1.7E-03
miR-17–5p	0.021
miR-197–3p	0.035
miR-223–5p	1.7E-04
miR-23a-3p	4.0E-03
miR-24–3p	1.7E-03
miR-32–5p	0.039
miR-320b	3.1E-03
miR-34a-5p	0.043
miR-374a-5p	3.1E-03
miR-451a	1.1E-04
miR-486–5p	3.8E-03
miR-660–5p	0.013
miR-92a-3p	3.7E-03
miR-93–5p	0.012
Oxoproline	0.013
Phosphatidyl choline 36:4	0.038
Serine	0.037

^*Features that were also significant in a comparison between Control-Pre and Control-Post are removed, as they are likely related to time in service rather than deployment. These were; Acetyl carnitine, L-selectin, miR-21–5p, miR-326, miR-590–5p, platelet factor 4, and Serum amyloid A.

We also used the SVM model to identify features that accurately classified Case (deployed) and Control subjects. The first run, using all 236 biomarkers, was 92% accurate but with only 66% cross-validation accuracy. To improve cross-validation accuracy we performed recursive feature elimination. At the early stages, RFE can significantly improve validation accuracy as it removes the least-weighted classifiers, which may include biomarkers that are unrelated to deployment and constitute noise rather than signal. The feature set with the best combination of training accuracy and cross-validation accuracy contained 70 features, and achieved 86.3% training accuracy and 83.5% cross-validation accuracy (Table S6, Supplemental Digital Content). To arrive at a feature set of more manageable size, we arbitrarily selected the smallest data set that achieved both >80% training and cross-validation accuracy. This set contained 37 members, as shown in Figure 2, and had 82.0% training accuracy and 80.3% cross-validation accuracy. Features are ranked by weight assigned by the SVM model. The highest weighted classifier was oxoproline; miR-144–3p, kynurenine and indoleacrylic acid were also present as before. Several markers of inflammation and cardiovascular effect were also found, including IL-6, IL-8, serum amyloid A and α2-macroglobulin.

Figure 2. Classifiers of deployment identified by supervised vector machine (SVM) analysis, with their relative weights.

miRNA are highlighted in gray, other biomarkers in white.

Finally, we compared the results of the regression model with the SVM model (Figure 3). Although the results do not overlap completely, 11 features were common to both models, suggesting that they are strongly associated with deployment and represent biological changes that occurred as a result of deployment. That the two feature sets do not completely overlap is not surprising as the two methods ask different questions. The regression model asks which features are significantly different between groups on an individual basis; the SVM asks, when considering all the features together, which are most useful in correctly categorizing the subjects. This is discussed further in Kahn et al., this issue.

Figure 3. Comparison of biomarkers that were significantly associated with deployment by regression model and biomarkers identified as classifiers of deployment by SVM model.

A. Venn diagram. B. List of biomarkers in each category.

Which biomarkers are associated with both deployment and airborne pollution?

The original goal of the overall study was to identify biomarkers that are associated with burn pit exposure during deployment; here we have so far identified biomarkers associated with deployment (Figure 3), and with certain chemicals (PCDDs and PCDFs) produced by burn pits (Table 1). We compared the lists of significant associations to identify biomarkers present in both groups. 26 biomarkers were commonly associated with both deployment and serum levels of PCDDs, PCDFs, or BDPE-protein adducts (Figure 4). These included a number of metabolites including choline and acetyl phosphate; markers of cardiovascular risk including fetuin A and serum amyloid A; and a number of microRNAs previously reported to be involved in inflammation or biological responses to pollutant exposure.

Figure 4. Comparison of biomarkers associated with burn pit exposure and deployment.

A. Comparison of biomarkers that were associated with exposure to dioxins, furans or BDPE-protein adducts; and biomarkers associated with deployment by either the regression or SVM models. We used the most expansive definition; biomarkers were included if they were associated with PCDDs or PCDFs or BDPE-adducts (Figure 1B); and associated with deployment by regression analysis or SVM (Figure 4A). B. List of biomarkers associated both with environmental exposure and deployment. ^†found in communities associated with the exposure metabolome; _§response to exposure hazards evaluated in Woeller et al, this issue.

