Associations between microbial community succession and volatile flavor dynamics during traditional Huizhou Mao-tofu fermentation

Abstract

Mao-tofu is a traditional fermented soybean product, yet the relationships between microbial succession and flavor evolution remain poorly understood. This study systematically characterized bacterial and fungal succession alongside volatile flavor changes during traditional Huizhou Mao-tofu fermentation. The initial Suanjiangshui (SW) was dominated by Lactobacillus (91.37%) and Pichia (77.81%), whereas Enterococcus and Cutaneotrichosporon became enriched in the later stages. Relative odor activity value analysis revealed a shift from a damascenone-dominated aroma background in SW towards profiles dominated by 1-octen-3-one and 2,4-decadienal, accompanied by changes in esters and aldehydes. Rate-based two-way orthogonal partial least-squares, multiple factor analysis and Spearman correlation analyses showed that Comamonas covaried with lipid-oxidation-related volatiles, whereas Enterococcus, Lactococcus and Cutaneotrichosporon covaried with selected esters and aldehydes. These associations identify priority targets for strain-level functional and sensory validation and may inform future strategies for flavor control and quality improvement in fermented tofu.

Graphical abstract

Highlights

• •Bacterial and fungal communities showed stage-dependent succession in Mao-tofu
• •Lactobacillus and Pichia dominated Suanjiangshui before tofu fermentation
• •ROAV profiles shifted from damascenone to 1-octen-3-one and 2,4-decadienal
• •Enterococcus and Cutaneotrichosporon were enriched at later stages
• •Multi-omics integration revealed exploratory microbe-volatile covariation

1. Introduction

Mao-tofu is a traditional fermented soybean product from the Huizhou region of China. During fermentation, a dense white mycelial layer develops on its surface, giving the product its distinctive appearance (Guan et al., 2024). With its long history and distinctive production process, Mao-tofu is regarded as a representative fermented soybean food in Chinese culinary culture. Fermentation improves the quality attributes of Mao-tofu by enhancing nutrient digestibility, promoting the release of bioavailable components, and inducing textural modification of the soybean curd matrix through enzymatic processes (Li et al., 2025; Surya et al., 2025; Wu et al., 2024; Zhao & Zheng, 2009). Microbial metabolism further promotes the accumulation of free amino acids, amino nitrogen, and volatile compounds, which collectively impart the characteristic umami taste and ripened aroma of Mao-tofu (Guan et al., 2024; Yan & Dong, 2019; Zhao & Zheng, 2009).

The volatile profile of Mao-tofu is highly complex, comprising alcohols, aldehydes, esters, and other key odorants that collectively dictate its sensory properties (Guan et al., 2024; Li et al., 2025; Yan & Dong, 2019). In fermented soybean systems, the aroma compounds are generally derived from three major metabolic pathways: protein degradation followed by amino acid catabolism, lipid hydrolysis and fatty acid oxidation, and microbial conversion of residual carbohydrates (Guan et al., 2024; Yan & Dong, 2019). Consequently, flavor development cannot be fully explained by substrate chemistry alone but must be understood in conjunction with the succession of the fermentation microbiota.

Suanjiangshui (SW) is a naturally fermented acidic whey derived from soybean pressing and produced through repeated backslopping followed by spontaneous acidification at room temperature (Yan et al., 2020). This process facilitates the accumulation of organic acids, predominantly lactic acid, reducing the pH to strongly acidic levels (typically below 4.0 in mature SW) and selectively enriching acid-tolerant microbial populations such as lactic acid bacteria (Yan et al., 2020). Together, these physicochemical and microbial features suggest that SW provides an important ecological foundation for subsequent Mao-tofu fermentation. Despite its role in this traditional process, however, the ecological dynamics of Mao-tofu fermentation remain insufficiently understood. Previous research has predominantly focused on isolating functional strains for artificial inoculation or providing descriptive physicochemical profiles (Guan et al., 2024; Yan et al., 2020; Yan & Dong, 2019). An integrated characterization of bacterial and fungal succession alongside the temporal development of aroma-active compounds is therefore still lacking.

To address this gap, the present study systematically investigated bacterial and fungal succession alongside changes in volatile flavor compounds during traditional Huizhou Mao-tofu fermentation. High-throughput 16S rRNA gene and ITS sequencing were employed to track microbial community shifts across fermentation phases. In parallel, volatile flavor profiles were characterized using headspace solid-phase microextraction coupled with gas chromatography–mass spectrometry (HS-SPME-GC–MS). Relative odor activity values (ROAV) were then calculated to identify the compounds that contributed most strongly to aroma. Furthermore, O2PLS, multiple factor analysis (MFA), and Spearman correlation were executed to establish statistical linkages between specific microbial taxa and key flavor metabolites. Together, these analyses aimed to establish an integrated ecological framework for traditional Huizhou Mao-tofu fermentation.

