Discovery and Validation of a Volatile Signature of Eosinophilic Airway Inflammation in Asthma

Abstract

Rationale

Volatile organic compounds (VOCs) in asthmatic breath may be associated with sputum eosinophilia. We developed a volatile biomarker signature to predict sputum eosinophilia in asthma.

Methods

VOCs emitted into the space above sputum samples (headspace) from patients with severe asthma (n = 36) were collected onto sorbent tubes and analyzed using thermal desorption gas chromatography–mass spectrometry (GC-MS). Elastic net regression identified stable VOCs associated with sputum eosinophilia ⩾ 3% and generated a volatile biomarker signature. This VOC signature was validated in breath samples from: 1) patients with acute asthma according to blood eosinophilia ⩾0.3 × 10⁹cells/L or sputum eosinophilia of ⩾3% in the UK EMBER (East Midlands Breathomics Pathology Node) consortium (n = 65) and 2) U-BIOPRED-IMI (Unbiased Biomarkers in Prediction of Respiratory Disease Outcomes Innovative Medicines Initiative) consortium (n = 42). Breath samples were collected onto sorbent tubes (EMBER) or Tedlar bags (U-BIOPRED) and analyzed by GC-MS (GC × GC-MS for EMBER or GC-MS for U-BIOPRED).

Measurements and Main Results

The in vitro headspace identified 19 VOCs associated with sputum eosinophilia, and the derived VOC signature yielded good diagnostic accuracy for sputum eosinophilia ⩾3% in headspace (area under the receiver operating characteristic curve [AUROC] 0.90; 95% confidence interval [CI], 0.80–0.99; P < 0.0001), correlated inversely with sputum eosinophil percentage (r_s = −0.71; P < 0.0001), and outperformed fractional exhaled nitric oxide (AUROC 0.61; 95% CI, 0.35–0.86). Analysis of exhaled breath in replication cohorts yielded a VOC signature AUROC (95% CI) for acute asthma exacerbations of 0.89 (0.76–1.0) (EMBER cohort) with sputum eosinophilia and 0.90 (0.75–1.0) in U-BIOPRED, again outperforming fractional exhaled nitric oxide in U-BIOPRED (0.62 [0.33–0.90]).

Conclusions

We have discovered and provided early-stage clinical validation of a volatile biomarker signature associated with eosinophilic airway inflammation. Further work is needed to translate our discovery using point-of-care clinical sensors.

At a Glance Commentary

Scientific Knowledge on the Subject

Indirect biomarkers of eosinophilic airway inflammation, such as fractional exhaled nitric oxide and blood eosinophils, are routinely used in practice but are not directly related to sputum eosinophils.

What This Study Adds to the Field

This study has developed and provided initial clinical validation for a direct biomarker of sputum eosinophilia in severe asthma using headspace and exhaled volatile organic compound analysis and could offer the opportunity for future point-of-care testing to directly monitor eosinophils in the airway.

Asthma is a complex chronic disease characterized by type-2 inflammation in 40–60% of patients, associated with cognate elevation of relevant airway immune cells (1). According to international guidelines, approximately 4% of patients with asthma have severe disease, characterized by treatment failure to high-dose inhaled steroids and current add-on therapies and requiring high-cost targeted biologic therapies, such as eosinophil-depleting therapies targeting IL-5 and its receptor (2). Furthermore, several orally active biologic agents are now in development for type-2 asthma in patients who have a high unmet need but are “prebiologic” care.

Concomitant with the development of high-cost therapies in moderate to severe asthma is the need for biomarkers that can be readily translated for disease stratification, monitoring, and deeper understanding of disease mechanisms. Although sputum phenotyping has proven to be invaluable in the characterization of airway inflammation in severe asthma (3), sputum sampling is time consuming, biased toward patients who can generate a sample at any given time, and costly, limiting its application across centers. Fractional exhaled nitric oxide (Fe_NO) has emerged as useful biomarker in asthma and can be used to identify steroid-responsive airway disease (4), as well as to stratify biologic therapy (5). However, Fe_NO concentrations are largely driven by inducible nitric oxide synthase activation as a consequence of IL-13 induction (6) and are modified by corticosteroids, as well as other confounders such as diet, smoking, and coexisting rhinitis (7). Consequently, the development of a noninvasive exhaled breath biomarker of eosinophilic airway inflammation in patients with severe asthma would offer the potential for noninvasive and potential future near patient phenotyping of airway inflammation in severe asthma and may identify novel mechanisms of disease beyond T2 activation.

We recently systematically reviewed the evidence for the association of volatile organic compound (VOC) biomarkers with type-2 inflammatory clinical biomarkers (including sputum eosinophilia) across 44 studies in asthma and chronic obstructive pulmonary disease (COPD) (8). Notably, Schleich and colleagues recently identified hexane and 2-hexanone as exhaled breath VOCs able to discriminate between eosinophilic and paucigranulocytic airway inflammatory phenotypes in asthma (9), and Ibrahim and colleagues identified several exhaled VOCs able to discriminate eosinophilic from noneosinophilic phenotypes with a high classification accuracy (10). In addition, several studies have sought to determine the utility of VOC biomarkers for airway inflammation by examining the headspace volatilome of cultured primary cells (11–13) to identify VOCs potentially associated with a putative cellular source.

