Abstract
The aroma models of California Chardonnay wines were developed and evaluated using multivariate statistical procedures to investigate the sensory significance of odor-active (OA) compounds previously screened by gas chromatography/olfactometry. Partial least squares regression (PLSR) analysis was used to find the relevant combinations of OA compounds, representing aroma properties of wines determined by descriptive analysis. To test the ability of combinations of OA compounds to reproduce the aromas of wines, the four wines with the most different aromas by descriptive analysis were examined. Two combinations of OA compounds determined by PLSR and sensory testing were added to a neutral base wine at the concentrations found in the original wines. By similarity rating, the aromas of the original wine, the base wine, and two spiked aroma models were compared pair-wise. Similarity data were analyzed by multidimensional scaling. For one wine intense in fruit-related aroma attributes, addition of OA compounds produced an aroma more similar to the original wine than to the base wine. For the other three wines, although the spiked aroma models differed from the neutral base wine, none was more similar to the original wine than to the base wine, suggesting the importance of unidentified compounds.
Wines typically have more than 700 volatiles, of which only a small subset contributes significantly to the perceived aromas of wine. Gas chromatography/olfactometry (GC/O) has been most widely used technique for the detection of potent odorants, which have a high probability of contributing to wine aroma (Aznar et al. 2001, Escudero et al. 2004, Ferreira et al. 2002, Guth 1997, Moio et al. 1994). However, because of limitations of GC/O such as matrix effect and sensory adaptation (Acree 1993), the information from GC/O cannot be directly used to interpret aroma profiles or patterns of given food systems. Thus, multivariate statistical methods are often used to determine which volatiles best model the sensory profiles (Noble and Ebeler 2002). The ultimate aim is to understand how differences in sensory properties among a range of samples are caused by variations in chemical composition. Among various multivariate techniques, partial least squares regression (PLSR) analysis has been widely used to investigate the relationship between sensory and GC data sets (Fischer et al. 1996, Noble and Ebeler 2002). This technique has been previously used in the modeling of the aroma of wines of the Traminer variety (Bertuccioli et al. 1997), the measurement of the influence of the dealcoholization in the aroma of the wine (Fischer et al. 1996), and the prediction of sensory properties of aged Spanish red and white wines (Aznar et al. 2003, Campo et al. 2005).
Although these statistical techniques identify compounds that mathematically model the sensory profiles, sensory tests are necessary to determine whether these compounds contribute significantly to the characteristic aromas of the systems being studied. After the potent odorants were selected by GC/O and then quantified, aroma models duplicating the composition of a given sample were prepared and compared to the originals by various sensory tests, such as difference test, simple similarity scaling, or descriptive analysis. For the preparation of aroma models, odor activity values of odorants were used as major criteria (Grosch 2001). This approach was successfully applied in the study of aromas of orange juice (Buettner and Schieberle 2001), bread crust (Zehentbauer and Grosch 1997), and roasted coffee (Grosch and Mayer 2000). However, in studies of aged Spanish red wines and white Maccabeo wines, the aroma models did not show apparent similarities to the original wines (Ferreira et al. 2002, Escudero et al. 2004). For Maccabeo white wines in particular, aroma compounds with high odor-activity values did not have a major impact on the aroma of a given model system. A different strategy using PLSR analysis was applied for the construction of aroma models in this study.
In a previous study, 81 odor-active (OA) peaks were detected in 19 California Chardonnay wines by GC/O-mass spectrometry (MS) (Lee and Noble 2003). The objective of the present study was to determine the sensory significance of these OA volatiles in two steps. First, different subsets of the OA compounds that significantly modeled the aroma profiles of the wines were selected using PLSR analysis. Second, in similarity tests, the aromas of model wines spiked with compounds from the subsets were compared by multidimensional scaling (MDS) in four wines.
Materials and Methods
Chemicals and aroma compounds.
