Inflammatory Endotype–associated Airway Microbiome in Chronic Obstructive Pulmonary Disease Clinical Stability and Exacerbations: A Multicohort Longitudinal Analysis

1Chronic obstructive pulmonary disease (COPD) is heterogeneous. Increasing evidence shows that the airway microbiome is related to COPD clinical phenotypes, severity, and long-term mortality. Understanding the role of the airway microbiome in COPD neutrophilic and eosinophilic inflammatory endotypes may help develop microbiome-based approaches for patient selection for targeted therapeutic intervention. There is a paucity of data examining the dynamic relationships between the airway microbiome and COPD inflammatory endotypes across stability and exacerbations.

2This study reports an integrated analysis on the airway microbiome in COPD neutrophilic and eosinophilic endotypes using 1,706 sputum samples collected longitudinally from 510 participants with COPD in three UK cohorts in 2008–2015. We showed that two primary types of airway ecology existed in neutrophilic COPD, which differed by the predominance of Haemophilus, inflammatory mediators, temporal stability, and interchangeability with eosinophilic inflammation and could therefore justify different therapeutic approaches. There were specific, nondominant microbiome genera associated with eosinophilia. Monitoring temporal variability of these features tracked patient inflammatory status over time, suggesting a potential need for point-of-care diagnosis using sputum microbiome biomarkers. These results highlight the importance of the airway microbiome in the inflammatory endotype–based patient management in COPD.

3Chronic obstructive pulmonary disease (COPD) is a heterogeneous disease underpinned by diverse clinical characteristics and pathophysiological mechanisms, with episodes of exacerbations leading to significant mortality worldwide. A better understanding of clinical phenotypes and biological endotypes for COPD is crucial for developing precision-medicine strategies that enable patient-tailored treatment according to clinical characteristics coupled with biomarkers of underlying disease mechanisms.

4Inflammatory patterns observed in individuals with COPD have been referred to as inflammatory endotypes. Neutrophilic inflammation is a hallmark of COPD and contributes to key pathological features, including emphysema and mucociliary dysfunction. Bacterial infection is associated with neutrophilic inflammation, but its role in driving inflammation is uncertain, given that increased neutrophilia is observed in both colonized and noncolonized patients with COPD. Eosinophilic inflammation is also present as a stable endotype in a subgroup of patients with COPD and is associated with less bacterial infection and a favorable response to inhaled corticosteroids (ICSs). Individuals with COPD have been broadly classified into neutrophilic, eosinophilic, mixed, and paucigranulocytic inflammation according to differential sputum cell counts, yet it is currently unknown whether such a definition sufficiently captures the underlying disease heterogeneity.

5Recent studies have demonstrated a diverse airway microbiome associated with COPD severity, exacerbations, clinical phenotypes, and long-term mortality. The airway microbiota differs between bacteria-associated and eosinophilic exacerbations, with a lower diversity and increased Proteobacteria in the former, suggesting airway ecology may be correlated with their underlying inflammatory processes. Although the airway microbiome was recently shown to differ between neutrophilic and eosinophilic inflammation in stable asthma, the dynamic relationship between the microbiome and inflammatory endotypes across stability and exacerbations in COPD remains unclear and warrants investigation in large, longitudinal cohorts.

6We hypothesized that the airway microbiome is differentially associated in neutrophilic and eosinophilic inflammations in COPD and that such associations may be dynamic across stability and exacerbations. We also hypothesized that different airway ecology reflects distinct pathophysiology and may capture additional heterogeneity within the broadly defined inflammatory endotypes and assist in patient stratification. We tested these hypotheses using an integrated analysis on three large-scale longitudinal cohorts established at four clinical sites in the United Kingdom in 2008–2015: BEAT-COPD (Biomarkers to Target Antibiotic and Systemic COPD), COPDMAP (COPD Medical Research Council/Association of the British Pharmaceutical Industry), and AERIS (Acute Exacerbation and Respiratory Infections in COPD). The participants in these cohorts were well characterized and followed at stability and during exacerbations for up to 2 years. A total of 1,706 sputum samples from 510 patients were included in this analysis, which, to our knowledge, represents the largest COPD microbiome analysis to date. Importantly, all samples in COPDMAP and AERIS were processed using the same procedure and platform in the same genomic facility, making them essentially one centralized data resource for collective analysis. In this study, we analyzed the combined COPDMAP and AERIS cohorts as the main discovery data set. Whenever applicable, the results were independently validated in BEAT-COPD, in which the microbiota was characterized using a different platform. The aim of this analysis was to systematically assess the relationship between the airway microbiome and COPD inflammatory endotypes across stability and exacerbations.

