Breathomics for Assessing the Effects of Treatment and Withdrawal With Inhaled Beclomethasone/Formoterol in Patients With COPD

Background: Prospective pharmacological studies on breathomics profiles in COPD patients have not been previously reported. We assessed the effects of treatment and withdrawal of an extrafine inhaled corticosteroid (ICS)-long-acting β2-agonist (LABA) fixed dose combination (FDC) using a multidimensional classification model including breathomics. Methods: A pilot, proof-of-concept, pharmacological study was undertaken in 14 COPD patients on maintenance treatment with inhaled fluticasone propionate/salmeterol (500/50 μg b.i.d.) for at least 8 weeks (visit 1). Patients received 2-week treatment with inhaled beclomethasone dipropionate/formoterol (100/6 μg b.i.d.) (visit 2), 4-week treatment with formoterol alone (6 μg b.i.d.) (visit 3), and 4-week treatment with beclomethasone/formoterol (100/6 μg b.i.d.) (visit 4). Exhaled breath analysis with two e-noses, based on different technologies, and exhaled breath condensate (EBC) NMR-based metabolomics were performed. Sputum cell counts, sputum supernatant and EBC prostaglandin E2 (PGE2) and 15-F2t-isoprostane, fraction of exhaled nitric oxide, and spirometry were measured. Results: Compared with formoterol alone, EBC acetate and sputum PGE2, reflecting airway inflammation, were reduced after 4-week beclomethasone/formoterol. Three independent breathomics techniques showed that extrafine beclomethasone/formoterol short-term treatment was associated with different breathprints compared with regular fluticasone propionate/salmeterol. Either ICS/LABA FDC vs. formoterol alone was associated with increased pre-bronchodilator FEF25−75% and FEV1/FVC (P = 0.008–0.029). The multidimensional model distinguished fluticasone propionate/salmeterol vs. beclomethasone/formoterol, fluticasone propionate/salmeterol vs. formoterol, and formoterol vs. beclomethasone/formoterol (accuracy > 70%, P < 0.01). Conclusions: Breathomics could be used for assessing ICS treatment and withdrawal in COPD patients. Large, controlled, prospective pharmacological trials are required to clarify the biological implications of breathomics changes. EUDRACT number: 2012-001749-42.

The second e-nose, which detects frequency variations, contains an array of eight quartz microbalance (QMB) gas sensors coated by molecular films of metallo-porphyrins (Montuschi et al, 2010). Sensors detect the amount of molecules absorbed in a sensitive film through the changes of resonant frequency that is proportional to the absorbed mass (Montuschi et al, 2010).