2. Materials and methods

2.1. Sample preparation and fermentation process

Mao-tofu samples were obtained from Chen's Old Lady Mao-tofu Workshop in Xiuning County, Huangshan City, Anhui Province, China, a traditional producer with a history of more than one hundred years. The samples were collected from the same traditional Huizhou Mao-tofu production system as previously described (Yan et al., 2020). After coagulation and pressing, the tofu was cut into cubes (approximately 4.0 × 4.0 × 3.0 cm), inoculated with a spore suspension of Mucor rouxianus (CICC 3154), and placed on bamboo trays to facilitate mycelial growth. Fermentation proceeded for 4 days in a traditional workshop fermentation room maintained at 20 °C and approximately 90% relative humidity. Samples were collected at five temporal points: the initial SW medium and tofu fermented for 1, 2, 3, and 4 days (designated as M01, M02, M03, and M04, respectively). Sampling was conducted across five independent production batches to account for inherent batch-to-batch variation. All samples were immediately frozen in liquid nitrogen to arrest microbial activity and stored at −80 °C until DNA extraction and volatile compound analysis.

2.2. DNA extraction and amplicon sequencing

Total genomic DNA was extracted from the SW and Mao-tofu samples using the E.Z.N.A.® Soil DNA Kit (Omega Bio Tek, Norcross, GA, USA) according to the manufacturer's protocols. DNA quality was assessed by 1.0% agarose gel electrophoresis, while DNA concentration and purity were quantified using a NanoDrop 2000 spectrophotometer (Thermo Scientific, Wilmington, DE, USA). The V3-V4 region of the bacterial 16S rRNA gene was amplified with primers 338F (5’-ACTCCTACGGGAGGCAGCAG-3′) and 806R (5’-GGACTACHVGGGTWTCTAAT-3′). The fungal ITS1 region was amplified employing primers ITS1F (5’-CTTGGTCATTTAGAGGAAGTAA-3′) and ITS2R (5’-GCTGCGTTCTTCATCGATGC-3′). Each 20 μL PCR mixture contained 10 μL of 2× Pro Taq master mix, 0.8 μL of each primer (5 μM), 1 μL of template DNA (10 ng/μL), and ddH₂O to reach the final volume.

Amplification was performed on an ABI GeneAmp® 9700 thermal cycler (Thermo Fisher Scientific, Waltham, MA, USA). The thermal program comprised an initial denaturation at 95 °C for 3 min, followed by 27 cycles of denaturation at 95 °C for 30 s, annealing at 55 °C for 30 s, and extension at 72 °C for 30 s. A final extension was performed at 72 °C for 10 min, followed by a hold at 4 °C. Amplicons were excised from 2% agarose gels, purified using a commercial PCR Clean-Up Kit (YuHua, Shanghai, China), and quantified via a Qubit 4.0 fluorometer (Thermo Fisher Scientific, Waltham, MA, USA). Sequencing libraries were prepared using the NEXTFLEX® Rapid DNA-Seq Kit (PerkinElmer, Waltham, MA, USA) and sequenced on an Illumina NextSeq 2000 platform (Illumina, San Diego, CA, USA).

2.3. Bioinformatics analysis

Raw reads were demultiplexed and quality-filtered using fastp (version 0.23.4), merged using FLASH (version 1.2.11), and denoised into amplicon sequence variants (ASVs) with DADA2 in QIIME 2 (Chen et al., 2018). Taxonomy was assigned with the classify-sklearn Naive Bayes classifier at a confidence threshold of 0.7. Bacterial and fungal sequences were referenced against SILVA 138.2 and UNITE 9.0, respectively, yielding 1103 bacterial and 182 fungal ASVs.

