Introduction

There are many types of wine brandies, spirits distilled from white grape wine. All of them should then be aged in wooden barrels for at least 6 months, although some exclusive brandies can be aged for decades. Wine spirits differ from brandies in the minimum alcoholic strength, which is 37.5% and 36% for brandies and wine spirits, respectively. In addition, the aging of wine spirits in wooden barrels is only voluntary (Regulation (EC) No110, 2008). Sweet wines include a wide variety of styles with one common feature, sweetness. They are produced using various technologies based on the grape dehydration. The most famous are those made from white grapes infected with Botrytis cinerea mold. The botrytized wines are especially suitable for long-term maturation due to their high sugar content (30–150 g/L), which helps their preservation (Jackson, 2008).

An antioxidant is a substance that significantly suppress or prevents the oxidation of the respective substrate. The most common water-soluble antioxidants in sweet wines and brandies are the phenolic compounds including phenolic acids (hydroxybenzoic acids and hydroxycinnamic acids) and their derivatives, and flavonoids (flavonols and flavan-3-ols) (de Beer et al., 2002; Oliveira-Alves et al., 2022; Canas et al., 2008). Botrytized wines are more or less rich in phenolic compounds with corresponding antioxidant activity depending on the grape variety, vinification and ageing (Pour Nikfardjam et al., 2003; Jakubíková et al., 2022). In contrast, pure wine distillates are rich only in volatile compounds, and therefore the antioxidant activity of brandies is assigned to phenolic compounds released into the wine distillates from the wood during the ageing process (Alonso et al., 2004; Oliveira-Alves et al., 2022).

Several analytical methods, including 2,2’-diphenyl-1-picrylhydrazyl (DPPH) radical scavenging assay (Brand-Williams et al., 1995), were used to evaluate antioxidant activity of wines and brandies. DPPH assay is the official method of analysis of the Association of Official Analytical Chemists for the determination of antioxidant activity in foods and beverages (Plank et al., 2012) and the standard procedure recommended by International Union of Pure and Applied Chemistry (IUPAC) (Apak et al., 2013). The most common ways of expressing antioxidant activity are the inhibition percentage of the DPPH solution (DPPH %) (Mrvcic et al., 2012; Nocera et al., 2020; Canas et al., 2008; Bajčan et al., 2017/2018; Fikselová et al., 2010) and Trolox Equivalents Antioxidant Capacity (TEAC) expressed as mmol Trolox per L (mmol TEAC/L) (Mrvcic et al., 2012; Oliveira-Alves et al., 2022; Fernandéz-Pachón et al., 2006). Researchers reported diverse values of inhibition of DPPH in wine spirits, e.g. from 20 to 60% (from 0.05 to 0.15 mmol TEAC/L) in grape brandies (Mrvcic et al., 2012), 50% vs 36% in wine spirits aged using stainless steel tanks with micro-oxygenation and staves vs. new barrels, 63% vs. 21% using chestnut wood vs. Limousin oak wood (Nocera et al., 2020), and increasing inhibition of DPPH from 24 to 50% with the ageing time from one to four years (Canas et al., 2008). In varietal Tokaj wines forming the basis for botrytized wines inhibition of DPPH ranged from 27 to 67% (from 0.3 to 0.8 mmol TEAC/L) (Bajčan et al., 2017/2018). Botrytized wines generally contain a larger amount of polyphenols (Pour Nikfardjam et al., 2003; Jakubíková et al., 2022; Sádecká et al., 2018), which leads to their higher antioxidant activity (Pour Nikfardjam et al., 2003). And so, Fikselová et al. (2010) observed that botrytized Tokay wines showed a DPPH inhibition from 62 to 78%, which ranks them between white and red wines.

