Introduction

Facilitation of grape quality evaluation is a significant concern, especially for large-scale winemakers dealing with multiple vineyards, varieties and targeted products and quality levels. Planning harvest and sorting grapes for varying target products call for a rapid and accurate methodology. Whereas High-Performance Liquid Chromatography (HPLC) is recognised as the primary reference method for wine phenolic analysis, it may be too expensive, time-consuming and complex for effective implementation scale-up. Hence secondary methods, including Fourier Transform Infrared (FTIR), Near and Mid-Infrared (NIR and MIR), Nuclear Magnetic Resonance (NMR) and UV-VIS spectroscopies, have been examined with respect to their ability to be calibrated against HPLC and other reference chemistry methods (Aleixandre-Tudo et al., 2017; Harbertson et al., 2003; Harbertson and Spayd, 2006; Miramont et al., 2020; Rouxinol et al., 2022; Sen et al., 2016). The consensus is that while these secondary spectroscopic techniques provide generalised compound class information, they do not exhibit highly correlated molecular specificity for the key phenolic compounds generally regarded as quality parameters. However, in several reports, as reviewed by Ranaweera et al. (2021a), UV-VIS, FTIR and NIR, as well as NMR spectroscopies, have provided significant classification results for grape and wine varieties. Interestingly reports of successful data fusion for UV-VIS and FTIR for effective wine classification have also been reported (Geană et al., 2019; Sen et al., 2016). HPLC has been used for arguably the most highly effective varietal classification, albeit as mentioned above, the method is relatively slow and laborious compared to spectroscopic methods (Peng et al., 2002; de Villiers et al., 2005; Rouxinol et al., 2022; Urcan et al., 2016).

Recent developments in the areas of phenolic quantification and varietal and regional classification, focusing only on fermented wine products, have been documented using a new patented technique known as Absorbance-Transmittance fluorescence Excitation-Emission Matrix (A-TEEM) spectroscopy (Gilmore and Tong, 2019). A-TEEM is typically coupled with machine learning methods for both effective wine classification (Niimi et al., 2020; Ranaweera et al., 2021a; Ranaweera et al., 2021b; Ranaweera et al., 2022) and specific phenolic quantification with repeatability statistics on par with HPLC and other basic chemistry methods (Ranaweera et al., 2021b; Ranaweera et al., 2022; Schober et al., 2022). The A-TEEM method allows rapid (s-min) acquisition rates and requires a very small amount of test material. The A-TEEM's potential to identify and quantify specific chemicals is reinforced by several key features, including primarily the automated correction of fluorescence inner-filter effects (IFEs), which is critical for linear quantification (Gilmore, 2014; Gilmore and Tong, 2019). However, there is also the fact that the A-TEEM, when applied under Beer–Lambert linear concentration conditions, simultaneously evaluates five optical molecular parameters, including 1) the absorbance extinction coefficients, 2) the fluorescence quantum efficiencies, 3) absorbance spectral shape and the fluorescence, 4) excitation- and 5) emission-spectral shapes. Importantly, simultaneous analysis of the absorbance and fluorescence variables using a ‘multi-block' model or data-fusion approach facilitates an effective means to incorporate and leverage all five parameters in machine learning applications which can be significant regarding the ability to identify and quantify specific quality marker compounds in complex grape extract matrices (Ranaweera et al., 2021b; Schober et al., 2022).

In this study, we select a set of the major compounds that are important to measure industrial grape quality (Cleary et al., 2015). These include 1) Polymeric tannins which in wine are mostly grape-derived, are extracted from skins and seeds during fermentation and are essential because of their astringent properties; 2) Quercetin glycosides, including quercetin-3-O-glucoside and quercetin-3-O-glucuronide, which are flavonols extracted from skins, and contribute to wine's velvety astringency (Hufnagel and Hofmann, 2008a; Hufnagel and Hofmann, 2008b); 3) (+)-Catechin and (-)epicatechin, which are the monomeric fractions of procyanidins, and are associated with wine astringency and bitterness (Kallithraka and Clifford, 2007; Hufnagel and Hofmann, 2008a) and finally, 4) anthocyanins, including malvidin-3-O-glucoside, which are important for wine colour, and also for forming polymeric pigments and more stable pigmented tannins such as pyranoanthocyanins as wine ages (He, 2012).

