Prediction of mandarin juice flavor: a flavoromic approach.

Abstract

This thesis introduces an alternative approach to predict flavor and is referred to as flavoromics; this research adapts concepts and tools from metabolomics investigations. It is a non-targeted strategy combined with chemometrics which considers for study all (ideally) low molecular weight compounds in foods as candidate chemical stimuli in human flavor perception. The feasibility of flavoromics to predict the intensities of various flavor attributes was tested on mandarin juice. Forty-six juices were characterized by both instrumental and descriptive sensory analyses. Volatiles and non-volatiles were analyzed by headspace solid-phase micro extraction gas chromatography and solid-phase extraction ultra high performance liquid chromatography- time of flight mass spectrometry, respectively. The developed methods were a compromise between the number of compounds extracted and detected, throughput, and repeatability. The capability of distinguishing samples based on mass spectral information collected from the different instruments using chemometrics was confirmed. The descriptive sensory analysis of the mandarin juice samples revealed very different flavor profiles between and within hybrids and cultivars, and juices made from fruits with common genetic background tended to share some sensory characteristics. Compositional variations across mandarin juice samples and their sensory profile were correlated using partial least squares regression, from which different predictive models of sensory quality were developed. The explanatory and predictive performances of the models were improved when combining all instrumental data into one single data set as opposed to individual ones, indicating that each individual subset conveyed complementary information and that merging them improved the overall description of the sensory profile. The best PLS model was obtained with mid-level data fusion, for which a preliminary variable selection was done. The predictive power of the selected model was tested using a calibration and prediction sample sets. A fairly robust model was obtained and a strong relationship between instrumental and sensory measurements was observed. The resulting model showed that prediction of sensory scores was possible to a certain extent for a majority of the sensory descriptors, demonstrating the applicability of using a data-driven approach to predict flavor irrespective of whether the chemical identity of the instrumental signals was known or not.