Association of the exposome with microRNAs related to deployment

Another technique to identify biomarkers common to deployment and exposure to deployment-related hazards is to use metabolome-wide association studies (MWAS). We used high performance LC-MS-MS to determine over 3000 metabolic features in the Case deployment serum samples (see Go et al in this issue). These features include both endogenous metabolites as well as metabolites that originate from diet, lifestyle, environment and behavior, or from metabolic products from those external stressors known as the exposome (41, 42). Here, we investigated whether there were significant associations between a panel of 14 miRNAs that classify deployment (from Figure 3) and the exposome in the Case group using multi-omic integration. As shown in Figure 5, xMWAS analysis identified 3 communities of metabolites that had statistically significant associations with 6 miRNAs associated with deployment. The largest community (Community 2) shows positive associations between miR-103a-3p and multiple features of the exposome including 2,3,7,8-Tetrachlorodibenzo-p-dioxin (TCDD), Tetrachloro-biphenyldiol (a polychlorinated biphenyl or PCB), and iodofenphos, an organophosphate pesticide. This community also contains sphinganine and phosphatidyl cholines, which are known to be dysregulated in inflammation (43–45). Community 1 includes metabolites from community 2 which are positively associated with miR-103 but negatively associated with miR-342–3p, including tetrachloro-biphenyldiol and prometon, an herbicide used by the DoD. Community 3 consists of a number of metabolites that are also positively associated with miR-103a but negatively associated with miR-92a, miR-93a, miR-320 and let-7d. This shows that these miRNAs and metabolites interact in a coordinated network in which metabolic changes are associated with elevated levels of some miRNAs and depressed levels of others in individuals that were deployed. For a complete list of annotated metabolites, see Table S7, Supplemental Digital Content.

Figure 5. Association of the exposome with microRNAs related to deployment.

3011 metabolic features detected within the Case-Post deployment group were examined for their association with 14 specified microRNA’s using xMWAS. The miRNAs were selected from deployment-related miRNAs in Figure 4 (let-7d-5p, miR-103a-3p, miR-144–3p, miR-145–5p, miR-15b-3p, miR-32–5p, miR-320b, miR-326, miR-342–3p, miR-374a-5p, miR-409–3p, miR-486–5p, miR-92a-3p, miR-93–5p). Three major metabolic communities were detected. Community 1 (grey) associated miR-342–3p with 8 metabolites including tetrachloro-biphenyldiol (a PCB) and prometon/secbumeton/terbumeton (an herbicide used during deployment). In community 2 (white), miR103a-3p was found to be associated with 23 metabolites including triglycerides, sphinganine, phosphocholines, TCDD, and iodofenphos (an organophosphate pesticide). Community 3 (black) indicates associations between miR92a-3p, Let-7d-5p, miR320b, and miR93–5p with 5 metabolites which includes lactate. (|r| > 0.3 at p <0.05). The metabolites in community 1 and 3 form a complex network with community 2 via common association with miR-103a-3p. Grey dotted lines, black dotted lines, rectangles and circles indicate positive associations, negative associations, metabolites, and miRNAs, respectively. A color version of this Figure is available as Figure S2 in the Supplemental Digital Content.

