Investigating the flavor contributions of sulfur compounds in soy sauce flavor Baijiu using flavoromics and machine learning

Abstract

This study systematically investigated the flavor contribution of sulfur compounds to the soy sauce flavor Baijiu ( SSFB) by integrating traditional flavor analysis with machine learning methods. A total of 88 sulfur compounds were identified through the comprehensive two-dimensional gas chromatography-hydrogen sulfide chemiluminescence detector. The significant flavor contribution was confirmed by orthogonal partial least squares-discriminant analysis and odor activity value, and correlation analysis revealed their systematic associations with 8 aroma notes. Eight core sulfur compounds were screened in combination with 7 machine learning methods. The double matrix addition experiment demonstrated these compounds significantly enhanced the jiang and chen aroma. Notably, the study revealed the matrix-dependent flavor regulation of sulfur compounds. In young SSFB, these compounds simultaneously enhanced the jiang and chen aroma along with floral and fruity notes, while in aged SSFB, they acted more specifically to intensify the core jiang and chen aroma, with a reduced contribution to others.

Introduction

As a major category of distilled spirits in the world, Chinese Baijiu is renowned for its unique brewing technology and flavor characteristics¹. Among the 12 typical flavor types, soy sauce flavor Baijiu ( SSFB) has attracted much attention for its mellow and rich flavor. The national standard describes its typical sensory characteristics as “with prominent jiang aroma, harmoniously blended with fruity, floral and green aroma, accompanied by pronounced qu aroma, and a lingering fragrance in the empty cup”². Particularly noteworthy is the chen aroma formed by the prolonged aging of pottery jars, which is regarded as the core embodiment of the flavor of quality SSFB and serves as a key sensory indicator to judge its aging value³.

The chemical basis of Baijiu flavor has been systematically analyzed, and relevant studies have revealed the relationship between key aroma notes and volatile compounds. The flavor characteristics of Baijiu are mainly defined by trace substances, which only account for 1% to 2% of its total composition⁴. As of 2023, more than 3400 volatile compounds have been identified in Baijiu, including 729 esters, 419 acids, 304 nitrogen-containing compounds, and 217 sulfur-containing compounds⁵. Extensive research has confirmed that ester compounds constitute a significant portion of floral, fruity, and certain sweet aromas, while aldehydes and alcohols also contribute to floral notes^6–8. Moreover, acid compounds primarily impart acidic aromas^9,10. However, the key aroma compounds responsible for the iconic and most distinctive jiang and chen characteristic aroma of SSFB remain poorly elucidated.

In recent years, sensory analysis and flavoromics studies have increasingly demonstrated that sulfur-containing compounds are potential key contributors to the jiang and chen aroma profile. This is attributed to their extremely low odor thresholds and complex characteristic aromas, such as roasted, scorched, and even meaty notes^11–13. Research on sulfur compounds in Baijiu remains fragmented, primarily due to their extremely low concentrations, which poses persistent technical challenges for precise trace detection. Due to the limited sensitivity of traditional analytical techniques, early research primarily focused on sulfur compounds with relatively high concentrations. With advances in modern instrumentation, an increasing number of key sulfur compounds have been successfully identified. Li et al.¹³ achieved the measurement of benzyl mercaptan at concentrations as low as 1 µg/L through gas chromatography-mass spectrometry analysis (GC-MS) targeting the unknown aroma region in sesame flavor Baijiu, combined with standard substance validation. Song et al.^14,15 employed a comprehensive two-dimensional gas chromatography-sulfur chemiluminescence detector (GC×GC-SCD) to specifically identify key sulfides such as 4-methyl-4-mercapto-2-pentanone and 3-mercaptohexanol in various flavor types of Baijiu, suggesting their potential as flavor profile markers. Furthermore, studies have confirmed that certain sulfur compounds can modulate the release of fruity, floral, and acidic aroma notes through synergistic or masking effects^7,11,16, underscoring their critical role in flavor composition. However, existing research has primarily focused on individual compounds or specific categories, while a systematic analysis of their overall flavor contribution within the complex Baijiu matrix remains a significant challenge.

Traditional statistical methods and sensory description analysis have been maturely applied to Baijiu flavor analysis for many years, serving as reliable means for identifying key flavor compounds. However, they face certain limitations when handling high-dimensional and high-contribution data. Recently, machine learning algorithms have demonstrated significant advantages in analyzing complex systems, such as handling high-dimensional features and capturing nonlinear relationships. However, outcomes from purely data-driven models may lack direct interpretability within flavor science. Therefore, a hybrid intelligence strategy integrating traditional analytical methods with diverse machine learning algorithms is essential for achieving precise screening of flavor compounds^17–19.

Therefore, this study developed a hybrid strategy integrating traditional analysis with machine learning, which was systematically applied for the first time to comprehensively analyze sulfur compounds in SSFB. Overcoming the limitations of single methodologies through multi-model cross-validation and consensus screening enabled in-depth analysis of key flavor compounds, providing a methodological reference for achieving precise flavor profiling in complex systems. The specific research content includes: (1) establishing sensory characteristic profiles for SSFB across different aging periods (3, 17, 39, and 50 years); (2) comprehensive identifying and quantifying volatile sulfur compounds in the samples, along with calculation of their odor activity values (OAVs); (3) precisely screening key sulfur compounds by integrating three traditional flavor analysis methods with seven machine learning methods; (4) elucidating the direct contribution of target sulfur compounds to core aroma notes through dual-matrix addition experiments using authentic SSFB samples.

Results

Quantitative flavor profiles

Based on sensory quantitative description analysis, this study constructed radar charts of flavor profiles for samples aged for four different years, covering eight notes, namely, fruity, floral, sweet, acidic, qu, grain, jiang, and chen aroma (Fig. 1a). As shown in Fig. 1a, aged duration significantly influenced the overall flavor profile of SSFB. Specifically, the flavor profiles of the 3 N and 17 N samples exhibited relative harmony and balance. However, as aging time increases, the flavor profiles for the 39 N and 50 N samples showed pronounced distortion, accompanied by a noticeable absence of fruity, floral, and sweet notes, alongside a significant intensification of jiang and chen aroma notes.

Fig. 1. Dynamic evolution of sensory profiles across aging years.

a Radar charts of the overall flavor profiles; b Intensity changes of key aroma notes.

