VOLATILE FINGERPRINTING OF PSEUDOMONAS AERUGINOSA AND RESPIRATORY SYNCYTIAL VIRUS INFECTION IN AN IN VITRO CYSTIC FIBROSIS CO-INFECTION MODEL

Abstract

Volatile molecules in exhaled breath represent potential biomarkers in the setting of infectious diseases, particularly those affecting the respiratory tract. In particular, Pseudomonas aeruginosa is a critically-important respiratory pathogen in specific subsets of the population, such as those with cystic fibrosis. Infections caused by P. aeruginosa can be particularly problematic when co-infection with respiratory syncytial virus (RSV) occurs, as this is correlated with the establishment of chronic P. aeruginosa infection. In the present study, we evaluate the volatile metabolites produced by P. Aeruginosa (PAO1)-infected, RSV-infected, co-infected, or uninfected cystic fibrosis bronchial epithelial (CFBE) cells, in vitro. We identified a volatile metabolic signature that could discriminate between P. aeruginosa-infected and non-P. aeruginosa-infected CFBE with an area under the receiver operating characteristic curve (AUROC) of 0.850, using the machine learning algorithm Random Forest (RF). Although we could not discriminate between RSV-infected and non-RSV-infected CFBE (AUROC = 0.431), we note that sample classification probabilities for RSV-infected cell, generated using RF, were between those of uninfected CFBE and P. aeruginosa-infected CFBE, suggesting that RSV infection may result in a volatile metabolic profile that shares attributes with both of these groups. To more precisely elucidate the biological origins of the volatile metabolites that were discriminatory between P. aeruginosa-infected and non-P. aeruginosa-infected CFBE, we measured the volatile metabolites produced by P. aeruginosa grown in the absence of CFBE. Our findings suggest that the discriminatory metabolites produced likely result from the interaction of P. aeruginosa with the CFBE cells, rather than the metabolism of media components by the bacterium. Taken together, our findings support the notion that P. aeruginosa interacting with CFBE yields a particular volatile metabolic signature. Such a signature may have clinical utility in the monitoring of individuals with cystic fibrosis.

1. Introduction

Respiratory tract infections are among the leading causes of death worldwide, resulting in 3.2 million deaths in 2015 alone [1], and pneumonia remains one of the world’s leading causes of death for children under the age of five [2]. According to the Etiology of Pneumonia in the Community (EPIC) study conducted by Centers for Disease Control and Prevention (CDC), human rhinovirus (9%), influenza virus (6%), and Streptococcus pneumoniae (5%) are the most prevalent pathogenic agents in the adult population, while respiratory syncytial virus (28%) and human rhinovirus (27%) are the predominant pathogens in children [3,4]. In this study, polymicrobial infections (i.e., infections caused by more than one pathogen) were identified in 36% of adults and 81% of children, with the occurrence of polymicrobial infections resulting in substantially worse morbidity and mortality [5–8]. Of particular interest are viral-bacterial polymicrobial infections, as viral infections have been shown to promote bacterial colonization of the airway up to several weeks following initial infection [9–13].

Polymicrobial infections involving influenza virus and S. pneumoniae are commonly encountered in the human population [10]. However, other viral-bacterial interactions may occur in specific subsets of the population, such as in the setting of cystic fibrosis (CF). In CF, a complex and dynamic respiratory microbiota interacts and evolves in a way that is not fully characterized to-date [14]. In this population, co-infection of respiratory viruses and Pseudomonas aeruginosa is correlated with the establishment and exacerbation of chronic P. aeruginosa lung infection [15][16–20], with respiratory syncytial virus (RSV) as the etiological agent in 9–58% of viral infections in these patients [21]. Chronic P. aeruginosa infection has dire health consequences in these individuals, including decreased quality of life and reduced life expectancy [22]. It has been demonstrated that the antiviral interferon host response to RSV, along with the release of transferrin-bound iron that occurs during viral infection, promotes P. aeruginosa biofilm formation, and this, in turn, can be correlated with disease exacerbation [23].

