Exhaled breath biomarkers of influenza infection and influenza vaccination

Each subject was defined as his/her own control to minimize confounding biologic variables. Breath metabolites were measured and compared just before and after the vaccination: on day 0 prior to the flu-shot, and then on the days 1, 2 and 3 after the immunization. EBC samples were collected during two flu seasons from a total of 12 healthy human subjects (five subjects in 2018, seven subjects in 2019). Subjects were adults (over 18 years old), non-smokers with no recent history of viral infection or lung disease. The mean age from all subjects was 34 years (SD = 7.8) with an even number of male/females (6/6). All subjects were treated with a single dose of a quadrivalent inactivated flu liquid vaccine injected intramuscularly. No adverse effects were reported by any of the subjects, either of influenza vaccination or the breath collection procedure.

The patients included as eligible in this study were adults that were admitted into the hospital after physician's determination of upper respiratory infection symptoms. Information about race, ethnicity, tobacco and food/beverage consumption, asthma diagnosis and any medication consumed was collected. Patients had to avoid eating, drinking and smoking at least 1 h before the sample collection. In some of the cases, a preliminary POC flu test was previously used for a rapid evaluation of the patient. But, according to hospital policies, a nasopharyngeal swab was used after all the patient's admission to perform a respiratory viral panel (RVP; GenMarkDx®). RVP consisted of a PCR-analysis of nasal swab (BD Universal Viral Transport, Beckton, Dickinson and Co, Sparks, MD or Miraclean Technology Co, Ltd, Shenzhen, China) samples collected and transported by Universal Transport Media (BD Universal Transport Medium, Beckton, Dickinson and Co, Sparks, MD, or remel MicroTest™ M4RT® Transport, ThermoFisher Scientific Inc.) for specific viruses including influenza A, HMPV, and HRV/ENT. HRV/ENT were not distinguishable by the used RVP. All patients were asked to participate in the study and responded to a short questionnaire regarding asthma pre-diagnosis, other diseases, or medications, as well as recent food/drink consumptions.

A total of 15 subjects (4 males and 11 females) with respiratory infection symptoms admitted to the hospital and one control sample (female with no symptoms) were included in the study. The subjects' average age was 41 years old (SD = 19). Information is summarized in table 1.

From the 15 admitted subjects, only five were tested with POC flu test, resulting in four positives. All patients went through RVP and only seven were positive, being two of them characterized by Influenza A virus, two with HMPV and three HRV/ENT. Two of the patients with confirmed POC flu positive resulted negative in virus infection for RVP tests (more accurate). Nine of the subjects had diagnosed asthma and were nonsmokers, except for two patients.

Targeted analysis showed low concentrations of the compounds we were pursuing. Most of the compounds were detected closer to the limit of detection (LOD) defined by the method. Only 21 compounds out of 55 showed enough concentrations above LOD on more than 50% of the samples. 9(10)-EpOME and 12(13)-EpOME were found at higher concentrations, with an average amount of 0.18 and 0.25 ng ml⁻¹, respectively. 9-HODE and 13-HODE followed with amounts around 0.1 ng ml⁻¹. All detected compounds were corrected by IS and amount of EBC used in sample preparation step.

When data are compared by pre (day 0) and post (days 1, 2 and 3) flu shot days, we only found significant differences for 13-HODE (p < 0.05). Figure 1 show concentrations achieved each day of the study for this compound (figure 1(a)), where mean values show and increasing of 13-HODE the day after the vaccine, and then progressively decreasing until day 3. Boxplots explain differences pre vs. post, considering different days after flu shots. In general, there are differences between pre and post, when considered days 1–3 together, but no differences from prior concentrations (day 0) and days 1 and 2, individually. Day 3 show clear differences of 13-HODE decreasing compared to initial levels. 13-HODE is an intermediate product from linoleic acid lipoxygenation, and it prevents cell adhesion to endothelial cells and can inhibit cancer metastasis [44].

Zoom In Zoom Out Reset image size

Surprisingly when using a multivariate approach, good classifications were not possible. An initial PLS-DA show low discrimination between compounds detected in EBC before flu shot (pre) and after it (post). We considered classifications with three post days merges and individually, and AUC values lower than 0.6 and sensitivities and specificities around 50% (data not shown). Other classification methods were applied, such as SVM-C and XGB-DA, but AUC < 0.5 were achieved in both cases (figure 1(c)).

