Airway Microbiome Is Associated with Respiratory Functions and Responses to Ambient Particulate Matter Exposure

Abstract

Background

Ambient particulate matter (PM) exposure has been associated with respiratory function decline in epidemiological studies. We hypothesize that a possible underlying mechanism is the perturbation of airway microbiome by PM exposure.

Methods

During October 2016 – October 2017, on two human cohorts (n = 115 in total) in Shanghai China, we systematically collected three categories of data (1) respiratory functions, (2) airway microbiome from sputum and (3) ambient PM_2.5 (PM of ≤ 2.5 micron in diameter) – level in ambient air. We investigated the impact of PM_2.5 on airway microbiome as well as the link between airway microbiome and respiratory functions using mix linear regression models.

Results

The respiratory function of our primary interest includes forced vital capacity (FVC) and forced expiratory volume in 1st second (FEV₁). FEV₁/FVC, an important respiratory function trait and key diagnosis criterion of COPD, was significantly associated with airway bacteria load (p = 0.0038); and FEV₁ was associated with airway microbiome profile (p = 0.013). Further, airway microbiome was significantly influenced by PM_2.5 exposure (p = 4.48E-11).

Conclusions

To our knowledge, for the first time, we demonstrated the impact of PM_2.5 on airway microbiome, and reported the link between airway microbiome and respiratory functions. The results expand our understanding on the scope of PM_2.5 exposure’s influence on human respiratory system, and pointed to novel etiological mechanism of PM_2.5 exposure induced diseases.

Introduction

Air pollution is identified as a leading cause of global disease burden, especially in low-income and middle-income countries (Cohen et al., 2017). China is one of the regions in the world of severe ambient air pollution (Wang et al., 2017a; Xu et al., 2016). As a major component of airborne pollutants, particulate matter (PM) has been linked to pulmonary disorders (Bahr et al., 2013; Beelen et al., 2014; Bell et al., 2013), including chronic obstructive pulmonary disease (COPD) (Wang et al., 2017b). Particulate matter of ≤ 2.5 micron in diameter is termed as PM_2.5, and its adverse effect on respiratory functions has been reported by twenty years of epidemiology studies (Xing et al., 2016). The “Harvard Six Cities Study” demonstrated that PM_2.5 exposure was causally associated with non-accidental death (Schwartz et al., 1996). Respiratory diseases account for a large proportion of the non-accidental deaths caused by airborne pollutant exposure (Schwartz et al., 1996). Studies also indicated the causal role of PM_2.5 in asthma, respiratory inflammation, and decline in respiratory functions (https://developer.apple.com/healthkit/; Lewis et al., 2005; Ostro et al., 2006). It was estimated that the prevalence of respiratory diseases increased by 2.07% and hospitalization rate elevated by 8% when daily PM_2.5 concentration increased by 10 µg/m³ (Dominici et al., 2006; Zanobetti et al., 2009). Further, the impact of PM_2.5 exposure is more profound in the elderly, pregnant women, adolescents, infants, patients with a history of cardiopulmonary problems and other susceptible populations (de Oliveira et al., 2012; Huynh et al., 2006; Martinelli et al., 2012).