The DPPH method is fast and simple and can therefore be used successfully in routine laboratories. Nevertheless, it requires reagents, is laborious, and produces waste, therefore faster, reagentless, less laborious, and more environmentally friendly methods are continuously being developed as complementary or replacements for existing methods. Spectroscopic methods combined with chemometrics provide an alternative to conventional methods in the routine high-throughput determination of antioxidant activity of foods and beverages. Many antioxidants exhibit fluorescence, which has enabled the development of chemometric models based on fluorescence spectra to determine antioxidant activity in coffee and peppermint extracts (Orzel & Daszykowski, 2014), tomato paste (Orzel et al., 2015), cereal vinegar (Long et al., 2023), tea (Bilge & Özdemir, 2020) and apple juice (Włodarska et al., 2016, 2017). The oxygen radical absorbance capacity assay (ORAC) was used to determine the antioxidant capacity of the tomato paste, coffee and peppermint extracts, and partial least squares (PLS) and N-way PLS (N-PLS) regression models were developed and compared for the excitation–emission matrix (EEM) fluorescence spectra (Orzel & Daszykowski, 2014; Orzel et al., 2015). Considering tomato paste and peppermint extracts, PLS method gave smaller root mean squares error of prediction (RMSEP) value ​​than N-PLS, but N-PLS was preferred for coffee extracts. The RMSEP values ​​were 1.2, 648 and 448 µmol Trolox per g of sample for the tomato paste (Orzel et al., 2015), peppermint and coffee extracts (Orzel & Daszykowski, 2014), respectively. Recently, the antioxidant activity of Chinese traditional cereal vinegars was determined using EEM fluorescence spectra and variable-weighted least-squares support vector machine based on particle swarm optimization. A high coefficient of determination of prediction (R2P = 0.85) was observed between the antioxidant activity predicted by the chemometric model and that determined by the DPPH method (Long et al., 2023). When PLS and N-PLS models of EEM fluorescence spectra were compared for the total antioxidant capacity of apple juices, the best model with root mean squares error of cross-validation (RMSECV) value of 1.7 mmol Trolox/L and coefficient of determination of cross-validation (R2CV) value of 0.81 was obtained by applying N-PLS, a reference method was TEAC assay (2,2’-azinobis-(3-ethylbenzothiazoline-6-sulfonic acid) (ABTS)/Trolox-equivalent antioxidant capacity) (Włodarska et al., 2016). Similar results (RMSECV value of 1.8 mmol Trolox/L and R2CV value of 0.79) were obtained using the PLS method for the synchronous fluorescence spectra (SFS) recorded at the wavelength difference between excitation and emission monochromator (Δλ) of 70 nm. The PLS models based on unfolded SFS have a generally slightly poorer performance than those based on single-offset SFS (Włodarska et al., 2017). Combining SFS recorded at Δλ of 20 with PLS was a fast way to determine antioxidant activity of teas with R2CV values of 0.92 and 0.86 for fermented and non-fermented teas, respectively (Bilge & Özdemir, 2020).

In this context, the goal of this work was to develop an eco‑friendly method for the prediction of antioxidant activity in wine brandy and sweet wine. It was hypothesized for the first time that fluorescence spectra combined with PLS regression could potentially be used for this purpose. The validity of the hypothesis was investigated as follows: (1) antioxidant activity of wine brandies and sweet wines was determined using the DPPH assay recommended by IUPAC; (2) total SFS of bulk and diluted samples were recorded and processed - total SFS were preferred over EEM fluorescence spectra due to their advantages such as better selectivity and sensitivity, simplification of the spectral complexity, suppression of Rayleigh scattering, faster acquisition of spectra, and faster and easier chemometric processing (Andrade-Eiroa et al., 2010); (3) PLS regression was used to investigate the relations between the collected SFS and their corresponding DPPH assay values - PLS models were built on the SFS at each Δλ value individually, on the unfolded SFS, and on the variables selected by the variable importance in the projection (VIP) algorithm; (4) the performance of the PLS models was evaluated using R2, RMSE and relative predictive deviation (RPD) values to select the optimal procedure for the determination of antioxidant activity; (5) the greenness of the DPPH assay and SFS method was evaluated using Green Analytical Procedure Index (GAPI) and Analytical GREEnness (AGREE) tools.

Although not essential for finding the above PLS regression models, the antioxidant properties of beverages can be better understood by considering their phenolic composition. Therefore, selected phenolic compounds were determined in wine brandies and sweet wines by the HPLC method and the correlation of the content of individual phenolic compounds with DPPH assay values was assessed.

Materials and Methods

Materials

Sixty five sweet wine samples from the Slovak Tokaj region were obtained from the local producers and the retail stores. The content of residual natural sugar (87 − 206 mg/L) and content of sugar free extract (24 − 51 mg/L) was determined according to the Commission Regulation (EC) No 2676, 1990). Samples were stored at 4 °C and equilibrated at 20 °C before analysis. Sixty wine spirits and brandies produced in nine countries were obtained from the retail stores and stored at laboratory temperature in the dark. Ethanol content was in the range of 38.1 to 41.0% (v/v) (OIV-MA-BS-03, 2022a) and pH ranged from 3.7 to 4.9 (OIV-MA-BS-13, 2022b).

DPPH (2,2-diphenyl-1-picrylhydrazyl) was purchased from Sigma-Aldrich (Steinheim, Germany). A working solution of DPPH 0.11 mmol/L was prepared daily in methanol (HPLC gradient grade, Merck (Darmstadt, Germany)). Trolox (6-hydroxy-2,5,7,8-tertamethylchroman-2-carboxylic acid, 97% purity grade) was obtained from Sigma-Aldrich (Steinheim, Germany) and used as a standard. A working solution of Trolox 3.2 mmol/L was prepared in methanol.

