Validation on high variance metabolic profiles: Taste stratification in a free living population

In man, living environment and lifestyle influence host physiology and microbial populations, and these interactions determine the health of individuals. In particular, dietary preferences are partially triggered by genetically-mediated taste responses to specific foods and may be reflected in the metabolic phenotype. However, identification of dietary metabolic signatures in free-living populations is made difficult by the relatively low amplitude of nutritional modulations when compared to inherent intra- and inter-individual variability that results from numerous confounding factors, e.g. sampling, lifestyle, gender and ageing. The identification of dietary biomarkers lies in the application of chemometrics to recover key metabolic information from the high density and highly variable metabolic phenotypes. One preferred approach is the use of linear regression models such as Partial Least Squares – Discriminant Analysis, whose reliability depends on appropriate validation procedures. However, the high inter- and intra-individual variability in free-living populations makes difficult the direct application of consolidated validation methodologies such as n-fold cross-validation that often leads to unstable solutions. In the present manuscript, we propose a new validation strategy based on the combination of several methodologies, merging n-fold cross validation and bootstrapping procedures with the confirmation of the reliability of the model through the generation of a set of replicate models. The proposed method was employed to validate the identification of dietary specific metabolic biomarkers in 1H NMR and UPLC TOFMS urine metabolic profiles obtained from 21 free-living subjects stratified according to their sensitivity to propylthiouracil.