Further analysis of aroma intensity across individual notes (Fig. 1b) revealed that fruity, sweet, and floral aroma notes were less pronounced in the 39 N and 50 N samples compared to the 3 N and 17 N. Conversely, acidic reached higher levels in the 39 N and 50 N samples. This observation aligned with the universal principle of ester reduction and acidity increase during Baijiu aging. Previous studies have confirmed that esters are primary contributors to fruity, floral, and sweet aromas, while alcohols and aldehydes also participate in the formation of sweet note^6–8. During prolonged aging, the oxidation of alcohols and acids, along with the dynamic esterification between alcohols and acids, collectively led to ester hydrolysis and acid accumulation, as well as the conversion of alcohols and aldehydes into acids compounds^20–22. This process involved multiple dynamic equilibria, including the volatilization of compounds with low boiling points and the dissolution of trace ions. The evolution of acidic aromas stood in extremely contrast to that of fruity, sweet, and floral aroma notes, further supporting that esterification reactions and related oxidation processes were the primary pathways for the transformation of these compounds during aging.

As aging progresses, both the jiang and chen aromas exhibited an overall trend of increasing intensity. Although the aged aroma was less pronounced in the 39 N sample, potentially due to production process variations in the base Baijiu, its overall flavor evolution still aligned with the aging trend. In comparison, the variation in qu aroma demonstrated the most complex evolution, with no discernible similarity to the aging progression. This was highly likely due to the complex chemical basis underlying qu note. Research indicates that qu aroma exhibits strong correlations with esters, pyrazines, aldehydes, ketones, and even acidic compounds^23,24. Since the aging process affects these compounds inconsistently, the associated aroma profiles tended to show relatively random outcomes. Therefore, correlating the sensory differences among Baijiu samples of different aging years with their key compounds provided a reliable approach to elucidate the mechanism of flavor evolution.

Identification of volatile sulfur compounds

This study identified 88 volatile sulfur compounds in SSFB samples from different aged periods through GC×GC-SCD combined with reference standard comparison. As shown in Fig. 2a, a total of 50, 66, 74, and 60 compounds were respectively identified in the 3 N, 17 N, 39 N, and 50 N Baijiu samples. Notably, 33 compounds were consistently detected in all samples, forming the foundation of the sulfur compound profile in SSFB throughout aging.

Fig. 2. Evolution of sulfur-containing compounds during SSFB aging.

a and (b) represent variations in number; (c) and (d) represent variations in peak area.

The identified sulfur compounds included thiols, thiazoles, thiophenes, thioesters, and sulfides (Fig. 2b), demonstrating the diversity of sulfur compounds in SSFB. In terms of distribution, sulfides were the most abundant subclass across all aging years, followed by thiazoles, thiophenes, and thiols at similar quantities, while thioesters were the least abundant. Notably, the relative abundance of different compound categories (expressed as total peak area) showed a highly similar ranking across different aging years (Fig. 2c), which generally consistent with the compound category order. Specifically, sulfides were most abundant while thioester contributed the least. Only in the 39 N sample did the abundance ranking between thiophenes and thiols reverse compared to the others. These results indicated that although the absolute number of compounds fluctuates during aging, the relative content relationships among subclasses remained largely stable. Rarely did a dramatic change in the concentration of a single compound category led to a shift in the ranking order. This universal distribution pattern confirmed the existence of common chemical reaction mechanisms driving the evolution of sulfur-containing compounds.

Research evidence indicated that sulfur compounds undergo mutual conversion during the aging process of Baijiu. For instance, thiols could oxidize to form corresponding disulfides, such as 2-methyl-3-furanthiol forming bis(2-methyl-3-furyl) disulfide²⁵. They may also react with other thiols to generate asymmetric disulfides, such as the reaction with methanethiol yields methyl 2-methyl-3-furyl disulfide²⁶. Meanwhile, thioesters may degrade during aging, releasing thiols again, as illustrated by the conversion of S-methyl thioacetate and ethyl thioacetate into methanethiol and ethanethiol, respectively^27,28. These dynamic oxidation, coupling, and degradation reactions collectively constitute a complex transformation network. This network is likely the intrinsic cause of the regular variations in the types and abundance of sulfur compounds across different aging years, while maintaining stable proportional relationships between categories.

Heatmap analysis (Fig. 2d) revealed complex dynamic changes in the abundance of sulfur compounds during aging. Among the thiols, 50% of the components exhibited an initial decrease followed by an increase in concentration, including ethanethiol, furfuryl mercaptan, benzyl mercaptan, 2-phenylethanethiol, 2-methyl-3-furanthiol, and ethyl 3-mercaptopropionate. Sulfide compounds also exhibited a similar pattern. The concentrations of ethyl (methylthio) acetate, dibutyl sulfide, furfuryl isopropyl sulfide remained relatively stable throughout the aging process, while most other sulfides showed a trend of initial decrease followed by increase. This phenomenon is likely attributable to the volatilization of low boiling point compounds, which leads to an initial concentration decrease in the early aging stages, followed by the initiation of complex and slow transformation processes. The relevant reaction has been confirmed in previous studies²⁹. 4-Butyrothiolactone, propyl sulfide, diallyl sulfide, methyl propyl disulfide, propyl disulfide, propyl 2-methyl-3-furyl disulfide, 3-(methylthio) butanal, and 3-methylthiopropionic acid were detected only in the 17 N and 39 N samples. In contrast, isopropyl disulfide, allyl methyl sulfide, difurfuryl sulfide, 2-methyl-3-(methylthio) furan, 3-(methylthio) propionaldehyde appeared only in the 39 N and 50 N samples, indicating that these compounds were most likely formed during the aging process.

Quantification of volatile sulfur compounds

To thoroughly investigate the transformation patterns during the aging process of sulfur compounds, we further employed the external standard curve method combined with HS-SPME and GC×GC-SCD to achieve precise quantitative analysis of sulfur compounds present in all aged samples. As shown in Table 1, the concentrations of sulfur-containing compounds exhibited significant differences across the four distinct years (p < 0.05). Detailed analysis of the variation patterns reveals two primary trends: initial decrease followed by increase, and continuous decline. Specifically, methyl allyl disulfide, allyl propyl disulfide, diallyl trisulfide, and 3-acetylthiophene exhibit a stepwise decreasing trend during aging, reaching their lowest concentrations in the 50 N sample. This pattern suggests that sulfur compounds containing allyl structures may not undergo delayed formation during aging but instead continuously degrade or transform under varying storage temperatures and oxygen levels³. The concentrations of the remaining 24 sulfur compounds all exhibited a trend of decreasing first and then increasing, with 7 of them reaching their peak level in the 50 N samples. Notably, the concentrations of methyl 2-methyl-3-furyl disulfide and 2-methyl-3-furanthiol increased by 10-fold and 3-fold in the 50 N sample compared to the 3 N sample, respectively. The initial decrease in concentration may be attributed to volatilization or transformation under conditions such as elevated temperatures and oxygen enrichment. The subsequent recovery in concentrations further confirms that pathways for the generation or transformation of sulfur compounds may exist during prolonged aging processes. Changes in sulfur compound concentrations may be related to new substances dynamic equilibria established during the latter stages of aging. During prolonged storage, processes such as the gradual oxidation of alcohols and acids, along with the hydrolysis of esters, persistently occur. These reactions not only alter the skeletal composition of the SSFB but may also indirectly influence the formation of sulfur compound precursors or secondary transformations among sulfur species. For instance, the gradual degradation of esters and the concomitant increase of acids lead to changes in pH during aging, which may indirectly influence the formation of dimethyl sulfide and dimethyl sulfone³⁰. Meanwhile, variations in oxygen levels can selectively oxidize methanethiol to form dimethyl disulfide or dimethyl trisulfide, thereby modulating the concentration and distribution of certain sulfur-containing compounds³.

