Research Article Early Prediction of Shiraz Wine Quality Based on Small Volatile Compounds in Grapes

,


Introduction
Shiraz is the most popular grape variety in Australia, grown in nearly every wine-producing region, and the most exported variety of wine by volume and value [1]. Tere are nearly 2,500 wineries and over 6,000 wine grape growers across Australia, contributing 45.5 billion AUD to the Australian economy in 2019 [2]. Wine quality is fundamental to the reputation of wine producers and consequently infuences the price of wine [3], and as such, there is a strong need to understand the drivers of quality and to do this as early in the winemaking process as possible, potentially in the vineyard, when grapes are ripening.
Wine quality is often assessed in professional wine show events, where wines are scored based on their appearance, smell, and taste. For example, in a wine show, each accounts for 15%, 35%, and 50% of the total quality score, respectively [4]. Although the integrity of wine show assessment has been challenged by many researchers [5][6][7], it is still an ideal approach to obtain quality scores for the modelling purpose in this study as formal sensory evaluations using trained panels only provide data regarding perceived intensities of sensory attributes. In professional wine shows, judging is conducted under a formal environment where identities of wines are hidden except the variety and vintage; fnal quality scores are given based on the decision of multiple judging panels comprising experienced and trained judges in each panel to minimise bias of individual judges. Aromatic volatile compounds not only play an important role in the aroma perception process through orthonasal olfaction but also signifcantly afect favour perception through their retronasal detection when the wine enters the mouth [8]. Terefore, volatile compounds-the contributors to odour and favour perception-are essential determinants of wine show performance and infuencers of consumer preferences [9]. Furthermore, previous studies have demonstrated the potential to predict wine quality from wine volatiles for Cabernet Sauvignon [10], Chardonnay [11], and Sauvignon Blanc wine [12]. However, apart from our recent work [4], no such wine volatile-wine quality relationship has been established for Shiraz wines.
Noting the time and cost involved in the winemaking process, there is a desire for early prediction of wine quality from grapes. Current practices for early prediction involve sensory evaluation of grape appearance, texture, and favour by experienced staf accompanied by basic chemical analyses to infer likely characteristics and quality of resulting wines, with implications to wine production and pricing strategies [13]. However, sensory assessment by individuals is vulnerable to subjectivity, even for experienced vineyard managers or winemakers [14,15]. In addition, basic chemical analyses of grapes, such as sugar content and titratable acidity (TA), have limited predictive power. It was reported that the berry sugar content tends to function as an indicator of berry ripeness and wine alcohol content, with wine odour quality potentially compromised from increased berry sugar content due to reductions in aromatics associated with increased wine alcohol content [16]. Additionally, Luo et al. [17] identifed that the accumulation of aromatic compounds (terpenes) in Shiraz grapes did not reliably align with changes in sugar content, further highlighting the limitations associated with the prediction of wine quality from grape sugar content. Similarly, the impact of berry TA on the resulting wine sensory characteristic appears to be limited to the "sour" and "bitter" tastes and the "astringent" mouthfeel [18,19]. Accordingly, basic analytical measures have limited predictive power for overall wine quality, necessitating the exploration of the predictive capabilities of more advanced analytical measures.
While both grapes and wine contain complex volatile profles, the transformative process of fermentation results in a substantially varied profle in terms of chemical species present. Te aromatic volatile compounds in grapes are present in both free and glycosidically bound forms, which are transformed and hydrolysed into the exclusively free forms present in wines [20]. Gambetta et al. [13] demonstrated that complete grape volatile profles (free plus bound) have predictive capabilities for the quality of Chardonnay wines. Furthermore, Forde et al. [21] and Niimi et al. [22] demonstrated that the volatiles in Cabernet Sauvignon and Chardonnay grapes, respectively, had predictive power for the resulting wine sensory descriptors and characteristics. Tese results support the potential for analysis of grape volatile profles for the early prediction of wine quality.
Te aim of this project was to explore the statistical associations between Shiraz wine and grape volatile profles, with professional quality scores of the respective wines. Tis involved the chemical profling of free and bound Shiraz grape metabolites, followed by standardised vinifcation, chemical profling of produced wines, and professional scoring of wine quality. Te resulting datasets were explored for statistical associations, which allowed for the generation of 3 high-quality statistical models: (1) prediction of wine quality from wine volatiles, (2) early prediction of wine quality from free and bound grape volatiles, and (3) early prediction of wine quality from free grape volatiles only. Te models presented provide a valuable tool to Shiraz wine producers by allowing accurate early prediction of wine quality prior to investment of the time and costs associated with production.