Volatile molecules have been investigated as potential biomarkers for the diagnosis of both acute and chronic respiratory infections [24]. Numerous studies have focused on the volatile metabolic fingerprints generated by bacterial infections in vivo [25–31], but none have evaluated the ability of this approach to discriminate between virally- and bacterially-derived volatile compounds, nor deconvolute bacterial- and viral-associated signatures in the setting of a polymicrobial environment. Nevertheless, several transcriptomics studies have demonstrated that viral and bacterial infections induce distinct responses by host cells [32–42]. Such findings provide the rationale behind our hypothesis that volatile metabolic signatures can be used to differentiate between virally-infected, bacterially-infected, and co-infected cells. The aim of this study is thus to investigate the ability of volatile metabolites to discriminate between respiratory epithelium infected, in vitro, with P. aeruginosa and RSV singularly, and to evaluate the polymicrobial fingerprint generated by the infection of P. aeruginosa and RSV simultaneously. For this, we employ solid-phase microextraction (SPME) followed by comprehensive two-dimensional gas chromatography (GC×GC) with a time-of-flight mass spectrometer (ToF MS), a powerful technique for the analysis of complex mixtures.

2. Materials and Methods

2.1. Viral and bacterial infection sample preparation

Cell culture: Immortalized homozygous CFTR ΔF508 CFBE41o- human bronchial epithelial cells (CFBE) were maintained in a humidified incubator at 37°C 5% CO2 in Minimum Essential Medium (MEM) containing phenol red (Gibco, Waltham, MA, United States) supplemented with 10% fetal bovine serum (FBS, Gemini Bio-Products, West Sacramento, CA, United States), 0.5 μg/mL Plasmocin prophylactic (InvivoGen, San Diego, CA, United States), 2 mM L-glutamine, 5 U/mL penicillin, and 5 μg/mL streptomycin (Sigma-Aldrich, St. Louis, MO, United States) (Hendricks PNAS 2016, 113:1642–1647). Immortalized STATE^−/− NY3.2 mouse fibroblasts were maintained in MEM containing phenol red supplemented with 5% FBS, 5 U/mL penicillin, and 5 μg/mL streptomycin. Both cell types are not on the commonly misidentified list, and cells were tested quarterly for mycoplasma using a Southern Biotech mycoplasma detection kit. CFBE41o- cells were seeded at near confluency on 12 mm diameter, 0.4 μm pore size Transwell permeable polycarbonate membrane supports (Costar, St. Louis, MO, United States). After attachment and confluency, epithelial cells were differentiated at air-liquid interface (ALI) for one week. Respiratory syncytial virus (RSV) line A2 was propagated in NY3.2 cells. RSV was extracted and stored in high salt media (HSM) at −80°C. Plaque-forming units were determined by immunoreactive staining assay in NY3.2 cells. P. aeruginosa strain PAOl was grown in Lysogeny Broth (LB) at 37°C on a roller.

P. aeruginosa-only samples: Bacteria were prepared by washing an overnight culture in MEM supplemented with 2 mM L-glutamine and culturing in MEM supplemented with 2 mM L-glutamine in a 25 cm² cell culture flask (Coming, Coming, NY, United States) at 37°C 5% CO2 for 24 h. Media was then transferred to a tube and centrifuged to remove bacterial aggregates. Clarified media was filtered through a 0.22 μm filter (Merck Millipore, Burlington, MA, United States) and stored at −80°C. The same set up was used without P. aeruginosa inoculation for media-only controls.