Examples of A-TEEM finished wine analysis using the machine learning method based on a decision-tree algorithm known as Extreme Gradient Boosting (XGB) (Chen and Guestrin, 2016), including Discrimination Analysis (XGBDA) and regression, have shown the potential to classify wine grape varieties and regions accurately as well as accurately quantify phenolic compounds, respectively (Ranaweera et al., 2021a; Ranaweera et al., 2021b; Ranaweera et al., 2022; Schober et al., 2022). In this study, we investigated the capacity of A-TEEM using multi-block data organisation using both Agglomerative Hierarchical Cluster Analysis (HCA) and XGBDA for grape varietal classification and XGB regression for estimating key chemical compound quality markers using unfermented grape extracts. The efficiency of grape berry phenolic extraction used in this paper is evaluated and compared to earlier finished wine extraction studies to evaluate whether berry extractions lead to more variance and/or lower spike-recovery rates for key phenolic quality-marker compounds. The main sources of variance in the compound model predictions in this paper are evaluated and discussed with respect to their potential impact on field and laboratory implementation for grape inspection and quality assessment.

Materials and methods

Results

Figure 1 shows typical example absorbance profiles (A) and CIE 1931 x-y colour coordinates (B) from A-TEEM samples for Cabernet-Sauvignon, Merlot, Pinot noir and Zinfandel grapes. The absorbance profiles all exhibit the peak signal below 250 nm, a major peak around 280 nm attributed to phenolic and flavan-3-ols compounds, a minor shoulder in the 300-400 nm range associated mostly with flavonols and a second major peak around 520 nm associated with anthocyanins. There were slight differences noted among the different varieties in Panel A that were associated with the different colour mapping coordinates shown in Panel B. Importantly, although these typical examples were clearly unique, the varietal differences among all samples in this study were not clearly discerned solely by the absorbance profiles or CIE coordinates when plotted together (not shown); hence additional evaluation of the EEM data is pursued below.

Figure 1. Typical Absorbance spectra (A) and CIE 1931 x,y coordinate indices (B) for Cabernet-Sauvignon, Merlot, Pinot noir and Zinfandel grape extract samples. The data in Panel A represent the extracts diluted 50 fold in 50 % EtOH pH 2 solvent whereas the CIE indices in Panel B were adjusted for the dilution factor.

Save View full size Expand inline Collapse inline

Figure 2 shows the EEM contour plots for the same typical samples as shown in Figure 1, and within the peak-normalised scaling shown, each variety exhibited a unique fingerprint. Cabernet-Sauvignon exhibited the strongest peak intensity (8.1), followed by Pinot noir (6.7), Zinfandel (6.3) and Merlot (5.24). All four varieties exhibited two contour peaks with major excitation/emission coordinates at < 240/315 nm and 280/315 nm. Interestingly, both the Merlot and Pinot noir also exhibited distinct minor contour peaks with excitation/emission coordinates at 260/375 nm that were much less pronounced in the Cabernet-Sauvignon and Zinfandel varieties. Supplementary Figure S1 shows the same EEM data as Figure 2 scaled to a peak value of 1 for all four varieties to illustrate that multiple minor contours, including the anthocyanin compounds in the red, can also be observed beyond the limited scaling shown in Figure 2 to distinguish the four varieties further visually.

Figure 2. Typical fluorescence Excitation-Emission matrix contour plots for the same Cabernet-Sauvignon (A), Merlot (B), Pinot noir (C) and Zinfandel (D) grape extract samples shown in Figure 1 each scaled to the peak EEM contour values.

