Abstract
The current study explores the relationship between sensory characteristics and wine composition of Cabernet Sauvignon wines in relation to Australian geographical indications (GIs). Descriptive sensory analysis was conducted to characterize the sensory attributes of commercially produced Cabernet Sauvignon wines from the Barossa Valley, Clare Valley, Coonawarra, Frankland River, Langhorne Creek, Mount Barker, Margaret River, McLaren Vale, Padthaway, and Wrattonbully GIs. Canonical variate analysis using the significant sensory attributes demonstrated that each GI could be distinguished from the others. A recently developed analytical method was used to analyze over 350 volatile compounds in the wines assessed, and measures of the major nonvolatile components were also determined. Compositional results were analyzed using partial least squares discriminant analysis to identify candidate components that were unique to certain GIs, including 2-isobutyl-3-methoxypyrazine, menthone, isomenthone, carvacrol, δ-octalactone, p-methylacetophenone, m-dimethoxybenzene, protein-precipitable tannin, and monomeric anthocyanins. Results demonstrate that Australian Cabernet Sauvignon wines have common sensory attributes related to geographic origin. The work also identifies a number of candidate components that are related to individual GIs which warrant further investigation. The study is the first to explore the concept of regionality in Cabernet Sauvignon wines from Australia.
Cabernet Sauvignon originates from Bordeaux, France, and is now planted worldwide, including Europe, the United States, South Africa, Chile, New Zealand, and China. In Australia, Cabernet Sauvignon is the second highest planted red grape variety, following Shiraz, accounting for 28% of the total tonnage of red winegrapes crushed in 2009 (ABS 2009), and is important to red wine blends and premium, regional, and varietal wines.
Given the numerous wine products and wine brands currently available on the international market, product differentiation is vital to remain competitive, especially in markets where price discounting is ineffective. An important aspect of product differentiation, second to variety, is the region of origin, which is holistically referred to as terroir and legally delineated as geographic indication (GI). The French term terroir is traditionally used to encompass the notion that agricultural sites within the same geographic area share similar climatic, soil, and managerial practices that all contribute to the unique characteristics of the product produced (Van Leeuwen and Seguin 2006).
The World Trade Organization recognizes GIs defined in Article 22 of the Agreement on Trade-Related Aspects of Intellectual Property Rights as “indications which identify a good as originating in the territory of a member, or a region or locality in that territory, where a given quality, reputation or other characteristic of the good is essentially attributable to its geographical origin.” It is understood that consumers associate quality, reputation, and other intrinsic characteristics of a product with its place of origin. GIs are similar to the European appellation systems, but are less prescriptive with respect to viticultural and winemaking practices.
Numerous studies have demonstrated differences in the sensory characteristics of wines from different geographic origins, including Cabernet Sauvignon, Pinot noir, and Chardonnay in California (Heymann and Noble 1987), Riesling in Germany (Fischer et al. 1999), Malbec in Argentina (Goldner and Zamora 2007), Albariño in Spain (Vilanova and Vilariño 2006), and Sauvignon blanc in New Zealand (Lund et al. 2009). However, to date there have not been similar studies conducted in Australia, where there are about 60 wine regions and over 100 GIs that are climatically and geographically diverse. Furthermore, Australian viticulturists are not constrained with regard to what varieties they can plant, with the major grape varieties, Chardonnay, Cabernet Sauvignon, and Shiraz, commonly planted in most if not all of these wine regions. In addition, there is an increasing volume of research exploring food traceability/authenticity, including wines, which uses chemometric techniques to establish relationships between product composition and geographic origin (Gonzalvez et al. 2009).
This study combines descriptive sensory analysis with the compositional results of a new HS-SPME GC×GC-TOFMS methodology. As the wines were made solely from Cabernet Sauvignon grapes under commercial winemaking conditions, the differences in composition and sensory characteristics should reflect the regional styles available to the consumer. Discriminant multivariate statistical techniques including canonical variate analysis (CVA) and partial least squares discriminant analysis (PLS-DA) have been used to categorize the wines according to their geographic origins. The current study investigates the role of geographic origin in defining the sensory characteristics of commercially produced, single-vineyard Cabernet Sauvignon wines from Australia and attempts to identify candidate compositional characteristics of Cabernet Sauvignon wines unique to the GIs studied.
Materials and Methods
Experimental design.