Bacterial-epithelial co-cultures: ALI-differentiated CFBE41o- cells were washed with MEM lacking phenol red (Gibco) supplemented with 2 mM L-glutamine, and basolateral media was replaced with MEM containing phenol red supplemented with 10% FBS and 2 mM L-glutamine. Cells were inoculated with RSV at a multiplicity of infection (MOI) of 1 in MEM supplemented with 2 mM L-glutamine in the apical compartment. Control cells were inoculated with MEM containing an equal volume of vehicle (HSM). Apical media was removed after infection and cells were returned to ALI and maintained in antibiotic-free basolateral media. RSV-infected cells were then inoculated with P. aeruginosa that was prewashed in MEM lacking phenol red supplemented with 2 mM L-glutamine at an MOI of approximately 25. After 1 h attachment, non-attached bacteria were removed and apical media was adjusted to 0.4% L-arginine. After an additional 5 h, apical media was transferred to a tube and epithelium-associated biofilms were disrupted in 0.1% Triton X-100 (Bio-Rad, Hercules, CA, United States) and analyzed by serial dilution and plating on LB agar for colony-forming units. Apical media was centrifuged to remove bacterial aggregates, passed through a 0.22 μm filter, and stored at −80°C. RNA was collected from a separate set of Transwells by RNeasy with QIAshredder columns (Qiagen, Hilden, Germany) according to manufacturer recommendations. Transcript levels for GAPDH, the RSV N gene, and IFNL1 were assessed by converting RNA into cDNA with iScript cDNA Synthesis Kit and performing RT-qPCR using iQ SYBR Green Supermix (Bio-Rad) on a StepOne Real-Time PCR System (ThermoFisher Scientific, Waltham, MA, United States). Primers for GAPDH (5’- CGACCACTTTGTCAAGCTCA-3’, 5’-AGGGGAGATTCAGTGTGGTG-3’), for RSV N gene (5’-CGCCTTGGAAGAGTCACTCA-3′, 5’-GAAGCCTCAGGTCCCAATTC-3’), and for IFNL1 (5‘-GCTCTTAGCAAAGTCAAGTTGAATGA-3‘, 5‘-TGCTCCGTTGGATGGTGTATT-3‘). Conditions analyzed were epithelial cells alone, RSV-infected epithelial cells, P. aeruginosa-infected epithelial cells, and co-infected epithelial cells.

2.2. Sample preparation for volatile metabolites analysis

Five-hundred μL of filtered apical media was collected into a 10 mL air-tight glass vial sealed with a PTFE/silicone cap (Sigma-Aldrich) frozen at −20°C until analysis (within one month of collection). Samples were incubated for 15 min (the equilibration phase) at 37°C before fiber exposure for 45 min at the same temperature (the extraction phase). Samples were agitated at 250 rpm during both the equilibration and extraction phases. A divinylbenzene/carboxen/polydimethylsiloxane (DVB/CAR/PDMS) d_f 50/30 μm, 2 cm length fiber (Supelco, Bellefonte, PA, USA) was used to pre-concentrate the volatile molecules. The fiber was then introduced into the GC injector for thermal desorption for 1 min at 250°C in splitless mode. The fiber was conditioned before use and baked-out after each run for 7 minutes at 250°C. Absence of carryover was verified by desorbing the fiber into the GC injector after the baking time.

2.3. Analytical instrumentation

The analyses of volatile molecules were carried out using a Pegasus 4D (LECO Corp., St. Joseph, MI, United States) GC×GC time-of-flight (TOF) MS instrument with an Agilent 6890 GC, and equipped with an MPS autosampler (Gerstel, Linthicum Heights, MD, United States). The first dimension column was an Rxi-624Sil (60 m × 250 μm × 1.4 μm (length × internal diameter × film thickness)) connected in series with a Stabilwax secondary column (1 m × 250 μm × 1.4 μm), both from Restek (Bellefonte, PA, USA). The carrier gas was helium, at a flow rate of 2 mL/min. The primary oven temperature program was 35°C (hold 1 min) ramped to 230°C at a rate of 3.5°C/min. The secondary oven and the thermal modulator were offset from the primary oven by +5°C and +25°C, respectively. A modulation period of 2.0 s (alternating 0.5 s hot and 0.5 s cold pulses) was used. The transfer line temperature was set at 250°C. A mass range of m/z 30 to 500 was collected at a rate of 200 spectra/s following a 2.5 min acquisition delay. The ion source was maintained at 200°C. Data acquisition and analysis were performed using ChromaTOF software, version 4.50 (LECO Corp).

2.4. Processing and analysis of chromatographic data

Chromatographic data were processed and aligned using ChromaTOF. A signal-to-noise (S/N) cutoff was set at 50:1 in at least one chromatogram and a minimum of 20:1 S/N ratio in all others. For the alignment of peaks across chromatograms, maximum first and second-dimension retention time deviations were set at 6 s and 0.2 s, respectively, and the inter-chromatogram spectral match threshold was set at 600 (out of 1000). Compounds eluting prior to 4 min and artifacts were removed prior to statistical analysis. Artifacts were identified by running a series of instrument blanks (i.e., runs without SPME fiber injections) and fiber blanks (i.e., runs with SPME fiber injections but without fiber headspace exposure) to identify the compounds deriving from column and fiber bleeding (e.g., siloxane, silanol, etc.). Furthermore, the most common and ubiquitous plasticizers (i.e. phthalates) were removed as well.