Initial raw datasets contained 3875 and 6176 features for negative and positive modes, respectively. After data processing, the number of features was reduced around 50%, and samples were normalized by the volume of EBC used and responses of IS. Both ionization modes were then merged into a dataset with 5448 features. First, a PCA was applied to check and visualize data distribution (figure 2). It could be observed that main differences were clearly explained by sample flu season (figure 2(a)), which makes sense since as the components of the flu vaccine change each season. This seasonal variation was corrected by subtracting median values per each season. After season correction, PCA showed high reduction of seasonal differences (figure 2(b), (I)) and, when coloured by pre- and post-vaccination, it started showing a slight differentiation of some of the samples prior the shot (figure 2(b), (II)).

Zoom In Zoom Out Reset image size

Corrected data were then filtered by an initial variable selection based on differences from prior flu shot (pre/day 0). For that, volcano plot was used and variables with p-values higher than 0.05 and FC < 2 were removed. Noise and unwanted features were reduced around 90%, and these new datasets were used to predict specified classes using PLS-DA. A second feature selection was applied by selecting variables with VIP values higher than 1. This new data were then used to build final PLS-DA models and find main differences between pre vaccination and the posterior days after the shot (table 2). To get accurate results, 25 PLS-DA iterations were run using random split data between calibration and validation sets (66% samples in calibration set). Model iterations got average and standard deviation values for AUC, sensitivity, and specificity, both for cross-validation (CV) and prediction (Pred).

AUC values were higher than 0.9 in all models, with low variability for models including all days or individual days for post vaccination samples. Sensitivities and specificities were also over 0.8 in all cases, with low standard deviations.

Figure 3 shows four PLS-DA models grouping of samples that have differences by injection day considering: all post vaccine days together (figure 3(a)), and individual-day differences between day 0 (before flu shot), and day 1 (figure 3(b)), day 2 (figure 3(c)), and day 3 (figure 3(d)). A clear separation was observed in all models, with samples before flu shot (figure 3, green dots) clustered and separated from samples collected after the flu shot (figure 3, reddish dots). These optimal models allowed a final identification of the features that explain these differences in EBC metabolomic profile between pre and post flu vaccines. A total of 220 features were determined as important (VIP > 1), being 98 compounds the ones for the pre vs. post model (model A), then, 58, 54 and 99 features were found relevant for the differentiation between day 0 and day 1 (model B), day 2 (model C), and 3 (model D), respectively. Table 3 summarize 25 these main biomarkers found in all the prediction models. It describes the ESI mode detected, molecular formula and compound tentative identification with a score value, that corresponds to a percentage of matching between experimental spectra and the one from METLIN database. Only formulas and IDs were listed with score values higher than 70%. Some characterizations gave high scores for multiple compounds. The regulation based on post vaccine effect is also described, as well as the model where these biomarkers were highlighted as remarkable for the differentiation.

Zoom In Zoom Out Reset image size

Figure 4 shows up- and down-regulation profiles of some of the compounds from table 3. It presents the evolution of the compound abundances (after normalization) through each day after the vaccine (left graphs) and through all responses after vaccination together (right graph, pre/post). From all the 220 compounds, only 20% were up-regulated, meaning that their abundances increased after getting the vaccine (figure 4(a)). Some were identified as fatty acyls or prostaglandins (compounds 2 and 22), molecules commonly related to metabolomic processes associated to pro- and anti-inflammation and oxidative stress. However, most of the compounds were down-regulated, with signal reduction after the flu shot was administered (figure 4(b)). That was the case of known oxylipins, such as DiHOMEs or DiHODEs (compounds 4 and 21), associated to lipoxygenase and cytochrome P450 pathways. Amino acids (compounds 6 and 16) and other prostaglandins or thromoboxanes (compound 25) also showed decreasing in EBC content. Although most of these compounds showed same up or down-regulation for each of the three days after the shot, some compounds only showed significant differences just one of the days. For example, compound 25 showed VIPs > 1 and reduced signals for days 2 and 3, but and increasing signal (although non-significant with VIP < 1) on day 1, making it not remarkable for the general mode defining pre vs. post differences.