However, no study to date has sought to validate in vitro–detected volatile biomarkers in exhaled breath samples acquired from relevant clinical target asthma populations, with a view to maximizing their potential for translation into the clinic.

This study aimed to: 1) discover a panel of in vitro volatile biomarkers associated with sputum eosinophilia (⩾3%) in headspace (region directly above the sputum samples entraining emitted VOCs) of sputum from patients with severe asthma and develop a VOC biomarker signature from the VOCs; and 2) validate the VOC biomarker signature in exhaled breath samples (in vivo) within two independent clinical cohorts in the context of acute asthma exacerbations (East Midlands Breathomics Pathology Node [EMBER] consortium) (14) and stable severe asthma (Unbiased Biomarkers in Prediction of Respiratory Disease Outcomes [U-BIOPRED] consortium) (15), characterized according to blood or sputum eosinophilia.

Methods

Patient Cohorts

All patient cohorts for biomarker discovery and exhaled breath validation are summarized in Figure 1; their clinical, breath sampling, and biomarker discovery and replication is summarized in Table 1.

Figure 1.

(A) Graphical summary of the discovery cohort and identification of canonical eosinophil-associated volatile organic compounds (VOCs) using elastic net (eNET) regression and generation of volatile biomarker score using the regression coefficient and VOC concentrations in tissue headspace. (B) Summary of EMBER (East Midlands Breathomics Pathology Node) and U-BIOPRED (Unbiased Biomarkers in Prediction of Respiratory Disease Outcomes) replication cohorts for exhaled breath validation of eosinophil-associated VOCs identified in the discovery cohort. GC-MS = gas chromatography–mass spectrometry; TD = thermal desorption.

Table 1.

Summary of the Methodological Aspects of the Three Datasets

Sample Type	Discovery Cohort, Headspace	EMBER Cohort, Exhaled Breath	U-BIOPRED Cohort, Exhaled Breath
No. of subject/samples	N = 36	N = 65 (cohort with sputum eosinophils, 24/65)	N = 42 (cohort with sputum eosinophils, 22/42)
Asthma conditions GINA 2023 treatment step in the sputum cohorts	Stable severe asthma GINA* 5 (anti-IL-5 ±  mOCS): 14/36 GINA 5 (mOCS): 6/36 GINA 4-5 (ICS/LABA only): 16/36	Acute asthma exacerbation GINA 1: 2 GINA 2–3: 9 GINA 4–5 (none on mOCS): 13 Asthma biologic: 0	Stable severe asthma GINA 4–5: 8 (not on mOCS) GINA 5: 14 (all on mOCS) Asthma biologic: 0
Asthma phenotype according to sputum eosinophils ⩾ 3%	Eosinophilic: 22/36 Noneosinophilic: 14/36	Eosinophilic: 9/25 Noneosinophilic: 16/25	Eosinophilic: 14/22 Noneosinophilic: 8/22
Fraction of exhaled breath collected	Not applicable—headspace of native sputum eosinophils	Mixed expiratory breath	Mixed expiratory breath
Exhaled breath container	Dedicated headspace collection systems, reported in Ref. 17	ReCIVA Breath Sampler, 1 L of breath sampled from the phase II/III slope of the CO2 sensor trace	10 L Tedlar bag (SKC)
Preconcentration method	Carbograph 1TD and Tenax TA 60/40 (Markes International)	Carbograph 1TD and Tenax TA 60/40 (Markes International)	Tenax (Tenax GR SS 6 mm × 7 inch; Gerstel)
Sample storage	Into sorbent tubes, at 4°C for no longer than 15 d	Into sorbent tubes, at 4°C after dry purge, for no longer than 72 h	Up to 39 d on sorbent tubes at 4°C, breath samples were purged onto to sorbent tubes within 10 min of acquisition.
Analytical platform	TD-GC-MS	TD-GC × GC-FID-MS	TD-GC-MS
Data processing platform	AnalyzerPro (Spectral Works, v.5.7)	MassHunter GC-MS Acquisition B.07.04.2260 (Agilent Technologies Ltd) GC Image v.2.6 GC Project and Image Investigator (JSB Ltd)	Masshunter Quantitative Analysis (Agilent Technologies)
Multivariate statistical method	Elastic net regression fitted using the cv.glmnet function from the glmnet package in R.3.6.1 (R Core Team, https://www.R-project.org).	Elastic net regression fitted using the cv.glmnet function from the glmnet package in R.3.6.1 (R Core Team, https://www.R-project.org).	Elastic net regression fitted using the cv.glmnet function from the glmnet package in R.3.6.1 (R Core Team, https://www.R-project.org).

Definition of abbreviations: EMBER = East Midlands Breathomics Pathology Node; FID = flame ionization detector; GC-MS = gas chromatography–mass spectrometry; GINA = Global Initiative for Asthma; ICS = inhaled corticosteroids; LABA = long-acting β antagonist; mOCS = maintenance oral corticosteroids; TD = thermal desorption; U-BIOPRED = Unbiased Biomarkers in Prediction of Respiratory Disease Outcomes.

^*Global Initiative for Asthma 2023, treatment intensity step (17).