EBC sampling
The condenser has a saliva trap to reduce the chance of salivary contamination Montuschi et al, 2012). Under our experimental conditions, no effect of salivary contamination or cleaning solution on EBC profiles has been observed ( ). Subjects were asked to wash their mouth thoroughly before collecting EBC, to breathe tidally through a mouthpiece into a two-way non-rebreathing valve for 15 minutes wearing a noseclip, and to stop breathing into the mouthpiece and swallow every time they felt salivation. An average of 1.5 ± 0.2 ml (mean ± SD) of EBC was collected in 15 minutes of tidal breathing. EBC sampling was performed as previously described (Montuschi et al, 2012). EBC was immediately transferred into 10 ml glass vials, closed with 20 mm butyl rubber lined with polytetrafluoroethylene septa and crimped with perforated aluminium seals (Montuschi et al, 2012) . Before sealing, volatile substances were removed from samples by a gentle nitrogen gas flow for 3 min (Montuschi et al, 2012).[ Previous experiments showed no difference with spectra acquired after a variable time of nitrogen exposure (1, 3, 5, 10, 15 and 20 min). [9] However, as 1-min interval appeared to be too short to avoid systematic errors, a 3-min interval was chosen. Samples were not dried out to avoid their precipitation, with a possible loss of nonvolatile compounds, and/or formation of aggregates upon dissolving the dried condensate for NMR measurements.
Sealed samples were then frozen in liquid nitrogen, so as to immediately "quench" metabolism and 1 H-NMR spectra were acquired using a Bruker 600 MHz spectrometer (Bruker BioSpin) operating at 600. 13 MHz proton Larmor frequency and equipped with a 5 mm CPTCI 1 H- 13 C/ 31 P-2 H cryo-probe including a z-axis gradient coil, an automatic tuning-matching (ATM), and an automatic sample changer (Bertini et al, 2014).
A PT 100 thermocouple served for temperature stabilization at the level of approximately 0.1 K at the sample. Before measurement, samples were kept for at least 3 min inside the NMR probe head for temperature equilibration (300 K). For each sample, a one-dimensional NMR spectrum was acquired with water peak suppression using a standard pulse sequence (NOESY presat), using 512 free induction decays (FIDs), 64 k data points, a spectral width of 12019 Hz, an acquisition time of 2.7 s, a relaxation delay of 4 s, and a mixing time of 100 ms (Bertini et al, 2014).
Free induction decays were multiplied by an exponential function equivalent to 1.0 Hz linebroadening factor before applying Fourier transform. Transformed spectra were manually corrected for phase and baseline distortions, and calibrated (TMSP peak at 0.00 ppm) using TopSpin (Bruker). Each one-dimensional spectrum in the range of 0.02-10 ppm was segmented into 0.02-ppm chemical shift buckets, and the corresponding spectral areas, after scaling, were used as input variables for the subsequent statistical analysis (Bertini et al, 2014). Regions between 6.0 and 4.5 ppm, containing residual water signals, were removed. Probabilistic quotient normalization (PQN) algorithm (Dieterle et al, 2006) was chosen to normalize the spectra, because it is a more accurate and robust alternative to the total area scaling, and PQN performs well even when large variations of dilution occur. Signals were assigned on template one dimensional NMR profiles by using matching routines of AMIX 3.8.4 (Bruker BioSpin) in combination with the BIOREFCODE reference database (Bruker BioSpin), and publicly available databases like HMDB (Human Metabolome DataBase) (Wishart et al, 2007).
All data analysis were performed using R, an open source software for the analysis of statistical data (Ihaka and Gentleman, 1996). Multivariate data analyses were conducted on processed data by combining established methods. Data reduction was obtained by means of multilevel Partial Least Square Analysis (mPLS) using the algorithm implemented in the R-library "plsgenomics" and the standard R function "cancor".
For the purpose of classification, we used the K-nearest neighbors (k-NN) method (k = 3) applied on the multilevel PLS scores (Cover and Hart, 1967). All the accuracies reported and the confusion matrix for the different classifications were assessed by means of 100 cycles of a Monte Carlo cross-validation scheme (MCCV, R script in-house developed) (Bijlsma S et al, 2006;Boulesteix, 2015). Briefly, 90% of the data were randomly chosen at each iteration as a training set to build the model. Then the remaining 10% was tested and sensitivity, specificity and accuracy for the classification were assessed.
For the univariate analysis, after PQN normalization of the spectra, each spectral region related to the metabolites assigned in the 1 H NMR profiles was aligned, by a simple horizontal rigid shift, in a way that the corresponding peaks of interest resulted superimposed. Then, they were integrated reducing the selected ppm intervals, and improving the precision in the integration by compensating the variations in peaks position in the different EBC samples. The obtained integrals represent the metabolite concentrations in arbitrary units. The whole procedure was performed using a R script developed in house. EBC metabolite concentrations at each visit were determined. Statistical significance was assessed using the univariate non-parametric Wilcoxon signed-rank test (Wilcoxon, 1945). A p-value < 0.05 was deemed statistically significant.
To evaluate the discrimination power of the integrated model containing all the different analytical approaches used in this study, paired visit comparisons were performed using data matrices created by binding together all the different variables obtained from the various techniques (EBC NMR-spectroscopy, spirometry, carbon polymer sensor e-nose, quartz crystal sensor e-nose, eicosanoids and F E NO measurements, sputum cells). On this dataset, the same multivariate statistical approach previously described was applied (multilevel PLS and crossvalidation using k-NN). Furthermore, the accuracies calculated were assessed for significance against the null hypothesis of no prediction accuracy in the data by means of 10 2 randomized class-permutations test (Alonso et al, 2015).

Sputum cell analysis
Sputum induction, processing and analysis were performed according to the European Respiratory Society (ERS) guidelines Paggiaro et al, 2002). Baseline FEV 1 was recorded before sputum induction. Subjects were pre-treated with inhaled salbutamol (400 µg), and, after 10 min, a spirometry was repeated. (Paggiaro et al, 2002) Subjects were asked to inhale hypertonic saline (3%) for 5 min and, then, to rinse their mouths and try to expectorate into a sterilized box (Paggiaro et al, 2002). Five-minute inhalation sessions were repeated four times for a total of 20 min (Paggiaro et al, 2002). A spirometry was performed after each inhalation session to detect significant fall of FEV 1 . The procedure was stopped when approximately 1 g of plugs was collected, or if patients had symptoms or/and if FEV 1 was reduced more than 20% over baseline values (Paggiaro et al, 2002). Sputum was processed within 2 h to ensure optimum cell counting and staining, with the sample always kept in ice (Efthimiadis et al, 2002). 100-500 mg sputum was selected for sputum analysis. Dithiothreitol (DTT) 0.1% was added to sputum samples which were kept in a shaking rocker at room temperature for 20 minutes for sample homogenization Cytospins were stained for differential cell counts using Giemsa staining (Efthimiadis et al, 2002).
The differential cell counts was performed by counting a minimum of 400 nonsquamous cells and reported as the relative numbers of eosinophils, neutrophils, macrophages, lymphocytes, and bronchial epithelial cells, expressed as a percentage of total nonsquamous cells (Efthimiadis et al, 2002). The percentage of squamous cells was reported separately. Slides with squamous cells > 30% of total cells were discarded. Slides were read blindly by two qualified and fully trained physicians. Monthly quality control was performed including internal slide reading and equipment calibration (Efthimiadis et al, 2002).