The exact mechanism that PM_2.5 exposure reduces respiratory function is unknown, and there are multiple molecular etiology pathways possible: (1) Injury from free radical peroxidation. Components of PM_2.5 can induce free radical production, oxidize lung cells and directly lead to injury (Donaldson et al., 1996; Greenwell et al., 2002; Kelly, 2003; Rahman and MacNee, 1996). The reactive oxygen species (ROS) generated by particles, particularly by water soluble particles, produce hydroxyl radical (• OH) by activating metals (Valavanidis et al., 2005). The PM_2.5 surface is rich in iron, copper, zinc, manganese, and other transition elements, as well as polycyclic aromatic hydrocarbons (PAH), which are able to increase free radical production in the lung, consume antioxidant ingredients and cause oxidative stress (Donaldson et al., 1996). (2) Injury from inflammatoryresponse. PM_2.5 is related to inflammatory cytokines as it upregulate multiple inflammation-inducing transcription factor (TF), which leads to increase in the number of neutrophils, eosinophils, T cells and mastocytes in bronchoalveolar lavage (BAL) fluid (Gripenback et al., 2005; Sigaud et al., 2007). The interactions between inflammatory cells and cytokines can damage lung cells synergistically (Xing et al., 2016). COPD is a leading cause of morbidity and mortality worldwide (http://www.goldcopd.org; Wang et al., 2017b). COPD has a complex etiology, where inflammation plays an important role (Barnes, 2014). Oxidative stress may further amplify the inflammation (Rahman, 2005), and it is established that PM_2.5 contribute to the oxidative stress (Wong, 2017). (3) Changes in airway microbiome. Humans and microbes have evolved together for millions of years. Humans harbor 10 times more microbes inside or on their bodies than the total number of their somatic and germ cells (Wang et al., 2017b). Human Microbiome Project (HMP)(Blaser, 2010) showed that we greatly underestimated the diversity, prevalence, and persistence of the bacteria in our body, partly because majority of the bacteria cannot be cultured in the laboratory (Hayashi et al., 2002; Suau et al., 1999). The development of culture-independent techniques for microbiological analysis (e.g., 16S rRNA sequencing) showed the airway and bronchial tree is not sterile even in healthy individuals (Huang et al., 2013). Instead, the lung is a complex microbial ecosystem, the microbiota resides in the bronchial tree and parenchymal tissues, which could be important for colonization resistance, epithelial integrity, and immune-regulation. Different taxonomic composition was reported in sputum microbiome samples from eight COPD patients and ten “healthy” smokers (Cameron et al., 2016). The changes of microbiome composition and functional capacity were associated with COPD severity (FEV₁% of predicted) (Cameron et al., 2016).

To date, the link between airway microbiome and respiratory functions, particularly in non-COPD individuals, has not been fully studied. There are no empirical data assessing the impact of PM_2.5 exposure on airway microbiome. As results, the roles of airway microbiome, in terms of mediating the adverse effect of PM_2.5 exposure on respiratory functions and biomarkers for airway micro-environment damage caused by PM_2.5 exposure, remain unknown. In order to fill in such knowledge gaps, three categories of data were collected on human cohorts of this study: (1) respiratory functions, (2) airway microbiome from sputum and (3) levels of PM_2.5 exposure. We investigated the association between airway microbiome and respiratory functions, as well as, the perturbation caused by PM_2.5 exposure.

Materials and Methods

COPD risk cohort

Seventy-four participants from a community-based epidemiological study, termed as COPD Risk Cohort (Fig 1 and Table 1), were examined at the Department of Respiratory Medicine, Shanghai Tenth People’s Hospital, Tongji University, Shanghai, China. All subjects reside in Shanghai, China, and provided written informed consent. None of the subjects were diagnosed with COPD, but satisfying the criteria: history of cigarette smoking or a history of exposure to second-hand smoke; post-bronchodilator FEV₁/forced vital capacity (FVC) ratio < 70% and FEV₁% predicted value >80% in the absence of a bronchodilator or an inhaled corticosteroid treatment. The exclusion criteria included asthma, bronchiectasis, infectious disease, psychiatric disease, cognitive disorders, malignancies, severe kidney or liver dysfunction. An annual on-site follow-up were conducted on June 4th and 11th, 2017. We collected sociodemographic characteristics, such as age, gender, height, weight, smoking status, street address and comorbidities. All subjects completed the COPD Assessment Test (CAT), the modified British Medical Research Council (mMRC) questionnaire, and the Clinical COPD Questionnaire (CCQ); in addition, all subjects conducted the spirometry test and contributed sputum sample. Sputum is the coughed-up material (phlegm) from the lower airways (trachea and bronchi), and contain large number of white blood cells. The study was approved by the Institutional Review Boards at Shanghai Tenth People’s Hospital, Tongji University, Shanghai, China.

Figure 1. Study design and flow chart.

Table 1.

Demographical and respiratory function characteristics of two human cohorts.

Variable	COPD risk cohort (n = 74)	Normal healthy cohort (n = 41)
Agea	59 ± 8	26 ± 3g
Gender (Female)b	47 (63.5%)	19 (46.3%)
Height (m)	1.63 ± 0.08	1.69 ± 0.09 (1)
Weight (kg)	63.7 ± 8.3	60.7 ± 9.7 (1)
Smoke (Yes)	14 (18.9%) (7)	2 (4.9%)
FEV1 (L)	2.28 ± 0.40 (3)
FEV1% predicted	91.2 ± 12.3 (3)
FVC (L)	2.72 ± 0.50 (3)
FVC% predicted	88.2 ± 14.8 (3)
FEV1/FVC ratio (%)	84.1 ± 6.4 (3)
CATc	11.0 ± 3.4 (5)
mMRCd	0.3 ± 0.5 (6)
CCQe	3.0 ± 4.8 (5)
Airway bacteria loadf	119.5 ± 188.2

^a:Values shown for continuous variables are mean ± S.D. The number of missing data is shown in parentheses if there is any.

