In this issue of Liver International, Barba and colleagues have applied a ‘metabonomic’ approach to investigate the metabolic effects of hepatic encephalopathy in intact rat brain tissue. Multivariate models have been created to distinguish grades of encephalopathy (1). They have also highlighted the protective effects of hypothermia, as evidenced by diminished metabolic abnormalities, in this rat encephalopathy model. In this group, tissue alanine and lactate levels were reduced and N-acetylaspartate and myo-inositol levels were elevated compared with brain tissue from normothermic animals. In addition, the study demonstrated regional variation in metabolic deterioration in the encephalopathic brain.
Data were acquired using 1H nuclear magnetic resonance (NMR) spectroscopy, a technique that allows the simultaneous acquisition of multiple biochemical parameters from biofluids and tissues (2). Specific metabolites are represented as a characteristic pattern of chemical shifts. The signal intensity depends on the metabolite concentration and hence the technique is quantitative. Magic angle spinning NMR methodologies have been increasingly employed to obtain biochemical information from intact tissue without the need to use tissue extracts for analysis (2, 3). This technique, as used by Barba and colleagues, has the advantage that the tissue is available for histology or other parallel investigations once the NMR experiment has been performed (3).
The emerging technique of ‘metabonomics’ (4) applies sophisticated multivariate pattern-recognition strategies to the analysis and interpretation of complex data, such as that acquired with NMR. The entire NMR spectrum is typically divided into smaller regions or ‘bins’, representing individual metabolites, and the integral of each bin is normalized to the sum of the total spectral integral. Thus, each metabolite in the spectrum is expressed relative to all of the others. Recently, the possibility of using full-resolution spectra without binning has been demonstrated (5). Chemometric analysis may then be carried out (6).
The initial method of choice for gaining an overview of the samples, demonstrating any inherent clustering and identifying outliers, is ‘principal components analysis’ (PCA). This is an unsupervised method, in that there is no prior knowledge of the category to which each sample belongs. ‘Principal components’ are identified: linear combinations of variables (metabolites), accounting for the greatest variation within a whole dataset. A scores plot is generated, plotting each individual sample in the new coordinate space. The corresponding loadings plot is then examined for each principal component to elucidate the combination of metabolites responsible for the patterns in the scores plot.
Using this relatively simple analytical technique, Barba and colleagues have demonstrated distinct clustering of their samples along their principal components, 1 and 3. The patterns in the scores plot showed differences between the sham rats and those with liver failure, and also distinctions dependent on the origin of the tissue studied (brainstem or frontal cortex). Inspection of the corresponding loadings allowed the specific metabolite profile responsible for the separation to be identified, including increased glutamine and histidine, decreased myo-inositol and creatine.
A classification model was then constructed using supervised ‘partial least squares discriminant analysis’ (PLS-DA). This is a linear regression technique in which the NMR spectroscopic variables, corresponding to metabolites, are related to the class membership of the sample. Again, scores plots are generated and inspection of the loadings and the regression coefficients allows identification of the spectral regions (and hence metabolites) responsible for separation between the groups. Rigorous validation of the PLS-DA model constructed is all-important, as the method is ‘eager to please’ in finding separation (7). Both cross-validation (in which a specified number of samples is sequentially excluded from the analysis and then predicted back into a model created from the rest of the samples) and full external validation (in which a model is created from a ‘training’ set of samples and the class membership of a separate ‘test’ set is predicted) may be employed. The predictive ability of a PLS-DA model is usually expressed either as the prediction error measure Q2 (with values >0.05 considered a statistically significant difference) or as the number of misclassified samples (7–9).
The validation method used by Barba and colleagues in the calibration of the PLS-DA model of their four experimental groups is not clear, although they state, perhaps confusingly, that they tested the classification model that was developed in normothermic rats with two test sets of hypothermic animals. Rather than quoting Q2 values or showing a misclassification matrix, the average principal component scores for each group were compared using analysis of variance.
Given the excellent separation between the study populations observed through the unsupervised PCA on the basis of certain specific metabolites, it is not surprising (and indeed may be seen as corroborative) that the same metabolites contributed to the PLS-DA model. The authors make the point that although the majority of the discriminatory metabolites identified have been reported in previous studies of liver failure, it is the pattern-recognition analysis of the spectroscopic profile as a whole that allows the grading of encephalopathy and the monitoring of the response to hypothermia. The technique has also facilitated the novel observation that, as hepatic encephalopathy progresses, metabolic deterioration in the brainstem is more rapid than in the cortex. Further functional investigation of regional differences in the development of encephalopathy may prove fruitful in improving understanding of the disease process.
Barba and colleagues also highlight a potential problem with the method of expressing one metabolite level relative to another, in this case creatine. In order to interpret an increase in the metabolite ratio as an increase in the numerator, it must be assumed that creatine levels remain constant. The tendency of creatine to decrease during disease progression would clearly influence metabolite ratios. This issue could, perhaps, be addressed by relating specific metabolites of interest to the total spectral integral, as the authors have done for their multivariate analysis.
This paper demonstrates the applicability of an NMR-metabonomic approach to the investigation of disease pathogenesis using in vitro biological samples. The authors comment on the eminent feasibility of such techniques being employed for in vivo studies. This emerging technology may certainly offer a valuable resource in the elucidation of metabolic differences that occur with distinct disease processes, whether in vivo or in vitro, using biofluids or tissue samples, from either animals or humans.