Determination of Antioxidant Activity by DPPH Assay

Brandy samples were analyzed undiluted, but some wine samples were diluted in the ratio 1:1 with methanol before analysis. Antioxidant activity was determined using modified DPPH method described by Brand-Williams et al. (1995). Firstly, 0.25 mL of methanol was added to 4.75 mL of a 0.11 mmol/L DPPH working solution to provide an absorbance of approx. 1.0 at 515 nm (A0). Secondly, 0.25 mL of drink sample was added to 4.75 mL of a 0.11 mmol/L DPPH working solution and left for 30 min in the dark. Then an absorbance was recorded at 515 nm (As) using a UV 1800 Spectrophotometer (Shimadzu, Japan), equipped with a quartz cell with an optical path of 10 mm. In both cases, pure methanol was used as the reference solution.

The percentage of DPPH inhibition was calculated according to the following equation: % DPPH inhibition = 100 × (A0 – As)/Ao. Twelve Trolox standard solutions ranging from 0.08 to 1.04 mmol/L and the same procedure as for the samples were used for the calibration. The p-value for the linear model (the regression of the % DPPH inhibition versus Trolox concentration) from the ANOVA was estimated to be 2 × 10−12, indicating a highly significant linear fit (p < 0.01). Coefficient of determination R2 = 0.998 also suggested statistically valid fit. The results were expressed as mmol Trolox/L of beverage (Table 1). All analyzes were performed in triplicate.

Table 1 Antioxidant activity (% inhibition of DPPH, mmol TEAC/L) and phenolics content of wine brandy and sweet wine
Full size table

Determination of Phenolic Compounds by HPLC

HPLC method of Sádecká et al. (2023) was used to analyze wine brandies in which ellagic, gallic, syringic and vanillic acids were determined. HPLC analysis of sweet wines was done according to Jakubíková et al. (2022) to determine phenolics as caffeic acid, caftaric acid, (+)-catechin, gallic acid and p-coumaric acid.

Synchronous Fluorescence Spectroscopy

SFS were recorded on the bulk and 100 times diluted wine spirits, as well as on the bulk and 500 times diluted sweet wines. Water purified by a Milli-Q system (Millipore, USA) was used for all dilutions.

SFS were collected using a Perkin-Elmer LS 50 Luminescence spectrometer equipped with a Xenon lamp (Perkin-Elmer, USA), a quartz cell with an optical path of 10 mm and FL Data Manager Software (Perkin-Elmer) for spectral acquisition and data processing. Excitation and emission slits were both set at 5 nm. SFS were recorded in the excitation wavelength (λex) range of 250 to 600 nm and 250 to 500 nm for bulk wine spirits and others, respectively, using Δλ set at 20, 40, 60, 80 and 100 nm. The spectra were depicted as the contour plots of total SFS (Fig. 1). The contour plots were constructed in such a way that x-axis shows the λex (nm), y-axis represents the Δλ, and z-axis is plotted by linking points of equal fluorescence intensity. Three repeated measurements were made for each sample.

Fig. 1
figure 1

The representative total synchronous fluorescence spectra in the form of contour plots. (a) bulk wine brandy, (b) diluted wine brandy, (c) bulk sweet wine, (d) diluted sweet wine

Full size image

PLS Regression

The raw spectral data was converted to ASCII format and processed with the Microsoft Excel 2010 (Microsoft Office, USA, 2010), OriginPro 2018 (Microsoft Office, USA, 2018), and STATISTICA version 12 (StatSoft, USA, 2017) software. SFS data of all relevant samples was arranged in the two-dimensional matrices (number of samples × number of λex) for particular Δλ values. In addition, the respective matrices were unfolded into one two-dimensional matrix (number of samples × number of λex multiplied by number of Δλ) (low-level data fusion) (Biancolillo et al., 2014). Mean-centering algorithm was used to pre-processing SFS data.