Save View full size Expand inline Collapse inline

As with the absorbance and colour profiles in Figure 1, the EEM contours in Figure 2 and Figure S1 can only elicit a limited amount of varietal information by simple visual inspection of one sample at a time. Therefore, we analysed the complete sample set first by Agglomerative Hierarchical Cluster Analysis (HCA) in Figure 3; the results from Figure 3 were evaluated from the concatenated multi-block absorbance and unfolded EEM variables for the combined calibration and validation data. Here the dendrogram represents the variance-weighted distribution between clusters for the nine varieties identified in the legend. Each variety was exclusively clustered together to support that the A-TEEM multi-block data set contained enough unique chemical fingerprint information to distinguish the different varieties clearly. The apparent varietal linkages clearly show that Pinot noir is the most unique, with Cabernet-Sauvignon initiating the second tier and exhibiting relatively closer linkages to all the other varieties. Supplemental Figure S2 shows that the mean sum (A) and normalised (B) distributions of the key quality marker compound concentrations were uniquely distributed among the tested varieties. The data in Figures 3 and S2 are consistent with other published observations that Pinot noir exhibits a unique phenolic composition, being notably enriched in flavan-3-ols, including catechin and epicatechin, and with generally reduced total anthocyanins (Cliff et al., 2007; Rouxinol et al., 2022; Urcan et al., 2016).

Figure 3. Agglomerative Hierarchical Cluster Analysis (HCA) dendrogram for the multi-block A-TEEM data representing the varieties shown in the legend. The data included both the calibration (n = 390) and validation (n = 62) file sets.

Save View full size Expand inline Collapse inline

To further investigate the A-TEEM’s capacity to classify the varieties in a quantitative manner, we subjected the same data shown in Figure 3 to XGBDA with the results shown in Table 1. Using the methods described above, Table 1 reports no errors in the identification of the calibration set (not shown) or the validation data set for each of the nine varieties analysed as indicated by True Positive Proportion (TPR) and True Negative Proportion (TNR) values of 1 and False Positive Proportion (FPR) and False Negative Proportion (FNR) values of 0 which correspond to Error (Err) values of 0, and Probability (P) and F1 values of 1, respectively. Hence, the HCA (Figure 3) and XGBDA (Table 1) data both support that the A-TEEM accurately qualified the nine grape varieties in the tested samples.

The quantification of key quality chemical markers for grapes was another key topic of investigation for the A-TEEM multi-block model approach. In Figure 4, we show the XGB regression prediction data set values for (A) Polymeric Tannins, (B) Malvidin-3-Glucoside and Total Anthocyanins, (C) Flavan-3-ols, Catechin and Epicatechin and (D) Quercetin Glycosides. Polymeric Tannins ranged from 150 to 350 mg/L, Total anthocyanins ranged above 185 mg/L, Catechin ranged up to around 55 mg/L and Quercetin Glycosides were observed up to around 14 mg/L. For all six target compounds, the predicted linear and the quantitative performance prediction statistics are described in Table 2. Table 2 data are sorted from top to bottom by the maximum range for each compound. The R²P was > 0.917 for all compounds with an average of 0.9411 ± 0.0247 which equates to an average REP of < 5.9 %. For all models in Table 1, the R² for calibration was greater than or equal to the R²P. It is notable that the target for an ideal secondary method for this primary HPLC method should be around 5 to 8 %, given the previously described typical spike recovery and Inter- and Intra-day repeatability statistics (RSD) of this HPLC method with grape berry extracts.

Figure 4. Extreme Gradient Boosting regression plots for the prediction data set for key phenolic and anthocyanin compounds. All lines were constrained to a slope of unity and intercept of 0 mg/L. The compound identities are listed in the legends. The regression statistics are contained in Table 2.

