Volatile Compound Abundance Correlations Provide a New Insight into Odor Balances in Sauce-Aroma Baijiu

Sauce-aroma Baijiu (SAB) is one of the most famous Baijius in China; SAB has more than 500 aroma compounds in it. However, the key aroma compound in SAB flavor remains unclear. Volatiles play an important role in SAB aroma and are highly correlated to SAB quality. In the present study, 63 volatile compounds were quantified among 66 SAB samples using gas chromatography with flame ionization detector (GC-FID). The authors analyzed odor contributions and volatile compound correlations in two quality groups of SAB samples. Moreover, an odor activity value (OAV) ratio-based random forest classifier was used to explain the volatile compound relationship differentiations between the two quality groups. Our results proved higher quality SABs had richer aromas and indicated a set of fruity-like ethyl valerate, green- and malt-like isobutyraldehyde and malt-like 3-methylbutyraldehyde and sweet-like furfural, had closer co-abundance correlations in higher quality SABs. These results indicated that the aroma and contributions of volatile compounds in SABs should be analyzed not only with compound odor activity values, but also the correlations between different aroma compounds.


Introduction
Chinese Baijiu is one of the six most famous distilled spirits in the world, and among all types of Baijiu, sauce-aroma Baijiu (SAB; also called Maotai-flavor Baijiu) is one of the most famous Baijiu in China [1]. SAB is made from sorghum and wheat using sophisticated fermentation techniques with multiple rounds of fermentation/distillation [2].
The complexity of SAB in flavors and compositions has been broadly realized, and a recent study showed that more than 500 volatile compounds can be detected in SAB. More than 100 important aroma compounds obtained via odor activity values (OAVs) were analyzed, and the ten aroma compounds which had the highest OAV values in SAB (ethyl butanoate, ethyl hexanoate, isopentyl hexanoate, ethyl isobutyrate, ethyl octanoate, ethyl isovalerate, ethyl propanoate, ethyl 2-methylbutanoate, butyl butyrate, propyl acetate) contributing in fruity, apple, sweet, grape, banana, pear, pineapple, floral and caramel odors [2]. However, the key aroma compound or a key set of aroma compounds representing SAB flavor or quality has not yet been revealed [3].
Different fermentation methods and varied environmental conditions result in the variation of flavors [4]. The discrimination between different types of Baijiu [5], the same A total of 66 SAB samples were collected from different distilleries in Maotai Town, Renhuai City, Guizhou Province, China. All 66 SAB samples were made via traditional soy sauce aroma-type solid-state fermentation with sorghum as raw material, in which 21 samples could be purchased in local markets and 45 samples were collected directly from factories. All samples had 53% alcohol by volume and were stored in a storage room under 15 • C for further analysis.

Sensory Evaluation
Ten well-trained panelists including 4 national Baijiu panelists judged all SAB samples. Samples were then grouped into excellent grade, grade I, and grade II according to the China National Institute of Standardization GB/T 26760-2011. Briefly, SAB's aroma, taste, and aftertaste were evaluated. For excellent grade SAB, the Baijiu body had an elegant fragrance with the sauce flavor taking the prominent position, the Baijiu body had plentiful integrity, the feeling in the mouth had mild-mannered coordination, and the feeling in the finish was prolonged. For grade I SAB, the Baijiu body had a pure and natural sauce flavor, the feeling in the mouth was well coordinated, and the finish was prolonged. For grade II SAB, the Baijiu body had a significant sauce flavor, the feeling in the mouth was mild, and the finish was relatively long.