All reagents were of analytical grade. For the OA compound stock solution, potassium hydrogen tartrate (KHT), tartaric acid, and deactivated charcoal were purchased from Sigma-Aldrich Chemical Co. (St. Louis, MO). Reference compounds obtained from Sigma-Aldrich included ethyl butanoate, ethyl isopentanoate, 2-methyl-1-propanol, isoamyl acetate, isoamyl alcohol, ethyl hexanoate, acetoin, 1-hexanol, acetic acid, furfural, 2-ethyl hexanol, 2-acetyl furan, linalool, butanoic acid, 3-methyl-butanoic acid, furfuryl alcohol, α-terpineol, 2-phenylethyl acetate, 2-methoxy phenol, oak-lactone (mixture of cis and trans forms), ethyl cinnamate, eugenol, 4-vinylguaiacol, and vanillin. 4-Ethylguaiacol and γ-nonalactone were from Pfaltz & Bauer, Inc. (Waterbury, CT). All reference compounds were of greater than 99% purity.
Wines.
Nineteen commercial 1997 Chardonnay wines from California (Table 1⇓) were evaluated by instrumental and sensory tests. Fractionated volatile extracts of the wines were evaluated by gas chromatography/olfactometry (GC/O) and gas chromatography-mass spectrometry (GC-MS) (Lee and Noble 2003). The aromas of the same 19 wines were profiled by descriptive analysis in which the intensities of eight aroma terms (peach/apricot, citrus, floral, caramel, butter, vanilla, spice, and oak) were rated in triplicate by 14 trained judges (Yegge and Noble 2001). The four wines (JL, CDB-C, CAL, and DEL) that had the most different aromas of the set of 19 wines were selected for the similarity testing (Table 1⇓). Their sensory plots are shown in Figure 1⇓. The neutral white wine used as a base wine for evaluation of selected OA volatiles in similarity tests was an unoaked Chardonnay made from grapes from the University of California, Davis, Tyree vineyard, with pH 3.43 and 13% v/v ethanol.
Selection of odor-active volatiles.
Initial screening by PLSR.
Of the 81 OA peaks detected by GC/O in 19 wines (Lee and Noble 2003), a subset of 63 identified compounds was examined for the selection of OA compounds for sensory evaluation. The initial PLSR was performed using the peak areas of all 63 identified compounds as the independent variables (X matrix), with the mean intensity ratings of eight aroma attributes used as the dependent variables (Y matrix). In the initial PLSR, the first dimension explained 17% and 64% of the variation in GC peaks and aroma terms, respectively. However, only 28 of the 63 peaks were significantly modeled, as determined by the uncertainty estimation from jack-knifing that compared the partial model parameter estimates from cross-validation with estimates from the full model (Westad et al. 2003, Martens and Martens 2000). A second PLSR was performed using these 28 GC peaks to model eight sensory terms (Figure 2⇓). Different iterations excluding the least important variables in the second PLSR model were further run to identify the model with the best prediction ability measured by cross-validation. Compounds that varied significantly across the 19 wines or had relevant odor qualities by GC/O were also retained in the model. A third PLSR model (Figure 3⇓) using 14 GC peaks was selected by excluding the 14 compounds that were most poorly modeled in the second PLSR (Figure 2⇓). A fourth PLSR model (Figure 4⇓) was determined using 15 OA compounds selected from the initial 28 GC peaks in the second PLSR model. The sets of OA compounds from the three selected PLSR models, respectively, were considered for evaluation in similarity tests.
Preliminary sensory evaluation.