Save View full size Expand inline Collapse inline

Discussion

Both the A-TEEM varietal classification and phenolic regression results with unfermented grape extracts presented here are consistent with previously published work with finished wines (Ranaweera et al., 2021a; Ranaweera et al., 2021b; Schober et al., 2022). In contrast, however, this work is a special industrially focused application exclusively with vineyard and maturation-stage specific commercial grape extracts with significance for rapidly and accurately establishing both grape quality and authenticating varietal characteristics. Both the varietal classification and phenolic quantification capabilities are important for advising vineyard management and berry harvesting and sorting for targeted product quality control and formulation.

The combined use of the multi-block data organisation and use of the XGB algorithm in this paper yielded satisfactory regression results, which were consistent with the established performance of the primary HPLC method with respect to observed inter- (8 %) and intra-day (5 %) RSD associated with recovery of the measured compounds from raw berry extracts. Importantly, the A-TEEM repeatability for any given sample was measured to be < 2 % Relative Standard Deviation (RSD) as evidence that the mean 5.89 % REP for the phenolics in Table 2 was mostly associated with error propagation associated with the overall berry extraction and HPLC methodology. It is also notable that this reference HPLC method required approximately 20 min, while the A-TEEM analysis was limited to less than one. Importantly, the REP (≅ 6 %) for the overall phenolic analysis of finished wines in Schober et al. (2022) compared quite closely to the phenolic REP (≅ 5.89 %) in this grape extract study.

Here the significant resolution of each of the nine grape varieties with HCA and XGBDA supports the potential for the A-TEEM method to be employed for authentication of grape lots, especially for labelling of pure varietal products as well as for blending as had been shown earlier for finished wine products (Ranaweera et al., 2022). It is evident from the unique A-TEEM fingerprint characteristics that the successful discrimination is primarily based on the unique phenolic compositions of each variety. It is also important to consider the samples evaluated here for each variety, and samples were sourced from different vineyards and harvest dates (all from the same year) which lends credence to the concept that the varietal discrimination is primarily based on the genetic characteristics as opposed to being more strongly associated with vineyard or growth (maturation). This is especially evident with the previously referenced uniqueness of the Pinot noir variety that consistently exhibits much higher concentrations of the flavan-3-ols Catechin and Epicatechin as well as other significant differences in the polyphenolic profiles (Yang et al., 2009), including lack of acylated anthocyanins (Boss et al., 1996; Mazza et al., 1999), compared to the other varieties in this study.

In conclusion, this study helps to establish further the A-TEEM coupled with multi-block data fusion and machine learning, especially XGB, as an effective tool to monitor incoming grape samples rapidly. This can facilitate analysis faster than HPLC with similar degrees of precision for phenolic concentrations and, more importantly, with the added capacity for rapid, accurate grape variety identification. Future work may extend to the rapid measurement of phenolic compounds in grapes and authentication by the A-TEEM method on a wider geographical scale. Importantly, the individual regression and discrimination models, once calibrated, can both be applied with commercially available software in an automated fashion that is amenable to field or remote-laboratory operations for near real-time monitoring (Schober et al., 2022).

Acknowledgements

We thank our colleagues and interns from Winegrowing Research at E&J Gallo Winery who collected and helped to process the grapes for this study. We would also like to show our gratitude to Dr. Nick Dokoozlian, Vice President, E&J Gallo Winery for his support to advance the research in grapes and wine with A-TEEM technology.