A mixture of normal alkanes (C₆-C₂₀) was analyzed every 20 runs to calculate the linear retention index [43] and evaluate the instrument performance. The SPME and GC methods are as described above, except for the SPME exposition time, which was shorter (5 min) to avoid excessive overload of the fiber.

In agreement with the minimum reported standard guidelines suggested by the Metabolomics Standard Initiative, the most discriminatory features were putative annotated (Level 2) or putatively assigned to a chemical class (Level 3) [44] based on mass spectral similarities to the NIST 2011 mass spectral library, with a match score ≥ 800 (out of 1000) required for putative identifications. In addition to the MS similarity, the putative identification was also supported by the linear retention index (LRI) in agreement (i.e., in the ±15 range), with data reported using the same stationary phase [45] (if available). Unless confirmed by the LRI value, hydrocarbons were generally assigned as “alkylated hydrocarbons”, as only the chemical class of these compounds can be assigned considering their mass spectral fragmentation pattern and their location in the two-dimensional chromatogram.

2.5. Statistical analysis

All statistical analyses were performed using R v3.3.2 (R Foundation for Statistical Computing, Vienna, Austria). Prior to statistical analyses, the chromatographic area of compounds across chromatograms was normalized using Probabilistic Quotient Normalization [46], Variables were log-transformed, mean-centered, and unit-scaled prior to all statistical analyses. Random forest (RF) [47] was used to identify the most highly discriminatory volatile molecules and predict the class to which a leave-one-out cross validation (LOOCV) sample belonged. LOOCV was repeated 31 times, leaving out each sample (e.g., 8 PA⁺/RSV⁺, 8 PA⁻/RSV⁺, 8 PA⁺/RSV⁻, and 7 PA⁻/RSV⁻) once. The class probabilities were used to generate receiver operating characteristic (ROC) curves, and from these ROC curves, sensitivities, specificities, and area under the ROC curve (AUROC) were calculated. The optimal thresholds for class probabilities were calculated using Youden’s J statistic [48], rather than the 0.5 cutoff that is traditionally applied to two-class classification problems. Mean decrease in accuracy (MDA) was used as the measure of variable importance. The Mann-Whitney U test [49] was used to compare the normalized area of volatile molecules between experimental groups (Table 1), while a Student’s t-test was used to compare class probabilities between groups.

Table 1.

List of volatile metabolites significantly different between experimental groups, along with their putative identification. The direction of the arrow indicate the group in which the compound was significantly more abundant, “↑” indicates that the compound was significantly more abundant in the first group listed in the column heading, and “↓” indicates that the compound was significantly more abundant in the second group.