Multivariate data analysis
To investigate the correlations among the different analytical approaches used in this study, a matrix with data from all techniques was created and then its correlation matrix was calculated using the algorithm implemented in the R-library "psych" (Hahsler et al, 2008;Revelle, 2016).
The heatmap of the correlation matrix was built using the R-function "heatmap.2", implemented in the "gplots" package (Warnes et al, 2016). This function reordered the rows and columns of the correlation matrix according to the restrictions imposed by the dendrogram, calculated on the basis of the Euclidean distance among the data and using the R-function "hclust".

Statistical analysis
Data were expressed as mean ± SEM or medians and interquartile ranges (25 th and 75 th percentiles), after assessing for normality with the D'Agostino-Pearson omnibus normality test.
Depending on data distribution, repeated-measures ANOVA or Friedman test were used for assessing within-group pharmacological treatment effect. If overall statistical significance was observed, unpaired t test and Mann-Whitney U test were used for comparing groups for normally distributed and nonparametric data, respectively. Correlation was expressed as a Pearson coefficient. Significance was defined as a value of p < 0.05.

Study subjects
The mean age of the subjects (2 females and 12 males) enrolled in this study was 73.6 ± 1.8 years.
Patients smoke a mean of 64.3 ± 1.8 pack-year and, according to the airflow limitation severity, were classified as GOLD I (5 patients), II (5 patients), and III (4 patients). Information about common aeroallergens, history of atopy, sputum eosinophils at visit 1 are detailed in Table 1.

Pulmonary function testing
All patients with COPD performed spirometry at all visits. There was no missing data.
An Excel table with lung function test values for individual patients with COPD is provided as online supplementary material. Results of pulmonary function tests show higher mean absolute and percentage of predicted pre-bronchodilator FEF 25%-75% values after treatment with inhaled beclomethasone/formoterol FDC in microparticle formulation (visit 2) compared with inhaled formoterol alone in microparticle formulation (visit 3) (Figure 2A and 2B). These data suggest that inhaled ultrafine beclomethasone/formoterol FDC improves small airway function as reflected by FEF 25%-75% values. This might be particularly relevant as this functional effect is observed after only 4 week treatment with inhaled beclomethasone/formoterol, a relatively short duration of treatment for COPD trials, and in view of the fact that all patients with COPD included in our study had normal sputum eosinophils, negative reversibility test to bronchodilators, negative skin prick tests, and no history of atopy, thus, excluding an asthma component, on which ICS are generally more effective.
An effect of ICS on small airway function is also suggested by a higher mean pre-bronchodilator FEV 1 /FVC ratio at visit 1 (COPD patients were on inhaled fluticasone/salmeterol FDC at full doses) compared with visit 3 (post-treatment with inhaled formoterol alone) ( Table 2).
Higher mean absolute and percentage of pre-bronchodilator PEF predicted values were observed after 4 week treatment with inhaled beclomethasone/formoterol (visit 2) compared with visit 1 (Figure 2C and 2D, Table 2). These data might be the consequence of a higher lung deposition of the ultrafine drug formulation delivered through pMDI, resulting in a greater antiinflammatory and/or bronchodilating effect of inhaled beclomethasone/formoterol combination in the small airways compared with inhaled fluticasone/salmeterol in standard formulation delivered through DPI (Table 2). No within-group differences in post-bronchodilator functional parameters were observed (Table 3).