References

• Aleixandre-Tudo, J.L., Nieuwoudtc H., Aleixandre, J. L., & du Toit, W. (2017) Chemometric compositional analysis of phenolic compounds in fermenting samples and wines using different infrared spectroscopy techniques. Talanta, 176, 526–536. http://dx.doi.org/10.1016/j.talanta.2017.08.065
• Boss, P.K., Davies, C., & Robinson, S.P. (1996). Anthocyanin composition and anthocyanin pathway gene expression in grapevine sports differing in berry skin colour. Australian Journal of Grape and Wine Research.,2, 163–170, DOI: https://doi.org/10.1111/j.1755-0238.1996.tb00104.x
• Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, USA. DOI: https://doi.org/10.1145/2939672.2939785
• Cleary, M., Chong, H., Ebisuda, N., Dokoozlian, N., Loscos, N., Pan, B., Santino, D., Sui, Q., & Yonker, C. (2015). Objective Chemical Measures of Grape Quality. ACS Symposium Series, 120, 365–378. DOI: 10.1021/bk-2015-1203.ch023
• Cliff, M.A., King, M.C., & Schlosser, J. (2007) Anthocyanin, phenolic composition, colour measurement and sensory analysis of BC commercial red wines. Food Research International, 40, 92–100. DOI: https://doi.org/10.1016/j.foodres.2006.08.002
• de Villiers, A., Majek, P., Lynen, F., Crouch, A., Lauer, H., & Sandra, P. (2005) Classification of South African red and white wines according to grape variety based on the non-coloured phenolic content. Eur Food Res Technol, 221, 520–528. DOI: 10.1007/s00217-005-1169-5
• Geană, E-I., Ciucure, C.T., Apetrei, C., & Artem, V. (2019) Application of Spectroscopic UV-Vis and FT-IR Screening Techniques Coupled with Multivariate Statistical Analysis for Red Wine Authentication: Varietal and Vintage Year Discrimination. Molecules, 24(22), 4166. DOI: https://doi.org/10.3390/molecules24224166
• Gilmore, A. (2014). How to Collect National Institute of Standards and Technology (NIST) Traceable Fluorescence Excitation and Emission Spectra. In Y. Engelborghs & A. J. W. G. Visser (Eds.), Fluorescence Spectroscopy and Microscopy: Methods and Protocols (pp. 371–417). https://doi.org/10.1007/978-1-62703-649-8
• Gilmore, A, & Tong, X. (2019) System and Method for Fluorescence and Absorbance Analysis. United States Patent No. US8901513B2
• Harbertson, J. F., & Spayd, S. (2006). Measuring Phenolics in the Winery. American Journal of Enology and Viticulture, 3(57), 280–288.
• Harbertson, J. F., Picciotto, E. A., & Adams, D. O. (2003). Measurement of Polymeric Pigments in Grape Berry Extracts and Wines Using a Protein Precipitation Assay Combined with Bisulfite Bleaching. American Journal of Enology and Viticulture, 54(4), 301–306.
• He, F. (2012) Anthocyanins and their variation in red wines I. Monomeric anthocyanins and their color expression. Molecules, 17(2), 1571-1601. DOI: https://doi.org/10.3390/molecules17021571
• Hufnagel, J.C., & Hofmann, T. (2008a). Orosensory-Directed Identification of Astringent Mouthfeel and Bitter-Tasting Compounds in Red Wine. Journal of Agricultural and Food Chemistry, 56(3), 1376–1386. DOI: https://doi.org/10.1021/jf073031n
• Hufnagel, J. C., & Hofmann, T. (2008b). Quantitative Reconstruction Of the Nonvolatile Sensometabolome of a Red Wine. Journal of Agricultural and Food Chemistry, 56, 9190–9199. DOI: https://doi.org/10.1021/jf801742w
• Iland, P., Bruer, N., Edwards, G.,Weeks, S., & Wilkes, E. (2004) Chemical Analysis of Grapes and Wine: Techniques and Concepts; Patrick Iland Wine Promotions PTY Ltd.: Campbelltown, Australia.
