Bone marrow characterization in COPD: a multi-level network analysis

We included in this prospective and controlled study 35 clinically stable COPD patients defined by GOLD criteria [11] and 25 healthy controls (10 current smokers and 15 never smokers with normal spirometry) who required planned cardio-thoracic surgery for clinical reasons (COPD and smoker controls had indication for lung resection due to lung nodule or cancer by thoracic surgery and never-smoker controls had indication for valve replacement by cardiac surgery).

Forced spirometry, static lung volumes and carbon monoxide diffusing capacity (DLCO) were determined following international recommendations, using Mediterranean reference values [19, 20] in the pre-anaesthetic visit before surgery.

Sternum fine-needle (18G) BM aspiration (3–5 ml) was performed by an expert haematologist in the surgical theatre, once the patient was under general anaesthesia and before the sternotomy. Fresh BM smears were used for direct cell count (only in COPD patients). A 50 μl BM sample was diluted 1:4 in 1× PBS (pH 7,4; PAA) and analyzed using CELL-DYN Sapphire (Abbot, USA) to determine cell types distribution. HSC and EPC were isolated from fresh BM aspirates using a RosetteSep Kit (STEMCELL Technologies Inc., Canada) following manufacturer’s instructions to obtain a more representative sample. Cells were quantified by flow cytometry on a FACScan (Coulter Epics XL-MCL; Beckmann Coulter) by the presence of antigens CD34 (BD Pharmingen, NJ, USA), KDR (R&D Systems, Minneapolis, USA), CD133 (Miltenyi Biotec, Germany), c-kit, Ki67 (Abcam, Cambridge, UK). Gates were set to detect CD34+ cells and to evaluate co-expressions of CD34+ cells with CD 133, KDR, c-kit and Ki67. At least 20.000 events were acquired for each sample. Following previous publications [2‐4], EPCs were also identified by FACS as CD34+ CD133+ KDR+ cells. The remaining BM sample were centrifuged at 1800 rpm/5 min/4 °C to separate cells from plasma and the supernatant was stored (100 μl aliquots) at − 80 °C until cytokine analysis.

Circulating blood (10 ml) was collected in tubes with or without EDTA by venipuncture before anaesthesia. Total and differential leukocyte counts were determined automatically (Sysmex K-4500, Toa Medical Electronics Co Ltd., Kobe, Japan). Blood was allowed to clot and serum was separated by centrifugation at 3000 rpm for 15 min at 4 °C and stored at − 80 °C until analysis. The concentration of IL-6 (Labclinics, Barcelona, Spain), IL-8 (Invitrogen, California, USA), VEGF (Ebioscience), TGF-β (Ebioscience), IGF (R&D Systems) and HGF (Invitrogen) in plasma (and BM supernatant) were determined by ELISA. Assay sensitivities, as expressed by the manufacturer, were: 0.03 pg/ml for IL-6, < 100 fg/ml for IL-8, 7.9 pg/ml for VEGF, 8.6 pg/ml for TGF-β, 0.056 ng/mL for IGF and 20 pg/ml for HGF. C-reactive protein (CRP) concentration was determined by Nephelometry.

Results are expressed as mean ± standard deviation, or median [interquartile range] values. ANOVA (or Kruskall-Wallis test) and unpaired t-test (or Mann-Whitney U-tests) were used to compare normally (and abnormally) distributed variables. Bivariate correlations were tested with Spearman test. Differences were considered statistically significant at 2-tailed p < 0.05 and the stronger of these Spearman relationships (threshold p-value < 0.01) between clinical and biological variables were plotted together as multi-level networks by representing variables as nodes and correlations between them as edges as previously described [21‐23]. These correlation networks were constructed with custom R scripts and Cytoscape [24]. The multi-level network includes variables from multiple levels (or categories), including clinical, lung function, BM and circulating blood information. Each node in the network corresponds to a specific variable (see names) whose colour denotes the level to which it belongs (see legends). Nodes are linked by edges only if their Spearman correlation was statistically significant (p < 0.01), either positively (continuous lines) or negatively (dashed lines). The thickness of the edge is proportional to the correlation coefficient (Rho) (see legends). Grey shaded areas highlight 5 isolated modules (or sub-networks) that are named arbitrarily based on the assessment by the investigators of the type of variables included in each of them. In order to compare different biological effects, we built different networks comparing COPD and controls, COPD with DLCO< 60% or ≥ 60%, COPD with blood eosinophils< 300 or ≥ 300 cells/μL, and COPD with a smoking history< 45 or ≥ pack-years.

Table 1 presents the main demographic and clinical characteristics of the three groups studied. By design, smoking history and lung function values were different between them. BMI was higher in never smokers (p < 0.05). COPD patients had moderate-severe airflow limitation, gas trapping and reduced DLCO.