Discovery cohort

Thirty-six patients were recruited from the adult severe asthma service at the Glenfield Hospital (United Kingdom) between 2018 and 2019. All patients had a diagnosis of severe asthma according to the American Thoracic Society/European Respiratory Society 2014 consensus criteria (16), were recruited at least 6 weeks after exacerbation, and were required to have been on stable treatment for at least 4 weeks before study entry. Clinical characteristics of the discovery cohort are provided in Table 2. Participants were asked to provide a spontaneous sputum sample, which was immediately transferred to the lab for the collection of 1 L of headspace onto sorbent tubes (Carbograph 1TD and Tenax TA 60/40; Markes International), using a previously validated and reported protocol (see online supplement section 1.3) (17).

Table 2.

Clinical Characteristics of Donors of Sputum Samples with Severe Asthma in the Discovery Cohort

	Patients with Eosinophil-enriched Sputa* (n = 22)	Patients with Non–Eosinophil-enriched Sputa† (n = 14)	P Value
Age, yr	67 (58–71)	58 (56–66)	0.23
BMI, kg/m2	29.6 (26.4–37.6)	31.1 (29.1–35.2)	0.49
Sex, % females	45	28	0.31
Age of asthma onset, yr	39 (7–53)	35 (29–49)	0.72
Asthma duration, yr	24 (8–33)	23 (12–38)	0.95
Smoking status, current–never–ex	1/20–13/20–8/20	0/12–7/12–7/12	0.56
Post-bronchodilator FEV1, % predicted	75 (53.8–92.4)	53 (49.5–81)	0.17
Post-bronchodilator FEV1/FVC, % predicted	89.1 (76.3–99.3)	73.9 (66.7–99.1)	0.47
FeNO, ppb	37 (23–52)	26 (14–68.5)	0.40
Blood eosinophils, ×109/L	0.21 (0.08–0.46)	0.10 (0.05–0.3)	0.30
ACQ 6 score	1.67 (1–3.17)	2.67 (0.67–4.08)	0.85
Atopy,‡ % yes	53	60	0.70
No. of patients on daily dose of maintenance OCS	13/22	7/14	0.31
Maintenance OCS dose (mg or prednisolone/24 h)	5 (5–10)	7.5 (5–15)	0.31
No. of patients on high dose of ICS§	20/22	11/14	>0.99
No. of patients on LABA	21/22	12/14	>0.99
Number of patients on LAMA	11/22	5/14	0.72
Anti-leukotrienes (montelukast)	5/22	5/14	0.24
No. of patients on concurrent anti-IL5 therapy (mepolizumab)	8/22	5/14	0.71
Anti–IL-5 (mepolizumab) treatment duration, wk	8 (6–12)	12 (4–16)	0.98

Definition of abbreviations: ACQ 6 = Asthma Control Questionnaire; BMI = body mass index; LABA = long-acting β antagonist; Fe_NO = fractional exhaled nitric oxide; LAMA = long-acting muscarinic antagonist; OCS = oral corticosteroid.

Data are summarized as median (Q1–Q3) unless otherwise noted. A nonparametric test (Mann-Whitney) was performed for continuous variables.

^*Sputum eosinophils ⩾ 3%.

^† Sputum eosinophils < 3%.

^‡ Positive skin prick test.

^§ Beclomethasone dipropionate (hydrofluoroalkane [HFA]-propelled pressurised metered-dose inhaler) > 1,000 μg/d.

Subsequently, within 1 hour of collection, sputum was processed for the differential cell counting using the protocol described in online supplement section 4.3.2 to group sputum headspace samples as eosinophil-enriched (n = 22) and non–eosinophil-enriched (n = 14) samples according to the sputum eosinophil threshold of 3% (see Table E5 in the online supplement).

Both control and sputum headspace samples were then analyzed by thermal desorption gas chromatography coupled to mass spectrometry (TD-GC-MS) as previously reported (17) and summarized in Table 1.

Exhaled breath clinical validation cohorts

Separate exhaled breath clinical validation studies were performed to evaluate the biomarker signature accuracy to predict eosinophilic asthma exacerbations (EMBER cohort) (14) and severe eosinophilic asthma (U-BIOPRED cohort) (15), as summarized in Table 1.

EMBER cohort (n = 65): Participants from an acute breath phenotyping cohort (14) were admitted to the hospital for severe asthma exacerbations (Glenfield Hospital, United Kingdom). Breath samples were acquired within 24 hours of hospital admission and were collected onto sorbent tubes using a ReCIVA sampler (Owlstone Medical), as described in online supplement section 1.4 and Figure E2, and analyzed by two-dimensional gas chromatography with a flame ionization detector and mass spectrometry (GC × GC-FID-MS), as previously reported in the core EMBER study (18) and analytical validation study (19). Participants were then characterized according to blood or sputum eosinophilia (blood eosinophils ⩾ 0.3 × 10⁹/L [n = 20] and blood eosinophils < 0.3 × 10⁹/L [n = 45]; or spontaneous sputum eosinophils ⩾ 3% [n = 9] and sputum eosinophils < 3% [n = 15] in a subgroup of participants). Fe_NO data were not available in EMBER, as patients had by and large presented with severe exacerbations (Tables E6 and E7), with airflow obstruction and elevated respiratory rates, and consequently were unable to perform Fe_NO tests acutely.