Table 1.

The concentrations and OAVs of 28 sulfur-containing compounds

No.	Compounds	Concerntions/(μg/L)				OTs/(μg/L)	OAVs
		50 N	39 N	17 N	3 N		50 N	39 N	17 N	3 N
1	Ethanethiol	57.32 ± 5.04	9.00 ± 0.61	0.06 ± 0	32.55 ± 0.43	0.8a	71.65	11.25	0.08	40.69
2	Methyl allyl disulfide	0.08 ± 0	0.08 ± 0	0.17 ± 0.01	0.25 ± 0.02	0.0063c	11.98	12.55	27.31	39.53
3	Dimethyl disulfide	48.33 ± 1.37	7.19 ± 0.51	26.09 ± 1.43	132.52 ± 7.89	9.13b	5.29	0.79	2.86	14.51
4	Thiazole	62.54 ± 1.76	62.11 ± 6.12	53.08 ± 4.56	68.79 ± 1.11	740d	0.08	0.08	0.07	0.09
5	2,4-Dimethylthiazole	0.97 ± 0.07	0.36 ± 0.03	0.21 ± 0.01	2.41 ± 0.2	0.1c	9.74	3.57	2.13	24.15
6	Dimethyl trisulfide	21.51 ± 1.57	9.94 ± 0.54	11.96 ± 0.29	82.36 ± 5.54	0.36 b	59.75	27.60	33.22	228.77
7	Allyl propyl disulfide	0.23 ± 0	1.02 ± 0.09	0.24 ± 0.02	21.01 ± 1.49	0.22e	1.06	4.63	1.10	95.50
8	Ethyl (methylthio)acetate	3.93 ± 0.21	0.74 ± 0.05	2.44 ± 0.24	1.42 ± 0.04	120f	0.03	0.01	0.02	0.01
9	2-Pentylthiophene	4.69 ± 0.29	0.19 ± 0	2.88 ± 0.05	9.08 ± 0.88	0.1f	46.94	1.88	28.80	90.82
10	Diallyl trisulfide	0.20 ± 0	0.19 ± 0	0.20 ± 0	1.87 ± 0.18	120f	0.00	0.00	0.00	0.02
11	Dimethyl sulfoxide	104.17 ± 8.48	10.46 ± 0.33	53.10 ± 4.29	46.51 ± 1.86	-	-	-	-	-
12	Benzyl mercaptan	0.50 ± 0.04	0.19 ± 0	0.32 ± 0.01	0.90 ± 0.02	0.0025g	201.14	74.82	128.00	358.29
13	2-Acetylthiazole	6.78 ± 0.06	1.08 ± 0.05	0.85 ± 0.07	2.45 ± 0.2	20h	0.34	0.05	0.04	0.12
14	Methyl 2-methyl-3-furyl disulfide	14.06 ± 1.04	1.05 ± 0.08	0.10 ± 0.01	0.13 ± 0.01	0.3i	46864.51	3510.19	338.21	439.30
15	3-Thiophenecarboxaldehyde	1.48 ± 0.04	0.32 ± 0.02	3.01 ± 0.07	4.81 ± 0.42	-	-	-	-	-
16	2-Thiophenecarboxaldehyde	11.96 ± 0.72	0.29 ± 0	0.68 ± 0.06	18.15 ± 0.03	5000c	0.00	0.00	0.00	0.00
17	2-Phenylethanethiol	2.36 ± 0.17	1.87 ± 0.02	4.00 ± 0.29	12.59 ± 1.17	0.00005c	47169.82	37376.43	79961.20	251865.43
18	Ethyl 2-thiophenecarboxylate	1.45 ± 0.11	0.31 ± 0.03	0.24 ± 0.02	2.60 ± 0.07	-	-	-	-	-
19	3-Acetylthiophene	1.87 ± 0.05	2.20 ± 0.17	2.47 ± 0.16	4.13 ± 0.3	-	-	-	-	-
20	2-Acetylthiophene	3.15 ± 0.07	1.06 ± 0.05	3.98 ± 0.31	15.14 ± 0.73	0.00008c	39385.90	13286.77	49716.49	189198.93
21	5-Methylthiophene-2-carboxaldehyde	1.63 ± 0.1	0.82 ± 0.01	0.73 ± 0.06	6.71 ± 0.23	-	-	-	-	-
No.	Compounds	Concerntions/(μg/L)				OTs/(μg/L)	OAVs
		50 N	39 N	17 N	3 N		50 N	39 N	17 N	3 N
22	Methyl furfuryl disulfide	1.79 ± 0.05	0.10 ± 0	0.84 ± 0.04	1.45 ± 0.06	0.00017c	10526.50	562.65	4917.38	8534.10
23	5-Acetyl-2,4-dimethylthiazole	159.43 ± 10.17	34.31 ± 1.56	121.04 ± 0.94	179.92 ± 1.04	5c	31.89	6.86	24.21	35.98
24	Dimethyl sulfone	25.78 ± 1.93	2.81 ± 0.17	6.29 ± 0.44	93.75 ± 7.6	-	-	-	-	-
25	Benzothiazole	10.80 ± 0.75	5.89 ± 0.08	6.67 ± 0.33	11.31 ± 0.75	0.35c	30.84	16.83	19.06	32.30
26	2-Methyl-3-furanthiol	21.26 ± 1.17	5.21 ± 0.18	6.20 ± 0.07	9.34 ± 0.24	0.0048i	4428.19	1086.32	1291.52	1946.67
27	Furfuryl thiopropionate	6.45 ± 0.08	9.80 ± 0.51	4.35 ± 0.07	20.29 ± 1.31	0.1c	64.47	97.99	43.53	202.87
28	2-Methylmercaptobenzothiazole	7.03 ± 0.31	2.32 ± 0.19	1.24 ± 0	3.90 ± 0.18	0.0075c	937.20	309.73	165.40	520.26

The superscript represents the source of the olfactory threshold, which are respectively a,³⁴; b,³⁵; c,³⁶; d,³⁷; e, chemical book; f, Flavor and Extract Manufacturers Association database; g,⁷; h,³⁸; i,³⁹. - Indicates that the odor threshold of this substance has not been determined. OTs were determined in the following matrices: water (dimethyl disulfide, 2,4-dimethylthiazole, 2-acetylthiazole, 2-thiophenecarboxaldehyde, 2-acetylthiophene, 2-methylmercaptobenzothiazole); air (ethyl (methylthio)acetate, diallyl trisulfide, methyl furfuryl disulfide); sunflower oil (2-phenylethanethiol); and ethanol-water solution (all other listed compounds).

