Multivariate analysis

The NMR derived spectral dataset was further analyzed by multivariate analysis methods as provided by SIMCA V17 (Umetrics, Umeå, Sweden). Since metabolomic data, especially NMR spectral data, are characterized by a high degree of collinearity, we applied multivariate analysis methods, the principal component analysis (PCA) and the orthogonal partial least squares discriminant analysis (OPLS-DA). Those methods take into account correlations between metabolites and have been widely used to identify biomarkers in metabolomics studies [29, 30]. PCA was used to generate a first overview of information contained in the data, since it reduces the dimensionality of such datasets to increase interpretability and to minimize information loss. Thus, the original data can be described in a lower-dimensional space, defined by the principal components, which are ordered according to their ability to capture the total variance of the data. The score values represent the coordinates of the samples in the lower-dimensional space defined by the principal components. The principal components are displayed in a two-dimensional score plot, allowing visualization of the distribution and grouping of the samples in the new variable space [29]. Accordingly, by inspecting the score plot the homogeneity of the samples can be evaluated and any possible trends and outliers between the samples become visible. Thereafter, a supervised multivariate analysis OPLS-DA was performed to identify the discriminatory features for each comparison of the different assigned groupings. Significant metabolites were selected based on the p(corr) > 0.5 from the OPLS-DA models, where p(corr) is defined as the loadings rescaled as a correlation coefficient between the original data and the scores, thereby standardizing the range from − 1.0 to 1.0. There is no consensus on what p(corr) cutoff represents significance, but an absolute p(corr) > 0.4–0.5 is commonly used [31–33]. The quality of the OPLS-DA models was evaluated by using the default sevenfold crossvalidation in SIMCA and the built-in permutation plot (in short: permuting the y-variable 200 times and subsequently correlating these results with that of the original models). Analysis of variance of cross-validated predictive residuals (CV-ANOVA) was used to assess the significance of the OPLS-DA models, where a p-value lower than 0.05 is associated with a significant model.