All wines were made from Vitis vinifera L. Cabernet Sauvignon produced from the 2009 vintage and represented 10 GIs: Barossa Valley (BV), Clare Valley (CV), Coonawarra (CW), Frankland River (FR), Langhorne Creek (LC), Mount Barker (MB), Margaret River (MR), McLaren Vale (MV), Padthaway (PA), and Wrattonbully (WR). Three wines from each of the 10 regions were assessed. All wines were sourced from commercial producers, who were requested to provide samples that best reflected Cabernet Sauvignon from that winemaking region. Compositional analysis was conducted in Perth, Western Australia, while an additional six bottles of each product were air freighted, over two days, to California in wine bottle foam packaging for sensory analysis at the J. Lohr Sensory Laboratory, University of California, Davis. All wines were made entirely from Cabernet Sauvignon grapes using commercial winemaking practices. Wine alcohol, pH, titratable acidity, malic acid, glucose, fructose, and volatile acidity were determined using a Foss Winescan Auto (FOSS, Hillerød, Denmark) equipped with a CETAC autosampler (ASX-260). Monomeric anthocyanins, small polymeric pigments (SPP), large polymeric pigments (LPP), protein-precipitable tannin, and iron reactive phenolics were analyzed using published methodology (Harbertson et al. 2003).
HS-SPME GC×GC-TOFMS volatile compound analysis.
Samples were analyzed using a HS-SPME GC×GC-TOFMS methodology that has been previously described in detail (Robinson et al. 2011a, 2011b). A Pegasus 4D GC×GC-TOFMS (LECO Corp., St. Joseph, MI) coupled to a CombiPAL autosampler (CTC Analytics, Zwingen, Switzerland) with an agitator and SPME fiber conditioning station was used for all analysis. The GC primary oven was equipped with a 30 m Varian FactorFour VF-5MS capillary column, 0.25 mm i.d. and 0.25 μm film thickness, with a 10 m EZ-Guard column (Varian, Walnut Creek, CA). This was joined using a SilTite mini-union (SGE, Ringwood, Victoria, Australia) to a 1.65 m Varian FactorFour VF-17MS capillary column, 0.10 mm i.d. and 0.20 μm film thickness, of which 1.44 m was coiled in the secondary oven. The sample headspace was sampled using a 2 cm DVB/CAR/PDMS 50/30 μm SPME fiber (Supelco, Bellefonte, PA) for 120 min at 30°C and desorbed in the GC inlet at 260°C for 1 min. GC×GC-TOFMS interrogation and spectral deconvolution was conducted using ChromaTOF optimized for Pegasus 4D software (ver. 4.24; LECO Corp.). Compound mass spectral data were compared against the NIST 2008 and Wiley Mass Spectral Libraries (9th ed.). Retention index (RI) methods were used to calculate RI for each compound identified, which were compared to published RI for 5% phenyl polysilphenylene-siloxane capillary GC columns or equivalents (Adams 2007, Stein 2009) for identity confirmation. Peak area integration was conducted using the unique ion listed in Supplemental Table 1. Peak assignments, integration, and summation of modulations were automatically conducted by the software.
Descriptive sensory analysis.
The wines were evaluated by a trained panel of 18 volunteers (five men and 13 women) between 21 and 43 years of age. All panelists had previous winetasting experience and were selected due to interest and availability. Panelists were trained with the reference standards over 10 consecutive training sessions to align panelist terminology. Reference standards were presented in black wineglasses and are listed in Table 1. Panelists were also asked to evaluate all 30 products broken into blocks of six products over five consecutive sessions to familiarize the panel with the wine samples that constituted the study. Following training, panelists were asked to evaluate each of the 30 wine products in triplicate over the course of 18 sessions, equating to five wines per session presented in a randomized block design. Prior to each formal evaluation session, the reference standards described above were assessed to refresh each panelist’s memory. Wines were assessed monadically and panelists were asked to rate attributes using a continuous unstructured scale (10 cm). A 30-sec rest was included between each sample, during which the panelist was able to refresh his or her palate with water and an unsalted water cracker. FIZZ Software (ver. 2.31G; Biosystèmes, Couternon, France) was used for data acquisition and for generating a randomized presentation order using a modified Williams Latin Square design.
Composition of sensory reference standards used to define aroma and taste attributes.
Statistical analysis.