Gas Chromatography with Flame Ionization Detector (GC-FID)
The analyses were performed on a GC 7890B gas chromatograph (Agilent Technology, Santa Clara, CA, USA) equipped with a flame ionization detector (FID). A DB-WAX column (60 m × 0.25 mm i.d., 0.25 µm film thickness, Agilent Technology, Santa Clara, CA, USA) was operated under programmed temperature conditions: 35 • C held for 8 min, 35-40 • C at 2.5 • C/min, 40-100 • C at 5 • C/min, 100-200 • C at 10 • C/min, 200-220 • C at 20 • C/min and held for 10 min (detector temperature of 250 • C, injector temperature of 230 • C), and nitrogen gas carrier gas (1.6 mL/min). Three internal standards were used in this study: amyl acetate (IS1, 10 µL, 10.33g/L), 2-methyl hexanoic acid (IS2, 10 µL, 14.00 g/L), and tertiary amyl alcohol (IS3, 10 µL, 8.05 g/L). The mixed standard solutions (1 mg/mL) were prepared in absolute ethanol and diluted with 53% (v/v) ethanol to the following concentrations: 50,000, 10,000, 5000, 1000, 500, 100, 50, 20, 10, and 1 mg/L. Each working standard solution (1 mL) was spiked with 10 µL IS1-3. Afterward, a 1 µL sample was then directly injected in a split mode (split ratio was 50:1). The LOD and LOQ values of standards were determined at their concentrations when their signal-to-noise ratios (S/N) were 3 and 10, and calibration curves were drawn using the response ratio between the target aroma and internal standards versus the ratio between their concentrations. Detailed information is provided in Supplementary Table S1. Every 1 mL SAB sample was spiked with 10 µL IS1-3, and then GC-FID quantification was performed the same way as the working standards. The signal was recorded and processed with OpenLab CDS (Agilent Technology, Santa Clara, CA, USA) software.

OAV Ratio-Based Random Forest Classifier
To evaluate the correlations between different volatile compounds, a ratio-based random forest classifier was used in this study programmed with Python packages including sklearn 0.23.2, pandas 1.1.4, and numpy 1.19.4 . Three data tables are needed for the classifier including a table of compound quantities, a table of odor thresholds, and a table of odor descriptions. Briefly, (1) data were cleaned based on non-zero values and outliers; (2) the OAVs of metabolites were divided exhaustively and then logged by 2 to get the correlations for each set of compounds; (3) an optimized random forest classifier was used for group separations, and the parameters (number of trees, maximum depth, minimum of leaves) for the classifier were chosen with the GridSearchCV algorism together with 10-fold cross validation. The algorism used in this study can be found at GitHub with the following link: https://github.com/MiracleEaTu/ratio_based_corr_classifier, accessed on 5 October 2022.

Statistics
The significance between the two groups in this study was calculated using the Kruskal-Wallis test. A p-value less than 0.05 was considered significant. Principle component analysis (PCA) was analyzed with GraphPad Prism 9.0. A Pearson correlation Foods 2022, 11, 3916 4 of 14 coefficient greater than 0.6 of two odors was the criterion used for a connection (edge) between pairs of odors. The odor correlation network was visualized using Cytoscape software (v3.9.0, National Institute of General Medical Sciences, Bethesda, MD, USA) [19]. Other figures were visualized with the 'ggplot2' package in R 4.1.0.
Before the random forest classification, volatile compounds that had zero values in more than half of the samples were removed and outlier samples in each group were removed with local outlier factor algorism performed with the sklearn 0.23.2 package.

The Differences in Aroma Compounds between the Two Groups of SABs
According to the sensory evaluation, 39 samples were evaluated as excellent grade or grade I; in the present study, these samples were grouped as a relatively good (RG) group. Meanwhile, 27 samples were evaluated as grade II and these samples were grouped as a relatively poor (RP) group.
Based on the GC-FID result, 63 volatile compounds were quantified in at least 17 SAB samples, including 27 esters, 12 alcohols, 10 acids, 5 aldehydes, 4 alkanes, 3 ketones and 2 furans. It was reported that esters and alcohols contributed greatly to the aroma of SABs and there were 6 important skeleton aroma compounds (ethyl hexanoate, ethyl acetate, ethyl butyrate, 1-propanol, 2-phenylethanol, and 3-methylbutanol) in SAB [20]. In our study, all 6 important aroma compounds were quantified. However, only 3-methylbutanol had a statistical difference between the RG and RP groups and was higher in the RG group. This result indicated that although the SAB skeleton aroma compounds are important for SAB flavor, their concentration alone could not comprehensively present the quality of SABs.
To illustrate the aroma compound differentiations between the RG and RP groups, PCA was performed. The PCA result ( Figure 1A,B) indicated a high similarity of aroma compound constructions between the RG and RP groups. Among 63 quantified compounds, 14 compound abundances were statistically different between the RG group and the RP group ( Figure 1C). It was reported that two fruity-like odors (ethyl isobutyrate, ethyl isovalerate) gave additive or synergistic odor effects for Baijiu [21]. The result was consistent in our study; the concentrations of ethyl isobutyrate and ethyl isovalerate were both significantly higher in the RG group. Except for a fatty-like compound (ethyl linoleate), the other 13 compounds all had higher abundances in the RG group; these compounds mainly contributed in fruity and malt aromas. The higher abundance of aroma compounds detected in the RG group indicated that SAB in the RG group had a richer aroma and that could possibly be the reason for the sensory evaluation judgment. Interestingly, isobutyraldehyde, 3-methylbutanol, and 3-methylbutyraldehyde had a contribution to the pungency perception [22], although pungency evaluation was generally perceived as a challenging task, owing to its low reproducibility, subjectivity, and taste fatigue of assessors [23]. It is interesting to notice that RG grade is associated with a relatively higher pungency character.