Notably, due to the inherent instability of thiols, prolonged thermal desorption in the GC injection port can lead to oxidation reactions, potentially causing underestimation of the detected concentrations of thiols such as 2‑methyl-3-furanthiol and ethanethiol. Consequently, the quantitative values should be regarded as conservative estimates. In addition, although calibration curves were prepared using a 12% (v/v) ethanol-water solution, which closely approximates the sample matrix, the absence of non-volatile components may affect the headspace partitioning behavior, thereby influencing the quantitative accuracy. Despite these limitations, the relative differences among samples remain reliable, and the subsequent conclusions regarding the contributions of these compounds to key flavor attributes are not affected. Future studies could mitigate thiol degradation by employing derivatization strategies or cold injection techniques and more accurately reflect their true concentrations in Baijiu by combining approaches such as standard addition, stable isotope dilution analysis, or matrix-matched model solutions to minimize matrix effects.

Screening key sulfur compound based on OPLS-DA, OAVs and sensory correlation analysis

Based on the above analysis, it was found that the aged SSFB exhibit distinct differences in both sensory perception and substance composition. To deeply analyze the potential substance basis influencing flavor changes during the aging process of Baijiu and clarify the flavor contribution of sulfur compounds to SSFB across different aging years, this study employed three methodologies (OPLS-DA, OAVs, and correlation analysis between sulfur compound concentrations and SSFB aroma intensity) to identify key sulfur compounds closely associated with flavor evolution. The OPLS-DA and correlation analysis focused on all sulfur compounds identified in samples from four different aging years, rather than only the compounds observed in all samples.

PCA was first performed on the peak area data from 88 sulfur-containing compounds to explore the overall distribution differences among Baijiu samples under an unsupervised model. As shown in Fig. S1, the first two principal components (PC1 and PC2) collectively accounted for 69.1% of the variance, which sufficiently reflects the overall variability in the dataset. Samples were distributed across four distinct quadrants in the PCA score plot, indicating significant differences in sulfur compound composition among SSFB samples, which further confirms that sulfur-containing compounds serve as crucial chemical markers for distinguishing aging characteristics.

To further screen for key compounds responsible for inter-group differences, OPLS-DA was employed. As illustrated in Fig. 3a, the model clearly distinguished samples from different aging years. To ensure model reliability, rigorous cross-validation and 200 permutation tests were performed (Fig. 3b). In the model, R²X = 0.904, indicating strong explanatory power for independent variables; R²Y = 0.992, reflecting exceptional fit for year grouping; and Q² = 0.98, demonstrating excellent predictive ability. All permutation test points fell below the original model parameters, and the Q² regression line exhibited a negative intercept with the Y-axis, confirming the absence of overfitting and demonstrating that the results are robust and reliable.

Fig. 3. Results of OPLS-DA screening for key sulfur-containing compounds.

a, score plot; (b), permutation test; (c), 49 compounds with VIP > 1.

Based on this reliable model, Variable importance projection (VIP) values were calculated to assess each compound’s contribution to aged years discrimination (Fig. 3c). Using VIP > 1.0 as the criterion, 49 sulfur compounds were selected, indicating these substances underwent significant changes during aging and substantially contributed to the flavor differences in SSFB. This result further confirms the pivotal role of sulfur compounds in the evolution of aged flavor and highlights their potential as characteristic flavor markers for SSFB.

Meanwhile, the OPLS-DA screening results also revealed the limitation of this method in precisely identifying key flavor compounds. While effectively extracting intergroup differential compounds, the method yielded a large number of candidate substances and could not directly distinguish which compounds actually drive sensory attributes. Therefore, relying solely on VIP is insufficient to fully elucidate their flavor contribution mechanisms. Further integration of sensorily guided screening methods is necessary to achieve precise identification of key compounds.

The OAVs of the sulfur-containing compounds were calculated based on the quantitative results, as detailed in Table 1. The OAVs calculation results revealed that only five compounds had OAVs<1, while the vast majority exhibited OAVs substantially greater than 1. This confirms that sulfur compounds make significant flavor contributions in SSFB and are potential characteristic flavor compounds. Notably, five compounds, which include methyl 2-methyl-3-furyl disulfide, 2-methyl-3-furanthiol, 2-phenylethanethiol, 2-acetylthiophene, and methyl furfuryl disulfide, exhibited OAVs >2000, making them the most prominent contributors among all analyzed compounds. These high-OAV compounds primarily exhibit meaty, rubber-like, and roasted aromas. Their strong sensory impact highlights their crucial role in forming the overall flavor profile of SSFB. Furthermore, considering their marked concentration increase during aging, these compounds are likely core flavor compounds driving the formation of the jiang and chen characteristic aromas in SSFB. Importantly, these values may vary depending on the matrix, and therefore the calculated OAVs should be interpreted as relative rather than absolute indicators.

In flavor research, OAVs > 1 are commonly used as a criterion for screening key flavor compounds. However, in this study, the vast majority of sulfur-containing compounds met this criterion, indicating that relying solely on the OAVs is insufficient for the precise screening of key flavor compounds. Therefore, it is necessary to further integrate a multi-method strategy combining sensory attribute correlation analysis and machine learning to identify, from the many high-OAV compounds, the key compounds that actually drive the overall flavor.

Flavor compounds constitute the essential chemical foundation of SSFB’s flavor characteristics, directly or indirectly participating in the modulation of the Baijiu flavor profile. Changes in the concentration of key flavor compounds typically occur in parallel with variations in sensory perception intensity. Therefore, correlating compound concentrations with the corresponding aroma attribute intensities and evaluating their contribution based on the strength of the correlation represents an effective and reliable strategy for screening key flavor compounds. This study correlated the peak areas of 88 sulfur compounds in samples from different years with intensity data for various aroma notes. Through correlation analysis, key sulfur compounds significantly influencing flavor perception were identified, with detailed results presented in Fig. 4.

Fig. 4. Correlation heatmap between 88 sulfur-containing compounds and aroma notes.