For each of the four wines evaluated in the similarity testing, two combinations of OA compounds were selected for the evaluation in the base wine (Table 2⇓). Initially, for each parent wine, the base wine was spiked with the compounds used in the second, third, and fourth PLSR at the concentrations in which they occurred in each wine. The aromas of the spiked wines were informally evaluated along with the original wine by moderately experienced Chardonnay drinkers who smelled each model system blind and described the sensory characteristics. Although 28 GC peaks were modeled in the second PLSR as described above, in JL and CDB-C, one of the peaks was a mixture of 3-methyl-butanoic acid and furfuryl alcohol. In CAL and DEL wines, this peak only contained 3-methyl-butanoic acid. Informal evaluation of the model wines spiked with 28 or 29 OA compounds revealed little resemblance to the four parent wines. Moreover, the spiked solutions had strong, unpleasant off-odors. Therefore, this model was excluded from the formal sensory similarity testing. For two wines (JL and CDB-C) the aromas of solutions spiked with 14 compounds in third PLSR were more similar to those of the original wines than the aromas spiked with 15 compounds in fourth PLSR. Thus, those solutions with 14 compounds were selected for the formal sensory tests of JL and CDB-C wines. However, for CAL and DEL, the solutions spiked with these 14 compounds were too intense in oak and spice aromas. Instead, the aromas of the solutions spiked with 15 compounds in the fourth PLSR were considered to be closer to those of the originals; thus this combination was selected for the further study in CAL and DEL wines.
In additional informal sensory evaluations, a second combination of volatiles was chosen for each parent wine (Table 2⇑). For the fruitier wines, DEL and CAL, six and seven compounds, respectively. with fruity and floral aromas were chosen. For the oaky wines, CDB-C and JL, three and seven oaky, spicy compounds, respectively (selected from initial 28 OA peaks) were added to the initial 14 compounds.
Similarity testing.
Samples.
The composition of the spiked samples evaluated in similarity tests is shown in Table 2⇑. All spiked aroma models, the base wine, and the four parent wines were stored at 10 to 11°C before use, then held at room temperature for one hour before serving.
Sensory protocol.
Ten females and two males (between 21 and 50 years) were recruited for the similarity tests. All were students and staff at the University of California, Davis and were naive to sensory experiments. Each panelist completed nine sessions, including one training session over the course of two weeks. Panelists participated in the training session to familiarize themselves with similarity rating methodology. Panelists were presented with all possible paired combinations of the original wine, the base wine, and the two spiked combinations in order to compare the similarity of the aromas. In each session, the six pairs of samples for one parent wine were presented in a random order, with the order of samples within a pair also randomized. Panelists rated the similarity of the aromas for each pair using a 9-point scale (1 = identical, 9 = completely different). Panelists were asked to sniff a cup of deionized water between pairs. The stimuli were presented in coded 30-mL aliquots in clear tulip-shaped wineglasses covered with petri dish lids. Evaluations were made in sensory booths at room temperature under white light. A new bottle of the parent wine was opened for each session. All comparisons were made in duplicate.
Data analysis.
PLSR.
In all regressions, the OA volatiles were standardized (mean divided by standard deviation), while the aroma intensity ratings were assigned a weighting of 1.00 (e.g., not standardized). All PLSR analyses were conducted using PLSR2 algorithm with full cross-validation using Unscrambler version 7.6 (CAMO A/S, Trondheim, Norway).
MDS of similarity tests.
MDS was used to investigate overall similarities of aroma models to the originals. After overall similarity of each pair of samples was rated by panelists, MDS represented the perceived proximities among the samples as a geometric configuration of points on a two- or three-dimensional map or space (Schiffman and Knecht 1993). Similarity ratings were considered to be less biased then rating specific attributes, since there is no need to predetermine attributes that could discriminate between samples in a set (Lawless and Heymann 1998).
For each original wine, the similarity data were collected and results were summed over replications and panelists. Averages were submitted as similarity square matrices for analysis by metric MDS using SAS version 6.12 (SAS Institute, Cary, NC). The badness-of-fit criterion was used to determine the aptness of the resulting consensus configuration. A badness-of-fit value close to 0 is an indication of a good fit (Wilkinson 1990).
The MDS data set was also submitted to a three-way analysis of variance (ANOVA) for the main effects of judge, replication, and similarity rating and a two-way interaction (judge x rating) using generalized linear models (PROC GLM procedure in SAS). All similarity ratings in MDS data sets varied significantly across samples (p < 0.01 to 0.001), while there were no significant differences for replications or interactions. A T-test was also performed on the scaling data to see if there were any significant differences among the distances in the tested samples.