PLS regression (Wold et al., 2001) was applied to determine antioxidant activity in beverages using individual SFS data at particular Δλ values alone (model named as PLS in Tables 2 and 3) and their low-level fusion (model named as unfolded PLS, UPLS, in Tables 2 and 3). Full PLS models were built on the full λex ranges from 250 to 600 nm for bulk brandies and from 250 to 500 nm for diluted brandies and bulk/diluted wines. The entire sample sets were divided into calibration (70% of the samples) and prediction (30% of the samples) subsets using the Kennard-Stone algorithm (Kennard & Stone, 1969). The calibration subsets were used in the calibration and leave-one-out cross-validation steps of PLS model development, the prediction subsets enabled verification and comparison of the quality of PLS models. The optimal number of latent variables (LVs) was chosen to achieve the minimum RMSECV. Other useful statistics such as RMSEP, R2P, R2CV and relative predictive deviation (RPD), derived from PLS models, were calculated to assess and compare the predictive abilities of the PLS models. Small RMSE values ​​and R2 values ​​close to 1 are typical of good predictive PLS models. RPD value between 1.5 and 2 indicates that the model only discriminates between small and large values, a value between 2 and 2.5 is related to an approximate prediction, a value between 2.5 and 3 indicates a good prediction, and finally, a value above 8 indicates excellent prediction accuracy (Williams, 2001).

Table 2 Comparison of results of different calibration models for wine brandy
Full size table
Table 3 Comparison of results of different calibration models for sweet wine
Full size table

Many studies have confirmed that the performance of PLS ​​models can be improved by removing noisy, irrelevant and redundant variables, leaving only variables containing valuable information. In order to determine the contribution of each wavelength to the information content with respect to antioxidant activity, the VIP algorithm (Andersen & Bro, 2010) was applied on the unfolded SFS data. The resulting VIP scores characterized the relative importance of each independent variable in the PLS model. Since VIP scores were calculated for individual variables, they can be plotted in spectrum-like graphs (Fig. 2). A general criterion for the selection of variables was the VIP scores being greater than 1 (Andersen & Bro, 2010). It was applied in the development of PLS ​​models for sweet wines. However, for brandy samples, too many variables showed VIP > 1 and therefore significant variables were selected using larger threshold VIP values, resulting in different local PLS models that were compared using RMSE. The best sets of VIP and the corresponding UPLS-VIP results are presented in Tables 2 and 3.

Fig. 2
figure 2

Variable importance in projection (VIP). (a) bulk wine brandy, (b) diluted wine brandy, (c) bulk sweet wine, (d) diluted sweet wine a b

Full size image

Results and Discussion

Determination of Antioxidant Activity by DPPH Assay and Phenolics by HPLC

Table 1 shows an overview of the antioxidant activity and phenolic compounds content in wine brandy and sweet wine. The statistics were expressed as the range, mean and standard deviation (SD). In addition, to give insight into the relationship between the contents of individual phenolic compounds and antioxidant activities, the results of correlation analysis were presented for each type of samples. The hypothesis test of the significance of the correlation coefficient was performed using two-tailed testing at the significance level α = 0.05.

Antioxidant activity in brandies was in the range of 9.5 to 89.0% inhibition of DPPH (of 0.19 mmol to 0.81 mmol TEAC/L). Phenolic acids content determined by HPLC ranged from 7.6 to 98.2 mg/L for ellagic acid, from 3.2 to 105.8 mg/L for gallic acid, from 0.9 to 10.6 mg/L for syringic acid and from 0.35 to 13.1 mg/L for vanillic acid. High correlations were found between % DPPH inhibition and ellagic acid content (R2 = 0.948), gallic acid content (R2 = 0.875), vanillic acid content (R2 = 0.798) and syringic acid content (R2 = 0.694), indicating that these compounds can significantly contribute to antioxidant activity of the brandies (α = 0.05). Similar results were obtained in previous studies (Canas et al., 2008; Mrvcic et al., 2012; Nocera et al., 2020; Oliveira-Alves et al., 2022). Analysis of published data shows an effect of the wood botanical species on antioxidant activity of brandies, for example, brandy aged in chestnut wood had a higher antioxidant activity ((93.6% DPPH inhibition) than brandy aged in Limousin oak wood (45.8% DPPH). In addition, antioxidant activity of brandies increased with the duration of ageing (Canas et al., 2008). Correspondingly, there was a wide range of phenolic compounds concentrations in the brandies analyzed – for example, 2.3 − 104.0, 1.1 − 123.1, 0.5 − 13.1, 0.1 − 17.5 mg/L for ellagic, gallic, syringic and vanillic acids, respectively (Canas et al., 2008; Nocera et al., 2020; Oliveira-Alves et al., 2022). % DPPH inhibition was significantly correlated with ellagic acid content (R2 = 0.954), gallic acid content (R2 = 0.842), vanillic acid content (R2 = 0.845) and syringic acid content (R2 = 0.725) in wine spirits (α = 0.05) (Nocera et al., 2020).

