Plasma metabolomics for the diagnosis and prognosis of H1N1 influenza pneumonia

Forty-two patients with confirmed influenza virus A (H1N1) pneumonia were included in this study from multiple Canadian centers by the University of Manitoba. Only patients ≥18 years of age were included in the study.

To diagnose H1N1, 31 noninfected ventilated control ICU patients were selected on the basis of age- and sex-matching to the patients with H1N1 (Table 1). Ventilated ICU control subjects were patients admitted to the ICU postoperatively after an elective procedure (e.g., posterior cranial fossa or spinal surgery) if there was no suspicion of infection and plasma was collected on day 1 of ICU admission while these patients were ventilated in the ICU. Moreover, 29 culture-positive samples from patients with bacterial CAP were selected from a multicenter study at the University of Pittsburgh for diagnostic comparison of bacterial pneumonia with the patient cohort with H1N1 virus (Table 1). Bacterial CAP samples were identified on the basis of clinical and radiologic criteria as described previously [11]. The bacterial sources of infection included different species, such as Streptococcus pneumoniae, Staphylococcus aureus, Pseudomonas aeruginosa, and Escherichia coli.

Of 42 patients with H1N1, 21 patients consisting of 7 nonsurvivors and 14 survivors were used for the mortality prognosis training set in H1N1 pneumonia. Table 2 shows the demographics of the nonsurvivors and survivors used for the training set for the mortality prognosis portion of the study.

This case-control study was nested within three cohorts enrolled at the universities of Calgary, Manitoba, and Pittsburgh. To determine whether metabolomics can be used to diagnose H1N1 pneumonia, we compared 42 patients with H1N1 pneumonia with 31 age- and sex-matched ICU patients who required mechanical ventilation and 29 patients with H1N1 pneumonia with 29 sex-matched patients with bacterial CAP. To determine whether metabolomics could be used to predict mortality among 42 patients with H1N1, we compared 14 survivors and 7 age- and sex-matched nonsurvivors (ratio 2:1) at 90 days as a training set. A total of 42 patient samples were examined, with 21 patient samples used as a “discovery” cohort for the mortality study. The other 21 patient samples were used as a validation cohort; however, all of these were survivors. All tested patients with H1N1 infection had no initial detected bacterial coinfection.

¹H-NMR spectroscopic analysis was performed in one-dimensional mode for all samples using a 600-MHz Bruker Ultrashield Plus NMR spectrometer (Bruker BioSpin Ltd., Milton, ON, Canada). Details of the pulse sequence can be found in the data acquisition section of the supplement (see Additional file 1). Chenomx NMR Suite 7.1 software (Chenomx Inc., Edmonton, AB, Canada) was used to profile the ¹H-NMR spectra for metabolite identification and quantification using a nontargeted profiling approach in the profiler module [12]. We used 4,4-dimethyl-4-silapentane-1-sulfonic acid as an internal standard for metabolite quantification [13].

GC-MS analysis was also performed on all samples using an Agilent chromatograph 7890A (Agilent Technologies, Santa Clara, CA, USA) coupled with a Waters GCT mass spectrometer (Waters Corp., Milford, MA, USA), using a gas chromatography time-of-flight mass spectrometry technique. The mass spectrometer was programmed in the range of 50–800 mass-to-charge ratio. Using Metabolite Detector software (version 2.06; Institut für Biochemie & Biotechnologie, Technische Universität Carolo-Wilhelmina zu Braunschweig, Braunschweig, Germany), mass spectra were processed and analyzed to detect compounds. The Golm Metabolome Database [14] and National Institute of Standards and Technology 2008 library [15] were used to identify compounds. For sample preparation information, please see the online supplement (see Additional file 1).