The Correlation between the Grade of SABs and OAV of Volatiles
The OAVs of the identified compounds were calculated to evaluate the odor contributions between the two groups of SABs. Among 64 volatile compounds, 63 compounds reported odor descriptions in distilled liquor and 36 odor descriptors were summarized from these compounds. Forty-one compounds had an average OAV higher than 1 in the RG group (Table 1). These volatile compounds were regarded as contributors to the overall aroma profiles of SAB samples according to the classical Baijiu aroma characterization methods [24,25]. Here in this study, we used an accumulated OAV (OAV sum ) for each odor descriptor as the profiles of odor compositions in SAB samples, and the overall OAV sum contribution is shown in Supplementary Figure S1. The PCA result of OAV sum (Figure 2A,B) illustrated that the odor composition in the RG group and the RP group had a high similarity, which was consistent with the PCA result calculated with compound quantities. Among 36 odors, 8 odors were statistically different between the RG group and the RP group ( Figure 2C). Not surprisingly, all OAV sum of eight odors in the RG group were higher than those in the RP group. The richer aroma in the RG group not only contained bread-, sweet-, malt-, green-, and creamy-like odors, but also contained pungent-, rancid-, and fusty-like odors. Interestingly, the OAV sum of the fruity-like odor did not show any significance between the two groups, whereas fruity-like aroma compounds were an important part in the SAB flavor profiles [26,27]. Aroma compounds in Baijiu had aroma compound interactions; thus, the authors believed odor interactions should be considered as an important factor to the perceptual quality, and therefore, the aroma compound balances should be taken into consideration for Baijiu grading and these interactions may explain the differentiations between the sensory evaluation and OAV sum results.      (C) Boxplots of odors having statistically different OAV sum between the RG and RP groups. The orange boxes represent the OAV sum in the RG group, and the yellow boxes represent the OAV sum in the RP group. The significances were calculated via the Kruskal-Wallis test and a p-value less than 0.05 was considered significant. *: p-value ≤ 0.05, **: p-value ≤ 0.01, ns: not significant.

