Patient characteristics
Patient clinical and demographic characteristics of the patients are listed in Table 1. 18 (72 %) had non-small cell lung cancer, while 7 (18 %) had small cell lung cancer. The majority of patients had stage III disease and underwent treatment with concurrent chemoradiotherapy. The 2-year overall survival for the group was 53 %.
Table 1 Patient demographics and clinical outcomes of patients with small cell and non-small lung cancer
Metabolomics analysis
Temporal metabolomic analysis distinguish sera across treatment
Serum samples from each patient collected pre-treatment, during therapy, and post-treatment were subject to both NMR and GC–MS analysis. A total of 56 features were quantitatively assessed by NMR, and 106 features identified by GC–MS. Figure 1a is a box and whisker plot of the multivariate scores of OPLS-DA analysis of GC–MS metabolomic data from the three time-points. In this analysis, the discriminant analysis was naïve to the temporal nature of the data, however a defined temporal decrease in the scores from the first predicative component is apparent. Analysis of the OPLS-DA loadings reveals that the differences in timepoints was based on the differential abundance of glucopyranose, citric acid, butanoic acid, erythritol and ribitol between the three groups (N = 61, CV-ANOVA p = 0.0046, R2 = 0.186, Q2 = 0.128) (Table 2a.VIP > 1). Figure 1b presents a similar analysis of 1H-NMR metabolomic data. The differences were based on the changing level of metabolites 2-aminobutyrate, 2-oxoglutarate, threonine, methionine, creatinine and citrate between the three groups (N = 64 CV-ANOVA p < 0.001, R2 = 0.223, Q2 = 0.143) (Table 2b.VIP > 1). Interestingly metabolites such as threonine and citrate were part of up the 1H-NMR metabolite data; were not picked by the GCMS.
Table 2 Metabolites from the GCMS and 1H-NMR data involved in discrimination between the three time points of pre-therapy, therapy and post-therapy
In order to further probe the temporal nature of the data, we employed an analysis specific for time course data (SAM), which utilized the repeated measurements within individual patient samples. From the GCMS data, SAM analysis identified 8 differentially abundant metabolites across all treatment points (Fig. 1c). Of these named metabolites included 2-hydroxybutanoic acid, glucopyranose, citric acid, erythritol and ribitol. 1H-NMR metabolite data was able to identify three significant metabolites across the designated time points of sample collection. These included taurine, threonine and creatinine (Fig. 1c).
Relationship of metabolite data to clinical outcomes
GCMS metabolomic profiles facilitating prognostic evaluation of survival and progression from pre-treatment samples
From our preliminary O2-PLS models including all clinical covariates, we established that progression and survival were the two variables reflected by the GC–MS data (Fig. S1A). We then constructed individual OPLS-DA models for these covariates (Survival, Fig. S2B; Progression, Fig. S2C) based on the metabolic samples of pre-treatment samples only, i.e. naïve to any chemotherapy or radiotherapy. The survival model distinguishes the metabolite profile of patients who had died during the course of treatment from those who had survived treatment (N = 25, CV-ANOVA p = 0.0335, R2 = 0.388, Q2 = 0.285). This distinction was based on relative differences in metabolites such as tridecan-1-ol, octadecan-1-ol and hydroxylamine in that were abundant in patients who did not survive. Figure 2a is a box and whisker plots that show the distribution of scores as a function of deceased vs. survived therapy. A heatmap was constructed using hierarchical clustering (Fig. 2b). This data illustrates of metabolites such as glutamine, proline, valine, threonine and tyramine being differentially abundant in the population of patients who survived therapy where metabolites such as hydroxylamine and octadecan-1-ol being at a higher concentration in patients who did not survive therapy.
Similarly, a significant OPLS-DA model was built demonstrating a relationship between the GCMS metabolic profile and disease progression (Fig. 2c; Fig. S2C; N = 25, CV-ANOVA p < 0.05, R2 = 0.397, Q2 = 0.333). This difference was based on metabolites such as tagatose, hydroxylamine, glucopyranose, and threonine. A heat map illustrating the features involved in disease progression using SAM analysis to distinguish groups (Fig. 2d) demonstrates metabolites such as hydroxylamine are at relatively higher concentration in the group of patients who were noted to progress with the disease despite therapy, however glucopyranose and threonic acid were the metabolites was found to be at a relatively higher level in patients in whom the disease did not show signs of progression.
Clinically, progression and survival are related events, and thus in order to further understand the relationship between metabolite reflective of survival and disease progression a shared and unique structure (SUS) plot was constructed (Fig. 2e). The metabolites that line up along the diagonal running from the lower left corner to the upper right corner are common to both the patient progression and survival model. These included metabolites such as hydroxylamine (down-regulated), glucopyranose, tagatose, glutamine, tyramine, and proline (up-regulated). Metabolites such as phosphoric acid, glycine and octadecanoic acid were unique to the disease survival model. Our results thus indicate that a unique biomarker profile is possible which distinguishes progression from survival.
NMR metabolic profiles facilitates evaluation of cancer staging and cancer type from pre-treatment samples
Preliminary modeling of clinical covariates with quantitative NMR-derived metabolic profiles indicated that cancer stage and type were well reflected in the serum profiles. We were able to facilitate discrimination in cancer staging between cancer stages 1 and 2, versus stage 3 using 1H-NMR metabolomic data (Supplemental Fig. 3A) (N = 24, CV-ANOVA p < 0.05, R2 = 0.474, Q2 = 0.314). This was based on 8 metabolites such as 2-hydroxybutyrate, 2-oxoisocaproate, acetate, carnitine, 3-hydroxyisovalerate, 2-hydroxyisovalerate, glycerol and glycine. Summary of the scores from this analysis (Fig. 3a) in which the distribution is plotted according to the class i.e. stages 1 and 2, versus stage 3; patients who were staged lower had a lower score in comparison to patients with higher staging.
In a further subgroup analysis, we investigated the metabolomic profiles of non-small cell lung (NSCLC) cancer patients. Using 1H-NMR data we were able to distinguish sera based on cancer pathophysiology. Patients with NSCLC could be discriminated into subtypes of squamous and adenocarcinoma. OPLS-DA modeling was able to discriminate between the two cancer sub-types based on 19 differentially abundant spectral features (Supplemental Fig. 3B) (N = 18, CV-ANOVA p < 0.01, R2 = 0.677, Q2 = 0.536). Figure 3b shows the box and whisker plot plotted using metabolite scores based between two classes of non small cell cancer i.e. squamous cell and adenocarcinoma cell carcinoma respectively. Metabolites such as 2-oxoisocaproate, 4-hydroxybutyrate, lysine, arginine, dimethylamine, isobutyrate, 3-hydroxybutyrate, acetate, asparagine, phenylalanine were relatively higher in patients with adenocarcinoma. However, metabolites such as pyruvate, lactate, valine, proline, isoleucine, histidine, 2-aminobutyrate, leucine and alloisoleucine were relatively lower in the patients with adenocarcinoma.