7The procedure for patient recruitment was described in detail previously. Participants were recruited if they had a physician diagnosis of COPD with a post-bronchodilator FEV1/FVC ratio <70% at screening and no previous asthma diagnosis. BEAT-COPD participants were recruited from the Glenfield Hospital in Leicester, United Kingdom. COPDMAP participants were recruited at Imperial College London, the University of Leicester, and the University Hospital of South Manchester, United Kingdom. AERIS participants were recruited at the University Hospital Southampton, United Kingdom. The stable visits were defined as visits during stable disease, with at least 4 weeks free from a prior exacerbation. Exacerbations were defined according to Anthonisen criteria and/or healthcare use. Exacerbation samples were collected before treatment with antibiotics or steroids.

8Sputum samples were classified on the basis of established criteria for sputum differential cell counts into four groups: neutrophilic (eosinophils < 3%, neutrophils ≥ 61%), eosinophilic (eosinophils ≥ 3%, neutrophils < 61%), mixed-granulocytic (eosinophils ≥ 3%, neutrophils ≥ 61%), and paucigranulocytic (eosinophils < 3%, neutrophils < 61%). A panel of inflammatory mediators was measured in the sputum and serum for a subgroup of patients in COPDMAP (N = 157) and BEAT-COPD (N = 113) using the Meso Scale Discovery platform (Meso Scale Diagnostics). The measurements were quality controlled as described previously.

9All COPDMAP and AERIS samples were processed in a single, centralized laboratory at GlaxoSmithKline Research and Development) according to the same protocol as described previously. The V4 hypervariable region of the 16S rRNA gene was sequenced using Illumina MiSeq with proper reagent controls (see online supplement). For BEAT-COPD, the 16S V3–V5 region was sequenced using the 454 Genome Sequencer (454 Life Sciences). All sequencing data were deposited in the National Center for Biotechnology Information Sequence Read Archive (BEAT-COPD: SRP065072, COPDMAP: SRP102480, AERIS: SRP102629). The computer codes for data analyses were provided in the online supplement or deposited in GitHub under https://github.com/wangzlab/AERIS_MAP_BEAT_analysis.

10All 16S rRNA gene data sets were processed using a standardized pipeline in QIIME 2.0 (Quantitative Insights Into Microbial Ecology 2.0). The demultiplexed sequencing reads were denoised to generate amplicon sequence variants using DADA2 (Divisive Amplicon Denoising Algorithm 2). Additional denoising parameters were used for 454 data. A custom Naive Bayes classifier was trained on Greengenes Database (Second Genome, Inc.) 13_8 99% operational taxonomic units to assign a taxonomy for each data set. The COPDMAP and AERIS samples were rarefied to 29,117 reads. The BEAT-COPD samples were rarefied to 2,207 reads. The sequencing batch information for COPDMAP and AERIS were used to adjust batch effects for microbiome data using Combat, according to the method Gibbons and colleagues.

11The detailed procedure for statistical analyses is provided in the online supplement. Microbiome community types were identified using unbiased clustering by Wald linkage on the basis of the Jensen-Shannon divergence index. The optimal number of clusters was determined by using the Silhouette measure of the degree of confidence. Cluster memberships were validated using a partition around the medoids, with the optimal number of clusters determined by using the Calinski-Harabasz index. Changepoint detection analysis was performed using the pruned exact linear time algorithm (33) in the changepoint package in R (R Foundation for Statistical Computing) to search for temporal change points on the relative abundance of microbiome genera. We calculated cross-covariance scores between the relative abundances of microbiome genera and sputum neutrophilic and eosinophilic percentages for all longitudinal visits of each patient using the ccf function in R. Patients with at least five visits were included. A microbiome cooccurrence network was established using SparCC (Sparse Correlations for Compositional data) and visualized using Gephi. The correlation between microbiome genera and mediators was assessed first by residualization using a general linear mixed model to adjust for demographic covariates and then by hierarchical all-against-all association testing using HAllA (Hierarchical All-against-All association testing).