Electronic nose
E-nose analysis with carbon polymer sensor e-nose and quartz crystal sensor e-nose was performed in all 14 subjects at all 4 visits for a total of 56 breathprints for each e-nose. Results are shown in Table 4  Consistent with carbon polymer sensor behaviour, two quartz crystal sensors showed significant differences between visit 1 and visit 4, whereas between-visit difference in sensor 3 (P = 0.056) and 5 (P = 0.059) response was close to statistical significance. Sensor 3 and 4 showed significant differences between visit 1 and visit 2, whereas between-visit differences in sensor 2 (P = 0.054) and 5 (P = 0.093) approached statistical significance (Table 5).
These results are consistent with and confirm those obtained with the carbon polymer sensor array, an e-nose based on a different technology (see above), indicating a robust methodology.

Metabolomic analysis of EBC with NMR spectroscopy
EBC samples obtained from 14 patients with COPD at visit 1 to visit 4 for a total of 56 EBC samples were analyzed with NMR spectroscopy as described in Methods. There was no missing data.
When EBC spectra were analyzed with multilevel PLS, that is analysis of pairs of spectra from the same subject at two different visits, comparison between visit 1 and visit 4 showed a classification accuracy of 72% (Supplementary Figure 2).
These data indicate that regular treatment with inhaled fluticasone/salmeterol (COPD patients were on this FDC at a constant full dose, delivered through DPI, for at least 8 weeks) is associated with a different EBC metabolite profile compared with that observed after 4-week treatment with inhaled beclomethasone/formoterol in microparticle formulation delivered through pMDI. The pathophysiological meaning and possible implications of these results require further research.
Regarding the other paired comparisons, classification accuracy was below 70% (s- Table 1). The lack of between visit differences in EBC metabolite profiles, apart from visit 1 vs visit 4 (formate) and visit 3 vs visit 4 (acetate), could be due to high intragroup inter-individual variability in EBC metabolites which might mask pharmacological treatment-induced differences.
Among metabolites responsible for discrimination between visit 1 and visit 4, formate was the only one with a P-value < 0.05. EBC formate levels were lower at visit 4 (409.16 ± 224.76, relative intensity) compared with those observed at visit 1 (694.95 ± 360.47, relative intensity; P = 0.029) (Supplementary Figure 3A and 3B). EBC formate levels were also lower at visit 4 (409.16 ± 224.76, relative intensity) compared with those observed at visit 2 (695.14 ± 519.50, relative intensity; P = 0.049) ( Table 6), although accuracy of classification between visit 2 and visit 4 based on EBC NMR spectroscopy did not reach the significant threshold of 0.7 (0.63) (s- Table 1).
Moreover, at visit 4 EBC acetate levels were lower than those measured at visit 3 (P = 0.01) (s- Table 3). The pathophysiological meaning and possible implications of these findings require further research.

Measurement of fraction of exhaled nitric oxide (F E NO)
Measurement of F E NO was performed in all subjects at each visit. There was no missing data.
There was no within-group difference in F E NO concentrations in patients with COPD (overall P = 0.3529) (s- Table 2).

Measurement of PGE 2 in sputum supernatants
PGE 2 concentrations were detected in 49 out of a total of 56 sputum samples. In 7 samples, sputum PGE 2 concentrations were undetectable. One patient with COPD had undetectable sputum PGE 2 concentrations at all visits (4 samples) and was excluded from statistical analysis. At the other 3 undetectable samples, an arbitrary concentration value of 1 pg/ml, corresponding to 50% of the analytical technique detection limit (2 pg/ml), was assigned.
Median sputum PGE 2 concentrations in patients with COPD at visit 1 to visit 4 are shown in s- Table 3. These data suggest that treatment with both ICS combinations, containing either fluticasone propionate or beclomethasone, reduce sputum concentrations of PGE 2 , an eicosanoid which has pro-inflammatory effects in the airways.

Measurement of 15-F 2t -isoprostane in sputum supernatants
15-F 2t -isoprostane concentrations were detected in 30 out of a total of 56 sputum samples. In 26 samples, sputum 15-F 2t -isoprostane concentrations were undetectable. Five patients with COPD had undetectable sputum 15-F 2t -isoprostane concentrations at all visits (20 samples) and were excluded from statistical analysis. At the other 6 undetectable samples, an arbitrary concentration value of 1 pg/ml, corresponding to 50% of the analytical technique detection limit (2 pg/ml), was assigned.
Median sputum 15-F 2t -isoprostane concentrations in patients with COPD at visit 1 to visit 4 are shown in s- Table 3.
There was no within-group difference in sputum 15-F 2t -isoprostane concentrations (P = 0.84) (s- Table 3) suggesting that this biomarker of oxidative stress is resistant to ICS treatment in line with previous studies (Montuschi et al, 2000). undetectable samples, an arbitrary concentration value of 1 pg/ml, corresponding to 50% of the analytical technique detection limit (2 pg/ml), was assigned.