^b:Values shown for categorical variables are the number of samples along with percentage shown in the parentheses. The number of missing data is shown in the subsequent parentheses if there is any.

^c:CAT: COPD Assessment Test score.

^d:mMRC: modified Medical Research Council dyspnea scale.

^e:CCQ: Clinical COPD Questionnaire score.

^f:defined in Formula 1.

^g:Since the sputum sample for 16S sequencing in normal healthy cohort were collected at multiple time points during 2016–2017, the age for each subject is calculated as the mean value across time points.

Normal healthy cohort

During October 2016 – October 2017, 41 healthy volunteers were examined at the Department of Respiratory Medicine, Shanghai Tenth People’s Hospital, Tongji University, Shanghai, China (Fig 1 and Table 1). The inclusion criteria were: age ≥ 18 years, no use of antibiotics or immunosuppressive medications in the past two months, and no evidence of acute upper respiratory symptoms for the previous four weeks. The exclusion criteria were: history of asthma, other lung disease, heart failure, stroke, pregnancy or lactation, and infectious disease. Each subject was followed up every two months, contributed sputum sample and completed a questionnaire including anthropometric characteristics, education, street address, smoking status, medications used, and outdoor exercise time. The CAT, mMRC, and CCQ (see Abbreviations) are instrument for assessment of COPD status and severity, therefore, not applied to the health volunteers. All subjects provided written informed consent, and the study was approved by the Institutional Review Boards at Shanghai Tenth People’s Hospital, Tongji University, Shanghai, China.

Assessment of respiratory function

Spirometry was performed by clinicians according to standards established by the American Thoracic Society (ATS), using the Jaeger Spirometry Analyzer (Master Screen PFT, Hoechberg, Germany). The respiratory function parameters measured were: FVC and FEV₁ (see Abbreviations).

Sputum sample collection and DNA extraction

We collected sputum samples following previous described protocol (Brightling et al., 2001). Oral washes (OW) were performed with 10 ml sterile 0.9% saline before sputum induction. Sputum was induced as given inhaled sequences of 3%, 4%, and 5% saline for 7–15minutes via an ultrasonic nebulizer. The sputum induction procedure was terminated after the expectoration of sputum. Sputum is collected into sterile Petri dishes with phosphate-buffered saline, transferred to sterile collection tubes, and then stored at −80ºC for further analysis. DNA was extracted from sputum using PowerSoil® DNA Isolation Kit (Qiagen, Germantown, MD, USA) according to manufacture provided protocol. In each batch of sputum collection, we placed negative control, which was the sterile Petri dish with phosphate-buffered saline. The negative controls underwent identical procedure as patients’ sample, including storage, DNA extraction, and PCR.

Assessment of Bacteria abundance in sputum

Quantitative polymerase chain reaction (qPCR) using SYBR Green qPCR kit (MilliporeSigma, Burlington, MA, USA) was conducted to assess the abundance of bacteria in sputum samples. Quantifying the absolute count of bacteria is difficult and labor-intensive; instead, we assess the abundance of bacteria relative to the human cells. Two pairs of primers were added into the qPCR system, one pair specific to bacteria 16S rRNA gene, and the other pair specific to human β-actin gene. The ΔΔCt method (Livak and Schmittgen, 2001) was used to calculate the ratio of bacteria 16S rRNA genes relative to human β-actin gene in the DNA extracted from sputum.

Profiling airway microbiome by 16S rRNA sequencing

The 16S rRNA gene amplicon sequencing of the V4 hypervariable region was performed using Illumina HiSeq Fast Run mode (paired-end 250bp). For each PCR reaction, 4 ng of purified template DNA was amplified in a reaction volume of 20ul. Primers were designed to target the V4 region of the 16S rDNA (position 515 of the bacterial 16S rRNA gene to position 806). Each reaction was conducted with denaturation at 98°C for 30 sec, 26 cycles of (98°C × 10 sec, 50°C × 30 sec, 72°C × 30 sec), and a final extension at 72°C for 2 min. PCR amplicon were purified from agarose gel and sequenced on an Illumina HiSeq instrument at a loading concentration of 10pM.