12A total of 1,366 sputum samples were collected from 423 patients with COPD in the COPDMAP and AERIS cohorts, spanning between 1 and 13 visits over up to 2 years (time-span range, 4–658 d; mean, 301.5 d) at clinical stability (N = 920) and during exacerbations (N = 446) in London (N = 300), Leicester (N = 303), Manchester (N = 180), and Southampton (N = 583) in the United Kingdom (see Figure E1 in the online supplement and Table 1). As a validation data set, 340 samples from the BEAT-COPD cohort from 87 participants at a stable state (N = 203) and during exacerbations (N = 137) were included (1–9 visits per participant, time-span range, 5–881 d; mean, 238.4 d; Figure E1 and Table 1).

13Chronic use of macrolides was reported for antiinflammatory purposes for 16 stable visits from 13 participants in COPDMAP and AERIS. No macrolide use was reported for BEAT-COPD participants. No significant associations were found between the microbiota and chronic macrolide usage (Figure E2). No significant differences in the microbiota were observed between induced and spontaneous sputum samples at a stable state (Figure E3).

14We classified 1,366 samples in the discovery data set into neutrophilic (N = 551), eosinophilic (N = 189), mixed (N = 187), and paucigranulocytic (N = 302) subgroups. Of these, 137 samples could not be assigned to a subgroup because of missing data. The 340 samples in BEAT-COPD were also classified accordingly. The distribution of sputum differential cell counts was generally comparable across cohorts and sites (Figure E4), suggesting a degree of consistency between the independent assessments. The distribution of each subgroup was comparable across cohorts and sites, except for a relatively higher proportion of paucigranulocytic samples in AERIS (Figures 1A and E5A).

15Conducting α-diversity analysis on the discovery data set revealed an overall significantly decreased Shannon index in the neutrophilic group compared with the other three groups (Wilcoxon P ≤ 8.4 × 10−3, Figure 1B). However, there was a broad range of Shannon index values in the neutrophilic group, with a subset of samples without a notable reduction in α diversity. This was also manifested by a significantly greater deviation from the centroid in the neutrophilic group compared with the other groups in principal coordinate analysis of microbiome β-diversity (Bray-Curtis dissimilarity, Figures 1C and E5B). A similar pattern was observed in BEAT-COPD (Figures E6A and E6B). A similar pattern was observed when analyzing stable and exacerbation samples separately (Figure E7). These results suggested that there were heterogeneous patterns of the microbiome within the neutrophilic group.

16To dissect the heterogeneity of the microbiome in the neutrophilic group, we performed an unsupervised clustering of microbiome profiles into community types. Hierarchical clustering on all 1,366 microbiome profiles revealed an optimum of four community classes. The main community type was composed of a “balanced” microbial composition, whereas the other three types had a “biased” composition dominated by Haemophilus, Moraxella, and Streptococcus, respectively (Figure E8A). These three community types were overall in agreement with H. influenzae, M. catarrhalis, and S. pneumoniae colonization on the basis of quantitative PCR and culture results, except for a relatively low culture positivity for S. pneumoniae (Figure E9). Similar community clusters were also identified using the partition-around-the-medoids clustering method and in BEAT-COPD (Figures E8A–E8C), suggesting robustness of the community types. The four community types persisted when analyzing stable or exacerbation samples alone (Figure E10A).

17The relative abundance of Haemophilus and the quantitative-PCR load of H. influenzae had areas under the curve (AUCs) of 0.949 and 0.908 in distinguishing the Haemophilus-predominant subgroup from the remaining samples in the discovery data set (AUCs of 0.999 and 0.912 in the validation data set). The Haemophilus relative abundance of 0.41 and log10 H. influenzae load of 7.2 (copies/ml) reached the best discriminatory power for both data sets. In comparison, Veillonella and Prevotella were most significantly enriched in the balanced microbiome subgroup in both data sets. An added-up relative abundance of Veillonella and Prevotella of 0.26 reached the highest power in segregating the balanced microbiome subgroup in the discovery data set. In the validation data set, the optimal Veillonella+Prevotella relative abundance for the subgroup was 0.17.