Discussion

The critical problem in understanding the risks of burn pit exposure is a lack of personalized quantitative data on individual exposures. We may assume that personnel stationed close to the burn pits had higher exposure than other base personnel, but even this is affected by wind direction, location of duty stations, barracks, and other factors that are poorly documented. The Department of Defense is developing an Individual Longitudinal Exposure Record (ILER) that will keep track of an individual’s potential exposures and be accessible to future health care providers, but this has not been implemented yet. For now, the only exposure records available for warfighters during OIF and OEF are self-reported, either on form DD2966 “Post deployment health assessment”, form DD2900 “Post-deployment health reassessment” or reported to the VA Airborne Hazards and Open Burn Pit Registry (12, 13), and even these records are not quantitative and don’t identify the specific exposure. This study was undertaken to investigate whether burn pit exposure leaves a biological signal that can be identified in archived serum samples. We were unable to identify a statistically significant increase in any hydrocarbon pollutants associated with burn pit exposure other than 1,2,3,4,6,7,8-heptachloro-p-dioxin in post-deployment serum samples (15). However, we reasoned that such exposures might result in other biological changes that would have a serum signal, including changes in metabolism, inflammatory signaling pathways, and microRNA expression. We report here the identification of a group of biomarkers that were significantly associated with exposure to chemicals produced by burn pits and motor transportation (Figure 1) and another group of biomarkers that could be used to classify individuals as either deployed or not deployed (Figures 2 and 3). These two groups of biomarkers had 26 elements in common (Figure 4) which we suggest may represent a biological signal of exposure to burn pits during deployment. Given that specific environmental toxicants may be short-lived, or difficult to detect in serum, but nevertheless may have long-term health consequences, the potential to use miRNAs as a readily detectable source of information about past exposures is a significant advance toward the goal of retrospective identification of the health impacts of deployment to areas with environmental hazards.

The best way to confirm that these biomarkers are responsive to chemicals associated with burn pits would be to perform a second study on an independent cohort of service personnel deployed to a different environment where burns pits are in use. Unfortunately, we are constrained by sample availability and the movement of the DoD toward the use of regulated incinerators. Therefore, we have used two different alternative approaches to confirm our findings. For one, we have exposed primary human lung cells to selected PAHs and PCDDs, and confirmed that the exposures alter their expression of key miRNAs identified in this study (see Woeller et al. in this issue). We also took a bioinformatics approach in which we looked for correlations between key miRNAs and our metabolomic results, to see if we could confirm that these miRNAs were associated with environmental exposures (Figure 5). Six miRNAs that were associated with PCDDs, PCDFs, or PAHs, and deployment, formed 3 communities with a large group of metabolic features. Some of these features included important dioxins and PCBs, a pesticide, an herbicide, and sphingolipids, and phosphatidyl cholines, which are known to be dysregulated in inflammation (45). Go et al. (this issue) report that a similar MWAS analysis identified associations between sphingolipids and tetrachloro-biphenyldiol and serum levels of the PAHs benzo(ghi)perylene and fluoranthene. These results are supportive of our hypothesis that exposure to PAHs, PCDDs and PCDFs causes metabolic changes associated with miRNAs, and that these miRNAs could be signals of exposure to environmental hazards including airborne pollutants produced from burn pits.

The importance of these biomarkers for understanding potential long-term health risks of burn pit exposure needs further exploration. MicroRNAs are small RNA hairpin structures that act to repress their target genes. miRNAs frequently target multiple genes in the same metabolic pathway, resulting in coordinated changes in gene expression along that pathway (46, 47). Expression of miRNAs can be controlled epigenetically, meaning that alterations in miRNA expression can persist even after the original insult is no longer detectable (48), and recent research has identified specific miRNAs that are associated with specific diseases (49–51). The biologic targets of the miRNAs identified in Figure 4 and 5 include genes involved in lipid and glucose metabolism, diabetes and fatty liver disease, and cell proliferation and cancer (52–59). Conversely, miRNAs Let-7d and miR342 are considered tumor suppressors, so that loss of expression could contribute to increased cancer risk. (60–62). miR-320b showed a high correlation with PAH levels in coke oven workers in China (63), and was associated with increased cardiovascular risk, further supporting an association between miR320b and burn pit exposure. Finally, miR-103 was elevated in lung transplant patients with bronchiolitis obliterans syndrome, suggesting it could be an important biomarker for lung disease (64).