#	Compound	Class	Formula	CAS	Similarity	LRIEXP	LRILit	1tR (min:s)	2tR (S)	Reference	PA−/RSV−vs. PA+/RSV−	PA−/RSV− vs. PA−/RSV+	PA−/RSV− vs. PA+/RSV+	PA+/RSV− vs. PA−/RSV+	PA+/RSV− vs. PA+/RSV+	PA−/RSV+ vs. PA+/RSV+
1	2-Propenal	Ald	C3H4O	107–02-8	855	538	523	6:00	1.0		↓		↓
2*	Acetone	Ket	C3H6O	67–64-1	978	545	530	6:15	0.8	51,60	↓	↓	↓	↑		↓
3	2-methyl-pentane	Hyd	C6H14	107–83-5	941	570	564	7:12	0.6	58	↑	↑
4	Unknown	Aid				595		8:08	0.7		↓
5	Alkylated hydrocarbons	Hyd				627		9:27	0.7			↑
6	Butanal	Aid	C4H8O	123–72-8	863	627	629	9:28	0.8	51,57,60	↓
7*	2-Butanone	Ket	C4H8O	78–93-3	965	629	637	9:32	1.00	51–57,59,60	↓		↓			↓
8	Unknown	S-Com				650		10:28	1.5		↑
9	Tetrahydrofuran	Het-Cyc	C4H8O	109–99-9		651	655	10:31	0.7	58	↓	↓
10*	3-methylbutanal	Aid	C5H10O	590–86-3	924	693	694	12:19	0.8	54–56	↓			↑		↓
11*	2-methylbutanal	Aid	C5H10O	96–17-3	850	701	-	12:41	0.7	54,55,57	↓			↑
12	Alkylated hydrocarbons	Hyd				753		15:24	0.6				↑
13	Alkylated hydrocarbons	Hyd				760		15:46	0.6				↑
14	Alkylated hydrocarbons	Hyd				767		16:06	0.6		↑	↑	↑
15	4-methyl-2-pentanone	Ket	C6H12O	108–10-1	850	780	783	16:46	1.00
16	Alkylated hydrocarbons	Hyd				791		17:21	0.8			↓
17	Alkylated hydrocarbons	Hyd				822		19:02	0.6		↑	↑	↑
18	Alkylated hydrocarbons	Hyd				846		20:17	0.7		↑	↑	↑
19	Alkylated hydrocarbons	Hyd				858		20:56	0.6		↑	↑	↑
20	2-methyl-octane	Hyd	C9H20	3221–61-2	862	864	873	21:14	0.6	58	↑	↑	↑
21	Alkylated hydrocarbons	Hyd				874		21:48	0.6		↑	↑	↑
22	Cyclohexanone	Ket	C6H10O	108–94-1	910	955	957	25:58	0.9		↓				↑
23	Alkylated hydrocarbons	Hyd				977		27:05	0.6		↑	↑
24*	2-ethylhexanal	Ald	C8H16	123–05-7	918	993	-	27:54	0.8	29						↓
25	Alkylated hydrocarbons	Hyd				1088		32:35	0.6		↓
26	Alkylated hydrocarbons	Hyd				1091		32:44	0.6					↓
27	Alkylated hydrocarbons	Hyd				1097		33:02	0.6		↑	↑	↑
28*	Undecane	Hyd	C11H24	1120–21-4	872	1102	1100	33:14	0.7	56	↑
29*	Alkylated hydrocarbons	Hyd				1115		33:50	0.7		↑
30	Alkylated hydrocarbons	Hyd				1121		34:08	0.9		↓	↓
31	Unknown	Ket				1137		34:50	1.3		↓	↓	↓
32	Alkylated hydrocarbons	Hyd				1159		35:51	0.7		↓
33*	Alkylated hydrocarbons	Hyd				1175		36:33	0.6		↓			↑
34	Alkylated hydrocarbons	Hyd				1214		38:17	0.6					↑		↓
35	Alkylated hydrocarbons	Hyd				1294		41:42	0.6		↑	↑
36	Alkylated hydrocarbons	Hyd				1301		42:02	0.6		↑	↑
37	Alkylated hydrocarbons	Hyd				1319		42:44	0.6		↓
38	Alkylated hydrocarbons	Hyd				1325		42:58	0.6		↑	↑
39	Alkylated hydrocarbons	Hyd				1335		43:21	0.7		↑
40*	Alkylated hydrocarbons	Hyd				1340		43:34	0.6		↑
41	Alkylated hydrocarbons	Hyd				1496		49:32	0.6		↑
42	Alkylated hydrocarbons	Hyd				1541		51:08	0.6
43	Alkylated hydrocarbons	Hyd				1567		52:04	0.7		↑

^*MDA>0.01 for the comparison of P. aeruginosa-infected versus non-P. aeruginosa-infected cells

3. Results and Discussion

3.1. Respiratory syncytial virus and P. aeruginosa infections

Infections with RSV and/or P. aeruginosa were confirmed by RT-qPCR and dilution plating, respectively (Figure 1). RSV transcripts were readily detected in both RSV-infected and co-infected respiratory epithelium (Figure 1A), indicating that a productive RSV infection was achieved. In addition to RSV transcript detection, we also observed a strong antiviral response by the RSV-infected and co-infected respiratory epithelium, evidenced by upregulation of the gene encoding IFN-λΙ (IFNL1, Figure 1B). Similar to previous studies [23,50], RSV co-infection of the respiratory epithelium enhanced P. aeruginosa biofilm biogenesis, as compared to P. aeruginosa alone (Figure 1C). Together, these results indicate that the samples processed for volatile fingerprinting were suitable to serve as representative uninfected (PA⁻/RSV⁻), P. aeruginosa- infected (PA⁺/RSV⁻), RSV-infected (PA⁻/RSV⁺), and co-infected (PA⁺/RSV⁺) respiratory epithelium.