U-BIOPRED cohort (n = 42): The U-BIOPRED study design has been previously reported (20). Breath sampling in U-BIOPRED was as previously reported (described in online supplement section 1.5) into Tedlar bags, followed by trapping of volatiles onto Tenax sorbent tubes (15). For this secondary analysis (n = 42), patient data with nonsmoking severe asthma only were selected from the U-BIOPRED cohort. Participants had accessible raw data acquired using GC-MS for mass spectral analysis and generation of peak tables for the prespecified VOCs in the biomarker signature. Of these 42 participants, 12/42 had concurrent blood eosinophilia ⩾ 0.3 × 10⁹/L, and a subgroup (22/42) underwent sputum induction at the same visit as breath sampling; within this subgroup, 14/22 participants had concurrent sputum eosinophilia ⩾ 3%. A total of 40/42 participants in U-BIOPRED had concurrent Fe_NO measurements.

Sample Size

Sample size calculations for the biomarker discovery cohort are provided in online supplement section 3.1.

Volatile Biomarker Data Acquisition and Chemometric Analyses

Detailed methods for volatile data acquisition and chemometric analyses are provided in the online supplement (sections 2.1–2.3) and are reported in prior publications for EMBER (17, 18) and U-BIOPRED (15). A summary of the exhaled breath capture and storage and analytical methods for analysis and quantitation of VOCs is provided in Table 1.

Statistical Methods for Discovery and Validation Studies

Detailed statistical methods are summarized in Table 1 and provided in the online supplement (sections 3.2–3.5). We adopted a supervised machine learning method, as unsupervised principal components analysis of all 393 headspace features (Figure E4) did not separate clinically relevant groups. In brief, elastic net (eNET) regression models with cross-validation were used to identify a stable set of canonical VOCs (n = 19) in the discovery cohort associated with sputum eosinophilia ⩾ 3%. These VOCs were then aggregated into a nonweighted biomarker signature using the stable non-zero, eNET regression coefficients and z-transformed VOC concentrations of the biomarkers (Figure E5). The biomarker signature was then validated in EMBER and U-BIOPRED via identification of the same VOCs and the use of receiver operating characteristic (ROC) curves to examine the diagnostic accuracy for both blood (⩾0.30 × 10⁹/L) and sputum (⩾3%) eosinophilia within these respective cohorts. eNET regression models were used across all three cohorts to confirm the validity of biomarker diagnostic accuracy. eNET allows for grouped feature selection (i.e., the selection of features that form natural groups will be preserved); the eNET also performs variable selection (shrinkage to zero). We thus viewed the elastic net as a reasonable approach, given the correlation structure within the chromatographic feature matrix. For all eNETs, models were run 100 times with 10-fold cross-validation. eNETs were fitted to shuffled diagnostic labels to confirm that diagnostic predictions were not simply due to random noise and overfitting (Figure E5).

In the discovery cohort, eNET models were also run on the room air–acquired VOCs to establish the validity of the VOC model sputum headspace. In U-BIOPRED, sensitivity analyses were run with and without the VOC phenol, because of the potential for contamination of phenol from the Tedlar bag sampling. In both the discovery and U-BIOPRED cohort, breath biomarker signatures were compared with Fe_NO to establish the comparative diagnostic accuracy of Fe_NO and breath biomarker signatures. In U-BIOPRED, logistic regression analysis followed by ROC regression was used to evaluate the combined diagnostic accuracy of Fe_NO and the breath biomarker signature.

Results

Sputum Headspace Discovery Study

Figure 1 summarizes both the design of the sputum headspace volatile biomarker-based model discovery study to predict airway eosinophilia and the clinical validation studies in exhaled breath samples from EMBER and U-BIOPRED. Table 1 reports the clinical, breath sampling, and analytical procedures in all three cohorts, whereas Table 2 summarizes the clinical characteristics in the discovery cohort, and Table 3 shows the sputum differential cell count results in the discovery cohort. Sputum sample classification according to sputum eosinophil count threshold of 3% revealed 22 eosinophil-enriched sputa (⩾3%) and 14 non–eosinophil-enriched sputa (<3%). The comparison of the differential and total sputum eosinophil count (median, interquartile range) between the two groups showed, as expected, a significantly higher number of eosinophils in eosinophil-enriched sputa compared with the other groups (P < 0.0001; Table 3).

Table 3.