The percentages of cells expressing immune-phenotypic (HSC (CD34+) and EPC (CD34+ CD133+ KDR+)) or proliferation and differentiation markers (Ki67+, c-kit+) were similar pair wisely across groups (Table 2). Likewise, we did not find significant differences between groups in inflammatory or repair biomarkers, neither in BM or circulating blood, with the exception of increased CRP levels in COPD patients and elevated circulating neutrophils in smokers (Table 2).

Some previous reports have shown reduced BM-derived circulating progenitors (CD34+ cells) in patients with COPD, particularly in those with more severe airflow limitation, arterial hypoxemia and low-BMI [12, 14]. On the other, Broekman et al. recently described that the immune-phenotype, growth, differentiation and migration capacity of BM progenitor cells were not different between 9 COPD patients and 9 age-matched, never-smoking, controls [15]. The results of our pair wise analysis (Table 1) are in keeping with this latter report. However, our study extends all these previous studies by including a larger number of patients and controls and, specially, by using, for the first time, multilevel network analysis to get a holistic understanding of the system, both at the BM, circulating blood and physiologic levels.

Despite that, by and large, we did not find significant pair wise differences between groups in the different immune-phenotypic, proliferation and differentiation, inflammatory or repair markers, neither in BM or circulating blood (Table 1), the use of multi-level correlation network analysis provided several observations of interests. Firstly, it identifies a module in which B-cells, progenitor cells (CD34+) and eosinophils interact in the whole population, as well as a second module that relates lung function parameters with immunity and repair nodes. Secondly, the network was more interconnected, hence more complex and developed, in COPD patients than in controls (Fig. 2 and Additional file 1: Figure S1), suggesting chronic system activation in COPD. The difference might, however, be due to the higher cumulated smoking history of COPD patients, as we have observed a denser network when comparing networks of patients with Pack-Years≥45 versus Pack-Years< 45 (Additional file 5: Figure S4).Thirdly, the COPD correlation network was more fragmented in those patients with emphysema (DLCO< 60% reference) or ≥ 300 circulating Eos/μL, indicating reduced system resilience [21] and potentially impaired repair capacity. This is consistent with the accepted pathogenesis of emphysema [28, 29] but highlights a novel patho-biological effect of eosinophils in the lung maintenance program. In fact, supporting this hypothesis and in keeping with previous studies [30, 31], we found that COPD patients with ≥300 Eos/μL had worse lung function and less BM HSCs (Table S2), suggesting deficient BM repair response. In support of this hypothesis is the recent observation in the COPDgene cohort of a relation between biomarkers of extracellular matrix turnover, emphysema and eosinophilic-bronchitis in patients with COPD [27]. Moreover, other studies have found involvement of eosinophils in colonic inflammation and repair [26]. The finding of eosinophils as cells that are present in almost all the networks being one of the most connected hubs in each network and also the different number of compartments in which it was involved (as clinical variables, blood cytokines, bone marrow surface markers) made us to consider the role of eosinophil as a possible regulatory cell.

Finally, we did not find differences in the bivariate analysis between different growth factors but the network analysis showed relationships among HGF, VEGF, TGF, EPCs, eosinophils and lung function parameters. It has been recently described that human mesenchymal stromal cells exert HGF dependent cytoprotective effects in a human relevant pre-clinical model of COPD [6], that is consistent with the role of HGF in the network connected with DLCO, FEV1 and EPC. Moreover, impaired VEGF signaling has been associated with emphysema in animal models [7] and our network shows an inverse correlation with EPC which could be explained as a consequence of the impaired repair mechanisms in these patients.

The use, for the first time, of network analysis to understand the BM system in a large cohort of COPD patients and controls is a clear strength of our study. It also has, however, some potential limitations that deserve comment. First, because COPD patients (and controls) required cardio-thoracic surgery for clinical reasons (with different indication for cases and controls), results may not be automatically generalised. Second, as discussed above, we studied a higher number of COPD patients than controls, and this may influence the structure of correlation networks [25]. To address this, we increased the number of controls by merging smokers and never smokers, and yet found similar results. However, we still could not correct for the effect of active smoking due to the sample size. Third, network density is also directly proportional to the threshold chosen for statistical significance, the lower the p value, the higher network density. To get an appropriate signal to noise balance, as others have done before [16, 21], we established an arbitrary p threshold value of less than 0.01, so weaker relationships were not included in the network. Fourth, although the data point to eosinophil as a possible mediator of the immune and repair response, especially in patients with emphysema, the authors cannot exclude the presence of other co-mediators and further studies are needed to demonstrate the role of these cells in the repair process of the lung. In keeping with this, due to the sample size we could not explore the effects of inhaled corticosteroids, that some of these patients were receiving, on the BM characteristics. Finally, we studied only one source of stem cells, namely the BM, while there are other organs also able to produce these cells, such as the spleen or even the lung themselves [32].