In the expanding field of nutritional metabolomics the use of optimized and robust strategies are essential to find meaningful and significant results. Data from nutritional studies can be extremely large, multidimensional and complex. Moreover the data may hide just a subtle biological variation due to the nutritional treatment, and the effect may be highly variable across individuals. Usually the treatment effects are much smaller than the biological variations between individuals. A properly chosen experimental design and a well-adapted data analysis strategy are therefore essential to ensure that the intended information in the data can be assessed and that the biological question of interest can be answered.

Nutritional interventions studies, in which the bioavailability and/or the bioactivity of dietary ingredients are investigated, are usually designed as crossover studies. A major benefit of the crossover design is that each individual acts as his own control. This feature allows for the direct comparison of treatments, and is particularly efficient in the presence of large between-individual variation [1]. A specific method which is capable to cope with the crossover structure in metabolomics data is Multilevel Data Analysis (MLDA) [2]. This recently introduced method permits a separate analysis of the between-individual variation and the within-individual variation in the data (fig. 1). The basic principle of MLDA relies on the variation splitting property of ANOVA, and can be considered as the megavariate extension of a paired t-test. In crossover designed metabolomic experiments the use of MLDA is preferred over classical megavariate data analysis methods such as Principal Component Analysis (PCA) or Partial Least Squares Discriminant Analysis (PLS-DA). An important limitation of using these megavariate methods is that the paired structure in the data is not taken into account. Consequentially, the induced variation due to the nutritional treatment is often largely overwhelmed by the biological variation that exists between individuals.

We have demonstrated the MLDA method for the case of a placebo-controlled crossover designed human nutritional intervention study in which the metabolic impact of grape/wine extract consumption on the urinary ¹H NMR profiles was evaluated.

To investigate the underlying variation in the between-individual data, multilevel PCA (or MLCA [3]) was used. In the multilevel PCA two major sources of variation could be identified, i.e. biological variation between individuals (gender) and experimental variation (measurement batches). As shown in figure 2a the PC1-PC4 scores of the males could be distinguished from the females in the study group, whereas the PC2-PC4 score plot in figure 2b reveals a distinction between two different NMR runs.

To find systematic differences among the intervention groups, multilevel PLS-DA on the within-individual data was used. The results from the multilevel PLS-DA analysis on the within-individual differences show that the levels of a few phenolic metabolites were significantly increased in the urine after the treatment (fig. 2c). Remarkably, these metabolites could not be identified significantly in an original PLS-DA approach. Among the observed metabolites, we found hippuric acid as the strongest biomarker for the intake of the grape/wine extract. As shown in figure 2d, hippuric acid is represented by three signals in the aromatic region (δ 7.83 ppm, d, CH₂/CH₆; δ 7.64 ppm, t, CH₄; δ 7.55 ppm, t, CH₃/CH₅). The presence of hippuric acid is in agreement with previously reported studies where the metabolic impact of polyphenolic-rich diets was studied [4-7]. Hippuric acid is thought to be the metabolic end-product of flavonoid degradation by the gut microbiota. Besides hippuric acid also other phenolic compounds were significantly elevated in the urine, i.e. 4-hydroxyhippuric acid and 4-hydroxyphenylacetic acid. Like hippuric acid these phenolic acids are known gut microbial fermentation products of flavanoids.

Recently we were also able to integrate MLDA in a pharmacokinetic experiment. In a crossover designed, placebo controlled, intervention study, a human study group was investigated upon the temporal, urinary excretion of polyphenolic metabolites after the intake of black tea solids. Investigation of these urinary metabolites was performed using the ¹H NMR and GCMS urinary profiles. Due to the integration of experimental design information in the data analysis it was possible to uncover a considerable list of gut mediated metabolites. Among the metabolites we could identify phenolic intermediates such as 1,2,3-trihydroxybenzene, 1,3-dihydroxyphenyl-2-O-sulphate, 4-O-methylgallic acid, gallic acid, hippuric acid and 4-hydroxyhippuric acid. The excretion in the urine of 1,3-dihydroxyphenyl-2-O-sulphate4 over 48 hours is shown in figure 3a.