Antioxidant activity in botrytized wines ranged from 67.9 to 98.2% inhibition of DPPH (from 0.77 mmol to 1.12 mmol TEAC/L). The profile of phenolic compounds determined by HPLC analysis was composed mainly of: (+)-catechin (ranged from 12.5 to 64.2 mg/L), caftaric acid (from 1.2 to 41.5 mg/L), gallic acid (from 4.6 to 18.5 mg/L), caffeic acid (from 6.2 to 13.5 mg/L) and p-coumaric acid (from 1.6 to 3.9 mg/L). Of the listed components, only the content of caffeic acid and p-coumaric acid was highly correlated with antioxidant activity with R2 values of 0.782 and 0.725, respectively. Recently, a significant correlation was observed between the total phenolic content determined by Folin–Ciocalteu method and the content of caffeic acid (R2 = 0.742) as well as p-coumaric acid (R2 = 0.693) (α = 0.05) (Jakubíková et al., 2022). Many authors related the content of phenolic compounds and/or antioxidant capacity of wines to grape variety, geographical location, weather conditions, SO2 addition, mold pressure in the vineyard, and winemaking technology (Bajčan et al., 2017/2018; Fikselová et al., 2010; Pour Nikfardjam et al., 2003). Bajčan et al. (2017/2018) compared three varieties - Lindenblaetrige, Yellow Muscat and Furmint - originated from Tokaj wine region and found the highest value of antioxidant activity (67% inhibition of DPPH, 0.8 mmol TEAC/L) in variety Yellow Muscat. Botrytized Tokay wines showed a DPPH inhibition from 62 to 78% depending on the year of grape cultivation, temperature in the vineyard and winemaking technology (Fikselová et al., 2010). Pour Nikfardjam et al. (2003) determined antioxidant activity in botrytized Tokaj wines using ABTS/TEAC method and reported values ranged from 1.1 to 7.4 mmol/L. The phenolic compounds content was as follows: (+)-catechin (from 2.7 to 23.4 mg/L), caftaric acid (from 1.1 to 36.4 mg/L), gallic acid (from 0.96 to 13.9 mg/L), caffeic acid (below 1 mg/L) and p-coumaric acid (from 2.6 to 6.8 mg/L).

Synchronous Fluorescence Spectra

The contour plot of total SFS of bulk brandies generally spread in the λex range of 380 to 570 nm and in the Δλ range of 20 to 100 nm and were concentrated in the λex range of 450 to 460 nm and Δλ of 70 to 90 nm (Fig. 1a). The spectra obtained in this study were comparable to those published in the literature for brandies of different geographical origin, which were characterized by λex in the range of 410 to 460 nm and Δλ in the range of 60 to 100 nm (Sádecká et al., 2019).

In the contour map of total SFS of diluted brandy (Fig. 1b), which covered the λex range of 250 to 420 nm with Δλ set from 20 to 100 nm, two maxima were observed for Δλ about 80–90 nm, the first at λex around 280 nm and the second less intense at λex about 340 nm. The observed synchronous florescence bands corresponded well to those already assigned to gallic acid (Δλ/λex range at 80 − 90/270 − 280 nm), syringic acid (Δλ/λex at 80 − 85/270 − 280 nm), vanillic acid (Δλ/λex at 80/270–280 nm), ellagic acid (Δλ/λex at 90 − 100/330 − 340 nm) and ferulic acid (Δλ/λex at 100/320 nm) in previous works (Airado-Rodríguez et al., 2011; Sádecká et al., 2019; Žiak et al., 2014).

The contour map of total SFS obtained for bulk sweet wines showed synchronous fluorescence in the λex range of 380 to 500 nm for the Δλ ranged from 20 to 100 nm (Fig. 1c). Two spectral bands were observed with maxima in the following Δλ/λex ranges: (I) 80 − 90/420 − 430 nm (less intense) and (II) 70 − 80/450 − 460 nm, similar to those reported for botrytized wines previously (Δλ/λex ranges: (I) 80 –90/400 nm and (II) 70 − 90/450 − 460 nm) (Jakubíková et al., 2022; Sádecká et al., 2018).

Observing the contour plots of total SFS of diluted sweet wines (Fig. 1d), fluorescence was found in the λex range of 250 to 390 nm for the Δλ range of 20 to 100 nm with a characteristic maximum at 70 − 80/270 − 280 nm (Δλ/λex). In addition, two less intense bands located at 100/320 nm (Δλ/λex) and 50/350–360 nm (Δλ/λex), respectively, were observed. The bands were close to those described for phenolic acids in botrytized wines (Δλ/λex at 70 − 80/270 − 280 nm, λex/λem at 270 –280/350 nm; Δλ/λex at > 100/300–320 nm, λex/λem at 300 − 310/430 − 440 nm) (Jakubíková et al., 2022; Sádecká et al., 2018). Although the positions of the hydroxyl and methoxyl groups attached to the aromatic ring of phenolic acids are different, the basic conjugated structure of the benzene ring remains the same, resulting in overlapping fluorescence bands of benzoic-like acids observed at Δλ/λex of 70 − 80/270 − 280 nm. Similarly, cinnamic-like acids fluoresce at Δλ/λex of 100 − 120/310 − 330 nm and at Δλ/λex of 140 − 160/260 − 290 nm (Airado-Rodríguez et al., 2011; Sádecká et al., 2018; Tan et al., 2016). The relatively weak band at 50/350–360 nm (Δλ/λex) could be assigned to p-coumaric acid (Tan et al., 2016).