Unsupervised multivariate principal component analysis (PCA) was performed to assess the data acquired on plasma ¹H-NMR metabolites and GC-MS features from patients with H1N1 (n = 42) vs. ventilated ICU control subjects (n = 31) and from patients with H1N1 (n = 29) vs. patients with CAP with positive bacterial cultures (n = 29). PCA was also performed as an exploratory model on all identified plasma ¹H-NMR and GC-MS data from the training set of H1N1 samples (14 H1N1 survivors vs. 7 nonsurvivors matched by age and sex). The PCA model was used to identify data grouping and outliers and to examine the intrinsic differences between the two cohorts for each analytical technique. The number of quantified metabolites/features for PCA analysis of each study are depicted in Additional file 2: Table S14.

Orthogonal partial least-squares discriminant analysis (OPLS-DA) was then performed to build prediction models of H1N1 diagnosis and mortality. OPLS-DA was used to maximize covariance between the measured variables and the response variables (predictive classifications). The quality of the OPLS-DA model was verified using three performance indicators: cross-validation analysis of variance (CV-ANOVA), R²Y, and Q²Y (see Additional file 1 for more details). Additionally, the other parameters used to describe the predictive models were sensitivity, specificity, and AUROC. Potential confounders (e.g., age, sex, body mass index [BMI]) and comorbidities (e.g., asthma, chronic obstructive pulmonary disease [COPD]) were examined for their importance using orthogonal 2 partial least squares (O2PLS) modeling of mortality of H1N1 pneumonia.

Pathway analysis of potential biomarkers was performed using MetaboAnalyst software (http://​www.​metaboanalyst.​ca). We also used Ingenuity Pathway Analysis (IPA) software (version 3.1; Ingenuity Systems Inc., Qiagen Bioinformatics, Redwood City, CA, USA) to discover the most important biological networks using significantly altered metabolites between cohorts. The significantly changed metabolites were obtained through multivariate data analysis; OPLS-DA was used to discriminate metabolomic profiles between two groups.

To obtain sensitivity, specificity, and AUROC data, we performed a prediction test to create a misclassification table for all discriminant analysis models for diagnosis and prognosis studies. We repeated the process three times to create randomly the prediction test and to average sensitivity and specificity. Moreover, for the prognosis of mortality study, the prediction sets were created from current active models (work sets) by taking seven random samples (two nonsurvivors and five survivors) out of the OPLS-DA models three times and averaging sensitivity, specificity, and AUROC. In addition, another prediction model (validation) was tested using the cohort of 21 survivors who were not used in the former prediction (test) set and 2 randomly selected nonsurvivors. Misclassification tests showed 100% sensitivity and 100% specificity for this prediction group.

Univariate analysis was performed as a complementary method to multivariate statistical analysis for enhancing the amount of information from the study and to serve as a less complex way to understand the cohort differences. We used MetaboAnalyst software for the univariate analysis and Student’s t test for evaluation of each variable individually to determine whether the means of two groups were distinct for the diagnosis and prognosis studies. To perform univariate analysis, both NMR and GC-MS datasets were normalized, followed by log transformation and autoscaling. The important features in each dataset were selected by t test with a threshold of 0.05.

To assess the value of plasma metabolomics for the diagnosis of H1N1 pneumonia, patients with H1N1 pneumonia, ventilated ICU control subjects, and patients with CAP with positive bacterial cultures were used to explore and create prediction models based on 29 patients with H1N1 pneumonia vs. 29 sex-matched patients with CAP with positive bacterial cultures and based on 42 patients with H1N1 pneumonia vs. 31 age- and sex-matched ventilated ICU control subjects (Table 1).

A PCA scatterplot of the entire ¹H-NMR and GC-MS datasets demonstrated that the H1N1 cohort could be distinguished from patients with CAP with positive bacterial cultures (see Additional file 2: Figure S1a and b) and the ventilated ICU control subjects (see Additional file 2: Figure S2a and b). OPLS-DA models showed a clear discrimination of metabolomic profiles of patients with H1N1 from profiles of patients with CAP with positive bacterial cultures (Fig. 1a and b) and ICU control subjects (Fig. 2a and b).