All statistical analysis was conducted using JMP (ver. 8.0.2; SAS Institute Inc., Cary, NC). A one-way analysis of variance (ANOVA) of the normalized peak area was used to analyze the volatile composition results. A three-way ANOVA was conducted using the restricted maximum likelihood (REML) method to test the effects of Judge, Product, Replicate, and all two-way interactions for each sensory attribute using a pseudo-mixed model with the Judge x Product interaction as a denominator. Canonical variance analysis (CVA) was conducted using GI as the categorical factor and all significant sensory attributes were the variables used to describe the sensory differences between the GIs. Bartlett’s Chi-square approximation was used to determine the number of significant canonical dimensions. Partial least squares discriminant analysis (PLS-DA) was used to classify the regions (Y-variables) as a categorical variable using the normalized mean values for significant volatile components (X-variables). The Y-variable matrix was produced using a binary response (region = 1 and nonregion = 0). X-variables were normalized against the maximum value for any one product so that each variable had an equivalent influence on the PLS model. Both univariate response (PLS1) and multivariate response (PLS2) PLS models were used to explore the data. The inclusion of additional latent vectors to the PLS model progressively increases the accuracy of the model in describing the observed data. However, that may lead to overfitting, which may reduce the predictive capacity of the model. Cross-validation is typically used to determine the minimum number of latent vectors that sufficiently explain the variance of both the X- and Y-variables without over-fitting the model (Westerhuis et al. 2008). This is particularly important where the PLS model is intended to predict future Y-variable observations. The current study was explorative by using only three products from one vintage to represent each of the 10 GIs. PLS-DA was used to distinguish the 10 GIs and not to generate a predictive model (Westerhuis et al. 2008), therefore cross-validation was not conducted. The regression coefficients were used to determine which predictive (X) variables were important in modeling the response (Y) variables. The regression coefficients were assessed through a two-way hierarchical cluster analysis using a minimal variance algorithm. Cluster membership was used to interpret the relationship between the X- and Y-variables by analyzing the regression coefficients, averaged within a cluster group, using principal component analysis (PCA).
Results
Sensory analysis.
A three-way ANOVA using a pseudo-mixed model of the sensory results showed that the bell pepper, black berry, butter, canned vegetable, dried fruit, earthy, eucalyptus, floral, leather, mint, oak, red berry, smoky, vanilla, alcohol, astringent, bitter, and sour sensory attributes were significantly different (p ≤ 0.05) due to product. Least squares (LS) means of each product, for each significant sensory attribute, were compared using the Tukey-Kramer HSD test, which identified that the LS means for the black berry, leather, and vanilla sensory attributes were not significantly different among the products. Canonical variate analysis (CVA) of the 15 significantly different sensory attributes, using the GI as the grouping factor, clearly differentiated the GIs (Figure 1). Bartlett’s Chi-square approximation showed that there were four significant dimensions (p ≤ 0.05); however, the fourth dimension provided little additional information and is not presented. The first three dimensions accounted for 83% of the cumulative variance. The first dimension accounted for 53% of the variance and primarily separated the FR GI, which was positively correlated with the canned vegetable, earthy, and smoky aroma attributes, from the BV, LC, and MV GIs, which were positively correlated with the astringent taste attribute. The second dimension accounted for 21% of the variance and characterized the difference in GIs by the bell pepper aroma attribute, which was negatively correlated with the astringent taste attribute. The first two dimensions separated the CV, MB, and MR GIs from the BV, LC, and MV GIs, with the CW, MR, PA, and WR GIs significantly different from the MB GI. The third dimension accounted for 8% of the variance and was characterized primarily by the mint aroma attribute, which was positively correlated with the LC GI and separating it from the MV GI. The CW, WR, and PA regions were not separated in any of the first three dimensions.
Canonical variate analysis of sensory data for 10 geographical indications (GIs). Scores are plotted to the left (S1 & S2). Circles represent the 95% confidence limits for the mean scores of the GIs: Barossa Valley (BV), Clare Valley (CV), Coonawarra (CW), Frankland River (FR), Langhorne Creek (LC), Margaret River (MR), McLaren Vale (MV), Mount Barker (MB), Padthaway (PA), and Wrattonbully (WR). GIs that are significantly different have circles that do not overlap. Loadings for sensory terms are plotted to the right (L1 & L2). Dimensions 1 and 2 are plotted above (S1 & L1) and dimensions 1 and 3 are plotted below (S2 & L2).
PLS-DA relating GI and chemical composition.