18Next, we compared microbiome communities across inflammatory endotypes (Figures 2A–2C). The majority of samples in the eosinophilic, mixed, and paucigranulocytic groups had a balanced community type (77.1–83.5%, Figure 2A). In comparison, 56.8% of samples in the neutrophilic group had a balanced community type, whereas 27.8% of the samples were Haemophilus predominant (Figure 2A). The Moraxella- and Streptococcus-predominant communities constituted 13.8% of all samples. The distribution of community classes across inflammatory endotypes was overall consistent between stable and exacerbation samples, with a relatively lower representation of the “balanced” community for exacerbations (Figure E10B). We therefore concluded that there were two major types of airway ecology for neutrophilic COPD, which were differentiated by the predominance of Haemophilus. They were considered as two neutrophilic subgroups in subsequent analyses (named as the neutrophilic Haemophilus and neutrophilic balanced subgroups).

19The two neutrophilic subgroups were overall comparable in demographic features and severity in terms of spirometric findings and symptom scores (Tables 2 and E1). There was a significantly lower neutrophilic percentage and a higher eosinophilic percentage in the neutrophilic balanced subgroup compared with the neutrophilic Haemophilus subgroup (Figure 2B; Wilcoxon P < 0.001). However, the neutrophilic percentage alone cannot differentiate between the two subgroups (AUC, 0.620). In COPDMAP, 37 sputum and serum inflammatory mediators were measured for 157 samples from 80 participants. Principal component analysis on the sputum mediators revealed distinct clusters for the neutrophilic balanced, neutrophilic Haemophilus, and eosinophilic subgroups (Figure 2C). In the eosinophilic subgroup, eotaxin-3, TARC (thymus- and activation-regulated chemokine), and IL-5 were elevated, indicating a T-helper cell type 2 (Th2) signature. In the neutrophilic Haemophilus subgroup, IL-1α, IL-1β, and TNFα were elevated (false discovery rate [FDR] P < 0.001). In the neutrophilic balanced subgroup, IL-17A was most elevated (FDR P = 0.04), followed by SAA (serum amyloid A), Flt-1 (fms-related receptor tyrosine kinase 1), and IL-16 (Figures 2D and E11 and Table 2). A similar clustering pattern was also observed in BEAT-COPD, with elevated IL-17A in the neutrophilic balanced subgroup (Figure E12). A similar mediator clustering pattern was further observed for stable or exacerbation samples alone (Figure E13). Among serum mediators, elevated IL-17A and GM-CSF (granulocyte-macrophage colony-stimulating factor) were observed in the neutrophilic balanced subgroup, and IL-5 and eotaxin-3 were higher in the eosinophilic subgroup, both in COPDMAP and BEAT-COPD (Figure E14 and Table E2).

20We assessed within-patient stability for the neutrophilic and eosinophilic endotypes over time by analyzing paired samples collected sequentially from the same patients. Samples were excluded if they did not have a paired sample or if their inflammatory states were not putatively neutrophilic or eosinophilic (i.e., mixed or paucigranulocytic). Within stable disease, 71.4% of neutrophilic Haemophilus states were followed by the same state in the next visit, suggesting relative stability (Figure 3). In comparison, 38.4% and 41.8% of neutrophilic balanced and eosinophilic states were succeeded by the same state during stability. In addition, 59.1%, 21.4%, and 33.3% of neutrophilic Haemophilus, neutrophilic balanced, and eosinophilic states at stability were maintained during exacerbations. For 47.6% of neutrophilic COPD cases with a balanced microbiome, their communities switched to those dominated by Haemophilus, Moraxella, or Streptococcus during exacerbations. Consistent with this finding, the stable samples in the neutrophilic balanced group had the greatest microbiome shifts during exacerbations compared with other groups, as measured by Bray-Curtis dissimilarity for paired stability–exacerbation samples, both in the discovery data set and in the validation data set (Figures 2E and E12E). These results suggested that the neutrophilic balanced subgroup was dynamic, and patients in this state at stability were most susceptible to airway microbiome shifts during exacerbations. We noted that the neutrophilic Haemophilus and eosinophilic states were rarely transited to each other at stability or during exacerbations, indicating their mutually exclusive nature (Figure 3), whereas the neutrophilic balanced state was interchangeable to both neutrophilic Haemophilus and eosinophilic states. Similar transition pattern was observed in BEAT-COPD (Figure E15). Seasonal changes and baseline ICS usage had nonsignificant positive associations with intrapatient switches between neutrophilic and eosinophilic states (stable–stable odds ratios [ORs], 1.60 and 2.22; 95% confidence intervals [95% CIs], 0.63–4.09 and 0.47–10.36; P = 0.32 and 0.29; stable-exacerbation ORs, 1.55 and 1.30; 95% CIs, 0.48–5.07 and 0.24–6.78; P = 0.46 and 0.78).

