Robust and confident predictor selection in metabolomics

Metabolomics is a proven tool to obtain information about differences in food stuffs and to select biochemical markers for sensory quality of food products. A valuable application of untargeted metabolomics is the selection of metabolites that are (highly) predictive for sensory or phenotypical traits for use as (bio) markers. This chapter demonstrates how to robustly select key metabolites and evaluate their predictive properties. The proposed approach constrains the number of selected metabolites, searching for an optimal number of predictive metabolites by cross-validation. This mitigates the problem of selection of spurious metabolites. It also enables straightforward use of linear regression. In the present implementation simple forward selection is used. In concert with a second cross-validation to assess the predictive power of the selected set of metabolites, the proposed method involves two leave-one-out cross-validations and will be referred to as LOO2CV. In the second leave-one-out cross-validation a multitude of regression models is generated. This offers additional information that is potentially useful for selection of key metabolites in the spirit of stability selection. The proposed LOO2CV approach is illustrated with sensory and large-scale metabolomics data from a set of 76 different cocoa liquors. The proposed approach is compared with conventional stepwise regression and stepwise regression in concert with cross-validation for evaluation of predictive power of the model