The inclusion of additional latent vectors to the PLS1 and PLS2 PLS-DA models resulted in a significant increase in the mean % Y-variable explained (Figure 2). However, the PLS1 models were superior to the PLS2 models at explaining the Y-variables. A scatterplot matrix compares the stability of the PLS1 and PLS2 model correlation coefficients for the CV GI PLS-DA models (Figure 3). The inclusion of additional latent vectors to the PLS1 model resulted in progressively greater stability in the correlation coefficients where the PLS2 models changed unpredictably, indicating they were unstable. The inclusion of four latent vectors to the PLS1 models accounted for 96 to 99% of the cumulative variance of the Y-variables. Comparison of the X-variable correlation coefficients between the four and five latent vector models had an average R2 of 0.99 (± SE 0.001). Subsequent analysis was conducted using the four latent vector PLS1 PLS-DA model, as they best explained the Y-variables with no significant change in the X-variable correlation coefficients. Two-way hierarchical cluster analysis of the correlation coefficients was conducted to identify which X-variables were unique to the PLS-DA models for the Y-variables (Figure 4). The X-variables were grouped into 30 cluster groups that were a mixture of small and large groups, ranging from clusters 5, 7, and 9 (with four X-variables each) to cluster 3 (with 29 X-variables).
Bar graph of mean % Y-variable explained (± standard error) for each additional latent vector. PLS1 (dark grey background), PLS2 (light grey background).
Scatterplot matrix of correlation coefficient values for latent vectors 1 to 5 (LV1 to LV5) for the Clare Valley (CV) GI. PLS1 and PLS2 models are compared. The scale for PLS1 is 10 times the scale for PLS2. Points represent correlation coefficients for all 303 significantly different volatile compounds and 12 wet chemistry values presented in Supplemental Table 1. Ellipses represent the 95% bivariate normal density of the points.
Two-way hierarchical cluster analysis of X-variable correlation coefficients from PLS-DA of regions. The two-way color map is scaled from blue (low) to grey to red (high) for each attribute. Dendogram scales are distance scales with X-variable clusters colored in the dendogram on the right to distinguish different groupings. Cluster groups are numbered 1–30 in descending order from top to bottom and correspond with clusters listed in Supplemental Table 1.
Principal component analysis of the average correlation coefficient within each cluster group was conducted to visualize which X-variable clusters were positively or negatively correlated with the Y-variables (Figure 5), but should be interpreted in conjunction with the PLS model correlation coefficients in Supplemental Table 1 (detailed information for the results below can be found in the table). The first four principal components (PC) accounted for 62% of the total cumulative variance, with the first PC accounting for 18% of the variance and primarily separating the CV GI, which was negatively correlated with the FR GI. 2-Methylthiolan-3-one (cluster 12) was important in the PLS-DA model of the CV GI, with variables in clusters 12, 13, 24, 25, and 26 positively correlated while variables in clusters 5, 10, 11, 20, 29, and 30 were negatively correlated. The PLS-DA model for the FR GI was positively correlated with variables in clusters 19, 20, 21, and 23, including decanal, anthocyanins (cluster 19), and isomenthone (cluster 21), and negatively correlated with variables in clusters 5, 7, and 9, including volatile acidity (cluster 9) and p-cresol (cluster 7).
Principal component analysis of the average PLS-DA correlation coefficient values for each cluster of X-variables (compositional attributes). PLS-DA Y-variable (region) scores are plotted on the left and cluster factor loadings are plotted on the right. Y-variable symbols represent two-way clusters as presented in Figure 4 and Supplemental Table 1.
The second PC accounted for 15% of the variance and differentiated the PA GI, which was positively correlated with variables in cluster 18, from the LC and WR GIs, which were positively correlated with variables in cluster 2. The PLS-DA model for PA was defined strongly by variables in cluster 30, including δ-octalactone, 1-pentadecene, and cumene, and was positively correlated with variables in clusters 24, 28, 29, and 30 and negatively correlated with variables in clusters 5, 7, and 20. The PLS-DA model for the LC GI was positively correlated with variables in clusters 1, 2, and 20 and negatively correlated with variables in clusters 26 and 27, including α-amorphene (cluster 27) and β-ionone (cluster 26). The PLS-DA model for the WR GI was positively correlated with variables in clusters 2, 7, 11, and 29, including ethyl 2-hexenoate and naphthalene and dill ether (cluster 11), and negatively correlated with variables in clusters 16, 24, 25, and 30, including ethyl phenylacetate (cluster 24).
The third PC accounted for 15% of the variance and differentiated the CW and MV GIs from the CV GI primarily due to the variables in clusters 5 and 15. The PLS-DA model for the CW GI was positively correlated with variables in clusters 4, 5, 8, and 16, including benzenepropyl acetate and phenethyl formate (cluster 5), and negatively correlated with variables in clusters 12, 13, and 22. The PLS-DA model for the MV GI was positively correlated with variables in clusters 14, 15, and 18 and negatively correlated with variables in clusters 9, 21, and 25, including (Z)-oak lactone and (Z)-β-damascenone (cluster 21).