21Thirty-one genera had a significantly higher relative abundance in the eosinophilic group versus the neutrophilic group (linear discriminant analysis [LDA] effect size, LDA > 4.0; FDR P < 0.05; Table E2). Twelve of the 31 genera remained significant between the neutrophilic balanced and eosinophilic groups (LDA > 4.0; FDR P < 0.05; Figures 4A–4C and Table E3), indicating that their enrichments in eosinophilia over neutrophilia were independent of Haemophilus predominance in the latter. All 12 genera remained eosinophilically enriched using two additional data-normalization approaches to control for Haemophilus overgrowth (online supplement; Figure 4C). Furthermore, all 12 genera except for TG5 and Aggregatibacter remained eosinophilically enriched when analyzed cross-sectionally within each sequential stable or exacerbation visit. All 12 genera except for S24-7 remained eosinophilically enriched within at least three of four sites. Leicester and London shared a more similar microbiota–eosinophilia correlation pattern than they did with the other sites, which may possibly be related to their relatively smaller proportion of samples in the eosinophilic subgroup (Table 2). We noted that Mogibacterium, TG5, and S24-7 had higher abundances in patients with baseline ICS usage; however, these changes were nonsignificant (Figure E16). Gemellaceae, Granulicatella, Campylobacter, Porphyromonas, Capnocytophaga, and Rothia further exhibited significant positive correlations with the sputum eosinophilic percentage in a multivariate linear mixed model for all 1,229 discovery samples (P < 0.05), which together explained 35.6% of the variation in sputum eosinophilia (R2 = 0.356). Patient smoking history was a significant covariate (P = 0.018, Table E4). These results suggested that there were nondominant taxa specifically associated with sputum eosinophilia, which were generally robust across sites, visits, and demographics. The enrichments of Granulicatella, Campylobacter, Gemellaceae, and Capnocytophaga in the eosinophilic group were validated in BEAT-COPD (Figure E17).

22Knowing that neutrophilic and eosinophilic inflammations were interchangeable within patients, we next assessed whether the airway microbiome altered alongside such changes in the patient inflammatory status. We performed a changepoint detection analysis (Figure 4D). As a time-series statistical approach, this analysis divides longitudinal data into segments and identifies points when the distribution of features (i.e., microbiome genera) changes significantly between segments. Ninety-one participants with at least five longitudinal visits were included. Across stability and exacerbations, temporal trajectories of Moraxella and Actinobacillus were associated with the occurrence of exacerbation events (ORs, 3.58 and 2.42; 95% CIs, 2.54–5.29 and 1.43–4.10; P < 0.05, Figure 4E). No genera were associated with switches between neutrophilic and eosinophilic states when viewed across stability and exacerbations, indicating exacerbation had a major impact on temporal variability of the microbiota. When assessing microbiome change points within stability, Haemophilus was associated with switches from the eosinophilic to the neutrophilic state (OR, 2.66; 95% CI, 1.41–5.66; P = 0.01), whereas Campylobacter, Gemellaceae, Capnocytophaga, and Granulicatella were associated with neutrophilic to eosinophilic switches (OR ≥ 2.27, P < 0.05, Figure 4E).

23To further assess whether the microbiome changed in concert with the extent of neutrophilic and eosinophilic inflammation, we calculated the cross-covariance between temporal measures of the microbiome genera and sputum neutrophilic and eosinophilic percentages for each patient. Given two temporal measures, the cross-covariance estimates the covariation of one measure against the other at pairs of time points (Figure 4D). Granulicatella and Campylobacter had the highest cumulative cross-covariance with eosinophilia across all participants, which was significantly higher against the null distribution generated by the permutation test (Wilcoxon P < 0.05, Figure 4F and Table E5). This was consistent with the markedly lower interpatient variations of their cross-covariance scores and suggested between-patient consistency in their temporal covariation with eosinophilia (Figure 4F). Moraxella and Haemophilus exhibited the highest cumulative cross-covariance with neutrophilia. However, neither of them reached statistical significance (Table E5).