Total and Differential Cell Count of Sputum Samples from Patients with Severe Asthma in the Discovery Cohort

	Patients with Eosinophil-enriched Sputa* (n = 22)	Patients with Non–Eosinophil-enriched Sputa† (n = 14)	P Value
Total cells, ×106/g	3.02 (1.03–5.51)	2.86 (1.46–3.84)	0,88
Eosinophils, %	5.37 (3.00–8.75)	0.37 (0.25–0.93)	<0.0001
Eosinophils, ×106/g	0.09 (0.08–0.20)	0.012 (0.007–0.03)	<0.0001
Neutrophils, %	74.62 (65.01–83.37)	77.52 (64.43–91.95)	0.39
Neutrophils, ×106/g	1.82 (0.63–3.72)	1.84 (0.97–2.85)	0.88
Macrophages, %	14.33 (6.87–23.12)	11.37 (5.5–30.25)	0.73
Macrophages, ×106/g	0.40 (0.13–1.10)	0.27 (0.14–0.82)	0.66
Lymphocytes, %	0.39 (0.00–0.75)	0.87 (0.31–1.75)	0.14
Columnar epithelial cells, %	3.75 (0.76–5.50)	3.25 (1.31–6.68)	0.57
Squamous cells, %	1.91 (0.03–5.35)	5.35 (2.43–8.51)	0.08
Viability, %	69.86 (61.45–83.50)	68.55 (55.28–74.85)	0.76
Plug weight, mg	189.5 (106.5–264.25)	141.5 (57.75–213.75)	0.51

Data are summarized as median (Q1–Q3). A nonparametric test (Mann-Whitney) was performed for continuous variables.

^*Sputum eosinophils ⩾ 3%.

^† Sputum eosinophils < 3%.

TD-GC-MS analysis of 36 sputum headspace (volatiles emitted directly above the sputa) samples detected 393 features after thresholding based on frequency of observation and removal of siloxane artifacts. The eNET regression selected 19 features (VOCs). The selected VOCs are listed and grouped according to their chemical class in Table 4, which also reports the results of a literature search to check whether the selected VOCs had previously been identified as biomarkers of asthma or airway eosinophilia. The chemical structures of the 19 VOCs within the biomarker score are highlighted in Figure 2 and were compared with VOCs associated with sputum eosinophilia in previous GC-MS–based human exhaled breath volatile studies in asthma.

Table 4.

Chemical Classification and Prior Reports from Biomarker Studies of Discovery VOCs

Chemical Class	NIST Library Match	MSI Levels	CAS Registry Number	Summary of Literature Findings	Citation
Ketones	Acetone	1	67-64-1	Predictive of preclinical asthma in preschool wheezers in children Discriminated asthma with other atopic diseases in women of childbearing age	29 30
Aromatic hydrocarbon	Benzene	1	71-43-2	Predictive of asthma diagnosis in children	31
	Toluene	1	108-88-3	Associated with asthma symptoms in children with mild asthma	32
	Phenol	1	108-95-2	Prenatal exposure to phenols can promote asthma development May be a contaminant from Tedlar bag sampling of VOCs	33
	p-Xylene	1	106-42-3	Predictive of asthma diagnosis in children with asthma Predictive of asthma exacerbations in school-age children with asthma	31 34
	Styrene	1	100-42-5	Occupational exposure has been shown to elicit airway hyperresponsiveness and eosinophilic inflammation in the context of occupational asthma Observed in the headspace of epithelial cells in the context of hypoxia	35, 36
	α-Methyl styrene	1	98-83-9	Not reported in the context of asthma. Biomarker of precancerous gastric lesions and gastric cancer	37
	Benzothiazole	1	95-16-9	Not reported in the context of asthma. However, potential for development of benzothiazole-derived antiviral therapeutics based on robust in silico virtual and high-throughput screening data	38
Aldehydes	Benzaldehyde	1	100-52-7	No clear association in the context of asthma in exhaled breath. Oral ingestion in murine allergic asthma models demonstrated elevated number of eosinophils and neutrophils and Th2 cytokines in BAL fluid significantly decreased after the treatment	39
	Decanal	1	112-31-2	Reactive aldehyde species, including decanal, were dysregulated in in exhaled breath of adults with asthma exacerbations	18
	Nonanal	1	124-19-6	Detected in breath and discriminated neutrophilic from eosinophilic asthma, when classified by sputum cell counts Predictive of asthma exacerbation risk in children with asthma.	9 40
	Hexanal	1	66-25-1	Reactive aldehyde species, including hexanal, were dysregulated in exhaled breath of adults with severe asthma exacerbations	9
	2-Ethylhexanal	1	123-05-7	Predictive of asthma exacerbation risk in children with asthma	40
Alkanes	Decane	1	124-18-5	Together with other VOCs, discriminated exhaled breath of patients with atopic asthma compared with healthy control subjects and patients with allergic rhinitis	41
	Isothiocyanato-cyclohexane	1	1122-82-3	Predictive of respiratory disease in neonates	42
	Tridecane	1	629-50-5	Detected in the breath and associated with lung disease, including asthma and lung cancer, but is considered as a lung cancer biomarker	43 44
Alcohols	1-Hexanol	1	111-27-3	Associated with obstructive lung function and Cladosporium exposure in a general population home dwelling study	45
	2-Butoxy-ethanol	1	111-76-2	Discriminated headspace of activated eosinophil vs. activated neutrophils in ex vivo cell cultures of healthy donors without asthma or atopy	11
Others	Methylene chloride	1	75-09-2	Found ubiquitously in exhaled breath of children with mild asthma	32

Definition of abbreviations: CAS = Chemical Abstract Service; MSI = Metabolomics Standards Initiative; NA = not applicable; NIST = National Institute of Standard Technology; VOC = volatile organic compound.

Figure 2.