The fourth PC accounted for 14% of the variance and characterized the MV GI, which was negatively correlated with the BV and CW GIs. The PLS-DA model for the BV GI was positively correlated with variables in clusters 23, 26, and 27, including 4-hydroxy-3-methylacetophenone and p-methylacetophenone (cluster 27), and negatively correlated with variables in clusters 1, 12, and 14. The MB and MR GIs were not well described by the first four PCs; however, they were differentiated from the other GIs. The PLS-DA model for the MB GI was positively correlated with variables in clusters 7, 8, 9, and 10, including benzaldehyde and m-dimethoxybenzene (cluster 10), and negatively correlated with variables in clusters 19 and 24, although other individual variables had larger negative correlation coefficients, such as geraniol (cluster 12). The PLS-DA model for the MR GI was positively correlated with variables in clusters 6, 12, 23, and 24, including carvacrol, undecane, and m-dichlorobenzene (cluster 6), and negatively correlated with variables in clusters 4, 21, and 25. Variables in clusters 3 and 17 either were not all positively or negatively correlated with any of the Y-variables or had small correlation coefficients. Most variables within cluster 3 were positively correlated with the PLS-DA model for the LC GI and negatively correlated with the PLS-DA model for the MB and CV GIs, while most variables in cluster 17 were negatively correlated with the LC and MR GIs and positively correlated with the MV and MB GIs.
Discussion
Cabernet sensory attributes.
The concept of regionality and terroir implies that wines from a common GI should have similar sensory characteristics. The current study assessed the sensory characteristics of 30 single-vineyard wine products that originated from 10 GIs to determine if there are limitations to this concept in Australia. While these observations are initially limited to the number of samples used from a single vintage, there are some clear compositional and sensory differences among the GIs assessed. The FR GI was clearly discriminated from the other GIs (Figure 1), while separation of the other GIs from one another was somewhat convoluted. The FR GI was positively correlated with the canned vegetable, earthy, and smoky aroma attributes, which were driven by two of the FR samples that were highest in these attributes. The BV, LC, and MV GIs were positively correlated with astringency and negatively correlated with MB and FR GIs, which were low in astringency. As factors that influence astringency were not controlled in the current study, it is likely that the variations in astringency observed among GIs were influenced by climate and/or viticultural management practices together with enological practices, including fermentation temperature and skin contact time. It is possible that the protracted drought during the last 12 years in southeastern Australia (Webb et al. 2010) (an area that includes the BV, LC, and MV GIs) could have increased the wine phenolic content, resulting in an increase in perceived astringency, perhaps through an increase in relative skin mass that can occur in berries that have experienced water deficit (Roby and Matthews 2004) or phenolic content could have increased due to reduced canopy vigor (Dry and Loveys 1998) coupled with increased fruit light exposure (Cohen and Kennedy 2010). Further assessment of differences in phenolic content among GIs is warranted to discern what unique environmental or managerial aspects drive differences in astringency due to grape phenolic content and extractability.
The second dimension separated GIs due to bell pepper aroma, with the MB GI samples higher in this attribute than the MV and LC samples. Previous research has noted vegetative sensory attributes in Cabernet Sauvignon wines from cooler areas (Heymann and Noble 1987), where levels of 2-isobutyl-3-methoxypyrazine (IBMP), associated with bell pepper aroma in Cabernet Sauvignon, have been positively correlated with lower mean January temperatures in Cabernet Sauvignon wines (Allen et al. 1994). These previous studies support the observation that the MB GI samples exhibited more bell pepper characteristics due to elevated levels of methoxypyrazines which may also contribute to the canned vegetable characteristic of the FR GI samples. However, this conclusion would imply that the CW, PA, and WR GIs would have also exhibited this character; the observation that they did not may be attributed to the 2009 vintage, during which South Australia (BV, CV, CW, LC, MV, PA, and WR GIs) had two periods of exceptionally high temperatures, 28–31 Jan and 6–8 Feb, while Western Australia (MB, MR, and FR GIs) had average weather conditions (Webb et al. 2010). It is highly likely that these extreme temperature events resulted in changes to fruit composition (Downey et al. 2006, Jackson and Lombard 1993), subsequently reflected in the sensory characteristics of the wines produced in the South Australian GIs.