Table 3 shows the prediction models using OPLS-DA to discriminate plasma H1N1 samples from ICU control subject and positive bacterial culture CAP samples on the basis of NMR and GC-MS datasets and number of metabolites that contributed in the separation between groups. Q ² indicates an excellent separation between the plasma metabolic profile of H1N1 and ICU control subjects, whereas the prediction model for the NMR dataset is not as high as the GC-MS dataset to separate H1N1 from positive bacterial culture CAP cohort data. There are a large number of metabolites contributing to the separation of patients with H1N1 from patients with CAP with positive bacterial cultures and ventilated ICU control subjects, suggesting that H1N1 pneumonia could create a disease-specific metabolic profile quite distinct from the two other cohorts (Table 4). These data show that plasma metabolomics using ¹H-NMR and GC-MS analytical platforms could be applied as a diagnostic tool with high predictability, sensitivity, and specificity for identification of H1N1 influenza pneumonia. The data show that H1N1 pneumonia is accompanied by metabolomic changes in the concentration of some common metabolites (amino acids and ketone bodies) and some specific metabolites compared with patients with CAP with positive bacterial cultures and ventilated ICU control subjects. We also observed that some metabolites had similar patterns of changes in H1N1 in comparison to samples of patients with bacterial CAP and ICU control subjects. H1N1 pneumonia samples showed increased concentrations of dimethylamine, β-alanine, formate, and quinic acid and a decreased  concentration of alanine vs. the two other cohorts. Overall, plasma metabolic profiles by important metabolites/features showed clear differences between diagnosis of H1N1 compared to positive bacterial culture and diagnosis of H1N1 compared to ventilated ICU control subjects (Table 4).

Coefficient plots revealed the relative correlation of metabolites and features that showed decreased and increased concentrations in nonsurvivors vs. survivors and in patients with H1N1 vs. ICU control subjects on the basis of NMR and GC-MS datasets. Table 5 shows the most important biological pathways for diagnosis and prognosis based on multivariate data analysis using coefficient plot and S-plot analysis. S-Plot analysis was used to identify putative biomarkers on the basis of related OPLS-DA models to choose metabolites/features with high magnitude and high reliability (see Additional file 2: Figures S6, S8 and S10).

Univariate analysis was performed to show the mean, SD, and P values of metabolites/features obtained by NMR and GC-MS. Additional file 2: Tables S1 and S2 show the significantly different metabolites (P < 0.05) between H1N1 pneumonia and positive bacterial culture pneumonia samples for NMR (n = 13) and GC-MS (n = 98). Additional file 2: Tables S3 and S4 also show the significantly different metabolites (P < 0.05) between patients with H1N1 pneumonia and ventilated ICU control subjects for NMR (n = 27) and GC-MS (n = 57). Multivariate data analysis revealed more metabolites that were significantly changed between the H1N1 pneumonia cohorts and the two other cohorts on the basis of NMR data, but the GC-MS findings did not show a large difference in the number of metabolites between multivariate and univariate data analyses. Although the multivariate and univariate methods revealed high numbers of overlapping metabolites/features, we found a different pattern for more significant metabolites/features between multivariate and univariate methods in the diagnosis of H1N1 (Figs. 3 and 4).

For the prognosis of 90-day mortality, prediction models were built on the basis of the training set. There were no statistically significant differences (P < 0.05) on tested demographic variables between the survivor and the nonsurvivor cohorts, except for the presence of fever, which was higher in the survivor cohort (see Additional file 2: Table S15).

A PCA score plot of the entire ¹H-NMR dataset based on the first and second principal components demonstrated that the 90-day nonsurvivor group could be distinguished from the survivor group; that is, there was data clustering) (see Additional file 2: Figure S3a). The PCA score plot of the features detected by GC-MS revealed similarly separated clusters for H1N1 nonsurvivors and survivors (see Additional file 2: Figure S3b).