Figure 1.

RSV and P. aeruginosa infection of CF respiratory epithelial cells. A) ΔC_T values for RT-qPCR of infected CFBE cells with RSV N gene primers, normalized to GAPDH. Y axis inverted to represent that lower numbers (including more negative) indicate greater abundance of RNA. Data represent the mean and standard deviation, statistical comparison by paired ANOVA with Sidak correction for multiple comparisons. B) Fold change in IFNL1 expression for infected CFBE cells as determined by RT-qPCR, normalized to GAPDH and compared to uninfected control cells. Data represent the mean, statistical comparison to uninfected control by paired ANOVA with Dunnett correction for multiple comparisons. C) CFU of P. aeruginosa growing in mucosal biofilms associated with infected CFBE cells. Data represent the mean, statistical comparison by paired t-test. *: p < 0.05, ***: p < 0.001, ****: p < 0.0001.

3.2. Volatile fingerprints of P. aeruginosa and respiratory syncytial virus infections

Raw chromatographic data consisting of 762 features across the 31 samples of interest (8 PA⁺/RSV⁺, 8 PA⁻/RSV⁺, 8 PA⁺/RSV⁻, and 7 PA⁻/RSV⁻) were pre-processed to remove artifacts, resulting in a final data matrix consisting of 230 compounds plausibly derived from the biological specimens themselves.

We first sought to determine whether there was a subset of volatile metabolites that were significantly different in abundance between experimental groups. Forty-three features were identified as significantly different in at least one comparison and are reported in Table 1, along with their putative identification (if possible). The abundance of each feature in the different pairwise comparisons is reported in Table 1. Arrow directions indicate the group in which the compound was more abundant, with “↑” indicating that the compound was significantly more abundant in the first group listed in the column heading, and “↓,” indicating that the compound was significantly more abundant in the second group. Thirty-four of these 43 features were significantly different between PA⁺/RSV⁻ and PA⁻/RSV⁻, 19 between PA⁻/RSV⁺ and PA⁻/RSV⁻, and 16 between PA⁺/RSV⁺ and PA⁻/RSV⁻, of which only three (all alkylated hydrocarbons) were not overlapped with the features reported from the comparison between PA⁺/RSV⁻ and PA⁻/RSV⁻. Only six features were significantly different between PA⁺/RSV⁻ and PA⁻/RSV⁺, all more expressed in the former group, except one alkylated hydrocarbon. Notably, only one feature was significantly more expressed in PA⁺/RSV⁻ versus PA⁺/RSV⁺ (cyclohexanone), whereas five were significantly different in the comparison of PA⁻/RSV⁺ versus PA⁺/RSV⁺, all more expressed in the latter group Seven of the putatively identified metabolites have been previously identified in the headspace of P. aeruginosa grown under in vitro conditions [51–57], three metabolites were previously identified in the headspace of cell line infected with human respiratory viruses [58], and one, 2-ethylhexanal, has been detected in the headspace of several cancer cell cultures [28].

3.3. Discriminatory ability of the volatile metabolite signature

We next sought to determine whether volatile metabolic fingerprints could discriminate between infected and uninfected cystic fibrosis bronchial epithelial cells (CFBE). Random forest (RF) was used to assess the capability of volatile fingerprints to discriminate between: I) P. aeruginosa-infected (including both PA⁺/RSV⁻ and PA⁺/RSV⁺) versus non-P. aeruginosa-infected (PA⁻/RSV⁻ and PA⁻/RSV⁺) cells, and II) RSV-infected (including both PA⁻/RSV⁺ and PA⁺/RSV⁺) versus non-RSV-infected (PA⁺/RSV⁻ and PA⁻/RSV⁻) cells. Receiver operating characteristic (ROC) curves were generated using class probabilities for leave-one-out cross-validation (LOOCV) samples (Figure 2).

Figure 2.