24To assess the microbiome community structure across inflammatory subtypes, we performed a cooccurrence network analysis. In the neutrophilic Haemophilus subgroup, Haemophilus was the predominant hub node exhibiting coexclusive relationships against 35 genera (Figure 5A). In the neutrophilic balanced subgroup, Veillonella had the highest degree of connectivity, followed by Serratia, Acinetobacter, and Mycoplasma (Figure 5A). The network in the eosinophilic group was featured by the mutual cooccurrence relationships among Fusobacterium, Granulicatella, Capnocytophaga, and Campylobacter that were specifically enriched in this group (Figure 5A).

25We further performed an all-against-all correlation analysis between microbiome genera and inflammatory mediators for each subgroup, adjusting for patient demographic cofactors. In the neutrophilic Haemophilus subgroup, Haemophilus and Moraxella had the highest number of correlations with mediators, in particular IL-1β and TNFα (Figure 5B). In the neutrophilic balanced subgroup, there were multiple, moderate, negative correlations of Serratia, Oribacterium, and Lactobacillus with IL-17A, IL-1α, and IL-1β (Figure 5B). In the eosinophilic subgroup, positive correlations were identified between Lautropia, Campylobacter, and Granulicatella and eotaxin-3, IL-5, and TARC (Figure 5B). Analysis of BEAT-COPD revealed overall similar patterns (Figure E18); 83.2% of microbiome–microbiome correlations and 77.6% of microbiome–mediator correlations that were significant in the discovery data set were validated in BEAT-COPD (Tables E6 and E7).

26Here, we have shown that the airway microbiome is differentially associated with neutrophilic and eosinophilic COPD. It is well established that neutrophilic inflammation is associated with reduced microbial diversity and increased Proteobacteria in asthma and COPD. Taylor and colleagues showed that neutrophilic asthma had greater variability in the airway microbiota than eosinophilic asthma, which is consistent with our observation. However, the heterogeneity in neutrophilic airway ecology has not been examined in detail, given their relatively small sample size. Using unbiased clustering on over a thousand microbiome samples, we showed that neutrophilic COPD consisted of two main subtypes of airway ecology differentiated by the predominance of Haemophilus. Segregating these two subgroups could be important clinically, as they were associated with distinct inflammatory profiles and may justify different therapeutic approaches.

27The Haemophilus-predominant subgroup had decreased α diversity and enhanced proinflammatory mediators, IL-1β and TNFα, consistent with previous observations in both COPD and asthma. This group had high temporal stability, highlighting the persistence of Haemophilus colonization in the airways as an important pathogenic entity. For this subgroup, bacterial colonization is evident and likely amenable to targeted antimicrobial therapies.

28We showed that over half of neutrophilic COPD cases had a balanced microbiome profile. This group had lower sputum IL-1β and TNFα and higher sputum and serum IL-17A. In addition, it was temporally dynamic, and patients in this state at stability were not dominated by typical respiratory pathogens in their microbiota but may be susceptible to the greatest microbiome shifts during exacerbations, perhaps suggesting the need for monitoring airway ecology and pathogen acquisition in particular for this patient subgroup. The role of IL-17A has been increasingly recognized in COPD inflammation. IL-17A induces SAA, which in turn promotes neutrophilia by increasing IL-17A and Th17-regulating cytokines (i.e., IL-6). Therefore, IL-17A and SAA, both elevated in the neutrophilic balanced subgroup, may form a self-perpetuating axis precipitating neutrophilic inflammation.

29Our results suggested, albeit with distinct pathophysiology, that neutrophilic and eosinophilic COPD were temporally interchangeable in some patients, and the neutrophilic balanced subgroup served as an intermediate state between neutrophilic Haemophilus and eosinophilic states that were otherwise mutually exclusive. The biology underlying this dynamism is uncertain, but it has been shown that therapeutic blockade of Th2 cytokines enhances Th17 inflammation in asthma, suggesting that ICS treatment may contribute to the endotypic switches. Seasonality can be another factor contributing to the observed dynamics of the inflammatory pattern. It is important to note that our results are not necessarily contradictory to previous findings on the stability of eosinophilic COPD, in that some patients experienced transitions to a mixed granulocytic state or had decreased sputum eosinophilia slightly below a 3% threshold but may still have had eosinophilic inflammation. In our analysis, IL-17A was inversely correlated with the nonpathogenic taxa Oribacterium and Lactobacillus. Whether patients in this subgroup benefit from anti-IL-17 biologics or a microbiome modulation strategy (i.e., probiotics) warrants further investigation. However, the limited efficacy of targeting IL-17 and IL-23 in contrast to the beneficial effects of macrolide treatment in asthma suggests that targeting dysbiosis is likely preferable in this setting.