• Kallithraka, S., & Clifford, M.N. (2007). Evaluation of bitterness and astringency of (+)-catechin and (-)-epicatechin in red wine and in model solution. Journal of Sensory Studies, 12(1), 25-27. https://doi.org/10.1111/j.1745-459X.1997.tb00051.x
• Mazza, G., Fukumoto, L., Delaquis, P., Girard, B., & Ewert, B. (1999) Anthocyanins, phenolics, and color of Cabernet Franc, Merlot, and Pinot noir wines from British Columbia. J. Agric. Food Chem. 47, 4009–4017, DOI: https://doi.org/10.1021/jf990449f
• Miramont, C., Jourdes, M., & Teissedre, P-L. (2020) Development of UV-vis and FTIR Partial Least Squares models: comparison and combination of two spectroscopy techniques with chemometrics for polyphenols quantification in red wine. Oeno One, 4, 779-792. DOI:10.20870/oeno-one.2020.54.4.3731
• Niimi, J., Tomic, O., Næs, T., Bastian, S. E. P., Jeffery, D. W., Nicholson, E. L., Maffei, S. M., & Boss, P. K. (2020). Objective measures of grape quality : From Cabernet-Sauvignon grape composition to wine sensory characteristics. Food Science and Technology, 123, 109105. https://doi.org/10.1016/j.lwt.2020.109105
• Peng, Z, Iland, P. G., Oberholster, A., Sefton, M. A., & Waters, E. J. (2002). Analysis of pigmented polymers in red wine by reverse phase HPLC. Australian Journal of Grape and Wine Research, 8, 70-75. DOI: https://doi.org/10.1111/j.1755-0238.2002.tb00213.x
• Ranaweera, K. R. R., Capone, D.L., Bastian, S.E.P., Cozzolino, D., & Jeffery, D.W. (2021a). A Review of Wine Authentication Using Spectroscopic Approaches in Combination with Chemometrics. Molecules, 26, 4334. https://doi.org/10.3390/molecules26144334
• Ranaweera, K. R. R., Gilmore, A. M., Capone, D. L., Bastian, S. E. P., & Jeffery, D. W. (2021b). Spectrofluorometric analysis combined with machine learning for geographical and varietal authentication, and prediction of phenolic compound concentrations in red wine. Food Chemistry, 361, 130149. https://doi.org/10.1016/j.foodchem.2021.130149
• Ranaweera, K. R. R, Gilmore, A.M., Bastian, S.E.P, Capone, D.L., & Jeffery, D.W. (2022). Spectrofluorometric analysis to trace the molecular fingerprint of wine during the winemaking process and recognise the blending percentage of different varietal wines. OENO One, 56(1), 189-196. DOI:10.20870/oeno-one.2022.56.1.4904
• Rouxinol, M.I., Martins, M.R., Murta, G.C., Barroso, J.M., & Rato, A.E. (2022) Quality Assessment of Red Wine Grapes through NIR Spectroscopy. Agronomy, 12(3), 637. DOI:https://doi.org/10.3390/agronomy12030637
• Schober, D., Gilmore, A., Chen, L., Zincker, J., & Gonzalez, A. (2022). Determination of Cabernet-Sauvignon Wine Quality Parameters in Chile by Absorbance-Transmission and Fluorescence Excitation Emission MATRIX (A-Teem) Spectroscopy. Food Chemistry, 133101. https://doi.org/10.1016/j.foodchem.2022.133101
• Sen, I., Ozturk, B., Tokatli, F., & Ozen, B. (2016) Combination of visible and mid-infrared spectra for the prediction of chemical parameters of wines. Talanta, 161, 130–137. http://dx.doi.org/10.1016/j.talanta.2016.08.057
• Urcan, D.E., Lung, M-L., Giacosa, S., Torchio, F., Ferrandino, A., Vincenzi, S., Segade, S-R., Pop, N., & Rolle, L. (2016) Phenolic Substances, Flavor Compounds, and Textural Properties of Three Native Romanian Wine Grape Varieties. International Journal of Food Properties, 19, 76–98. DOI: https://doi.org/10.1080/10942912.2015.1019626
• Yang, J., Martinson, T.E., & Liu, R.H. (2009). Phytochemical profiles and antioxidant activities of wine grapes. Food Chemistry, 116, 332–339. https://doi.org/10.1016/j.foodchem.2009.02.021