Determination of Antioxidant Activity by PLS Models of SFS Data

PLS regression was used to model the relation between the antioxidant activity expressed as mmol TEAC/L and three types of spectral data: individual SFS recorded at particular Δλ values, unfolded array of total SFS, and variables (ranges of Δλ and λex) selected by the VIP algorithm. The performance of the PLS models was evaluated considering R2, RMSE and RPD values, the number of initial variables and the number of LVs to find the optimal procedure for the determination of antioxidant activity. Tables 2 and 3 show a summary of the results of the different PLS models developed for wine brandies and sweet wines.

In the first step, PLS was applied to the SFS at each Δλ separately. In general, a decrease in RMSECV and RMSEP values, an increase in R2CV and R2P values, and a corresponding increase in RPD were observed with increasing Δλ values. A similar conclusion about RPD values was reached by Włodarska et al. (2017) when investigating the effect of Δλ values ranging from 10 to 110 nm on RPD values ​​in PLS models, particularly of the total phenolic content (TPC) and the total antioxidant capacity (TAC) of apple juices. In our work, this phenomenon was most pronounced for PLS models of diluted sweet wines with an increase in RPD values from 2.2 to 4.2 for Δλ ranging from 20 to 100 nm.

Comparing wine brandies to sweet wines, more reliable PLS models were typical for brandies, as can be seen from the higher R2CV, R2P and RPD values. When deciding between bulk and diluted samples, PLS modeling of the spectra of diluted samples resulted in higher R2P, R2CV and RPD values. In addition, lower RMSE values were obtained in most PLS models for diluted samples. Thus, the best PLS models were obtained for the spectra of diluted samples recorded at Δλ of 100 nm (R2CV = 0.932, RMSECV = 0.015, R2P = 0.931, RMSEP = 0.016, RPD = 3.8 for brandy and R2CV = 0.929, RMSECV = 0.020, R2P = 0.915, RMSEP = 0.015, RPD = 3.4 for sweet wine) (Tables 2 and 3).

Fluorescent compounds, depending on their physicochemical properties, affect the shape of SFS at varied Δλ to a different degree. Based on this, SFS fusion at multiple Δλ is supposed to retain more fluorescence information and can improve the quality of chemometric models. Therefore, the respective SFS data at all Δλ were combined into a new matrix and then analyzed using UPLS. A comparison of the RPD values for the best individual PLS models obtained at Δλ of 100 nm with those for UPLS shows that the data fusion improved the prediction only for sweet wines. For example, RPD values of 3.4 and 3.8, R2P values of 0.915 and 0.932 and RMSEP values of 0.015 and 0.012 were achieved for diluted sweet wines using individual PLS and UPLS, respectively. Similarly, better characteristics of the UPLS model were also observed in cross-validation. When PLS models at Δλ of 100 nm and UPLS models were compared for bulk/diluted brandies, similar or identical R2, RMSE and RPD values were obtained, with no statistically significant difference (paired t-test, α = 0.05, p > 0.05) in the prediction of antioxidant activity by the respective individual PLS and UPLS models.

Also, for apple juices, the results obtained by individual PLS and UPLS models were similar (Włodarska et al., 2017). The best PLS model for TAC was based on individual SFS at Δλ of 70 nm with RPD = 2.1, R2CV = 0.787 and RMSECV = 1.8, while UPLS model using SFS at Δλ ranging from 70 to 110 nm was characterized by RPD = 1.8, R2CV = 0.708 and RMSECV = 2.1. The similarity of the models for TPC was even higher as RPD values were 1.9 and 2.0 at Δλ ranged from 70 to 100 nm and at Δλ of 80 nm, respectively (Włodarska et al., 2017). The results show that a simple fusion of all variables, including wavelengths, which introduce only noise into the data matrix, without rational variable selection, can negatively affect the resulting prediction obtained using UPLS models.