Results
PLSR models of OA compounds.
The second PLSR using 28 peaks to model the eight sensory attributes yielded a one-dimensional model that accounted for 27% and 60% of explained variance in the GC and the sensory data, respectively. The correlation loadings for the sensory and instrumental variables and the wine scores are shown in Figure 2⇑. The first dimension contrasts the wines (DEL, CAL, SH) that had high intensities of fruity/ floral aromas with the wines that had oak-related attributes (JL, CDB-C, CDB-A, VMEres, PR, TES). Compounds extracted from oak barrels during wine aging (the oak lactones, vanillin, 4-ethylguaiacol, and eugenol) and 2-acetyl furan are strongly associated with oak, spice, and vanilla descriptors.
The center ellipsoid in Figure 2⇑ indicates 50% explained variation, which means that all GC peaks located inside this circle were poorly modeled and mathematically do not explain variation in the sensory data. After excluding the poorly modeled peaks, the third PLSR using 14 peaks yielded a better model, explaining 49% of the variance in the GC peaks and 62% of the variance in the eight sensory terms in the first dimension. The loadings for the eight sensory and 14 instrumental variables and wine scores are shown in Figure 3⇑. All peaks except linalool, α-terpineol, and furfural were significantly modeled by the uncertainty test (Martens and Martens 2000) (p < 0.05). The PLSR configurations using 14 and 28 peaks were very similar. Both plots showed a distinct separation between the wines with fruity and floral notes (CAL, DEL, SH) and those with oaky and spicy notes (JL, CDB-C, CDB-A, VMEres) along the first dimension.
In the fourth PLSR model (Figure 4⇑), 2-methoxy phenol, which gave a strong phenolic odor, was excluded and ethyl hexanoate and ethyl butanoate were added. The one-dimensional model is very similar to that in third PLSR model (Figure 3⇑), with the first dimension explaining 45% and 63% of the variance in the GC and sensory data, respectively. Ethyl hexanoate and butanoate were not significantly modeled. However, these compounds were retained in the model because they showed the related aromas of investigated wines (Figure 1⇑). Linalool was highly associated with the “floral” sensory attribute. Fruity esters such as isoamyl acetate and 2-phenylethyl acetate were located close to the peach and citrus attributes. Similarly, oaky odor compounds (4-ethylguaiacol, eugenol, 2-acetyl furan, and [cis] oak-lactone) were associated with oak and spice sensory attributes. Expectedly, vanillin was more closely associated with the vanilla sensory attribute. All three models showed a good separation of wines according to their sensory characteristics (fruity/floral versus oak/spice wines) (Figures 2⇑, 3⇑, 4⇑).
Similarity testing.
From the MDS of the similarity ratings for the paired samples, two-dimensional consensus configurations with fair to good fits were generated for each wine. The consensus wine solutions for JL and CDB-C provided a fair fit to the similarity data with calculated badness-of-fit criteria of 0.08 and 0.1, respectively (Figure 5A,B⇓). As OA compounds were added to the base wine, the aromas of the spiked samples were different from that of the base wine for both wines. However, addition of either combination of OA compounds did not make the spiked wines produce an aroma that was closer to the original wines than the base wine. Interestingly, for JL, the sample spiked with fewer compounds (J14) was closer to the original wine than the sample spiked with more compounds (J21) (Table 2⇑). Both sets of OA compounds included many oaky, spicy, and smoky aroma chemicals such as furfural, furfuryl alcohol, cis and trans oak-lactones, eugenol, 4-vinylguaiacol, vanillin, and 4-ethyl-guaiacol. Although oaky, spicy, and smoked aromas in the test wines were created using these OA compounds, other compounds need to be added to more closely resemble the overall aroma of the original wine. As described by the panelists, the oaky and spicy aroma of JL was more balanced than the spiked wines.