The third dimension separated the LC GI from the MV GI due to the mint and eucalypt aroma attributes, which reinforces anecdotal evidence that Cabernet Sauvignon wines from the LC GI exhibit a distinctive mint/eucalypt characteristic. This separation explained the least of the variance of the three dimensions assessed, and it is worth noting that only one of the three LC GI samples exhibited a strong mint/eucalypt aroma compared to the other wines. Recent research has indicated that eucalypt aroma in red wine is contributed by eucalyptol (1,8-cineol) (Farina et al. 2005, Van Leeuwen et al. 2007), which is primarily attributed to airborne transmission of eucalyptol from Eucalyptus trees to grapevines (Capone et al. 2010). The mint/eucalypt sensory attribute of the LC GI wines did distinguish at least one of the Cabernet Sauvignon wines from the LC GI and warrants further investigation to elucidate the origin of this characteristic and to confirm the role of climate, viticultural management, and winemaking that may influence this characteristic.
The final observation from the sensory analysis was that the CW, PA, and WR GIs were not statistically differentiated in any of the first three dimensions of the analysis. The samples from these three GIs were clearly distinguished from the other samples, but there were commonalities among the three samples representing these GIs that made it difficult to differentiate them. There are several possible explanations. First, commonality between samples in other GIs, such as the FR GI, and/or differences among other GIs are greater compared to the CW, PA, and WR GIs, making them indistinguishable from one another. Second, the three regions are geographically close, which could suggest that the lack of geographic separation has resulted in more commonality among the CW, PA, and WR GIs than among the GIs that are geographically distant. Third, the products donated for the current study from these three GIs are less reflective of each respective GI and are more expressive of the markets to which the producers intend the wines to appeal. Subsequent canonical variate analysis of the CW, PA, and WR GIs only, using the same sensory attributes, demonstrated that the sensory results can discriminate the three GIs from one another (data not presented). This analysis identified that the CW GI was positively correlated with the mint and canned vegetable attributes, the PA GI was positively correlated with the red berry and sour attributes, and the WR GI was positively correlated with the smoky and butter attributes (data not presented). This supports both the first and second explanations that the differences among the other GIs is greater than the differences among the CW, PA, and WR GIs, which could be attributed to geographical separation. Similar observations can be made when drawing comparisons between other GIs assessed in the current study. For example, the BV, MV, and LC GIs that are geographically close are closely associated in the CVA analysis in Figure 1. Subsequent analysis of the BV, MV, and LC GIs indicates that they are clearly differentiated from one another, with the BV GI positively correlated with red berry and dried fruit attributes, the MV GI positively correlated with the smoky and oak attributes, and the LC GI positively correlated with the mint and eucalypt attributes (data not presented).
The results of this sensory analysis demonstrate that each of the GIs can be clearly discriminated from one another and that the statistical differences are more pronounced between GIs that are geographically removed from one another. This phenomenon may be specific to Australian GIs, as no comparison exists in the literature. A previous study identified that Cabernet Sauvignon wines from higher altitudes had higher bell pepper aroma while wines from lower altitudes had higher red fruit and jam aromas (Falcão et al. 2007). However, altitude was highly negatively correlated with temperature, suggesting that the observation from this study was temperature-dependent rather than altitude-dependent. This observation is similar to a previous study conducted in California where it was found that Cabernet Sauvignon wines originating from cooler growing areas were more vegetative compared to wines originating from warmer areas, which were more fruity (Heymann and Noble 1987). However, the results of the current study are limited to the samples assessed, which did not take into account vintage-related influences. Further assessment of the variation within a region is also vitally important in confirming whether particular sensory attributes truly define a GI. Previous research finds that these variations appeal to different market segments, and an awareness linking wine composition, sensory characteristics, and consumer preferences is important in producing products for specific markets (Lattey et al. 2010). Assessment of the commonalities within GIs and differences among GIs warrant further study to better establish what intrinsic sensory characteristics are indicative of wines produced from Australia’s unique GIs.
Relationship between sensory attributes and composition.