(A) An illustration of discovered eosinophilic breath volatile biomarkers, their chemical structures, and the metabolic pathways that could potentially be related to each chemical group. (B) Previously identified breath biomarkers in severe asthma studies characterizing patients according to sputum eosinophilia and arranged to show the possible chemical relationship and similarities with the reported discovery biomarkers in this study.

The normalized peak area value of each VOC was compared between eosinophil-enriched and non–eosinophil-enriched sputa, and significant differences between the groups were observed for 1-hexanol (P = 0.02), styrene (P = 0.017), phenol (P = 0.005), decane (P = 0.019), and benzothiazole (P = 0.03) (Figures 3A and E7). The majority of the other VOCs demonstrated higher normalized peak area values in the headspace of non–eosinophil-enriched sputa compared with headspace of eosinophil-enriched sputa.

Figure 3.

(A) Histogram of the normalized peak area values of the 19 discovered eosinophilic volatile organic compounds (VOCs) summarized as median and interquartile range and compared by Mann-Whitney test (*P < 0.05 and **P < 0.01). (B) Box plot of VOC biomarker score derived from elastic net regression in the eosinophil-enriched and non–eosinophil-enriched sputum samples (median, Q1, Q3, min and max) (Mann-Whitney test: **P < 0.0001). (C) Receiver operating characteristic (ROC) curve to evaluate the discriminatory performance of VOC scores in differentiating between eosinophil-enriched and non–eosinophil-enriched sputa. (D) Spearman’s correlation coefficient between VOC biomarker scores and the percentage of sputum eosinophils (r_s = −0.71; two-tailed t test: P < 0.0001). AUC = area under the ROC curve; CI = confidence interval; Fe_NO = fractional exhaled nitric oxide; GINA = Global Initiative for Asthma; ICS = inhaled corticosteroids; LABA =  long-acting β antagonist; mOCS = maintenance oral corticosteroids.

Biomarker signature generation is summarized with some worked examples in Figure S6 and in section 3.2 of the supplementary methods.

In Figure 3B, the biomarker signature values generated for each sputum headspace sample were compared between eosinophil-enriched and non–eosinophil-enriched sputa, showing a statistically significant difference (P < 0.01), with higher values for non–eosinophil-enriched sputum headspace samples compared with eosinophil-enriched ones. ROC curve analysis to estimate the ability of the VOC signature to discriminate between eosinophil-enriched and non–eosinophil-enriched sputa showed an area under the ROC curve (AUC) of 0.90 (95% confidence interval [CI], 0.80–0.99; P < 0.0001) and was substantially greater than either blood eosinophils (AUC, 0.61; 95% CI, 0.42–0.80) or Fe_NO (AUC, 0.61; 95% CI, 0.35–0.86) (Figure 3C). We observed moderate correlations between the VOC signature and sputum eosinophilia percentage (r_s = −0.71; P < 0.0001) (Figure 3D), as well as absolute sputum eosinophil counts (r_s = −0.54; P = 0.0006).

TD-GC-MS analysis of control background room air headspace samples was also performed; in this context, the AUROC (95% CI) was 0.5 (0.5–0.5). Further details are outlined in Figure E11a.

Clinical Validation Study for the Developed Volatile Biomarker Score

Demographics and clinical characteristics for the acute asthma EMBER and stable severe asthma U-BIOPRED cohorts are summarized Table 1 and detailed in in Tables E6–E9.

In the EMBER cohort, 16/19 VOCs within the biomarker signature were detected in exhaled breath sample GC × GC spectra; the missing VOCs were phenol, 2-butoxyethanol, and benzothiazole. In the U-BIOPRED cohort, 17/19 VOCs within the biomarker signature were detected in exhaled breath sample GC-MS spectra; the missing VOCs were 1-hexanol and 2-butoxyethanol.

The VOC signature demonstrated good diagnostic accuracies for sputum eosinophilia in both EMBER and U-BIOPRED exhaled breath cohorts (AUROCs, ∼0.90), with narrow confidence intervals for the given sample size, in addition to combined sensitivities and specificities of >0.80 (Figure 4 and Table 5). Models run on shuffled diagnostic labels in both EMBER and U-BIOPRED demonstrated significantly lower (P < 0.01) mean AUROCs (∼0.59), indicating that the models did not suffer from significant overfitting (Figure E11). In contrast and as expected for a VOC signature developed from sputum headspace, diagnostic accuracies for blood eosinophilia were modest and had wider confidence intervals (Figure 4 and Table 5).

Figure 4.

Visual summary of the diagnostic accuracy for both sputum and blood eosinophilia of eosinophil-associated volatile organic compounds (VOCs) in exhaled breath in the EMBER (East Midlands Breathomics Pathology Node) and U-BIOPRED (Unbiased Biomarkers in Prediction of Respiratory Disease Outcomes) cohorts. AUC = area under the receiver operating characteristic curve; CI = confidence interval.

Table 5.