30We identified microbiota features specifically enriched in eosinophilic over neutrophilic COPD and showed that their enrichments were not simply due to the reciprocal decrease of Haemophilus abundance and may not be fully explained by concurrent ICS usage. The role of the airway microbiome in eosinophilic inflammation remains uncertain. Taylor and colleagues found modest associations between sputum microbiota and eosinophilic asthma, although Gemella, Rothia, and Porphyromonas were enriched in the eosinophilic group, which was consistent with our results as well as with those from Millares and colleagues. We showed that Campylobacter and Granulicatella were associated with eosinophilia both cross-sectionally and longitudinally, indicative of patient endotypic switches and associated with Th2 mediators. The same two genera were also enriched in the gut microbiome of patients with eosinophilic esophagitis in response to allergenic foods. Campylobacter and Aggregatibacter were further shown to be capable of inducing eosinophilic chemotaxis and degranulation. Given these results, it is possible that there exists a specific human microbiome population favoring the ecological niches of eosinophilia. Moreover, we showed that monitoring temporal changes of these airway microbiome features tracked patient inflammatory status in real time, which supports the potential need for point-of-care diagnosis using sputum samples and based on microbiome biomarkers.

31An important strength of this work is the availability of over 1,700 sputum samples across multiple sites across stability and exacerbations, which, to our knowledge, is by far the largest in a single airway-microbiome study. The results in the combined COPDMAP and AERIS cohorts were subject to cross-validation between sites and independent validation in BEAT-COPD. This analytical strategy ensured the robustness of inflammation-associated microbiome signatures in terms of clinical sites, visits, and sampling procedures. Another strength of the work is the long-term patient follow-up and employment of novel time-series analytical techniques, as compared with previous studies with similar purposes that were mostly cross-sectional. The longitudinal analysis raised the confidence for the association of the airway microbiome with eosinophilia and revealed the dynamic pattern of the microbiome in relation to patient inflammatory status.

32There are several limitations to this study. First, the samples in this study included both induced and spontaneous sputum, which may have differentially impacted the microbiome profiling. However, previous analyses and our analyses suggested that this impact was nonsignificant. Although we have assessed contaminations from reagent controls, the extent of oral microbes in sputum needs to be evaluated further. It also remains uncertain how sputum can recapitulate the ecology in the lower airways, given its inherent mixture of variable elements from the upper and lower airways. Second, the longitudinal sampling remained sparse, which somewhat limited the power of time-series analyses. An intensive and regular participant follow-up would allow a finer-scale examination of the temporal variability of the airway microbiome. Third, a subset of participants in BEAT-COPD and COPDMAP had sputum and serum mediators profiled because of limited sample availability. No other host omics data were characterized, and this omission includes transcriptomics, which may have helped better define Th2 and Th17 inflammation according to gene signatures and immune processes. Fourth, all three cohorts are observational, and most patients had moderate to severe disease; therefore, multiple confounders such as ICS treatment or antibiotic treatment are present. Further prospective interventional studies (similar to the study by Segal and colleagues in COPD and the studies by Durack and colleagues in asthma) are required to explicitly understand their effects on the airway microbiome. In addition, the underrepresentation of female participants needs to be considered to assess the generalizability of our findings. Last, our results can only be viewed as associations that are subject to further experimental testing to explore their causality.

33In summary, the airway microbiome is differentially implicated in neutrophilic and eosinophilic COPD. The microbiome can stratify neutrophilic COPD into subgroups that justify different therapies. Neutrophilic and eosinophilic COPD are interchangeable in some patients, which is related to composition shifts of the airway microbiome. Monitoring temporal variability of the airway microbiome may capture key changes in patient inflammatory status and assist in therapeutic selection. Results in this study highlight the consideration of the airway microbiome in the inflammatory endotype–based patient management in COPD.