To assess data reproducibility, we created duplicate pairs. In detail, aliquots of the same sputum sample were separated into different tubes, and performed DNA extraction, PCR and sequencing independently, hence, created technical replicates. Importantly, aliquots of each duplicate pair were sequenced in different batches in order to factor in the uncertainty or systematical difference sequencing conditions in our reproducibility assessment.

Data processing and quality control

The raw bacterial 16S rRNA sequencing data (i.e. fastq files) was processed within each batch (i.e. sequencing library). The pair-end reads (i.e., R1 and R2 reads) were merged. The merged reads was then processed for microbiome profiling by QIIME 1.9.1 (Hiergeist et al., 2016). For each sample, the reads were clustered into operational taxonomic units (OTUs). A training dataset for taxonomic assignments was created using a modified NCBI taxonomy from the ‘Isolated named strains 16S’ in the Greengenes database (http://greengenes.secondgenome.com/). This dataset was manually curated by (1) removing strains in ‘Isolated named strains 16S’ that had non-standard taxonomy or that were not members of the domain bacteria, and (2) grouping strain level taxonomy from Greengenes assignments under a single NCBI species assignment. The Ribosomal Database Project (RDP) version 2.4 classifier (http://gordonlab.wustl.edu/SuppData.html) was used to assign taxonomy and OTUs.

We performed quality control (QC) to filter out OTUs and samples of poor quality. At OTU level, we excluded the OTUs with extremely low abundance, i.e. the median proportion across all samples is < 0.001%. As result, 78 OTUs were kept for downstream analysis, including a category of unassigned OTU, which cannot be assigned to any known taxonomic unit. At sample level, we performed principle component analysis (PCA) on the 78 OTU profile data and did a visual check on the first two PCs (Fig S1). A total of 6 outliers (1 in COPD risk cohort and 5 in normal healthy cohort) were excluded. We excluded an additional sample of COPD risk cohort, because it contained much more low abundance OTUs (ie, OTU proportion < 0.001%) than other samples and appeared as an outlier. We ended up with 16S sequencing data for 72 samples in COPD risk cohort, for which only one sample was collected from each subject at one time point, and 242 samples in normal healthy cohort, for which multiple samples were collected for each subject at multiple time points.

To evaluate the quality of 16 sequencing data, a subset (n=40) of the samples in the normal healthy cohort were technically sequenced multiple times with exactly the same protocol. We calculated the Spearman correlation coefficients of the OTU abundance profile for 40 replicated pairs. The correlations were computed across all the 78 OTUs, common OTUs (median proportion > 0.1%; n = 26) and rare OTUs (median proportion < 0.1%; n = 52).

Estimating PM_2.5 exposure level

We downloaded PM_2.5 data collected at 10 MEP (China Ministry of Ecology and Environment) stations in Shanghai, China, in 2016 and 2017 (http://beijingair.sinaapp.com/), which matched to cohort study period. The latitude and longitude of each station is also obtained. We calculated the mean of hourly measured PM_2.5 concentrations as the daily measurement at each station. The street address of the participant was retrieved from the questionnaire and translated into latitude and longitude using google map. PM_2.5 concentration of the closest station to each individual was treated as his/her exposure level. We considered the average daily PM_2.5 concentrations of the 14 days prior to the sputum collection date (including the collection day) in this study.

Association between OTU abundance and PM_2.5 exposure

To explore whether PM_2.5 exposure influenced the relative abundance (i.e. proportion) of the OTUs, we perform an association study on the pooled data of both COPD risk and normal healthy cohorts with the R package variancePartition (Hoffman and Schadt, 2016):

$$OTU=E+X_R+X_F+\epsilon ,$$ (model A1)

where E is a vector of 14 variables corresponding to PM_2.5 exposure of 1~14 days prior to sputum collection date and E is treated as fixed effect; X_R and X_F are vectors of covariates: X_R is the vector of covariates coded as random effects, including cohort, batch, individual, sex, smoking status, collection date, and PM_2.5 station; X_F is the vector of covariates coded as fixed effects including age, height and weight; ε is the error term. The variances explained by different factors were visualized by violin plot (Fig S2). For each OTU, the p-value of the overall influence of the PM_2.5 exposure during the 14 days prior to sputum collection was derived from a likelihood ratio test of 14 degree of freedom (LRT_14df) using nested models, where model A1 was compare to the model A0:$OTU~X_R+X_F+\epsilon .$