In each cell, the size of the colored rectangle represents the absolute value of the correlation coefficient, with larger rectangles indicating stronger correlations. Color direction denotes the sign of the correlation: blue indicates positive correlation and red indicates negative correlation, with deeper shading reflecting larger absolute values. ** indicates highly significant correlation (p < 0.01).

The correlation analysis indicated that 52 sulfur-containing compounds exhibited significant correlations with at least one aroma note. The acidic note showed a significant positive correlation with 7 sulfur compounds, including ethyl thioacetate (meaty coffee), 2-methylthiophene (sulfurous), methyl propyl disulfide (onion-like), furfuryl mercaptan (roasted coffee), furfuryl methyl sulfide (coffee), isopropylthiobenzene (spicy), and propyl 2-methyl-3-furyl disulfide (roasted meat meaty). The interactions between fruit, sweet, and floral aromas and sulfur compounds exhibited high consistency, showing significant positive correlations with 28 sulfur compounds and significant negative correlations with 7 sulfur compounds. This consistency likely arises because the chemical constituents of these three aroma notes may belong to similar substances, thereby exhibiting comparable interaction patterns with sulfur compounds. Previous studies have confirmed this consistent pattern of interaction^7,11. Notably, these findings are entirely opposite to the results for acidic notes. When sulfides are positively correlated with acidic notes, they show a negative correlation with fruity, sweety, and floral notes, and vice versa. This phenomenon may relate to the general trend of decreasing esters and increasing acidity during Baijiu aging. The dynamic changes in ester and acidity concentrations cause their associated flavor perceptions to exhibit opposite trends. Furthermore, 9, 12, 15, and 24 sulfur-containing compounds showed significant positive correlations with qu aroma, grain aroma, jiang aroma, and chen aroma, respectively. Among these, methyl propyl disulfide (onion-like), 4-methyl-5-vinylthiazole (musty, cocoa), and propyl 2-methyl-3-furyl disulfide (roasted, meaty) were associated with all four aroma notes, all exhibiting nutty, roasted, and meaty aromas. The compound recurrence rate significantly associated with jiang aroma and chen aroma reached 60%, including ethyl thioacetate (meaty, coffee), bis(methylthio)methane (spicy, mushroom), isopropyl disulfide (meaty), ethyl 2-mercaptopropionate (meaty, cooked), 2-methyl-3-(methylthio) furan (meaty), 4,5-dimethylthiazole (roasted), 3-(methylthio)propanal (potato), p-toluenethiol (burnt green fatty), and 2-thiophenethiol (roasted). These nine compounds showed significant correlation with both iang and chen notes, all exhibiting characteristic roasted and meaty flavor.

Aging is one of the most critical production processes for SSFB, aiming to reduce undesirable flavors in young Baijiu and enhance a mellow, aged character. The intensification of qu, jiang, and chen aroma plays an important role in improving the mellowness of the Baijiu. Therefore, to further analyze the contribution of sulfur-containing compounds to these characteristic notes, screening and analysis were conducted on sulfur compounds associated with these three aroma notes, with the condition that they must be detected simultaneously in at least three SSFB samples. Based on these criteria, 10 sulfur-containing compounds were ultimately identified, which made substantial contributions to the aged sensory profile. In detail, the compounds significantly positively correlated with jiang aroma and detected in no fewer than three aging-year samples included 2-isopropyl-4-methylthiazole, 4-methylthio-4-methyl-2-pentanone, methyl 2-methyl-3-furyl disulfide, and 2-ethyl-4-methylthiazole. Three compounds associated with chen aroma were ethyl 3-methylthiopropionate, 2-methyl-3-furanthiol, and 3-mercapto-2-pentanone, while three compounds related to qu aroma were allyl propyl disulfide, 2-(methylthio)pyrazine, and allyl methyl disulfide.

Consensus screening key sulfur compound based on multi-model machine learning

In this study, 7 complementary machine learning methods were employed to systematically screen for key sulfur-containing compounds associated with flavor evolution during the aging of SSFB. The methods included variance filtering, ANOVA F-value, correlation evaluation, random forest, SVM-RFE, elastic net, and XGBoost, each independently evaluating features based on different statistical assumptions, thereby avoiding bias from any single model. Each method first independently selected the top 30% of compounds based on importance, after which a multi-model consensus approach was applied to retain only those features that ranked in the top 30% in at least three methods, thereby enhancing the robustness and scientific reliability of the screening results. Furthermore, in tree-based models, overfitting was mitigated by employing shallow tree depths and a large number of trees, while filtering methods relied on variance and p-values, which remain robust under small sample conditions and thus further ensured the reliability of the screening. These strategies together ensured the robustness of the feature selection, providing a solid computational foundation for subsequent experimental validation.

Machine learning analysis was conducted based on concentration data of sulfur compounds. To ensure comprehensive and effective data, standard curves were also established for the key sulfur compounds identified in Section 2.4 to determine their concentrations. The results are presented in Table S1. The compound screening results from machine learning are shown in Fig. 5, where a total of 25 characteristic compounds were screened across the 7 methods. Among the filter methods, screening prioritized compounds exhibiting substantial variation across samples. In the variance filtering, the variance range for screened substances spanned from 41 to 3410 (Table S2). The significance levels of the ANOVA F-value analysis screening results all met p < 0.01 (Table S3). Combined correlation evaluation revealed that seven compounds were selected by at least two of the three methods, namely dimethyl disulfide, allyl propyl disulfide, ethyl 3-methylthiopropionate, ethanethiol, 2-thiophenecarboxaldehyde, methyl furfuryl disulfide, and benzyl mercaptan. This strongly indicates these compounds are likely key dynamic markers during the aging process.

Fig. 5. Screening results of seven machine learning methods.

a, feature consistency heatmap; (b), final stable feature selection results.

The random forest model consists of 200 decision trees and exhibits high stability (ROC curve shown in Fig. S2). Although small sample sizes may limit model evaluation capabilities, the feature importance rankings provided by random forests remain highly valuable for reference, as confirmed by previous studies^17,18. Therefore, to minimize potential uncertainties caused by the limited sample size, we adopted strict screening criteria, retaining only the top 5 most important feature compounds for subsequent analysis to maximize the reliability of model results. The compounds ultimately screened by this method were 2,4-dimethylthiazole, 2-acetylthiophene, 2-methyl-3-furanthiol, 2-methylmercaptobenzothiazole, and 4-methylthio-4-methyl-2-pentanone. The elastic net model identified 19 compounds, confirming that sulfur compounds consistently contribute significantly to flavor profiles across different aging stages. The SVM-RFE analysis indicated optimal model performance with eight features, achieving a validation accuracy of 66.7%. The features determined under this condition were dimethyl disulfide, allyl propyl disulfide, dimethyl disulfide, dimethyl sulfone, dimethyl sulfoxide, benzyl mercaptan, dimethyl trisulfide, and benzothiazole.