Receiver operating characteristic (ROC) curves for the discrimination between: A) P. aeruginosa-infected (PA⁺/RSV⁻ and PA⁺/RSV⁺) versus non-P. aeruginosa-infected (PA⁻/RSV⁻ and PA⁻/RSV⁺) cells, and B) RSV-infected (PA⁻/RSV⁺ and PA⁺/RSV⁺) versus non-RSV-infected (PA⁻/RSV⁻ and PA⁺/RSV⁻), generated using Random Forest. *Optimal sensitivity and specificity of each statistical model are calculated at the given class prediction cutoff (ranging from 0 to 1).

The results obtained from the first model (P. aeruginosa versus non-P. aeruginosa) yielded an AUROC of 0.850, with a sensitivity of 0.938 and specificity of 0.800 at an optimal class probability cutoff of 0.544 (2a, left). In contrast, the ability to discriminate between RSV-infected and non-RSV-infected cells was approximately random, with an AUROC of 0.431, and optimal sensitivity and specificity of 0.500 and 0.533, respectively (2b, right). Taken together, these findings suggest that under the experimental conditions employed in the present study, infection of CFBE with P. aeruginosa produces a detectable volatile molecular fingerprint, while infection of CFBE with RSV does not.

The class probabilities for each sample in the comparison of P. aeruginosa versus non-P. aeruginosa infected CFBE were plotted, with 0 indicating non-P. aeruginosa-infected, and 1 indicating P. aeruginosa-infected (Figure 3). Within the P. aeruginosa-infected group, there was no significant difference in class probabilities between PA⁺/RSV⁻ and PA⁺/RSV⁺ (p = 0.71). It is interesting, however, that the class probabilities for PA⁻/RSV⁺ were significantly different from both PA⁻/RSV⁻ (p = 0.03) as well as PA⁺/RSV⁻ (p = 0.02). This finding suggests that RSV singular infection results in a volatile molecular fingerprint that shares attributes with both PA⁻/RSV⁻ as well as PA⁺/RSV⁻, possibly resultant from a general host response to infection. In contrast to the comparison of P. aeruginosa-infected versus non-P. aeruginosa-infected cells, there were no significant differences in class probabilities between groups for the comparison of RSV-infected versus non-RSV-infected cells (Supplementary Figure S1).

Figure 3.

Class probabilities for LOOCV samples (n = 31) for the comparison of P. aeruginosa-infected versus non-P. aeruginosa-infected cells. Points towards the top of the plot indicate a higher probability of classifying as P. aeruginosa-infected, while points towards the bottom of the plot indicate a higher probability of classifying as non-P. aeruginosa-infected. The horizontal dotted line corresponds to optimal sensitivity/specificity cutoff (0.544). Wide bars represent median probability for each class, while narrow bars represent 25^th and 75^th percentiles.

3.4. Most discriminatory metabolites and their possible origin

For the comparison of P. aeruginosa-infected versus non-P. aeruginosa-infected cells, volatile molecules were ranked according to their discriminatory ability, as defined by their mean decrease in accuracy (MDA, a measure of variable importance). The intensity profile of the most discriminatory molecules (MDA > 0.01) are visualized in Figure 4. Six of these nine were putatively identified [two ketones (acetone and 2-butanone), three aldehydes (3-methylbutanal, 2-methylbutanal, and 2-ethylhexanal), one hydrocarbon (undecane)] and three compounds were generally assigned as alkylated hydrocarbons (Table 1).

Figure 4.

Boxplot of the most discriminatory features (MDA > 0.01) for discrimination between P. aeruginosa-infected versus non-P. aeruginosa-infected cells. *: p < 0.05. a: Data were mean centered and scaled before plotting. Identification as per Table 1.

The chemical classes (ketones, aldehydes, hydrocarbons) reported herein in culture headspace can be plausibly attributed to lipid oxidation pathways. Acetone and 2-butanone derive from the decarboxylation of ß-keto acids, and have been reported as important markers of P. aeruginosa elsewhere [51–57,59,60]. Aldehydes have been related to lipid peroxidation during inflammation [61], and 3-methylbutanal and 2-methylbutanal have been also reported as key intermediate in the catabolism of amino acids, such as leucine [62,63]. Aliphatic hydrocarbons (e.g. undecane) may be derived from the metabolism of fatty acids, which are converted to aldehydes due to the activity of acyl-CoA reductase, and enzymatically decarboxylated to hydrocarbons [64]. A preponderance of hydrocarbons were also highlighted in the headspace of human and murine epithelial cells infected with RSV and Influenza A, respectively [58]. Indeed, it has been shown that viral infection increases the production of reactive oxygen species, in part through a NAD(P)H oxidase-dependent mechanism [65], which is associated with the formation of hydrocarbons.