Overall significance of PM_2.5 exposure’s effect on OTUs was evaluated using a one-side Kolmogorov-Smirnov (K-S) test, where the distribution of the p-values generated for term E in nest models was compared to a uniform distribution. Given strong correlations between the proportions of the OTUs, the p-values were not independent. An iterative pruning procedure was performed to produce a list of independent p-values in order to ensure a valid K-S test. We compute the Pearson correlation between all pair of OTUs adjusted for covariates (using model A0), and then started the pruning. First, we define an empty vector, C. Second, the OTUs were sorted based on their LRT_14df p-values (derived from nested models: model A1 vs model A0) ascendingly, and the list is termed as S. Third, the actually pruning was executed in following steps: 1) picked up the OTU with smallest p-value in S, and term it as OTU_C; 2) append OTU_C to C; 3) remove all the OTUs in S (if any) whose Pearson correlation coefficient with OTUc ≥ 0.2; repeat 1) to 3) if S is not empty. After the pruning, C contains independent OTUs, whose LRT_14df p-values subsequently entered the K-S test.

Association between respiratory function traits and OTU abundance

In COPD risk cohort, we considered the effect of 78 OTUs on respiratory function traits, including FEV₁, FEV₁% predicted, FVC, FVC% predicted, FEV₁/FVC ratio, CAT score, mMRC score and CCQ score. In a linear mixed model (LMM), each trait was regressed on each OTU proportion.

$$Trai{t}_i=OT{U}_j+X_R+X_F+\epsilon ,$$ (model B1)

where Trait_i is the i^th respiratory function trait, OTU_j is the j^th OTU’s proportion. The denotation of X_R, X_F and ε are the same as in model A1. The p-value for each OTU was derived from a likelihood ratio test of 1 degree of freedom (LRT_1df) using nested models, where model B1 was compared to model B0: Trait_i = X_R + X_F + ε. For each respiratory function trait, the overall significance of the effect of all the OTUs was also evaluated by a K-S test (after OTU pruning).

Results

Demographic characteristics and bacteria load in bronchial tree

A total of 74 individuals and 41 individuals from the COPD risk cohort and normal healthy cohort, respectively was included in this study (Table 1). The average age of COPD risk cohort was 59 years old, much older than normal healthy cohort. Because of such systematical difference, we did not directly compare the two cohorts in this study in terms of sputum microbiome profile.

The abundance of bacteria in sputum relative to human cells (termed as R) was quantified by qPCR

$$R=\frac{copiesof16SrRNAgene}{copiesofhuman\beta -actingene}$$ (Formula 1).

The average R value was 119.5 (Table 1), and R was significantly associated with FEV₁/FVC ratio (p = 0.0038, Table 2), which is an essential respiratory function trait for COPD diagnosis. Interestingly, the association is positive, indicating individuals with better lung function (i.e., higher FEV₁/FVC ratio) had higher bacteria load in bronchial tree.

Table 2.

Association of respiratory function traits with airway bacteria load

	Crude model			Adjusted modela
Traitb	Estimate	Std. Err.	P-value	Estimate	Std. Err.	P-value
FEV1 (L)	0.00031	0.00026	0.23	0.00025	0.00015	0.11
FEV1% predicted	0.0026	0.0079	0.74	0.0081	0.0065	0.22
FVC (L)	−0.00043	0.0018	0.81	−0.00083	0.0019	0.65
FVC% predicted	−0.0040	0.0095	0.68	−0.0011	0.0089	0.91
FEV1/FVC ratio (%)	0.0076	0.004	0.061	0.010	0.0035	0.0038
CAT	−0.00071	0.0021	0.74	−0.00065	0.0019	0.74
mMRC	−0.00025	0.00034	0.46	−0.00015	0.00032	0.64
CCQ	−0.0028	0.003	0.37	−0.0032	0.0028	0.26

^aThe association test was performed in the COPD risk cohort (n = 74). In the linear mixed model, age, gender, height and weight were adjusted as fixed effect while collection date was adjusted as random effect.