2.4. Determination of volatile compounds

Volatile profiles were characterized using HS-SPME-GC–MS according to Wang et al. (2024), with minor modifications. Briefly, a homogenized sample (3.0 g) was transferred into a 20 mL headspace vial containing 5.0 mL of saturated sodium chloride solution and 2.0 μL of n-pentadecane-d₃₂ (50 μg/mL) as the internal standard. Volatile compounds were extracted using a Thermo Scientific SMART SPME Arrow coated with DVB/Carbon WR/PDMS (phase thickness, 120 μm; outer diameter, 1.1 mm; phase length, 20 mm; catalog no. 36SA11T1-SM; Thermo Fisher Scientific, Waltham, MA, USA). Samples were incubated at 60 °C for 20 min and subsequently extracted for 10 min. The SPME Arrow was then desorbed in the GC injector at 240 °C for 5 min. Chromatographic separation was performed on a VF-WAXms capillary column (25 m × 0.25 mm × 0.20 μm; Agilent CP9204) using high-purity helium as the carrier gas at a constant flow rate of 1.0 mL/min. Samples were injected in split mode at a split ratio of 10:1. The oven temperature was initially maintained at 40 °C, increased to 120 °C at 8 °C/min, and then raised to 230 °C at 20 °C/min, followed by a 4.5 min hold. The total chromatographic run time was 20 min. A homologous series of n-alkane standards (C₁₀–C₂₉) was analyzed under the same chromatographic conditions for calculation of the experimental retention indices (RI). Mass spectra were acquired in full-scan mode over an m/z range of 35–500. Electron-impact ionization was performed at 70 eV with an ion-source temperature of 250 °C and a resolving power of 30,000 full width at half maximum.

Following chromatographic peak detection, volatile compounds were annotated by comparing the acquired mass spectra with the NIST 2023 library, the GC-Orbitrap flavor and fragrance library, and an in-house database. Experimental RI were calculated from the retention times of the adjacent n-alkanes and compared with the corresponding library RI values to support compound identification. The final identification criteria included retention time, molecular formula, experimental RI, library RI, RI score, mass spectral match factor, and overall identification score. Semi-quantitative abundance was expressed as the peak-area ratio of each target compound to the internal standard.

2.5. Statistical analysis

Microbial-community bioinformatic and statistical analyses were conducted using the Majorbio Cloud Platform. To evaluate structural discrepancies within the microbial communities, beta diversity was assessed via Bray-Curtis dissimilarity. Principal coordinate analysis (PCoA), coupled with ANOSIM and Adonis permutations, was executed to quantify overall community variations across distinct fermentation phases. Taxonomic composition was visualized as stacked bar plots based on mean relative abundances, highlighting the top 30 taxa while aggregating the remainder as “Others”. To identify phase-characteristic microbial biomarkers, linear discriminant analysis effect size (LEfSe) was applied using an all-against-all multiclass comparison strategy across taxonomic levels from phylum to species, with an LDA score threshold of 3.0 and a significance level of p < 0.05.

For volatile metabolite profiling, principal component analysis (PCA) and partial least squares discriminant analysis (PLS-DA) were performed using the ropls package in R. prior to multivariate modeling, metabolite data were subjected to total sum normalization and mean-centering. The robustness of the PLS-DA model was evaluated using R²X and R²Y to assess goodness of fit, alongside Q² for predictive ability (with Q² > 0.5 indicating robust performance). Model validity was further verified via permutation testing (200 permutations, 0.95 confidence level).

Relative odor activity value (ROAV) analysis was used to identify potential aroma-active compounds based on semi-quantitative volatile abundance and odor thresholds in water. For each biological replicate, the compound exhibiting the highest OAV was assigned an ROAV of 100, and the ROAVs of the remaining compounds were calculated relative to this reference. The resulting ROAV values were subsequently averaged across the five biological replicates for each fermentation stage. Compounds with ROAV ≥1 were classified as potential key aroma-active compounds, whereas those with 0.1 ≤ ROAV <1 were considered aroma modifiers (Xiao et al., 2022). Because volatile compounds were semi-quantified using internal standard normalization rather than external-standard calibration, ROAV values were used only to estimate relative aroma potential. Because the reference compound was selected separately for each biological replicate, the stage-mean ROAV values remained stage-relative and were not interpreted as directly comparable across stages. Cross-stage changes were therefore evaluated using RC (%) and qualitative changes in the composition of potential aroma-active compounds, whereas ROAV values were used only for within-stage ranking and screening. The ROAV-based relative contribution of each selected compound was calculated as follows:

$${\mathrm{RC}}_{\mathrm{i}}\left(\%\right)=\frac{{\text{ROAV}}_{\mathrm{i}}}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\text{ROAV}}_{\mathrm{i}}}\times 100$$

where RC_i represents the relative contribution of compound i, and the denominator is the sum of the ROAV of all detected volatile compounds.