Ultimately, we integrated the screening results from 7 machine learning methods, defining a compound as a key sulfur-containing compound if it was selected by at least three methods as a criterion. This approach successfully identified 14 key sulfur-containing compounds (Fig. 5b). Among them, dimethyl disulfide and allyl propyl disulfide were selected were the most frequently selected compounds, identified by 6 and 5 methods, respectively. The highest number of compounds selected by 3 methods reached 9, namely 2‑methylmercaptobenzothiazole, benzyl mercaptan, methyl 2‑methyl‑3‑furyl disulfide, methyl furfuryl disulfide, 2‑acetylthiazole, 2‑thiophenecarboxaldehyde, ethanethiol, dimethyl sulfoxide, and dimethyl sulfone. The multiparametric consensus screening results not only validate the significance of these compounds at the algorithmic level, overcoming the limitations of a single approach. More importantly, they provide a set of highly viable target compounds for deciphering the complex flavor system of SSFB, confirming these substances as the fundamental drivers behind the systematic evolution of its overall flavor profile during aging.

Validation of key sulfur compounds

The screening of the aforementioned key sulfur-containing compounds scientifically validated their potential as characteristic flavor compounds in SSFB. This established a targeted compound database of key sulfur compounds and deeply explored components that significantly contribute to SSFB flavor. However, flavor research has demonstrated that the flavor contribution of compounds in complex systems cannot be directly equated with the aroma of individual reference standards. Therefore, the true aromatic contribution of sulfur compounds within the SSFB system must be revealed in a mixed system context. Nevertheless, the extreme complexity of the SSFB system and the trace concentrations of sulfur compounds make verification via methods like GC-O-MS challenging. Therefore, this study employed flavor addition experiments as the core validation approach. By adding target compounds to authentic Baijiu samples, we directly evaluated their influence on the overall flavor profile and their impact on characteristic aroma notes, thereby achieving a profound analysis of their flavor mechanisms.

This study systematically screened key sulfur-containing compounds in SSFB by integrating traditional statistical methods with multiple machine learning approaches. Multiple flavor contribution pathways of sulfur compounds in SSFB were elucidated, including serving as markers for distinguishing different aging stages and regulating the release of key aroma notes. To systematically identify compounds that substantively contribute to the aged flavor and the formation of characteristic flavors during the aging of SSFB, and to avoid the bias inherent in any single method, a multi-criteria cross-validation strategy was adopted to further refine the target compounds.

Specifically, candidate compounds were required to meet at least three of the following four criteria: (1) VIP > 1; (2) OAVs > 1; (3) significant correlation between concentration changes and qu aroma, jiang aroma, or chen aroma, along with stable detection in at least three aged-year samples; and (4) joint selection by no fewer than three machine learning methods. Based on these criteria, 8 target sulfur-containing compounds were ultimately selected as subjects for the subsequent addition experiments, namely methyl allyl disulfide, 2,4-dimethylthiazole, allyl propyl disulfide, benzyl mercaptan, methyl 2-methyl-3-furyl disulfide, 2-methyl-3-furanthiol, 2-methylmercaptobenzothiazole, and ethyl 3-methylthiopropionate.

In this study, the 17 N sample and a young base SSFB were used as matrices in addition experiments to evaluate the impact of key sulfur-containing compounds on 5 aroma notes. The addition concentration of each compound was adjusted to match its measured concentration in the 50 N sample. The resulting changes in the intensity of each aroma attribute are presented in Fig. 6.

Fig. 6. Intensity changes of aroma notes in Baijiu matrices after the addition of sulfur-containing compounds.

a Young base Baijiu matrix; (b) 17N Baijiu matrix.

The results indicated that the vast majority of the target sulfur-containing compounds significantly enhanced both jiang aroma and chen aroma in both matrices, directly confirming that sulfur compounds are key contributing compounds to the characteristic aged flavor of SSFB. In the young base SSFB matrix, methyl 2-methyl-3-furyl disulfide, methyl allyl disulfide, and allyl propyl disulfide showed relatively prominent boosting effects on jiang aroma, with ΔI values of 0.250, 0.313, and 0.188, respectively. Simultaneously, 2-methyl-3-furanthiol, benzyl mercaptan, and methyl allyl disulfide significantly enhanced the chen aroma, with ΔI values of 0.313, 0.125, and 0.125, respectively. In the 17 N matrix, sulfur-containing compounds exhibited more pronounced enhancement effects on both jiang aroma and chen aroma. Benzyl mercaptan, methyl 2-methyl-3-furyl disulfide, 2-methyl-3-furanthiol, 2-methylmercaptobenzothiazole, and ethyl 3-methylthiopropionate all effectively enhance the aged characteristics of Baijiu, with methyl 2-methyl-3-furyl disulfide and 2-methyl-3-furanthiol demonstrating the most pronounced effects.

Further comparison of the addition experiment results in both systems reveals that the effects of sulfur-containing compounds exhibit matrix dependence. In the young SSFB system, the target compounds not only significantly enhance the intensity of jiang and chen aroma but also substantially promote the release of floral and fruity notes. 2-Methyl-3-furanthiol achieved a Chen aroma enhancement value ΔI of 0.313, but its effect on fruity aroma was even more pronounced with a ΔI = 0.438. Similarly, compounds such as benzyl mercaptan and methyl 2-methyl-3-furyl disulfide also exhibited significant synergistic effects on floral and fruity aromas in the young system. However, within the 17 N aged SSFB matrix, the same group of compounds significantly amplified the enhancement of jiang and chen aromas while shifting their contribution to floral and fruity notes toward minimal promotion or suppression. This pattern indicates that sulfur-containing compounds play a more specialized role in reinforcing the core jiang and chen aroma within the aged system. This finding suggests that the role of sulfur-containing compounds in flavor regulation evolves with the Baijiu’s maturity, where they increasingly collaborate to shape the rich, mellow characteristics of aged SSFB.

In addition, 2,4-dimethylthiazole exhibited consistent effects across both systems, exclusively enhancing fruity notes while suppressing other aromas. This indicates its flavor influence acts relatively independently of the matrix background. Furthermore, the changes in qu and chen aroma following substance addition showed relatively minor matrix effects (similarity coefficients of 66.67% and 83.33%, respectively), whereas jiang aroma, floral notes, and fruity aromas demonstrated greater sensitivity to matrix composition.