To overcome this drawback, the VIP algorithm (Andersen & Bro, 2010) was applied on the fused SFS data. The VIP scores derived from the UPLS models are shown in Fig. 2 as the contour maps, where λex and Δλ with higher VIP values contributed more to the UPLS models. The generally accepted variable selection criterion, VIP > 1.0, was applied only to UPLS-VIP models for sweet wines. In the case of brandy, too many variables showed VIP > 1.0, and thus, the different local UPLS-VIP models based on higher VIP thresholds were compared. The best sets of VIP and the corresponding characteristics of UPLS-VIP models are shown in Tables 2 and 3.

For wine brandies, the highest VIP score values were observed at Δλ ranged from 80 to 100 nm. Applying the criterion VIP > 1.2, 82 variables covering the λex from 410 to 450 nm were selected for bulk brandy (Fig. 2a). The UPLS-VIP model using the first 3 LVs achieved a good performance expressed by R2CV = 0.933, RMSECV = 0.013, R2P = 0.929, RMSEP = 0.014 and RPD = 3.8, which was a slightly better result compared to respective UPLS. For diluted wine brandies, 62 variables were selected from the Δλ ranged from 80 to 100 nm and λex ranged from 290 to 320 nm based on a threshold of VIP > 1.5 (Fig. 2b). UPLS -VIP model for diluted samples had similar characteristics (R2CV = 0.932, RMSECV = 0.015, R2P = 0.933, RMSEP = 0.015 and RPD = 3.9) to the UPLS-VIP model for bulk samples but contained only the first 2 LVs. No statistical difference was found in the predicted values of antioxidant activity by UPLS-VIP models for diluted and bulk wine brandies (paired t-test, α = 0.05, p > 0.05). For fluorophores excitable at λex of 290 to 320 nm, emission can be expected at λem of 370 to 420 nm. Such fluorophores generally include benzoic-like phenolic acids with λex/λem at 270–280/320–380 nm, e.g. gallic, vanillic and syringic acids, for which a reasonable correlation with antioxidant activity was already observed in the previous section. Also, cinnamic-like phenolic acids with λex/λem of 260–280/320–430 nm can contributed to this fluorescence (Airado-Rodríguez et al., 2011; Sádecká et al., 2019; Žiak et al., 2014).

For sweet wine, the highest VIP scores ​​were at Δλ ranged from 60 to 100 nm (VIP > 1.0). A total of 63 variables in the λex range from 400 to 420 nm and of 108 variables in the λex range from 260 to 295 nm were selected for bulk (Fig. 2c) and diluted (Fig. 2d) sweet wines, respectively. After applying the variables in UPLS-VIP, calibration models with better characteristics were obtained compared to UPLS, while the best model with R2CV = 0.952, RMSECV = 0.011, R2P = 0.944, RMSEP = 0.011 and RPD = 4.2 was calculated by processing the data of diluted sweet wines. Cinnamic-like phenolic acids, e.g. caffeic acid (λex/λem at 262,325/426 nm) and p-coumaric acid (λex/λem at 290, 309/404 nm) (Jakubíková et al., 2022) can be related to selected optimal ranges, Δλ of 60 to 100 nm and λex of 260 to 295 nm (corresponding λem from 320 to 395 nm), as caffeic and p-coumaric acids were found to be significantly correlated with antioxidant activity.

After evaluating the proposed PLS models for wine brandies and sweet wines, it was interesting to compare them with the results of other authors. Unfortunately, RPD values ​​were only rarely available therefore Table 4 shows the reported values ​​of R2P in the prediction of antioxidant activity mainly of beverages obtained using different spectral data. To our knowledge, the determination of the antioxidant activity of brandy based on a combination of spectral and chemometric methods has not been published so far. In wine analysis, the FTIR method coupled with PLS was preferred (Preserova et al., 2015; Silva et al., 2014; Versari et al., 2010), leading to R2P ranging from 0.61 to 0.93 for dessert and rosé wines, respectively (Preserova et al., 2015; Silva et al., 2014;). Better R2P values ​​in the range of 0.93 to 0.96 were obtained by the fusion of IR and Raman spectroscopic data enhanced by the rather complicated chemometric algorithm PSO-VWLS-SVM not for grape wines but for rice wines (Wu et al., 2016). UV-Vis spectroscopy combined with PLS (González-Domínguez et al., 2021) and EEM fluorescence spectroscopy followed by SVM (Long et al., 2023) gave similar R2P values ​​to each other in the prediction of antioxidant activity of wine and cereal vinegars. Strictly, the data sets reported in Table 4 were very different in terms of sample type, number of samples, and chemometrics, so the results cannot be directly compared. Regardless, the SFS data combined with UPLS-VIP yielded antioxidant activity prediction better than IR and similar to the fusion of IR with Raman spectroscopic data.