Grape Sample Collection.
Shiraz grapes were collected from diferent blocks (n � 16) in 4 diferent commercial Shiraz vineyards in Geelong, Grampians, and Yarra Valley wine regions in Victoria, Australia, during commercial harvest in vintage 2018. Grape bunches (n > 30, approximately 8 kg) were randomly picked from diferent grapevines across each block. TA and total soluble solid content indicating grape berry maturity are provided in Table S1. After collection, the grapes were immediately stored and transported on dry ice and then kept frozen at −20°C until further usage.

Vinifcation of Experimental Wine.
Vinifcation was performed in triplicate for each collected grape sample (48 vinifcation events), following a standardised protocol. Briefy, before destemming and crushing, grapes were thawed at 4°C overnight. After that, 1.9 kg of the crushed grapes were transferred to a 2 L glass fermenter leaving approximately 20% of headspace in the container together with 40 ppm (40 mg/L) of PMS and 1 mL/L of DAP. After adjusting pH to 3.5 with 10% tartaric acid, yeast and malolactic fermentation bacteria were rehydrated and added following the instruction provided by the supplier at dosages of 200 mg/L and 10 mg/L, respectively. Fermenters were placed in a temperature-controlled incubator at 20°C with daily pressing down of grape skins and mixing to ensure that yeast and bacteria could properly interact with the grapes, and the placement positions of the fermenters were rotated every day in the incubator to ensure an even temperature of the fermenters. Te Baumé scale and pH were tested regularly to monitor the progress of fermentation. Once the Baumé scale reached 1, the ferment was pressed in a manual wine press to obtain 1 L of clear fuid. After transferring to a sterilised container preflled with argon gas, the wine was sealed with an airlock and transferred twice to fresh containers to clarify the wine further and held in an incubator at 18°C for 14 and 21 days for each transfer. Te wine was then bottled in a 750 mL standard wine bottle with preflled argon gas and capped with a screw cap. Wine samples were stored in a cool room at 18°C until further analysis. Before sending to the professional wine show for quality assessment, 3 replicates were combined in equal ratios (1 : 1 : 1, v : v : v) and bottled in a clean 750 mL standard wine bottle. A total of 16 wine samples were submitted to the wine show for quality assessment.

Wine Show
Judging. Te judging scheme was the same as recently reported by Luo et al. [4]. Briefy, 16 wine samples were assessed consecutively by 5 panels. Each panel consisted of 3 judges including an experienced judge as the chair. A specialised class in the wine show was created for experimental wines, so that judging was not comparative to commercial wines. Wines were scored out of 100 points based on appearance (15 points), aroma (35 points), and taste (50 points). Medals were given based on the following basis: gold medal (95-100 points), silver medal (90-94 points), bronze medal (85-89 points), and no medal (<85 points). Averages of the fnal scores from 5 panels for the same wine sample were used for modelling.

Determination of Basic Grape and Wine Parameters.
Grape TA was measured by using an HI84533 titrator (Hanna Instruments Inc., Woonsocket, RI). Te grape total soluble solid content was determined by using an HI96811 digital refractometer (Hanna Instruments Inc., Woonsocket, RI). Te colour intensity and hue were analysed following the Sudraud method [23]. Twenty microliters of wine was mixed with 180 μL of Milli-Q water in the 96-well plate. Absorbances at 420 and 520 nm were measured using a Multiskan ™ Go microplate spectrophotometer (Termo Fisher Scientifc Inc., Waltham, MA). Te wine colour intensity was calculated by summing up the absorbance of the two wavelengths, and hue was represented by the ratio of absorbances at 420 nm to 520 nm.