^bFEV₁, forced expiratory volume in 1 second; FVC, forced vital capacity; CAT, COPD Assessment Test; mMRC, modified British Medical Research Council; CCQ, Clinical COPD Questionnaire.

Sputum microbiome profile

16S rRNA-Seq data for 72 sputum samples from COPD risk cohort and 242 sputum samples from normal healthy cohort passed QC filters, and 78 OTUs were identified (Fig 2A and Table S1). Majority of the OTUs was of very small proportion (i.e., ≤1%) in airway microbiome, and the most abundant genera were Veillonella, Prevotella and Neisseria. Substantial inter-individual variation was observed in airway microbiome (Fig 2A). The OTU proportion of duplicate pairs were highly correlated, where median Spearman r²= 0.9295 (Fig 2B). On relative common OTUs (ie, median proportion ≥ 0.1%), the duplicate pairs were correlated with median Spearman r²= 0.9750 (Fig 2B).

Figure 2. Airway microbiome profile.

(A). OTU proportions in each sample. Each stacked color bar indicates the percentages of OTUs in each sample. Sample IDs from COPD risk cohort and normal healthy cohort are indicated by orange and blue labels beneath the color bars. The OTUs of low abundance (median percentage < 1%) are collapsed into one category “other.otus”. More detailed summary statistics of all the OTUs for the two cohorts are reported in Table S1. (B). Reproducibility of OTU abundance in technical replicates from normal healthy cohort. The reproducibility was measured as Spearman correlation coefficient between technical replicates across all OTUs (n = 78), common OTUs (median percentage > 0.1%; n = 26) and rare OTUs (median percentage < 0.1%; n = 52).

PM_2.5 exposure level

For each sputum sample, we quantified the subject’s exposure level for 14 days prior to the follow-up date. The PM_2.5 level in Shanghai varied greatly during the one-year study period (Fig 3A), where the average daily PM concentration was high in winter than in summer. The follow-up dates of normal healthy cohort (blue asterisks, Fig 3A) were well-spread, and both good and poor air quality periods were sampled. The follow-up of COPD risk cohort were conducted on two days (orange asterisks, Fig 3A). The PM_2.5 concentrations measured by MEP station in Shanghai were highly correlated (Fig 3B and 3C), for example, the correlation of Hong-Kou and Yang-Pu stations was as high as 97%. In consequence, the COPD risk cohort subjects were estimated to have very similar exposure level with very limited inter-individual variation, further the sample size of the COPD risk cohort was relatively small, hence, we did not examine the association between PM_2.5 exposure and respiratory functions.

Figure 3. PM_2.5 measured at 5 surveillance stations in Shanghai.

(A) Average daily PM_2.5 concentration at 5 surveillance stations in Shanghai, China, from October 1, 2016 to September 30, 2017. The dates on which sputum samples were collected for COPD risk cohort and normal health cohort are marked by orange and blue stars respectively. (B) The name, latitude and longitude of each station. (C) The Pearson correlation coefficients of PM 2.5 measurements between pairs of the surveillance stations across the time period shown in (A) are visualized by heatmap. (C) The name, latitude and longitude of each station.

Airway OTU profile was significantly associated with PM_2.5 exposure

Using variancePartition methods (Hoffman and Schadt, 2016), we evaluated the influence of PM_2.5 exposure (on each date up to 14 days prior to sputum collection) and covariates (Fig S2). The most influential factor was sequencing batch followed by exposure occurred 7 and 6 days prior to sputum collection. Interestingly, age, height, weight, sex, sputum collection date and smoking status had less influence (Fig S2). Focusing on each of the 14 days’ exposure level (Fig 4A), exposure occurred ≥ 6 days prior to collection had much stronger impact on airway microbiome that exposure occurred day 1 to day 3 prior to sputum collection. Further, the effect of PM_2.5 was long lasting. The exposure occurred 14 days before sputum collection still had substantial effect on airway microbiome, even controlled the exposure of day 1 to day 13 in the multi-variate model (model A1). Overall, airway OTUs were significantly associated with PM_2.5 exposure (K-S test p=4.48e-11, Fig 4B and Table S2), and the affected OTUs were not the most abundant one (Fig S3).

Figure 4. Association of OTU proportions with PM_2.5 exposure.