Discussion

This study comprehensively analyzed sulfur-containing compounds in SSFB of different aging years using HS-SPME combined with GC×GC‑SCD, identifying 88 sulfur compounds and systematically elucidating the pivotal role in flavors formation during SSFB aging. Based on the calculation of OAVs and VIP values, this study elucidated that sulfur compounds make significant flavor contributions in SSFB, while also revealing the limitations of relying solely on traditional methods for screening within the group of highly contributing compounds. Accordingly, this study further integrated sensory correlation analysis with multiple machine learning methods to identify the core sulfur compounds contribute prominently to characteristic aroma notes from multiple perspectives. The 8 core compounds were methyl allyl disulfide, 2,4‑dimethylthiazole, allyl propyl disulfide, benzyl mercaptan, methyl 2‑methyl‑3‑furyl disulfide, 2‑methyl‑3‑furanthiol, 2‑methylmercaptobenzothiazole, and ethyl 3‑methylthiopropionate.

The dual-matrix addition experiments further verified at the sensory level that these compounds significantly enhanced the aroma intensity of jiang and chen aroma notes in SSFB. This study further revealed that the flavor regulation of sulfur compounds exhibited pronounced matrix dependence. In young Baijiu, while enhancing the jiang and chen aroma, they also notably boosted floral and fruity notes. In contrast, their action became more specialized in reinforcing the jiang and chen aroma profile, with their contribution to floral and fruity notes diminishing or even shifting to suppression in aged Baijiu. This finding indicated that the role of sulfur compounds in SSFB flavor formation underwent dynamic transformation as the evolution of chemical composition.

Based on this, we propose a hypothesis to account for this pattern. The matrix dependence of flavor regulation in sulfur compounds is closely linked to the dynamic evolution of physicochemical properties during the aging process of Baijiu. In the young SSFB matrix, the degree of hydrogen-bonding between ethanol and water molecules is relatively low, resulting in a higher proportion of free ethanol and water. Consequently, aroma compounds exist predominantly in a dissolved state, and overall aroma release is somewhat suppressed. Meanwhile, esters are present at relatively high concentrations, and their OAVs dominate the aroma competition landscape, leading to a sensory predominance of fruity and floral notes. Under this background, the direct aroma contributions of sulfur compounds are limited. Instead, they primarily engage in intermolecular interactions within the liquid phase, where they exhibit pronounced perceptual synergistic effects with esters, further enhancing the expression of fruity and floral aromas. As aging progresses, the degree of ethanol-water association gradually increases, forming a more stable and ordered hydrogen-bond network. This structural change enhances the ability of sulfur compounds to migrate from the liquid to the gas phase, thereby increasing their aroma release. At the same time, the ester decreases and acid increases during aging, reducing the sensory dominance of esters and proportionally increasing the relative aroma contribution of sulfur compounds. In addition, the continuous Maillard reactions over prolonged aging leads to the accumulation of non-volatile macromolecules such as melanoidin-like compounds, which are closely associated with the characteristic jiang and chen aroma of aged SSFB and may modulate the volatilization behavior of sulfur compounds through complex interactions, ultimately reshaping the overall sensory profile. Essentially, the differential aroma contributions of sulfur compounds in young and aged SSFB fundamentally reflect the dynamic modulation of their sensory expression by the physicochemical environment of the matrix, rather than changes in the chemical properties of the compounds. This mechanistic hypothesis requires further validation through controlled model systems in which the degree of ethanol-water association, the acid-ester ratio, and the composition of macromolecules can be systematically manipulated.

In summary, this study overcame the limitations of traditional methods in analyzing high-contribution flavor data by establishing a research framework integrating traditional analysis, machine learning, and experimental validation. It clarified the pivotal role of sulfur-containing compounds in forming the key aromas of SSFB, providing explicit target compounds and theoretical foundations for directed flavor regulation. Future work could focus on deeply analysis of the transformation pathways of these key compounds and their interaction networks to unveil the underlying mechanisms of flavor formation in SSFB.

Methods

Chemical and sample

The Baijiu samples with aging periods of 50, 39, 17, and 3 years (50 N, 39 N, 17 N, 3 N) were sourced from a typical SSFB distillery in Sichuan Province for this study. The Baijiu samples were aged in sealed ceramic jars under relatively constant temperature and humidity conditions. Storage temperature was maintained between 19 and 21 °C, and relative humidity ranged from 60 to 70%.

All reference standards employed in the present work were purchased from Sigma-Aldrich (Shanghai, China) and had a minimum purity of >97%. Specific substance information can be found in Table S4. Pure water was obtained from a Milli-Q purification system (Millipore, Bedford, MA, USA). Sodium chloride (NaCl) and absolute ethanol (99.97%) were purchased from China National Pharmaceutical Group Corp (Shanghai, China).

Descriptive sensory analysis

The sensory evaluation panel consisted of 12 members (7 females and 5 males), aged between 25 and 32, all from the Baijiu Chemistry Laboratory at Beijing Technology and Business University, possessing a solid background in Baijiu flavor chemistry. All participants were in good health, free from smoking or other adverse habits. Written informed consent was obtained from all participants prior to the commencement of the study, which was approved by the Scientific Research Ethics Committee of Beijing Technology and Business University (Approval No. 231) and strictly adhered to the principles of the Declaration of Helsinki.

Prior to formal evaluation, the panelists underwent a week-long systematic training program. During the one-week training program, the panelists were first trained to recognize and evaluate the eight target aroma attributes, using a series of n-butanol solutions as reference standards for aroma intensity perception. Each training session consisted of independent scoring followed by group discussion and re-evaluation to calibrate individual intensity scales and align perceptual anchors across panelists.

Upon completion of training, inter-rater consistency was statistically assessed using the intraclass correlation coefficient (ICC (3, k)) based on a two-way mixed-effects model. The resulting ICC value of 0.77 indicated good agreement among panelists, confirming that the panel had achieved satisfactory calibration. Only upon meeting this threshold did the panel proceed to formal evaluation sessions.

During formal sessions, panelists performed independent olfactory evaluations of SSFB samples in a professional evaluation room. The intensity of eight key aroma notes (fruity, floral, sweet, acidic, grain, qu, jiang, and chen aroma) was scored on a 0 to 5 scale (0: not detectable; 5: extremely strong). Each sample was presented with a three-digit random code, and all assessments were replicated to ensure reliability. During the evaluation sessions, ICC was monitored after each round to ensure that panelist agreement remained stable throughout the study.

Headspace solid-phase microextraction (HS-SPME)

HS-SPME is a solvent-free sample preparation technique that integrates sampling, extraction, enrichment, and injection into a single step, offering the advantages of simplicity and high sensitivity. It is particularly well-suited for the analysis of trace volatile compounds in complex food matrices and has been widely applied to the detection and quantification of sulfur compounds, heterocyclic compounds, and aldehydes/ketones in food systems^31,32.