Determination of Wine Volatiles by Headspace-Solid-Phase Microextraction-Gas Chromatography−Mass Spectrometry (HS-SPME-GC−MS).
Te assay including HS-SPME-GC−MS conditions and compound identifcation procedures was performed as per Luo et al. [4] without modifcation for each wine sample prior to replicates being combined for judging (n = 48). Quantifcation was accomplished by using calibration curves of external standards. For compounds without corresponding external standards, semiquantifcation was facilitated based on the internal standard but without including SPME equilibrium factors, and results were expressed as μg/L 4-octanol equivalents.

Extraction of Free and Bound Grape Volatiles and Analysis by HS-SPME-GC−MS.
Te extraction of free grape volatiles and additional solid-phase extraction (SPE) followed by pectolytic enzyme hydrolysis processes for bound grape volatiles were performed according to [24]. Te same GC−MS conditions for analysing wine samples as reported by Luo et al. [4] were applied for the assessment of both free and bound volatiles. . To achieve normalisation of data, datasets underwent transformation and scaling prior to analyses. For the wine volatile dataset, values were square root transformed followed by Pareto scaling. For the free and bound grape volatile datasets, values were log transformed followed by Pareto scaling, and for the free grape volatile dataset, values were square root transformed followed by range scaling.
Predictive models were generated through general polynomial regression using Minitab 19 via the following method. Term selection (from the untransformed volatile datasets) within general regression involved a stepwise Australian Journal of Grape and Wine Research 3 method with an alpha value of 0.15 to enter and remove, initially limited to only frst-order terms to short-list potentially signifcant terms for model inclusion (p < 0.15). A second round of general regression was then performed utilising only these short-listed terms, allowing cross terms and higher-order terms up to and including 3 rd order. Term removal was then performed manually to achieve models with all terms with p < 0.1. Modelling was validated via kfold cross-validation with k assigned as 4 to achieve an even split of the data (n � 16). Tis extension of the "holdout" method involves model assessment for overftting via k rounds of training and testing of the model using random exclusive subsets of the data [25,26].

Results
Basic chemical parameters of the resulting wines of 16 Shiraz grape samples collected from 4 diferent commercial vineyards are summarised in Table S1, which showed that these experimental wines were diferent in pH and appearance. Overall, wines from Geelong had both higher pH and lower colour intensity values than Grampians samples. Except sample 1 from Grampians and sample 12 from Yarra Valley, variations in hue values by region were not observed. Te formation of ester, furan, and lactone compounds and a decrease in benzenoids, aldehydes, and ketones due to alcoholic fermentation were witnessed from the GC−MS analysis (Table S2). Of note, no wine samples were considered "faulty" by judges. Te frst two principal components (PCs) in the score plots explained a total of 79.8%, 44.6%, and 46.5% of the variance for wine (Figure 1(a)), free and bound grape (Figure 1(b)), and free grape (Figure 1(c)) volatile profles, respectively. Te absence of distinct spatial separation of grouping 95% confdence regions (Figures 1(a)-1(c)) identifed that the wine volatiles and the free and bound volatile profles of associated grapes are not distinguishable between wines that did (quality score ≥85%) and did not (quality score ≤84%) score medals.