To avoid misinterpreting concurrent increases in microbial relative abundance and volatile accumulation as evidence of a direct association, a rate-based multi-omics analysis of adjacent-stage changes was performed. For each biological replicate, the microbial change rate was calculated as ΔM/Δt, where ΔM represents the change in genus-level relative abundance between two adjacent stages and Δt is the corresponding fermentation interval. Similarly, the apparent volatile activity change rate was calculated as ΔV/Δt for the potential key aroma-active compounds identified by ROAV analysis. Here, ΔV represents the adjacent-stage difference in the semi-quantitative abundance of each compound. Four adjacent transitions were analyzed: M01–SW, M02–M01, M03–M02, and M04–M03. The SW–M01 interval (Δt) was likewise treated as 1 day, consistent with the sampling interval used for all other adjacent-stage transitions.

Rate-based multi-omics analyses were performed in R version 4.2.2. Separate bacteria–volatile and fungi–volatile O2PLS models were fitted using OmicsPLS version 2.1.0 after mean-centering and unit-variance scaling of each data matrix. For both models, two joint components and one orthogonal component for each data block were used (n = 2, nx = 1, ny = 1). MFA was implemented using base R functions by first centering and scaling each data block, weighting each block by the inverse square root of its first eigenvalue, and then performing global principal component analysis on the concatenated weighted matrix. The first two dimensions were retained for visualization. Pairwise associations were further assessed using Spearman correlation. The resulting p values were adjusted separately for the bacterial and fungal analyses using the Benjamini–Hochberg procedure, with q < 0.05 considered significant. All correlations were interpreted as exploratory covariation patterns rather than evidence of direct microbial production of these volatiles.

3. Results and discussion

3.1. Temporal succession of microbial communities during fermentation

The bacterial community showed clear temporal succession during traditional Huizhou Mao-tofu fermentation, accompanied by variation in α-diversity among fermentation stages (Fig. S1). As illustrated in Fig. 1A, bacterial communities from SW and Mao-tofu samples collected at different fermentation points were clearly separated, with PC1 and PC2 explaining 57.31% and 8.29% of the total variance, respectively. ANOSIM confirmed strong differentiation among stages (R = 0.9837, p = 0.001), indicating substantial changes in bacterial community structure during fermentation. At the genus level, SW was dominated by Lactobacillus (91.37%) and Limosilactobacillus (8.10%) (Fig. 1B). Lactobacillus remained highly abundant in M01 (82.47%), indicating continuity between the early Mao-tofu community and the initial SW microbiota. This early dominance aligns with the ecological role of Lactobacillus in acidic fermented food systems, where it promotes rapid acidification and imposes strong selective pressure on competing microorganisms (Chen et al., 2025). A marked transition occurred at M02, when the relative abundance of Lactobacillus decreased to 39.56% and that of Acinetobacter increased to 30.22% (Fig. 1B). This transition is consistent with the reported metabolic versatility of Acinetobacter, which can utilize a broad range of carbon sources, including amino acids, organic acids, and lipid degradation intermediates released during early proteolysis and lipolysis of the soybean curd matrix (He et al., 2025; Yang et al., 2024).

Fig. 1.

Succession of bacterial and fungal communities during traditional Huizhou Mao-tofu fermentation. Principal coordinate analysis (PCoA) of bacterial (A) and fungal (C) communities, and relative abundance profiles of bacterial (B) and fungal (D) communities, in SW and Mao-tofu samples collected at different fermentation stages. SW, Suanjiangshui; M01, M02, M03, and M04 represent Mao-tofu samples fermented with SW for 1, 2, 3, and 4 days, respectively. The same abbreviations and designations apply to all subsequent figures unless otherwise specified.

In the later stages (M03 and M04), the bacterial community became more diverse and comprised mainly Lactobacillus, Acinetobacter, Kurthia, Enterococcus, Weissella, Streptococcus, and Lactococcus (Fig. 1B). This diversification may reflect a relaxation of the strong acidic selection pressure that prevailed during early fermentation rather than continued acidification (Tan et al., 2022; Yan et al., 2020).

