Use of multivariate statistics in understanding wine flavor

Abstract

With the increase in sensitivity and selectivity of instruments, volatiles can now be analyzed under conditions very close to those by which humans perceive aroma. However, without sensory evaluation, even precise information about the volatile composition in the nasal passages cannot predict the flavor of the system as perceived by humans. To illustrate one facet of the complexity of flavor research, multivariate statistical methods that have been used in wine flavor studies to relate sensory descriptive analysis and volatile composition are reviewed. Principal component analysis of instrumental variables is a technique to find the fewest number of volatiles that match the configuration of the sensory profile data. Generalized procrustes analysis (GPA) and Partial least squares regression (PLS) are used to model data from different data sets, such as volatiles identified by GC-MS and intensity ratings for the sensory attributes. The GPA is used to determine if the two data sets "fit each other", whereas PLS determines underlying dimensions in each set of data that best explain the variation in the other set of data. Neither GPA nor PLS determine if any of the compounds associated with specific sensory attributes are actually responsible for specific aromas but they identify compounds on which the sensory studies should focus. These three methods are compared by the analysis of data from a study of Chardonnay wines that were profiled by descriptive analysis and the headspace volatiles identified by GC-MS, responsible for those aroma notes. However, they do 1

Relationships of Overall Estery Aroma Character in Lagers with Volatile Headspace Congener Concentrations

“…This strategy, an advance over previous static headspace analyses, seems fundamentally more similar to the mechanisms of aroma perception i.e. absorption at nasal olfactory sites 18 . Beer volatiles quantification by SPME has been validated for esters and higher alcohols with evidence of lower limits of detection, high repeatability and good linearity.…”

“…The limited range between lowest (sample 7, 40.94) and highest (sample 9, 64.13) showed a difficulty in discrimination of intensity of character in samples, assisted by the rank rating and Duncan's multiple range test (Table II) which yielded 3 groups: A, B, and C. Group A had least overall intensity of estery aroma and group C the greatest. Most (18) …”

“…However a wide range of other esters and higher alcohols in the volatiles also contribute to flavour 17 . Technical issues, notably construct definition, have limited modelling of origins of overall estery 18 .…”

“…Sensory and compositional data can be related using linear regression in simple model solutions, but in beers and wines 18 characteristic aroma characters may emerge from complex mixtures of volatiles in headspaces over solutions containing many less volatile solutes. Popular multivariate data analysis methods include partial least square regression 2,14,18,25 , principal component analysis of total 14,18,24 or edited data matrices 7 and generalised Procrustes analysis 5 .…”

“…Popular multivariate data analysis methods include partial least square regression 2,14,18,25 , principal component analysis of total 14,18,24 or edited data matrices 7 and generalised Procrustes analysis 5 . With partial least square regression (PLS) is it is possible to have more variables than samples and using a single Y variable and a matrix of predictors (X-variables), PLS has given good prediction of sensory bitterness for lagers 2 .…”