Discussion
Shiraz is not only popular within Australia but has also gained a reputation globally due to its iconic medium to fullbodied mouthfeel and diverse sensory characteristics [27,28]. While wine quality can be assessed when grapes are made into wine or alongside the fermentation process, it would be more advantageous to the wine industry if accurate quality prediction could be performed at harvest (early prediction). Quality prediction based on grapes can allow wine producers to focus their resources on high-potential grapes, which can improve resulting wine quality and perhaps pricing. However, as yet, no such tool for Shiraz, either based on wine or grape volatiles, is available for grape growers and winemakers in Australia.
Volatiles are important to wine aroma, which in professional wine scoring account for 35% of the overall quality score. Accordingly, potential clustering of wine scoring medals and wines not scoring medals based on volatile profles of the wine and associated grapes was explored (Figure 1). Results from PCA identifed that the scoring of a medal by a wine was not associated with a substantial or consistent shift in the overall profle of wine or grape volatiles, which would have been observed by spatial separations of the 95% confdence regions. Accordingly, the diferences in the volatile profles between wines receiving and not receiving medals are small compared to the overall variation in volatile profles across all wines.
Of note, all models generated demonstrated high k-foldR 2 values, which indicate minimal overftting and an  Figure 1: Principal component analysis (PCA) of wine and grape volatile compounds. Score plots comparing volatile profles of (a) wine volatiles, (b) free and bound grape volatiles, and (c) only free grape volatiles, associated with wines that did and did not receive medals during professional scoring. Shading indicates the 95% confdence regions of the wines awarded with a medal (quality score >85%; red, n � 8) and not awarded a medal (quality score <85%; green, n � 8).
Australian Journal of Grape and Wine Research associated high confdence in the generalisability of the model beyond the training data [25,26]. While Model 1 presented here is the frst model to accurately predict Shiraz wine quality from volatile profles, similar eforts have been undertaken for other wine varieties. Hopfer et al. [10] demonstrated the capacity to explain up to 31% of the variation of the Cabernet Sauvignon wine quality score from individual volatile concentrations, while Gambetta et al. [11] were able to account for 66% of the variation of wine quality scores via a PLS model which utilised concentrations from 6 volatile compounds. Additionally, a recent publication by Luo et al. [4] demonstrated the capacity to explain up to 18% of the variance in quality scores for Shiraz wine from individual volatile concentrations. By comparison to these previous works, Model 1 accounts for 99.97% of the variation in Shiraz wine quality scores based on wine volatiles, indicating a substantial improvement in precision and accuracy in comparison to previous models utilising correlation analysis. Model accuracy was particularly surprising within the context of professional scoring purporting to assign only 35% of the wine score to aroma [4], while almost the entirety of the score could be predicted from wine volatiles using Model 1.
Of note, Model 1 herein utilised 11 terms (equation (1), (R 2 ) = 99.97), the Gambetta et al. [11] model utilised 6 terms (R 2 = 66%), and the works of Hopfer et al. [10] and Luo et al. [4] utilised single terms identifying a maximum R 2 of 31% and 18%, respectively. Additionally, the artifcial neural networking results presented by Zhu et al. [12], who utilised 66 terms as inputs, were able to correctly categorise Sauvignon Blanc wine with 95.4% accuracy into 3 quality gradings, comparable to the 100% categorisation accuracy achieved by Model 1 presented in this study (Tables S5 and  S6). Accordingly, comparison between models highlights the utility of more complex statistical modelling to account for greater proportions of wine score variation and thereby provide greater utility to the wine industry via predictive capacity.
While prediction based on wine volatiles could be useful to wine producers, potentially afecting pricing and marketing decisions, it is more practical if a quality prediction can be made at harvest as resources and the vinifcation strategy can be adjusted accordingly. However, even with the standardised winemaking method utilised herein, signifcant changes in volatile profles from grape to wine were observed (Table S2). As such, early prediction of wine quality by investigating grape volatiles was explored, with Model 2 (R 2 � 99.89%) able to predict wine quality from free and bound grape volatiles and Model 3 (R 2 � 91.62%) able to predict wine quality from free grape volatiles alone. Similar prediction eforts have been explored by Gambetta et al. [13] for the early prediction of wine quality from Chardonnay grape volatiles, which identifed potentially informative correlations for the 5 compounds: hexyl acetate (R 2 � 73.0%, RPD � 1.8), linalool (R 2 � 79.0%, RPD � 2.1), 2-phenylethyl acetate (R 2 � 64.0%, RPD � 1.6), 3-methyl-1-butanol (R 2 � 60.0%, RPD � 1.5), and 2-phenylethanol (R 2 � 72.0%, RPD � 1.8). During method establishment within the applied sciences, minimum residual predictive deviation (RPD) for the classifcation of models as "excellent" is often assigned as 2 or 8 [29], but a more stringent minimum of 10 has been recommended by Williams and Sobering [30]. Compared to these recommended threshold values, Model 1 (RPD � 61.6) and Model 2 (RPD � 31.4) would be undoubtedly classifed as "excellent", while Model 3 (RPD � 3.6) and the linalool relationship (RPD � 2.1) presented by Gambetta et al. [13] would only ft this classifcation at the most lenient threshold. Tis aligns with the observation that while Model 1 and Model 2 (RPD > 10) could perfectly predict medal categorisation, Model 3 (2 < RPD < 10) prediction of medal categorisation demonstrated 81.25% accuracy (Table S3). Accordingly, Models 2 and 3 highlight the potential for early prediction of wine quality from grape volatiles, which predominantly utilised compounds which are not present in the associated wines (Table S2). As such, it is likely that the compounds in these models are-or correlate with-precursors to wine volatiles that are impactful to quality. While Model 3 showed the lowest accuracy of the models presented herein, R 2 of 91.62% is substantially higher than that of the previously published  Figure 2: Prediction of wine quality based on data for wine or grape volatiles. General regression analysis with 4-fold cross-validation for the prediction of the wine quality score (n � 16) based on wine volatiles (a) and early prediction of wine quality based on grape free and bound volatiles (b) or free grape volatiles only (c). early prediction model [13] for Chardonnay and presents a signifcant progression to Shiraz wine early prediction. Furthermore, as Model 3 utilises only free volatiles as input terms, and the associated chemical extraction and analytical protocol is faster and cheaper than that associated with Model 2, which requires additional steps to capture glycosidically bound volatiles, utilising Model 3 may be the preferred approach for implementation within the industry. It should be noted that rotundone, a signifcant compound for Shiraz that contributes to its desirable "peppery" attribute and is deemed related to Shiraz quality [31], was not analysed in this study. Although including rotundone could potentially enhance the models, its analysis requires extensive pretreatment and diferent GC confgurations compared to the analysis of other compound groups [32]. As the addition of rotundone is unlikely to contribute substantially to the accuracy of the present models, especially Models 1 and 2, there is not a strong need to incorporate rotundone into these models.