Once data clustering was revealed, we analyzed the metabolomic profiling data using supervised OPLS-DA for modeling. Of the NMR data (Fig. 5a), 27 different metabolites were used as potential variables (metabolites) to separate nonsurvivors from survivors with a R²Y = 0.831 and a Q²Y = 0.597, indicating very good separation between the two cohorts at the plasma metabolomic level for H1N1 mortality (Table 3). We used a statistical approach based on variable importance in the projection ≥1 to determine the number of analytes for the best predictive model using SIMCA-P Version 13.0, Umetrics AB, Umea, Sweden) [16].

As shown in Fig. 5b, the GC-MS OPLS-DA model showed samples from the survivor and nonsurvivor cohorts were well separated on the basis of 63 features (32 known and 31 unknown). The model characteristics were numerically significant, with a R²Y = 0.909 and a Q²Y = 0.829, indicating an excellent model. To assess the reliability of the OPLS-DA models for the NMR and GC-MS data, CV-ANOVA was performed. The P values for both models were 0.004 and 0.0001 for the NMR and GC-MS models, respectively.

Using O2PLS discriminant analysis [17], we tested the role of confounding factors in separation of nonsurvivors from survivors, including patient characteristics and comorbidities, but these did not affect the observed discrimination to prognosticate the mortality for both NMR and GC-MS analyses. This suggests that the differentiation between the survivor and nonsurvivor groups is based on disease metabolite changes rather than on the roles of age, sex, and BMI in this study (see Additional file 2: Figure S4a–c). Given these results, it can be concluded that ¹H-NMR and GC-MS are appropriate analytical tools to apply a metabolomic approach for prognosis of 90-day mortality of patients with H1N1 influenza pneumonia on the basis of samples taken within 24 h of hospitalization.

Univariate analysis showed that normalized concentrations of 7 metabolites detected by ¹H-NMR and the relative intensities of 19 features detected by GC-MS varied significantly among nonsurvivors and survivors. The summary of all significant metabolites/features (P < 0.1) from NMR and GC-MS are listed in Additional file 2: Tables S6 and S7 with their P values and mean (±SD) values of variables in each group. A comparison of multivariate and univariate approaches shows the differences in type and number of metabolites for both NMR and GC-MS datasets to separate nonsurvivors from survivors (see Additional file 2: Tables S5 and S6). It was interesting to note that using the multivariate statistical technique, we observed more potential metabolites/features than in univariate analysis to distinguish H1N1 nonsurvivor from survivor cohorts, with these two analytical methods showing a very different metabolic profile pattern. Univariate methods are used to simplify the interpretation of discriminating metabolites individually. Interestingly, this study showed that the univariate analysis could reject potential metabolites with only small changes that failed to have significant differences by t test, whereas they could be more important when they were analyzed simultaneously with the other metabolites; that is, the metabolites may not act independently on the outcome. Figure 6 shows all metabolites/features that are significantly different between nonsurvivors and survivors detected by NMR and GC-MS platforms using univariate analysis.

The important potential targets of metabolic pathway analysis obtained using MetaboAnalyst for NMR and GC-MS data showed an impact value ≥0.10. More detailed analysis of the most relevant pathways is listed in Addtitional file 2: Tables S7-S12 and Figures S11–S13. Table 5 shows the identified biological pathways involved, based on the most important metabolites found by NMR and GC-MS analyses, in the diagnosis of H1N1 pneumonia based on culture-positive bacterial pneumonia and ventilated ICU control subjects and prognosis of mortality in H1N1 pneumonia. The pathways have been ordered by their impact values from high to low in Table 5. Table 5 shows more differentiated biological pathways involved in the diagnosis of H1N1 from ICU control subjects compared with patients with culture-positive bacterial CAP. This evidence suggests that the difference in metabolomic profile between patients with H1N1 pneumonia and ICU control subjects is greater than that of patients with H1N1 pneumonia and culture-positive bacterial pneumonia, based on the OPLS-DA models (Table 4) as well as on the biological pathways. It is interesting that a variety of biological pathways were found in separation of H1N1 nonsurvivors from survivors on the basis of metabolite changes. More potential networks of biological pathways were generated through the use of IPA software in the diagnosis and prognosis of mortality studies for both NMR and GC-MS datasets (Table 6).