The CAL wine solution provided a good fit to the similarity data with a calculated badness-of-fit criterion of 0.07 (Figure 5C⇑). Samples C15 and C7 fell between the base wine and the original wine in this plot. Sample C15 seems a better aroma model to represent the aroma of the original wine than C7. In fact, C7 was only spiked with the fruity and floral odorants (ethyl butanoate, isoamyl acetate, ethyl hexanoate, linalool, α-terpineol, and 2-phenylethyl acetate) except vanillin. Sample C15 with fruity/floral and oak, spicy odorants overall produced a closer aroma to the original wine. T-test (Fisher’s LSD) showed that C7 was significantly closer to the base than to the original wine (p < 0.001). However, the original wine was significantly closer to C15 than the base (p < 0.05).
The DEL wine solution had the best fit with a badness-of-fit criterion of 0.04 (Figure 5D⇑). Spiking the base with either combination of OA compounds yielded aromas that were different from the base wine, but it did not successfully increase their similarity to DEL, as observed for CDB-C (Figure 5B⇑).
Discussion
In the present study, the addition of 15 fruity/floral and oaky volatiles in one wine (CAL) yielded an aroma more similar to the parent wine than the unspiked base wine. In contrast, for the other three wines, addition of OA compounds failed to produce aromas closer to the original wine than the base wine. Instead, the spiked volatiles increased the difference in aroma from the base wine. Similarly, in a previous study of Maccabeo white wines (Escudero et al. 2004), the aroma models determined from aroma extract dilution analysis were significantly different from the original wine by triangle test. Like California Chardonnay wines, Maccabeo is also a dry white wine often with an intense and complex aroma, not being excessively specific. In contrast to these results, informal sensory tests showed that removing the OA compound from solutions spiked with 42 OA compounds dramatically reduced the similarity of the mixture to the Scheurebe and Gewürztraminer wines (Guth 1997). Moreover, Scheurebe and Gewürztraminer are dominated by boxwood and floral notes, respectively, which are most potently elicited by the impact of compounds 4-mercapto-4-methylpentan-2-one and cis-rose oxide, respectively (Guth 1997). Interestingly, the aroma reconstitution for CAL wine, which had highest intensities of specifically fruit-related sensory attributes (peach/apricot and citrus), was more successful in this study than the other wines. Complexity of aromas in California Chardonnay and Spanish Maccabeo wines, with a lack of impact compounds that gave distinctiveness to the wine, could be one reason for the difficulty in reconstitution of the aroma models.
There appeared to be additional limitations for the construction of aroma models that were similar to these Chardonnay wines. The unidentified compounds that elicited strong aroma intensities in GC/O analyses might play an important role in Chardonnay aroma. For example, four unknown compounds were included among the 16 OA compounds that produced the best predictive PLS model of the wine aroma profiles (Lee and Noble 2003). Further concentration of the extract or more sensitive detection methods are needed to identify these sensorially important yet unknown compounds. In addition, while GC/O is useful in identifying OA volatiles, it cannot assess interactions among volatiles, such as synergy or blending effects (Acree 1993).
Conclusions
Researchers in disciplines such as flavor chemistry and sensory science have been working together to understand the sensory impacts of OA compounds, both singly and in combination, that have been identified in various varieties of wines. In this study, multivariate statistical procedures were performed on sensory evaluation data of aroma models of California Chardonnay wines. For one wine, addition of the OA compounds increased the similarity of the base wine to the original wine; for the other three wines, the spiked aroma models were closer to the base wine than the original wine. Lack of fit could be due not only to our failure to add potentially important but unidentified compounds to the spiked solutions but also to the possible dependence of interactions of many compounds to generate Chardonnay aroma. PLSR analysis was applied to find relevant patterns in GC and sensory data. In addition, an MDS similarity testing method showed a practical and rapid method for assessing overall similarities of aroma models to the originals.
Footnotes
Acknowledgments: The authors thank Dr. Jane Yegge, Kevin Scott, and the sensory panel for their contribution in this research.
- Received November 2005.
- Revision received March 2006.
- Copyright © 2006 by the American Society for Enology and Viticulture