Untargeted Flavoromics to Identify Flavor Active Compounds

Abstract

Flavor research has long utilized targeted methods to understand and identify flavor active compounds in food systems. While this has advanced the field, there are limitations to only investigating compounds that are directly flavor active. Flavor is a multi-model sensation consisting of a complex set of chemical stimuli that are perceived as a mixture, and more inclusive investigation can supplement targeted methods. Increasing the number of analytes allows for more comprehensive understanding of food systems which can lead to new discoveries. This dissertation illustrates the development of untargeted analytical methods to identify flavor active materials in citrus extracts that relate to aging. This dissertation uses an untargeted workflow to identify age related compounds, and establish their sensory significance through recombination modeling and structural elucidation. Preprocessing methods were optimized using an uneven multi-level two factor design of experiment and further untargeted modeling. Signal to noise ratio and thresholding were varied during pre-processing which produced data sets varying in size from >50,000 features to 500. Part per billion differentiation was achieved using both unsupervised (principle component analysis) and supervised (projection to latent structures) multivariate modeling, indicating a high quality analytical and statistical framework. To understand how a food system ages ethanol extracts from different citrus varietals (Navel, Mineola, and Valencia) were aged and chemically profiled using Ultra Performance Liquid Chromatography Mass Spectrometry. Machine learning (Random forest) and multivariate (projection to latent structures) models were implemented to model the age of the extracts. Varietal difference was leveraged and models were adapted to understand the age of the samples, rather than model the varietal differences. Statistically important compounds were isolated using food grade mass spectrometry directed fractionation. These isolated compounds underwent sensory evaluation using descriptive analysis in both a Solvent Assisted Flavor Evaporation (SAFE) extract of orange juice and a volatile orange flavor (VOF) model beverage. The isolated compounds showed significant impact in both tasting mediums. All compounds identified increased with sample age, and when evaluated in the SAFE extract showed significant decrease in orange character. Compound 413 showed a significant increase in cooked character and green bean character and a suppression of floral character. While compound 383 showed a significant increase in green bean character. In the volatile orange flavor (VOF) compound 413E2 and compound 457 showed an increase in sweetness over the control, which was the only noted change to the ‘taste’ attributes noted among the compounds isolated. Compound 383 showed a decrease in cooked character over the control. Compound 661 showed suppression of floral aroma over the sample blank. Compounds that positively correlated with age were reported to increase the cooked and green bean character and suppression of floral and orange character notes, which indicate degradation in flavor quality, or a deviation from fresh character. Structural elucidation using Nuclear Magnetic Resonance (1H, HMBC, TOCSY, HSQC) and accurate mass (TOF) revealed compound 693 was Nomilin 17-O-beta-D-glucopyranoside, while compounds 383 and 661 were shown to be novel compounds. The systematic name of compound 661 was (5-(((2R,3S,4R,5R)-4,5-dihydroxy-3-((3-hydroxy-3-methyl-5-oxohexanoyl)oxy)-6-(4-(3-hydroxypropyl)-2,6-dimethoxyphenoxy)tetrahydro-2H-pyran-2-yl)methoxy)-3,3-dimethyl-5-oxopentanoic acid. There are two 3-Hydroxy-3-methylglutaric acid units and a dihydrosinapyl alcohol moiety bonded to a sugar backbone. Compound 383 was identified as (3,5,5-trimethyl-4-((E)-3-oxo-4-(((2S,3S,4S,5S,6R)-3,4,5-trihydroxy-6-(hydroxymethyl)tetrahydro-2H-pyran-2-yl)oxy)but-1-en-1-yl)cyclohex-2-en-1-one) and suggested reaction product of a sugar and terpene moiety. The final piece of this research was to characterize aging in lemon extracts and identify contextual variable interactions associated statistically significant compounds. Six types of lemon extracts which were aged and profiled using Ultra Performance Liquid Chromatography Mass Spectrometry. The data was modeled using 9 different machine learning algorithms.. Random Forest produced the highest quality initial model, which went through further development and tuning. Within the random forest model number of trees, features tried at each node, max number of terminal samples were all optimized, as was the final model using both gini and entropy for decision criteria. The final model had a training fit of 0.951 and test score of 0.928(+/- 0.0049). Statistically important variables from this analysis were investigated for contextual data interactions that would illuminate data trends for further study. This approach aims to help provide additional value from untargeted flavoromics and better understand contextual interactions in data sets.