Zoom In Zoom Out Reset image size

In this study, only ten compounds showed concentrations detectable over the LOD of the method. From those compounds, lipoxin A4 (LXA4), prostaglandin J2 (PGJ2), 12(13)-EpOME and 9(10)-EpOME were detected at higher concentrations, around 2 ng ml⁻¹ of EBC, in some of the samples. 14(15)-DiETE and 9(10)-DiHOME reached amounts over 1 ng ml⁻¹. Compounds were studied by RVP results and type of virus detected in the viral panel (figure 5).

Zoom In Zoom Out Reset image size

Significative differences between negative and positive viral panel results were only found for LXA4 amounts (p = 0.018) (figure 5(a), left) [45]. However, when these differences were studied by the type of virus defined in the positive samples, we could observe that influenza A virus showed main differences compared to control or negative samples. That also happened for three of the compounds: LXA4, PGJ2 and 9,10-DiHOME (figure 5(a), right), suggesting that positive differences were mainly explained by the samples characterized by FluA. When all compounds were studied in a multivariate way we could confirm that trend (figure 5(b)). PLS-DA models separating positive and negative RVP samples were not able to achieve discriminating abilities, with AUC around 0.5 and sensitivity and specificity lower than 0.6 (dashed line). However, models that classified virus types, although the overall model ability was not reliable (non-dashed lines), were able to discriminate Flu A from the rest of samples with AUC around 0.9, but not the rest of the viruses.

Similar data pre-processing to pre/post vaccination was applied to untargeted data obtain from patients with respiratory infections. Almost half of the raw data were cleaned, and the merged dataset listed 8092 features, containing variables from positive and negative modes. In this case, the dataset was normalized just by the initial volume of matrix used for the analysis, including EBC and the ethanol rinse fractions in the same dataset. Variable selection was then applied based on the differences between RVP results, and between the specific virus detected in the clinical tests. P-values and fold change information was used to initially reduce features and remove noise from the datasets. RVP data (1723 features) was focused on differences between positive and negative results, and virus dataset (3029 features) was aimed on distinguishing influenza A, HMPV, rhinovirus/enterovirus, and non-virus. Relevant features were subsequently selected using VIP values from initial PLS-DA models. All PLS-DA iterations were built 50 times using random split between calibration and prediction datasets (table 4). Table 4 shows the increasing classification abilities, such as AUC, sensitivity and specificity values, once feature selection was applied using VIP scores.

RVP classifications achieve AUC around 0.86 but raise to 0.98 when VIP > 1 variables are selected (table 4). Sensitivities and specificities raised from 0.74 to 0.83 to more than 0.9 with low standard deviations. Figure 6 showed optimal PLS-DA models built with the final selected features.

Zoom In Zoom Out Reset image size

Clear differences were observed between RVP positive and negative samples (figure 6(a)). In that case, 90 features determined the classification (VIP > 1) with almost 50 of these with high score values for tentative identifications. Table S1 (available online at stacks.iop.org/JBR/15/046004/mmedia) (supplemental information) lists biomarkers identified for the RVP discrimination between positive and negative results. Each biomarker variation is defined by up-regulation depending on presenting significative higher intensities in RVP positive or negative. These differences are represented with some examples in figure 7(a), where most of the compounds were up-regulated by negative RVP, meaning that their abundances in breath were higher when negative test was obtained. Some of the identified molecules could be related to pro- and anti-inflammatory processes and oxidative stress, such as fatty acyls, phenols or sphingolipids. When we look at the variations between specific viruses (figures 6(b) and (c)) there is a clear difference between subjects diagnosed with influenza A and the others. Although the low number of samples for each positive virus type, these results direct promising expectations for future studies using a greater number of subjects. Also, the rest of the viruses were highly differentiated between them and negative results. A total of 146 features were defined as relevant for virus separation, being 75 of them tentatively identified. Table S2 (supplemental material) shows identified biomarkers for the virus type classification. In this case, each compound is up- or down-regulated considering the three viruses and non-virus class. Most of the compounds were up-regulated by virus presence, being characteristic markers of influenza A, HMPV and HRV/ENT present in higher intensities in 77%, 73% and 68% of the cases, respectively. Only 10% of the markers were up-regulated when no virus was detected. Similarly to the general RVP model, several fatty acyls were identified, as well as ceramides, sphingolipids and tripeptides (table S2). Some examples of specific biomarkers regulation are presented in figure 7(b). When we observe the specific markers, a total of 32 appear in both models, the one from general RVP response and from specific virus determination.

Zoom In Zoom Out Reset image size