Discrimination of white wine ageing based on untarget peak picking approach with multi-class target coupled with machine learning algorithms.

The complexity of the chemical reactions occurring during white wine storage, such as oxidation turns the capacity of prediction and consequently the capacity to avoid it extremely difficult. This study proposes an untarget methodology based on machine learning algorithms capable to classify wines according to their "oxidative-status". Instead of the most common approach in statistics using one class for classification, in this work eight classes were selected based on target oxidation markers for the extraction of relevant compounds. VIPS from OPLS-DA and mean decrease accuracy from random forest were used as feature selection parameters. Fifty-one molecules correlated with 5 classes, from which 23 were selected has having higher sensitivities (AUC > 0.85). For the first time to our knowledge hydroxy esters ethyl-2-hydroxy-3-methylbutanal and ethyl-2-hydroxy-4-methylpentanal were found to be correlated with oxidation markers and consequently to be discriminant of the wine oxidative status.