In addition to miRNAs, altered metabolism of several amino acids is also common to deployment and exposure to hydrocarbon air pollution. The importance of most of these is difficult to understand without further study. However, we are particularly intrigued by the presence of indoleacrylic acid, kynurenine and tryptophan among the list of biomarkers associated with both deployment and air pollution. The tryptophan pathway is closely tied to the aryl hydrocarbon receptor (AhR). While the AhR is best known as the key transcription factor responsible for sensing PAH exposure and upregulating Phase I and Phase II detoxifying pathways (65), it also plays important roles in downregulating immune responses to maintain homeostasis, and in cancer suppression. Kynurenine is an endogenous ligand for the AhR, and is produced from tryptophan by the indoleamine 2,3-dioxygenase (IDO) family of enzymes, while IDO activity is regulated by AhR signaling (66–68). Indoleacrylic acid is an endogenous inhibitor of tryptophan synthetase (69). Activation of AhR signaling due to exposure to hydrocarbon air pollution would be predicted to upregulate production of kynurenine to maintain immune homeostasis; loss of this response could contribute to the development of an inflammatory response, autoimmunity or cancer (70, 71). Thus, identification of kynurenine as an important classifier of both burn pit exposure and deployment is highly suggestive and worthy of further study.

An important limitation of this study is that it was performed on a small number of individuals (200) who were apparently healthy and did not show increased risk of any disease after deployment (1). Future studies might benefit from starting with a larger group of deployed personnel and then identifying subgroups for analysis who present with possible deployment-related illness. The availability and long-term stability of serum collected in the DoD Serum Repository makes it feasible to identify personnel with specific diseases decades after deployment, and compare their immediate post-deployment sera with post-deployment sera from personnel with similar deployment records who did not get sick. Future studies might also benefit by integrating the ILER database, such as by identifying a large group of deployed personnel and then stratifying their exposures using the ILER. Focusing on (for example) the top and bottom deciles (most and least exposed) may maximize the signal to noise ratio.

Another issue regarding biomarkers of deployment (Figures 2 and 3) is that deployed personnel are subject environmental factors beyond combustion sources like burn pits and motor transportation, including insects, sandstorms, change in diet and sleeping habits, combat stress, and many others. Out of 277 biomarkers studied, we identified 54 that were significantly associated with deployment, only 26 of which were also associated with exposure to combustion products like burn pits. Future study is needed to identify whether these other biomarkers are biologically important and what their triggers are. This could be addressed in part, by studying an independent cohort of personnel deployed to a different environment.

It is also important to recognize that we have not established a cause and effect relationship between miRNAs and the metabolome. It could be that exposure to air pollution associated with burn pits causes metabolic changes which then result in changes in miRNA expression; or that these chemicals cause direct changes in gene expression including miRNAs that are translated into alterations of metabolic pathways. Because PAHs, dioxins and furans were below the limit of detection in >50% of post-deployment serum samples, but the miRNAs were detected in 70–99% of samples, we think this supports the idea that miRNAs are suitable surrogate reporter of environmental exposures, especially when the exposures are difficult to detect in the small amounts of serum available in this study.

Finally, although we have taken substantial steps to show that biomarkers associated with deployment to burn pit areas in the machine learning and statistical models are also associated with PAHs and PCDDs using independent data sets and experiments, we are aware that association does not prove causation, and further work in this area is needed.

Taken together with our previous studies ((14–16, 20) and this volume) we conclude that archived serum samples from the DoD Serum Repository represent a viable source of material to study the health effects of deployment; that deployment leaves a biological signature in the serum; and that some features of the deployment signature are also consistent with exposure to hydrocarbon air pollution produced by open burn pits and transportation activities. This supports the concept that routine post-deployment serum samples can be used to obtain (indirect and retrospective) information about a warfighter’s exposure to environmental hazards. These results will form the foundation for future studies of potential deployment-related illnesses.

Clinical Significance:

This work demonstrates the feasibility of using post-deployment serum samples to retrospectively identify environmental exposures that occurred during deployment in warfighters. This has broad applicability to enable retrospective assessment of occupational exposures on an individual level when contemporaneous environmental and personal monitoring is inadequate.