3.4. Analysis of the headspace of P. aeruginosa during growth in MEM

Finally, we hypothesized that a subset of the nine metabolites identified as highly discriminatory between P. aeruginosa-infected versus non-P. aeruginosa-infected CFBE resulted from either the consumption or production of metabolites by the bacterium itself. To test this, P. aeruginosa was grown as a biofilm in MEM in the absence of respiratory epithelium. Of the nine compounds selected, six were found in the headspace of P. aeruginosa grown alone, namely acetone, 2-butanone, 2-methylbutanal, undecane, and two hydrocarbons (#29 and #40). However, none of these differed significantly in abundance between P. aeruginosa in MEM and sterile media (MEM alone), suggesting that the selected metabolites likely result from the interaction of P. aeruginosa with the CFBE cells, or are produced in higher quantity during the interaction between P. aeruginosa and CFBE cells. Three compounds, 3-methylbutanal, 2-ethylhexanal and a hydrocarbon (#33), were not found in the headspace of P. aeruginosa grown only in media, suggesting that they originate from the interaction of P. aeruginosa with CFBE cells, or from the response of CFBE cells to the presence of P. aeruginosa.

3.4. Study strengths and limitations

The present study represents a first attempt at evaluating the volatile metabolic signatures of both singular P. aeruginosa and RSV infection, as well as P. aeruginosa-RSV co-infection, in the setting of in vitro CFBE cell co-culture. The use of CFBE cell culture represents a more clinically-significant growth environment relative to growth in nutrient-rich in vitro environments. This, in turn, can provide insight for future in vivo experimental models and inform translational studies in patient populations. Furthermore, our evaluation of P. aeruginosa grown in MEM allows us to hypothesize about the origins of these metabolites, although further experiments are needed to determine whether these discriminatory metabolites arise from metabolism of alternate substrates by P. aeruginosa in the presence of CFBE, and/or the response of CFBE to the presence of P. aeruginosa.

With respect to study limitations, we acknowledge that we consider only a single, reference strain (PAO1), which may not be metabolically representative of all possible P. aeruginosa strains. Furthermore, we were able to analyze only a small volume (500 μL) per replicate, and that this may have contributed to a relatively weak signal intensity, possibly below the detection limit for a subset of analytes. This, in turn, may have contributed to our inability to identify a robust volatile metabolic signature associated with RSV infection. Moreover, the volatile profiles obtained do not necessarily mirror the real profile present in the headspace of the sample, since it is mediated by the specific selectivity of the SPME fiber (PDMS/Car/DVB) used. The identification of the discriminatory compounds was putative and the confirmation of their identity with analytical standards should be confirmed in future studies. Finally, it is important to consider that the CFBE cell system models the initial interaction (6 h) of P. aeruginosa with the respiratory epithelium. It is therefore possible that a longer interaction time might yield to a more robust signal and altered metabolite detection as the infection persists.

Conclusions

In the present study, we demonstrate that volatile metabolites can be used to discriminate between P. aeruginosa (PAO1) infected and non-P. aeruginosa infected respiratory epithelium. Although weak and non-discriminatory, a signature related to the RSV singular infection can be observed (see Figure 3), while the volatile signature of co-infected samples (PA⁺/RSV⁺) is dominated by the signature produced by P. aeruginosa. The contribution of the host response to the abundance of some of these discriminatory metabolites can be hypothesized, since their abundances were significantly different from the control when RSV singular infection was present (acetone, undecane, and a hydrocarbon (#29)). Furthermore, the formation of 2-butanone, 3-methylbutanal, and 2-ethylhexanal can be associated with the specific interaction between P. aeruginosa and the CFBE cells, as these metabolites were not detected in the headspace of P. aeruginosa grown as a biofilm in the absence of CFBE.