Compound Co-Abundance Correlation Network of SAB of Two Grades
The authors applied a correlation network analysis based on Pearson's correlation coefficient ( Figure 3A), and 24 compounds were included in the network. The main cluster in the network centered on furfural, an important flavor compound that contributed in nutty-, sweet-, and bread-like odors in SABs [35]. Fruity-like volatile compounds take an important part in the correlation network, including 2-butanol, ethyl propanoate, ethyl isobutyrate, ethyl hexadecenoate, isoamyl acetate, and isobutanol, which were all correlated to furfural, and the average abundance of these compounds in the RG group were all higher than that in the RP group. It is worth noting that due to the esterification and hydrolyzation balance of ester/acid in SAB [36], a close association between several pairs of ester/acid in the correlation network was noticed, for example, the correlation between acetic acid and ethyl acetate, and the correlation between butyric acid and ethyl butyrate.
isobutyrate, ethyl hexadecenoate, isoamyl acetate, and isobutanol, which were all correlated to furfural, and the average abundance of these compounds in the RG group were all higher than that in the RP group. It is worth noting that due to the esterification and hydrolyzation balance of ester/acid in SAB [36], a close association between several pairs of ester/acid in the correlation network was noticed, for example, the correlation between acetic acid and ethyl acetate, and the correlation between butyric acid and ethyl butyrate.  To better understand the differentiation of compound abundance correlation in the two groups, two correlation networks were made separately in the RG group and the RP group ( Figure 3B,C). The core of the main cluster in the RG group was a combination of furfural, isobutyraldehyde, 3-methylbutyraldehyde, and isobutanol, whereas the core of the main cluster in the RP group was made up of isobutyraldehyde, 3-methylbutyraldehyde, 2-butanol, and propyl acetate. Isobutyraldehyde and 3-methylbutyraldehyde had strong correlation in both the RG group and the RP group. Furfural had more correlations with other compounds in the RG group than in the RP group (14 compared to 5). The authors also calculated the OAV sum correlation network of two groups (Supplementary Figure S2) and the result was interesting. The main OAV sum cluster in the RG group was centered with a bread-like odor, however, the main OAV sum cluster in the RP group was centered with creamy-and fatty-like odors. In addition, furfural contributed to the nutty-, sweet-, and bread-like odors. These results together indicated that furfural might be an important compound that may affect the SAB sensory, and the co-abundance of furfural with several compounds (e.g., 3-methylbutanol, isobutanol) could have a positive impact on SAB flavor. As a supplement to the perceptual interactions, the co-abundance network analysis here provides a new approach to unravel the complexity of the Baijiu system.

Important Odor and Compound Sets Selected by OAV Ratio-Based Random Forest Classification between the Two Quality Groups of SABs
Based on the compound composition results, the authors hypothesized that odor balances play an important role in SAB quality, and the ratio of compounds with different descriptors might be able to reflect the odor balance and, therefore, could be applied as candidate input parameters for classifiers of SAB quality grading. It is reported that the aroma other than the saltness of Douchiba, a Chinese traditional soy-fermented condiment, depends not only on key compounds but also on a "critical balance" or a "weighted concentration ratio" of 60 compounds [37]. Here, we adapted the concentration ratios and calculated the balances of aroma compounds in SAB. A random forest classifier is suitable for non-linear classifications [38] and was used for grouping the samples.
Specifically, OAV ratios of aroma compounds in two sets of odor descriptions were generated and logged by 2, as well as the OAV sum of odor descriptions. A total of 1097 logged OAV ratios were generated, and the dataset was regarded as the classifier input.
Then, the OAV ratio-based random forest classifier was performed, and the results are shown in Figure 4; the classifier gave us a reasonable 83.33% accuracy. The top 5 features ordered by importance were OAV sum of green-like odor/OAV sum of sweet-like odor, OAV sum of green-like odor/OAV sum of strawberry-like odor, OAV of ethyl valerate/OAV of isobutyraldehyde, and OAV of isobutyraldehyde/OAV of ethyl valerate. These ratios logged by 2 were all statistically different between the RG group and the RP group.
The authors also used an OAV dataset (Supplementary Figure S3) and a combined dataset including an OAV table after z-score normalization and the 1097 logged ratios (Supplementary Figure S4) for the classification. In the OAV dataset classification, the top 5 features ordered by importance in the classifier were propyl acetate, isobutanol, furfuryl alcohol, 3-methylbutanol, and 2-butanol, and the top 10 features in the classifier contributed to 6 odors, including fruity-, sweet-, pungent-, rancid-, sweaty-, and malt-like odors. The result of the combined dataset classification was basically a combination of the OAV classification and the OAV ratio-based random forest classification.
The ratio between ethyl valerate and isobutyraldehyde had high feature importance in the classifier. Ethyl valerate did not have a statistical difference between the RG group and the RP group, it had fruity-, apple-and strawberry-like odors [16], and these odors did not show differentiation between the RG group and the RP group either ( Figure 2C). Isobutyraldehyde, which was significantly higher in abundance in the RG group, had pungent-, malt-, and green-like odors [29], and these odors were all included in the significantly increased odors in the RG group. Notably, the OAV sum of the green-like odor divided by the OAV sum of the strawberry-like odor was also an important feature in the classifier. Additionally, as shown in Figure 3, in the RG group, the green-like odor was positively correlated to the strawberry-like odor, but in the RP group, there was no correlation between the green-like odor and the strawberry-like odor. This result indicated a potential correlation between the green-like odor and the strawberry-like odor, which was mainly contributed by isobutyraldehyde and ethyl valerate.  These results together with the OAV-based random forest classifier results and the odor correlations gave us a new insight into odor balances in the two groups of SABs. The classifier is suitable for compounds with only relative abundances such as peak areas detected from mass spectrometry. There are still several limitations to our OAV ratio-based classifier. For example, some aroma compounds have multiple odor descriptions, such as ethyl valerate, isobutyraldehyde, and ethyl hexanoate. The OAVs of these compounds were used multiple times in our odor correlation and group classifier calculations, which may duplicate the important features in the results.