In this study, five milliliters of Baijiu sample were transferred into a 40 mL headspace vial and diluted with ultrapure water to adjust the ethanol content to 12% (v/v). The solution was then saturated with NaCl. A magnetic stir bar was inserted into the vial, which was subsequently incubated in a water bath at 40 °C for 30 min to reach equilibrium. A 50/30 μm DVB/CAR/PDMS fiber was exposed to the sample headspace for 30 min under continuous stirring. After extraction, the fiber was immediately inserted into the GC×GC-SCD injector and thermally desorbed at 250 °C for 5 min.

GC×GC-SCD method parameters

GC×GC-SCD analysis was performed using an integrated system, which primarily consisted of an Agilent 7890B gas chromatograph, a dual-stage quad-jet thermal modulator, and an Agilent SCD 355 detector. Separation was achieved using a two-column set: the first-dimension column was a DB-WAX (30 m × 0.25 mm × 0.25 μm), and the second-dimension column was a DB-5 (2 m × 0.1 mm × 0.1 μm), both from Agilent, USA. The temperature program for the first-dimension oven was set as follows: initial temperature 50 °C held for 2 min, increased at 3 °C/min to 150 °C, then immediately ramped at 5 °C/min to 230 °C and held for 10 min. The second-dimension oven temperature was maintained 5 °C above that of the first-dimension oven throughout the entire run. The modulator temperature was set 15 °C higher than the first one, with a modulation period of 5 s using liquid nitrogen for cooling. The injector was operated in splitless mode at 250 °C. High-purity helium (99.999%) was used as the carrier gas in constant flow mode at a rate of 1.0 mL/min.

The SCD burner and base temperatures were set at 800 °C and 250 °C, respectively. Hydrogen and air flow rates were set at 45 mL/min and 60 mL/min, respectively, and the ozone generator pressure was maintained at 0.3 Pa. The instrumental method was established based on previous research^14,33.

Qualitative and quantitative analysis of sulfur compounds with GC×GC-SCD

All sulfur-containing compounds were identified by comparison with commercial authentic standards. Under identical instrumental conditions, a standard solution of each compound at an appropriate concentration was analyzed by GC×GC-SCD, and its retention time was recorded to build a standard reference database. Compounds in the samples were identified by matching their retention times in both chromatographic dimensions against this database. The acceptable tolerance for retention time matching was set at ±20 s for the first dimension and ±0.05 s for the second dimension. All analyses were performed in triplicate.

Quantification of key target sulfur compounds was performed using an external standard calibration curve method. Individual standard compounds were dissolved in a 12% (v/v) ethanol-water solution to prepare a series of working standard solutions at 7 concentration levels. A volume of each standard solution equivalent to that of the Baijiu sample was transferred into a 40 mL headspace vial and processed following the HS-SPME and GC×GC-SCD procedures described in Sections “Quantification of volatile sulfur compounds” and “Screening key sulfur compound based on OPLS-DA, OAVs and sensory correlation analysis”. The standard curves were constructed by plotting the recorded peak areas against their corresponding concentrations. All samples were analyzed in triplicate.

Calculation of OAVs

The OAVs of the key target compounds were determined by dividing their measured concentrations in the Baijiu samples by their respective odor thresholds. The calculation was performed according to the following equation: $\mathit{OAV}=C∕O{T}_s$, where $C$ is the concentration of the compound, and $O{T}_s$ is its corresponding odor threshold.

Machine learning methods

This study employed seven machine learning methods to comprehensively screen key sulfur compounds in SSFB. Considering the limited sample size, the research emphasis was on achieving scientifically reliable feature selection through cross-validation based on multi-principle methods, rather than constructing predictive models. A complementary screening and evaluation system is systematically integrated by combining filtering methods, embedding methods, and wrapper/tree-based methods.

Initial screening was performed using variance-based filtering, ANOVA F-value, and correlation-based selection. These filtering methods are model-free and computationally efficient, enabling the rapid removal of compounds with minimal concentration variation across aging years or weak linear correlations with other variables. Elastic Net allows for sparse feature selection while effectively addressing multicollinearity, making it particularly suitable for small-sample, high-dimensional datasets and capable of stably identifying compounds with consistent and robust concentration change patterns. SVM-RFE optimizes feature selection based on model classification accuracy, recursively removing features that contribute least to the decision boundary, thereby identifying subsets of compounds most discriminative for differentiating aging years. Random Forest and XGBoost, as tree-based ensemble methods, do not rely on linear assumptions and can capture non-linear patterns in compound concentrations over aging years as well as interactions among compounds. Their feature importance is assessed based on impurity decrease or information gain within decision trees. A candidate was selected as a key compound only if (i) its average rank across the methods that evaluated it fell within the top 30% of all considered features, and (ii) it was concurrently selected by at least three methods.

Identifying sensory effects through addition

To validate the specific flavor contributions of the key sulfur compounds, a systematic addition experiment was conducted. First, sample 17 N, which has relatively low background concentrations, was selected as the matrix. Each target compound was added at a level corresponding to its concentration difference between the 50 N and 17 N samples. To further verify the contribution of the key sulfur compounds to the jiang and chen aroma notes, a young SSFB was also selected as matrix. The endogenous concentration of each target compound in this sample was first quantified following the methods described in Sections “Headspace Solid-Phase Microextraction (HS-SPME)” and “Qualitative and Quantitative Analysis of Sulfur Compounds with GC×GC-SCD”. The difference between this concentration and that in the 50 N sample was then calculated and used as the additive level. The specific addition concentrations can be found in Table S5.

All reconstituted samples were labeled with random three-digit codes. Trained panelists evaluated the samples with a randomized serving order to minimize sequence effects. Following the methodology outlined in Section “Descriptive Sensory Analysis”, the panelists sequentially rated the intensity of five aroma notes: fruity, floral, qu, jiang, and chen aroma. All experiments were performed in triplicate.

Statistical analysis

The statistical analysis was performed using SPSS v16.0 (SPSS Inc., Chicago, USA) for correlation test, while Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) was conducted with SIMCA version 14.0 (Umetrics, Umeå, Sweden). Machine learning analyses were conducted in R Studio (2025.09.2 + 418). All other graphical representations were generated using Origin 2022 (Origin Lab Corporation, Northampton, MA) and R (version 4.5.2) within the R Studio (2025.09.2 + 418).

Author contributions

Shiqi Yang and Hehe Li wrote and edited the main manuscript text. Shiqi Yang also contributed to conceptualization, methodology, software, and visualization. Xinyue Zhang and Jiao Niu performed data curation and formal analysis. Yi Shen carried out the investigation. Yonghua Wang, Jinyuan Sun, and Baoguo Sun supervised the study. All authors reviewed the manuscript.

Data availability

All data supporting the findings of this study are available within the paper and its Supplementary Information.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41538-026-00929-x.