Measurement of PGE 2 in EBC
Median EBC PGE 2 concentrations in patients with COPD at visit 1 to visit 4 are shown in s- Table   3.

Measurement of 15-F 2t -isoprostane in EBC
15-F 2t -isoprostane concentrations were detected in 46 out of a total 56 EBC samples. In 10 samples, EBC 15-F 2t -isoprostane concentrations were undetectable. At these samples, an arbitrary concentration value of 1 pg/ml, corresponding to 50% of the analytical technique detection limit (2 pg/ml), was assigned.
Median EBC 15-F 2t -isoprostane concentrations in patients with COPD at visit 1 to visit 4 are shown in s- Table 3.
There was no within-group difference in EBC 15-F 2t -isoprostane concentrations (P = 0.60) (s- Table   3) suggesting that this biomarker of oxidative stress is resistant to ICS treatment in line with previous studies (Montuschi et al, 2000).

Sputum cell analysis
Sputum was successfully induced and collected in all patients at all visits. Fifty-six sputum samples were collected. Corresponding cytospins were prepared and slides were stained with May-Grunwald-Giemsa as described Methods. Sputum slides were read as described in Methods.
Thirteen sputum slides were discarded because of high salivary squamous cells (> 30% of total sputum cells). No patient with COPD had sputum eosinophilia, as defined by sputum cell counts > 3%, at visit 1 (screening visit).
Medians and interquartile range of sputum cell types in the 8 patients with COPD who had a complete set of sputum slides are shown in s- Table 4. Neutrophils were the predominant cell types in sputum obtained from patients with COPD.
Within-group comparison was performed with Friedman test in the eight patients who had a complete set of sputum slides (visit 1 to visit 4).
There was no within-group difference in neutrophil, macrophage, eosinophil, basophil, bronchial epithelial cell counts (s- Table 5). Basophil and bronchial epithelial cell counts are not shown in Table 5 as counts were mostly 0.

Correlations
Correlations between multidimensional variables in 14 patients with COPD at visit 1 to visit 4 (n = 56) are shown as a heatmap (Figure 4).
The multidimensional models showed higher accuracy than the models based on spirometry alone (Table 5). The sensitivity and specificity reported in Table 5 follow the standard definitions of proportion of true positives and proportion of true negatives, respectively. It is worth noting that they appear equal for all the comparisons. Furthermore, score plots from PLS analysis are centrosymmetric. This seeming uncommon behaviour is fully understandable given the pairwise nature of the analyses performed. In a multilevel PLS, between subject variation is separated from within subject variation by subtracting the individual specific average. In the two class cases, this construction leads to a matrix with a two block structure with opposite signs, thus producing

Correlations
In the EBC, the strong correlations between PGE 2 and small molecular weight metabolites detected by NRM spectroscopy are unlikely to be explained by individual variability in aerosol particle formation (Effros et al, 2003) as there was no correlation EBC 15-F 2t -isoprostane and EBC metabolites nor between EBC PGE 2 and EBC 15-F 2t -isoprostane.
Persistent neutrophilic airway inflammation is a typical characteristic of COPD, which persists even after smoking cessation and sputum neutrophil cell counts are a direct measure of airway inflammation which can be used for assessing the anti-inflammatory effects of drugs (Simpson et al, 2014). We report a negative correlation between the response of some quartz crystal sensors, in terms of relative changes in sensor frequency [(ƒmax-ƒ0)/ƒ0] (Hz), and sputum neutrophil cell counts. These data suggest that the quartz crystal sensor e-nose is potentially useful for assessing neutrophilic airway inflammation and that the higher neutrophilic airway inflammation the lower enose response.
The response of 15 out of 32 carbon polymer sensors was correlated with EBC phenol concentrations. However, the biological meaning of these data has to be defined. There was no correlation between any carbon polymer sensor and the inflammatory outcomes measured in this study, including sputum neutrophil and macrophages cell counts, PGE 2 concentrations in sputum supernatants and EBC, 15-F 2t -isoprostane concentrations in sputum supernatants and EBC.
However, these findings do not exclude the potential utility of carbon polymer sensor e-nose for assessing airway inflammation as 1) inflammation that characterizes COPD is a complex, heterogeneous, pathophysiological process involving many mediators, whereas only a limited number of inflammatory outcomes was measured in the present study; 2) correlations were studied in 14 patients with COPD considering all 4 visits (n = 56) in a time frame of 10 weeks during which pharmacological treatment was changed three times: this might have caused different effects on enose sensors and inflammatory outcome measure, thus, resulting in non-significant correlations.