Correlations of Key Compound Sets in SABs
Based on our OAV ratio-based classification result and content OAV classification, we selected 25 possible compounds that either had differentiation in abundance or showed high feature importance in at least one classification (Supplementary File S1).
The differentiations of sets of volatile compounds between the RG and RP groups were identified ( Figure 5). The interesting compound set of ethyl valerate and isobutyraldehyde had a positive correlation in both RG and RP groups. Isobutyraldehyde also had a positive correlation with 3-methylbutyraldehyde and furfural in both the RG and RP group. It is worth noting that in the RG group, the correlation coefficient scores between isobutyraldehyde and 3-methylbutyraldehyde and the correlation coefficient scores between isobutyraldehyde and furfural were above 0.9, representing the balances between these three compounds in SABs. Furfural had nutty-, sweet-and bread-like odors, whereas isobutyraldehyde had pungent-, malt-and green-like odors, and 3-methylbutyraldehyde had grass-and green-like odors. It was reported that 3-methylbutyraldehyde was one of the key aroma compounds in SAB [2], as well as the furfural content in SABs, and was more than those in strong-flavor Baijiu and light-flavor Baijiu [35]. The co-abundance relationship of these four volatile compounds had a p-value less than 0.05 but different correlation coefficient values in all SAB samples, which gave us a new insight that the combination of ethyl valerate, 3-methylbutyraldehyde, isobutyraldehyde, and furfural may be an important aroma set for SABs. In addition, the relationship between furfural and the other two compounds also supports that the off-odor compounds also play important roles in SAB flavors.

Conclusions
In this study, the authors quantified 64 aroma compounds in 66 SAB samples using a GC-FID platform; samples were grouped into two quality groups based on their sensory evaluation result. The authors used aroma compound correlations and OAV ratio-based classification to illustrate the odor balance differentiations between the RG and the RP groups. The results showed the OAV-based random forest classification and the OAV ratio-based random forest classification had the same accuracy but different precision. According to the OAV ratio-based random forest classification and compound correlations in the two quality groups of SABs, the balance of fruity-like ethyl valerate, green-and malt-like isobutyraldehyde, and malt-like 3-methylbutyraldehyde and sweet-like furfural might be one of the most important aroma compound balances sets in SAB. Though this study illustrated a new insight into aroma compound correlation, future studies are needed to illustrate the mechanism of aroma compound interactions and assess the impact of compound balance via sensory tests. The classifier established here using the compound OAV ratio as an additional dataset can be improved to be more suitable in future flavor studies.
Supplementary Materials: The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/foods11233916/s1, Figure S1: A heatmap showing OAV sum of all odors in the two quality groups of SAB; Figure S2: Compound co-abundance correlation network calculated with PCC in two quality groups of SAB; Figure S3: Random forest classification using OAV dataset; Figure S4: Random forest classification using combined dataset including logged OAV ratio dataset and OAV dataset; Table S1: Chemical standards and their quantification parameters.; File S1: Correlations of OAVs in the RG and the RP groups.

Data Availability Statement:
The algorism used in this study can be found at GitHub with the following link: https://github.com/MiracleEaTu/ratio_based_corr_classifier, accessed on 5 October 2022.

Conflicts of Interest:
Author Lei Zheng, Guangqian Liu, Songtao Wang, and Caihong Shen were employed by the company Luzhou Laojiao CO., LTD. The company was not involved in the study design, collection, analysis, interpretation of data, the writing of this article or the decision to submit