(A) The variance of each OTU proportion (n = 78) explained by the PM_2.5 exposure of the 14 days prior to sputum collection day (inclusive) with a linear mixed model, which was fitted based on the combined data from both COPD risk cohort and normal health cohort (n = 314 samples in total after QC). The results are visualized by violin plot. A violin plot is a method of visualizing the distribution of numeric data, similar to box plot with a rotated kernel density plot on each side. Inside the violin plot is the regular box plot with outliers indicated by dots. (B) Q-Q plot of the p-values (on -log10 scale) in the association test between the OTU proportion and PM_2.5 exposure. The overall significance of the enrichment of small p-values was evaluated by a one-side Kolmogorov–Smirnov (K-S) test with respect to a uniform distribution on [0, 1] as null. The blue line has a slop of 1. The dashed line indicates the significance level after Bonferroni correction (0.05/78 = 6.41×10⁻⁴). The OTUs that passed the significance level are highlighted in red.

Respiratory function traits were associated with airway microbiome

We surveyed the association between each of the respiratory function traits and overall airway microbiome using K-S test. Firstly, we focused on key characters in COPD diagnosis: FEV₁% predicted and FEV₁/FVC ratio. FEV₁% predicted was significantly associated with airway microbiome (p = 0.013, Fig 5A and Table S3). In contrast, FEV₁/FVC ratio was not associated with OTU proportion (Fig 5B), although it was significantly associated with airway bacteria load (Table 2). Secondly, several additional respiratory function measures (Fig S4) were in significant association with airway microbiome, including FEV₁ (p = 0.017) and CAT (p = 0.0042).

Figure 5. Q-Q plot of the p-values in the association test between respiratory function traits and OTUs.

The x- and y-axis of the plot present the p-values of the association test between respiratory function trait and OTUs. The x-axis is the expected p-value under the null hypothesis that there is no association, and the y-axis is the p-value observed on the actual data. All p-values are –log₁₀ transformed. (A) FEV₁% predicted; (B) FEV₁/FVC ratio (%). In the linear mixed model, age, gender, height, weight, and smoke were adjusted as fixed effect while 16S sequencing batch and sputum collection date were adjusted as random effect. The overall significance of the enrichment of small p-values was evaluated by a one-side Kolmogorov–Smirnov (K-S) test with respect to a uniform distribution on [0, 1] as null. The blue line has a slop of 1. The dashed line indicates the significance level after Bonferroni correction (0.05/78 = 6.41×10⁻⁴). The OTUs that passed the significance level are highlighted in red.

Discussions

In the past two decades, multiple epidemiology studies have linked PM_2.5 exposure and declines in respiratory functions (Xing et al., 2016). Microbiome is a constitutive component of human’s airway micro-environment and closely interacts with host immune systems. To our knowledge, we are the first to demonstrate the impact of PM_2.5 on airway microbiome, and reported the link between airway microbiome and respiratory functions. Key findings of this study: (1) FEV₁/FVC, an important respiratory function trait and a diagnosis criterion of COPD, is positively associated with sputum bacteria load (p = 0.0038). (2) Multiple respiratory function traits are associated with airway microbiome profile, for example, FEV₁% predicted (Fig 5A, p = 0.013). (3) Airway microbiome was profoundly influenced by PM_2.5 exposure (Fig 4B, p = 4.48e-11). Importantly, the strongest impact attributed to exposure occurred ≥ 6 days ago (Fig 4A and Fig S2). Further, PM_2.5 exposure has long-lasting effect, and exposure occurred 14 days ago was still associated with the airway microbiome (Fig 4A and Fig S2).

We compared published respiratory tract microbiome to that found in our sample. On 64 individuals, Morris et al used 16S rDNA sequencing technology to profile bacteria from BAL which mainly represents lower lung (Morris et al., 2013). Eight genera (Streptococcus, Veillonella, Prevotella, Fusobacterium, Neisseria, Actinomyces, Porphyromonas, Gemella) were universally found in BAL sample (Dickson et al., 2015), and importantly, all these genera were detected in our data. Dickson et al studied BAL from 15 healthy individuals and detected 12 unique bacteria family, among which 9 were found in our sample. In this study, the 10 most abundant genera are Veillonella, Prevotella, Neisseria, Streptococcus, Haemophilus, Porphyromonas, Fusobacterium, Leptotrichia, Rothia, Actinomyces (Table S1), and all these 10 genera were reported by at one of the BAL microbiome studies (Dickson et al., 2015; Morris et al., 2013).