Summary of the Diagnostic Accuracy of Breath Biomarker Scores in Identifying Eosinophilic Inflammation/Eosinophilic Cohorts in Replication Cohorts

	Target Patient Profile	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	AUROC (95% CI)
EMBER: Blood eosinophilia	Acute asthma exacerbation	0.31 (0.06–0.56)	0.90 (0.82–0.98)	0.44 (0.12–0.77)	0.84 (0.74–0.94)	0.59 (0.40–0.78)
EMBER: Sputum eosinophilia	Acute asthma exacerbation	0.89 (0.68–1.00)	0.80 (0.60–1.00)	0.73 (0.46–0.99)	0.92 (0.78–1.07)	0.89 (0.76–1.00)
U-BIOPRED: Blood eosinophilia	Stable severe asthma	0.75 (0.51–1.00)	0.77 (0.62–0.92)	0.56 (0.32–0.81)	0.89 (0.76–1.01)	0.79 (0.64–0.93)
U-BIOPRED: Sputum eosinophilia	Stable severe asthma	0.93 (0.79–1.00)	0.88 (0.65–1.00)	0.93 (0.80–1.10)	0.88 (0.65–1.10)	0.90 (0.75–1.00)

Definition of abbreviations: AUROC = area under receiver operating characteristic curve; CI = confidence interval; EMBER = East Midlands Breathomics Pathology Node; NPV = negative predictive value; PPV = positive predictive value; U-BIOPRED = Unbiased Biomarkers in Prediction of Respiratory Disease Outcomes.

We did not find any correlation between blood c-reactive protein (CRP) level (n = 6/63 with an available CRP had a CRP > 50 mg/L) or individual VOC concentration/eNET signature values in EMBER. Furthermore, analysis of rhinovirus-19 using quantitative PCR (online supplement, section 4.7) in the sputum of the EMBER cohort (n = 17/24 with available sputum for PCR analyses), identified only weak but statistically significant unadjusted inverse correlations between tridecane (r_s = −0.55; P = 0.02) and 1-hexanol (r_s = −0.52; P = 0.02).

Further analyses of the U-BIOPRED cohort (Figure 5) identified that two VOCs (toluene and benzaldehyde) within the multimarker signature were numerically and significantly lower in exhaled breath in the patients with eosinophilia compared with the patients without eosinophilia (P < 0.05, unadjusted) (Figure 5A). In addition, the concentrations of several VOCs, in particular the reactive aldehyde species (RASPs) (nonanal, decanal, hexanal, and benzaldehyde) correlated significantly with the eNET score values for the multimarker VOC signature (Figure 5B), indicative of dysregulated RASPs in eosinophilic asthma. VOC biomarker score values were significantly higher in the patients with eosinophilia compared with the patients without eosinophilia (Figure 5C). Comparative analyses using ROC logistic regression evaluating the diagnostic accuracy of Fe_NO alone or Fe_NO in combination with the breath biomarker signature indicated that Fe_NO did not provide any added accuracy to the exhaled VOC signature and had modest diagnostic accuracy alone (AUROC, 0.62; 95% CI, 0.33–0.90) (Figure 5D). Diagnostic accuracy for the VOC score in predicting sputum eosinophilia when phenol (a potential Tedlar bag contaminant) was removed from the multimarker data was unaffected (AUROC, 0.90; 95% CI, 0.75–1.0).

Figure 5.

(A) Ln (x + 1) transformed exhaled volatile organic compound (VOC) biomarker concentrations (y-axis), exhaled VOC (eVOC), in patients with eosinophilia (red) and without eosinophilia (blue) U-BIOPRED (Unbiased Biomarkers in Prediction of Respiratory Disease Outcomes) characterized by sputum eosinophilia. *P < 0.05 (unadjusted) when comparing patients with or without eosinophilia. (B) Correlation (Pearson’s r) heatmap for the exhaled VOCs that individually constitute the eVOC score. *P < 0.05 for the correlation with the eVOC biomarker score. (C) Elastic net (eNET) biomarker score values for patients with or without eosinophilia according to sputum eosinophilia in U-BIOPRED. **P < 0.001. (D) Receiver operating characteristic (ROC) curves for classification of severe eosinophilic asthma, according to sputum eosinophilia, with fractional exhaled nitric oxide (Fe_NO) (orange), eVOC score (green), and combination of Fe_NO and eVOC score (blue). Area under ROC curve (AUROC) for eVOC score is presented in green. SpEos = sputum eosinophil count (%).

Discussion

This is the first report to identify a noninvasive breath biomarker signature of eosinophilic airways disease in asthma using exhaled breath volatilomics. Our study used robust and previously validated discovery and replication methods for headspace volatile detection and quantification (17, 18). Breath biomarker signature values were directly correlated with both the percentage and total eosinophil counts in sputum. In addition, we have identified that the breath biomarker signatures discovered in sputum headspace were predictive of eosinophilic asthma exacerbations identified by sputum in the EMBER cohort (14, 18) and of severe eosinophilic asthma in the U-BIOPRED cohort (20).

Our findings are important and indicate that the VOC biomarkers are associated with sputum eosinophilia via robust discovery and external validation in two independent disease cohorts, underscoring their validity as biomarkers of eosinophilic inflammation. In addition, discovery of the biomarker score in a cohort with severe asthma with eosinophilia in both the presence and absence of anti–IL-5 therapy raises the possibility that future studies could evaluate the utility of the VOC biomarker signature for monitoring mucosal eosinophilia in patients receiving biologics or potentially for stratifying biologic response; however, these assertions would need to be tested prospectively in appropriately designed studies.