Principal component analysis (PCA) is well suited to sample classification in scenarios where the variability between groups is vastly larger than the variability within groups; however, in some instances the group-to-group differences may not dominate the total variability as measured by the sample variance/covariance matrix (Barker and Rayens 2003). Linear discriminant analysis (LDA) is a suitable method for distinguishing the difference between among-group and within-group variability of samples; however, LDA cannot be applied in instances where the number of variables are greater than the number of samples (Barker and Rayens 2003). The solution is to use partial least squares discriminant analysis (PLS-DA), a linear multivariate model where continuous variables (X-variables) are related to a categorical membership (Y-variable) (Barker and Rayens 2003). PLS-DA is suitable for discriminant analysis where dimension reduction is required, where the number of variables are substantially greater than sample observations, and where data exhibits strong collinearity (Wold et al. 2001). The compositional results from the current study were assessed using PLS1 and PLS2 models where binary variables (1 and 0) were used to distinguish each GI. PLS1 models described a greater percentage of the Y-variable variance with the addition of fewer latent vectors (Figure 2). This was coupled with the observation that the correlation coefficients for PLS2 models were relatively unstable with the inclusion of additional latent vectors compared to those from PLS1 models (Figure 3). As the objective of this analysis was to explore how the GIs (Y-variables) were related to the compositional data (X-variables), subsequent analysis was conducted using PLS1 models. Statistical validation tools such as cross-validation, jack-knifing, and permutation tests are commonly used to assess the performance and stability of PLS-DA statistical models (Kjeldahl and Bro 2010, Westerhuis et al. 2008). However, it has been established recently that in studies where megavariate data sets are assessed (number of variables greater than 10 times the number of observations), these validation tools are less useful (Rubingh et al. 2006). As the current study used three wines to represent each GI, it is not possible to separate the data set into a training set, validation set, and a test set, which would be necessary to validate the PLS-DA model (Westerhuis et al. 2008). Therefore it should be made clear that the results of the current analysis are not predictive but explorative and aid in identifying compositional components that are candidates for classifying GIs within Australia for these wine regions and wines.
The PLS-DA correlation coefficients distinguished the GIs assessed. The BV GI was positively correlated with 4-hydroxy-3-methylacetophenone, which has been reported in green coffee bean extracts (Lee and Shibamoto 2002), and p-methylacetophenone, which has a hawthorn / sweet / mimosa / coumarin / cherry aroma (www.thegoodscentscompany.com). To the authors’ knowledge, neither of the compounds has previously been reported in Cabernet Sauvignon wines and it is unknown what role they may play. The MB GI was positively correlated with the m- and o-dimethoxybenzenes (Supplementary Table 1), which have aroma characteristics of sweet / cooling / medicinal / root beer-like and of sweet / creamy / vanilla / phenolic / musty, respectively (www.thegoodscentscompany.com). m-Dimethoxybenzene has been studied in Portuguese grapes and wines and it was suggested that the compound could contribute to the aroma of port wines at peri-threshold levels through interaction effects with other volatile compounds (Rogerson et al. 2002). Both the FR and MB GIs were positively correlated with menthone and isomenthone, which have aroma characteristics of minty and of mentholic / cooling / minty / camphoreous, respectively (www.thegoodscentscompany.com). The menthone and isomenthone X-variables were not clearly identified in the PLS-DA models, as these two compounds could not distinguish the FR from the MB GI. PLS-DA analysis, which combined the FR and MB GIs as one group, identified menthone and isomenthone as positively correlated and protein-precipitable tannin as negatively correlated with the FR and MB GI group (data not presented). This result suggests that the PLS-DA models were able to identify X-variables that were unique to each GI but reduced the importance of X-variables that were important to a group of GIs. This also supports the CVA analysis indicating that the FR and MB GIs were less astringent due to lower tannin levels in comparison to the other GIs. It was also interesting that the FR GI was positively correlated with the monomeric anthocyanins while also lower than other GIs in protein-precipitable tannins and iron-reactive phenolics. The MR GI was positively correlated with carvacrol (Supplementary Table 1), which is described as spicy / cooling / thymol-like / herbal / camphoreous (www.thegoodscentscompany.com), which to the authors’ knowledge has not been studied in wine. Carvacrol is a monoterpenic phenol found in essential oils made from Origanum, Thymus, Coridothymus, Thymbra, Satureja, and Lippia species, which are collectively referred to as oregano (Baser 2008). Thymol, the structural isomer of carvacrol, was tentatively identified but did not play a strong role in the PLS-DA models presented here. IBMP was positively correlated with the CV and MR GI models; however, it was also found at similar levels in wines from CW, FR, MB, PA, and WR GIs, suggesting that it was not indicative of any one GI in the current study. PLS-DA analysis, which combined the BV, MV, and LC GIs as one group, identified IBMP as the major X-variable that was negatively correlated with the BV, MV, and LC GI group (data not presented). IBMP is recognized as the major impact compound associated with bell pepper aroma in Cabernet Sauvignon wines (Allen et al. 1994) and, based on previous research, is expected to be negatively correlated with higher mean January temperatures characteristic of the BV, MV, and LC GIs (Robinson 2011a). This is another example where particular X-variables were considered less important in defining one specific GI, as they characterized multiple GIs that were close to one another geographically. δ-Octalactone, which is characterized as sweet / coconut / creamy / coumarin / lactonic (www.thegoodscentscompany.com), was positively correlated with the PA GI. Although δ-octalactone has not been discussed in the literature in relation to wine aroma, it has been previously identified as an aromatic constituent of Cabernet Sauvignon wines (López et al. 1999). δ-Lactones are generally discounted in importance compared to the λ-lactones, which tend to have odor thresholds an order of magnitude lower for compounds of similar molecular weight (Ferreira et al. 2000). δ-Decalactone and δ-dodecalactone did not have strong correlation coefficients for any of the other PLS-DA models in the current study, indicating that δ-octalactone was accumulated in the PA wines for reasons unknown to the authors.