Table 4 Comparison of the determination coefficient of prediction (R2P) in the prediction of antioxidant activity of different types of samples using spectrometric methods and chemometrics
Full size table

Green Profile Evaluation

The analytical method is considered green if minimizes the use of hazardous chemicals, energy consumption and waste production. In this study, the greenness of the proposed UPLS-VIP method and the reference DPPH assay was evaluated using two metrics. GAPI tool (Płotka-Wasylka, 2018) was used for the evaluation of the whole analytical method. Each step in the procedure (sample collection and preparation, chemicals and solvents, instruments, and technique type) was expressed by a coloured pentagram to represent its impact on the environment (green – environmentally acceptable step, yellow - moderate environmental effect, red - environmentally hazardous step). DPPH assay was characterised by three red parts in the first and the second pentagrams related to sample processing with DPPH reagent, use of methanol as solvent, and in the fourth pentagram related to instrumentation and no waste treatment (Fig. 3a). On the other hand, the proposed SFS method is greener due to the use of a green (water) solvent and minimal health and safety hazard (Fig. 3b).

Fig. 3
figure 3

Evaluation of the greenness of the DPPH assay (a,c) and synchronous fluorescence spectroscopic method (b,d) using GAPI (a,b) and AGREE (c,d) tools. 1 - Sample preparation placement, 2 - Hazardous materials, 3 - Sustainability, renewability, and reusability of materials, 4 - Waste, 5 - Size economy of the sample, 6 - Sample throughput, 7 - Integration and automation, 8 - Energy consumption, 9 - Post-sample preparation configuration for analysis, 10 - Operator´s safety

Full size image

The Analytical GREEnness (AGREE) tool evaluates the method in accordance with the Green analytical chemistry principles (Sample preparation placement, Hazardous materials, Sustainability, renewability and reusability of materials, Waste, Size economy of the sample, Sample throughput, Integration and automation, Energy consumption, Post-sample preparation configuration for analysis, Operator´s safety) on the scale of 0 to 1 to show strong and weak hazard effect (Pena-Pereira et al., 2020). SFS method has higher AGREE score of 0.74 (Fig. 3d) in comparison to DPPH method (AGREE score of 0.45; Fig. 3c). The AGREE pictogram for SFS method presents predominantly green colour in the evaluation of individual aspects, confirming high greenness and lower environmental impact of the SFS method.

‹Figure 3›.

Conclusions

As hypothesized, SFS combined with PLS provided a prediction of the antioxidant activity of wine brandies and sweet wines in the context of state-of-the-art methods.

Three powerful models were proposed for wine brandies showing no statistically significant difference (paired t-test, p > 0.05) in the predicted values of antioxidant activity. The first two models were calculated on the diluted wine brandies data. Individual PLS model at Δλ of 100 nm was characterized with R2CV = 0.932, RMSECV = 0.015, R2P = 0.931, RMSEP = 0.016, RPD = 3.8, and UPLS-VIP model based on 62 variables at Δλ from 80 to 100 nm had R2CV = 0.932, RMSECV = 0.015, R2P = 0.933, RMSEP = 0.015 and RPD = 3.9. The third UPLS-VIP model containing 128 variables at Δλ from 80 to 100 nm achieved R2CV = 0.933, RMSECV = 0.013, R2P = 0.929, RMSEP = 0.014 and RPD = 3.8, and was obtained for bulk wine brandies. Each of the three approaches has its pros and cons. The individual PLS model required registration of the spectrum at only one value of Δλ, but contained a higher number of variables. The second model was calculated from the smallest number of variables, but required sample dilution. The third model was for the bulk samples, but contained a higher number of variables. Taking into account the previous ones, the second model was chosen as the best, because it required the smallest number of initial and latent variables.

As for the sweet wines, the decision was easier and clearer. The best model was UPLS-VIP model with R2CV = 0.952, RMSECV = 0.011, R2P = 0.944, RMSEP = 0.011 and RPD = 4.2 calculated on 108 variables at Δλ of 60 to 100 nm registered on diluted sweet wines. In summary, RPD values ​​close to 4, R2 values higher than 0.9 and low RMSE values indicated very good prediction accuracy obtained by UPLS-VIP models.

Considering the results of the HPLC analysis, the antioxidant activity and variables selected in the VIP algorithm were related to ellagic, gallic, syringic and vanillic acids in wine brandies and to caffeic and p-coumaric acids in sweet wines.

On the AGREE scale of scores from 0 to 1, the SFS method achieved a score of 0.74, which confirmed the good greenness and low environmental impact of the proposed procedure. Therefore, SFS combined with PLS can be considered as a good alternative for the rapid determination of the antioxidant activity of wine brandies and sweet wines.