Diagnostic accuracies for sputum eosinophilia and combined diagnostic sensitivity and specificity were high, much more so than for blood eosinophilia, as would be expected for a biomarker developed from sputum headspace. Furthermore, in the discovery cohort and U-BIOPRED, the VOC signature substantially outperformed Fe_NO. Furthermore, in the EMBER cohort, we found only small but statistically significant correlations between rhinovirus-16 viral load in sputum and two of the VOCs in our marker score (tridecane and 1-hexanol) and no correlations between the VOCs and CRP as a marker of bacterial exacerbations. These observations would suggest that our biomarker score is specific for eosinophilic inflammation, rather than any specific etiotype of asthma exacerbation and may reflect the level of tissue eosinophilia; however, further research is warranted in this area to confirm and follow up on these findings.

Given the challenges of sputum induction and processing and obtaining Fe_NO in severe exacerbations of asthma, the identification of a canonical set of VOCs associated with sputum eosinophilia in both headspace and breath could pave the way for rapid point-of-care triage of eosinophilic disease in the future using point-of-care ion mobility or, potentially, printed and disposable colorimetric VOC detection arrays (21), as well as other types of sensors, such as electronic noses (22), specifically trained to identify the VOCs in this study.

The eosinophil-associated VOCs identified in sputum headspace samples were mostly represented by aromatic hydrocarbons, aldehydes, and alkanes and a small number of alcohols and ketones. Our previous systematic review of volatile biomarkers of type-2 inflammation in asthma and COPD (8) highlighted the potential importance of aldehydes and hydrocarbons; however, it also highlighted the variability of associated biomarkers across the literature and, in addition, methodological reporting of VOC biomarkers. Intriguingly, we identified several biomarkers that were derivatives of previously reported VOC biomarkers of eosinophilic inflammation. Specifically, oxygenation, hydroxylation, or hydrogenation of several of our reported biomarkers yielded biomarkers previously associated with sputum eosinophilia in adults with moderate to severe asthma (9, 10). For example, the eosinophil-associated VOC hexane, reported by Schleich and colleagues (9), is rapidly transformed to 1-hexanol by hydroxylation and subsequently hexanal through oxidation via alcohol dehydrogenase enzymes. Both of these VOCs were reported as eosinophil-associated VOCs in our discovery analyses. These observations potentially highlight the importance of alcohol dehydrogenase–based oxidation (i.e., of 1-hexanol to hexanal) in the context of eosinophilic airway diseases.

RASPs were the second most represented category among the detected volatile biomarkers of airway eosinophilia. Oxidative stress generates reactive oxygen species, which trigger lipid peroxidation (23), with the consequent production of several RASPs. Exhaled concentrations of hexanal have been reported as predictive of asthma exacerbations in childhood (24). In keeping with these prior observations, we identified several reactive aldehyde species in U-BIOPRED that correlated highly with the exhaled VOC signature. Given the potential for RASP inhibition (25) using oral small-molecule inhibitors, it is plausible that exhaled VOCs could be used as a biomarker of target engagement with RASP inhibitors.

In keeping with previous observations in cell culture studies of metabolically active cells (26), we observed that several VOCs were depleted in the headspace of eosinophilic sputa. For example, the aromatic hydrocarbon styrene was depleted in the headspace of eosinophilic sputa, relative to noneosinophilic sputum, while also being enriched in the headspace of sputum relative to background. Styrene is a volatile monomer widely used in the production of polymers and reinforced plastics; however, the relative depletion in eosinophilic sputum and enrichment in headspace compared with background would argue against external contamination as a causative factor in our data. Microbial and fungal degradation of plastic compounds containing styrene is well reported across a host of pathogens that could plausibly be resident within eosinophilic sputum samples (27).

Our research has several limitations. First, the derived biomarker signature of eosinophilic inflammation in sputum headspace was derived from spontaneous sputum samples and not samples obtained by induction. Consequently, our discoveries may be more applicable to the upper and larger airways, as opposed to lower airway eosinophils acquired by sputum induction of bronchoalveolar lavage analysis, although a previous report failed to identify relevant differences in sputum cellularity between induced and spontaneous samples, with the exception of better cell viability in induced sputa (28). Despite this limitation, the biomarker signature that we derived was validated in exhaled breath against the designated target in two independent populations characterized by eosinophilic exacerbations and stable severe eosinophilic asthma, underscoring the potential validity for future research. In addition, our replication cohorts were of small sample sizes, with diagnostic accuracy studies in subgroups of patients characterized by sputum eosinophilia, indicating the need for further research validating the VOCs identified as eosinophilic airway biomarkers.

In conclusion, we have discovered in tissue headspace and replicated in exhaled breath a biomarker signature of tissue eosinophil-associated VOCs, several of which could be plausibly linked to eosinophilic airway inflammation and have been reported in previous discovery VOC studies as derivatives of the biomarkers reported here. Future studies should target quantification and detection of the VOCs described in our report using point-of-care sensors, such as printed colorimetric arrays or electronic noses, in well-powered studies designed to develop point-of-care tests for eosinophilic airway inflammation in asthma.