The PLS-DA analysis presented here has been used to identify some candidate compositional components that were useful in discerning the origin of samples from the 10 GIs. A number of other candidate compounds were identified as important X-variables in defining various GIs, including 2-methylthiolan-3-one, p-cresol, 1-pentadecene, cumene, α-amorphene, α-cedrene, and β-ionone. However, the current analysis could not conclude that these candidates were particularly unique to any one GI, as a number of these X-variables were particularly high or low in one of the three samples representing the GI, suggesting that they might be important in discriminating the GI from other GIs. However, additional observations are necessary to confirm that these candidate components are not simply outliers. Observations from this analysis are limited to the samples assessed and clearly demonstrate the potential for assessing compositional characteristics that define particular GIs in Australia. However, the analysis also reveals limitations due to the number of samples compared to the number of variables. There were clear examples where PLS-DA models undervalue compositional components that are important in defining multiple GIs that are geographically close and overvalue compositional components that are particularly high in one of the three samples. Nevertheless, the current study will help direct future studies designed to confirm the importance of these components to their respective GIs through additional observations.
Conclusion
Discriminant analysis techniques were used to identify the underlying sensory and compositional attributes that define commercially produced Cabernet Sauvignon wines from 10 GIs within Australia. The work has clearly identified that each GI could be distinguished due to the sensory characteristics of the wines assessed. That was also the case for the PLS-DA models of compositional components where the GIs could be defined by some candidate X-variables that were unique for particular GIs and groups of GIs that were geographically close. Further work is required to capture and analyze the variation in wines reflective of multiple vintages and different stages of wine development or maturation. A limitation placed on discussing the results of the current study was that little was known or controlled in the enological production of the wines studied. Differences among GIs were observed despite this limitation, indicating that future studies using greater control over enological factors are likely to demonstrate the role of site on wine composition and sensory characteristics. Such findings will lead to a better understanding of what defines wines of a specific region and the stability of these relationships between seasons and over time. Future research should also consider the impact of climate change, including the impacts of extreme events, which may result in changes to these characteristics in the longer term. Nevertheless, this study provides a basis for future research to better understand the variation of compositional components and sensory attributes within Australian GIs and to develop models that better explain these variations due to geographic proximity that are likely to be associated with similar climatic conditions or soil characteristics.
Acknowledgments
Acknowledgments: This research project was funded by Australia’s grapegrowers and winemakers through their investment body, the Grape and Wine Research and Development Corporation, with matching funding from the Australian Federal Government. This work was conducted as a collaboration between Murdoch University, CSIRO Plant Industry, and the University of California Davis. The GC×GC TOFMS was purchased through an Australian Research Council Large Equipment Infrastructure and Facilities grant.
The authors thank the industry donors at Brown Brothers Milawa Vineyard, Casella Wines, Ferngrove, Treasury Wine Estates, Houghton Wines, Howard Park Wines, McWilliam’s Wines, Orlando Wines, and the Yalumba Wine Company for providing wine samples. The authors also thank Bruce Peebles (Murdoch University), Claire McAllan (Houghton Wines, Australia), Kim Mosse (Monash University, Australia), and Steven Nelson and Kevin Scott (University of California, Davis) for assistance in various elements of laboratory analysis.
Footnotes
-
Supplemental data is freely available with the online version of this article.
- Received February 2012.
- Revision received April 2012.
- Accepted June 2012.
- ©2012 by the American Society for Enology and Viticulture
Literature Cited
Vol 63 Issue 4