There are varieties of bacteria on ambient particulate matters (Cao et al., 2014). But the alternation of sputum microbiome we detected is unlikely be simply addition of PM-attached bacteria onto the indigenous airway microbiota. In fact, the PM-attached bacteria are distinct from those of any body site, as detailed by HMP (Blaser, 2010). The PM-attached bacteria are primarily from soil, followed by fecal, freshwater and marine sources (Cao et al., 2014), and not directly related to bacteria that habitat in human airway. PM in Chinese cities contains sulfate, ammonium, nitrate, and organic matter, and altogether they represent 80% of PM_2.5 (w/w) (Cao et al., 2014). A plausible mechanism that PM_2.5 exposure perturbs airway microbiome could be that the component of PM_2.5 (organic- and inorganic matters and bacteria) influence the micro-environment inside the bronchial tree, and in turn affect the microbiome.

A major confounder in human microbiome study is medication (especially antibiotic use). Microbiome study on COPD patients is particularly difficult, given majority of COPD patients receive medication (including antibiotics) to maintain and manage symptoms, and such medication profoundly impacts the microbiome ecosystems in the airway. Unfortunately, this confounder is also hard to control for, since the healthy controls have distinct medication profile comparing to patients. In other words, the confounder perfectly correlates with endpoint (case/control status), making it impossible to distinguish between the effect of disease and that of the confounder. Herein, we only studied non-COPD individuals who do not receive medication in a regular basis.

The study protocol showed high reproducibility (median r²=92.95%, Fig 2B). It should be noted, the association between airway microbiome and respiratory functions does not imply causality. Nevertheless, the microbiome at least can serve as biomarkers reflecting the health status of airway ecosystem. Most importantly, we showed the ecosystem can be perturbed by PM_2.5 exposure. Since microbiome is in close inter-talk with airway immune system, the perturbation induced by PM_2.5 exposure could be an etiological pathway towards respiratory function decline. Airway microbiome can be assess from sputum or BAL (Wang et al., 2017b). Sputum is the coughed-up material (phlegm) from the lower airways (trachea and bronchi), and may also contain materials from upper airways. The pattern of airway microbiome is “transient but not resident (TBNR)” (Saeedi et al., 2015), reflecting its complex and dynamic nature. The migration of microbes along the airway is bidirectional given there are no physical barriers. 16S rRNA profiling on sputum describes the bacteria ecosystem of the entire airway. In contract, BAL requires much more invasive procedure to collect (comparing to sputum) and has low bacteria content (Charlson et al., 2012; Erb-Downward et al., 2011), hence, is less suitable for airway microbiome study of moderate-to-large sample size.

The current study reported the short-term effect of PM_2.5 exposure. With the cohort established and still on-going, we are in a position to examine the long term impact of PM_2.5 on airway microbiome in the future. Besides PM, a variety of pollutants (e.g, O₃) exists in ambient air, but less studied in terms of their adverse effects on respiratory functions and influence on airway microbiome.

In summary, to our knowledge, this paper for the first time reported the impact of PM_2.5 exposure on airway microbiome, as well as the link between airway microbiome and respiratory functions. Given microbiota colonizes and inter-talks with the host cells in the bronchial tree, it may modulate to epithelial integrity and immune-regulation. The results advanced the understanding on the impact of PM_2.5 exposure on airway ecosystem, and shed light on possible biomarkers for pollution-induced airway injury.

Highlights.

• On human cohorts, we demonstrate the impact of PM_2.5 exposure on respiratory tract microbiome.

• The strongest impact is attributed to exposure occurred ≥ 6 days prior to microbiome collection.

• This study reports the link between airway microbiome and human respiratory functions.

Abbreviations

COPD

chronic obstructive pulmonary disease

ROS

reactive oxygen species

PAH

polycyclic aromatic hydrocarbons

TF

transcription factor

HMP

human microbiome project

BAL

bronchoalveolar lavage

FVC

forced vital capacity

FEV₁

forced expiratory volume in 1st second

OUT

operational taxonomic units

CAT

COPD assessment test

mMRC

British Medical Research Council questionnaire

CCQ

clinical COPD questionnaire

PCR

polymerase chain reaction