Conclusions
Grape growers and winemakers know that good grapes make good wine. While current practices rely on subjective and simple analytical assessments of grapes, those assessment tools have limited predictive capabilities. Tere is a great interest in the potential to accurately predict wine quality from grapes. Te results presented in this study not only demonstrate the capacity to accurately predict Shiraz wine quality from wine volatiles but additionally the capacity to predict wine quality from grape volatiles. Furthermore, much of the predictive capability was retained when only free grape volatiles were utilised as input terms for modelling, allowing for fast, cheap, and therefore high-throughput prediction of wine quality from grape analysis. Terefore, the models presented here provide Shiraz grape growers and winemakers with a potentially valuable tool to predict the quality of their wines prior to the investment of the time and costs associated with wine production. Given that results were based on grapes from 16 blocks in 4 vineyards, the inclusion of more grape samples from diferent wine regions in future model optimisation will improve the generalisation of these models.

Data Availability
Te chemical data used to support the fndings of this study are included within the supplementary information fles. Te small molecule and sensory data used to support the fndings of this study have been deposited in the FigShare repository: 10.26188/21747848.

Supplementary Materials
Te following information regarding sample and modelling details is provided in the supplementary information: Table  S1: experimental wine details. Table S2: presence of diferent classes of compounds in grapes in free and bound forms and in their resulting wine. Table S3: coefcients of predictive models and associated p values. Table S4: predicted quality scores and medal classifcations of wine samples. Table S5: accuracies of predicted medal classifcation by 3 models.