The fungal community also exhibited directional succession across the fermentation stages. As shown in Fig. 1C, fungal communities in SW and Mao-tofu samples were significantly separated, with PC1 and PC2 explaining 40.44% and 15.40% of the total variance, respectively (ANOSIM R = 0.7474, p = 0.001). SW samples exhibited relatively high within-group dispersion, whereas M01 samples formed a tighter cluster after transfer to the tofu matrix. This result suggests that transfer to the tofu matrix imposed a strong environmental filter on the incoming fungal community, restricting colonization to a consistent subset of adapted taxa. At the genus level (Fig. 1D), SW was jointly dominated by Pichia (77.81%), Trichosporon (11.15%), and Exophiala (5.47%). After inoculation onto tofu, Pichia declined rapidly, whereas Trichosporon became the predominant genus from M01 onwards. Pichia is commonly associated with spontaneous fermentation and may facilitate early colonization and aroma formation through the production of esters, higher alcohols and other aroma-related metabolites (Qiu et al., 2022; Zhang et al., 2023). However, its relative advantage may decline as microbial competition intensifies and the substrate environment shifts away from readily available nutrients (Chu et al., 2023; Padilla et al., 2016; Zhang et al., 2023). By contrast, Trichosporon appeared to be better adapted to the mature tofu environment. The enrichment of Trichosporon has also been reported in fermented tofu systems (Qiu et al., 2022). Similar succession involving Trichosporon in other fermented soybean matrices has been associated with environmental transitions and flavor development (Zhang et al., 2023).

Although Mucor remained detectable from M01 through M04, its relative abundance in the bulk ITS amplicon dataset was consistently lower than that of Trichosporon, Pichia, and other main taxa. This pattern is consistent with previous observations in traditional Mao-tofu, where Mucor showed low relative sequence abundance despite its macroscopically visible mycelial growth (Benucci et al., 2022). Amplicon sequencing detects taxon-associated DNA but cannot distinguish DNA originating from viable and non-viable cells. Therefore, the low relative abundance of Mucor in homogenized samples does not necessarily indicate low surface mycelial biomass. To complement the compositional ITS data, Mucor target-gene copy numbers were quantified using absolute qPCR. Detailed procedures and results are provided in the Materials and methods and Results and discussion sections of the Supplementary Material, respectively. Stage-resolved data are presented in Fig. S2 and Table S1.

3.2. Biomarkers identified across fermentation stages

To identify microbial taxa that discriminated among the fermentation stages of traditional Huizhou Mao-tofu, linear discriminant analysis effect size (LEfSe) was applied across taxonomic levels from phylum to species. An LDA score threshold of 3.0 was applied to identify robust temporal biomarkers and to mitigate the risk of overinterpreting weak signals (Yang et al., 2022). The phylogenetic distributions of stage-associated taxa are shown in the cladograms in Fig. 2A and C, while their effect sizes are presented in Fig. 2B and D. Overall, 25 bacterial biomarker taxa were identified across taxonomic levels, including 5 markers at the species level, whereas 45 fungal biomarker taxa were identified, including 14 markers at the species level.

Fig. 2.

Differential bacterial and fungal taxonomic features identified using linear discriminant analysis effect size (LEfSe) during traditional Huizhou Mao-tofu fermentation. Cladograms of bacterial (A) and fungal (C) communities showing taxonomic biomarkers enriched in Suanjiangshui (SW) and Mao-tofu samples collected at different fermentation stages. Histograms of the linear discriminant analysis (LDA) scores for bacterial (B) and fungal (D) biomarkers.

The bacterial biomarkers exhibited a clear transition from an acid-driven community characteristic of the initial stage to a progressively matrix-adapted consortium during tofu fermentation. In SW, the community was strongly characterized by Lactobacillus amylolyticus (LDA = 5.98), aligning with the strongly acidic environment that favors acid-tolerant bacteria with active carbohydrate utilization and lactic acid production (Fei et al., 2017; Fei et al., 2022). After transfer to the tofu matrix at M01, a marked ecological shift was observed, with Comamonas aquatica (LDA = 5.89) emerging as the key biomarker. Such ecological filtering is consistent with previous observations in fermented soy whey and other soybean fermentation systems, where matrix restructuring is accompanied by rapid microbial reassembly and related metabolic differentiation (Chen et al., 2023). Notably, no bacterial taxon exceeded the LDA threshold of 3.0 at M02, suggesting that this stage represented a highly transitional period of community reassembly. By M03, the Rothia kristinae (LDA = 5.41) and Corynebacterium flavescens (LDA = 5.75) were enriched. Among these, C. flavescens has been reported as a maturation-associated bacterium in foods, which supports its relevance to the advanced fermentation stage (Masoud & Jakobsen, 2003). At M04, Pseudoclavibacter sp. became a representative biomarker, suggesting further specialization of the mature bacterial consortium.

The fungal biomarkers displayed a similarly directional but more strongly differentiated succession pattern during fermentation. In the initial SW sample, the community was overwhelmingly characterized by Pichia kudriavzevii (LDA = 5.95). As a highly stress-tolerant yeast, P. kudriavzevii is well adapted to the harsh and acidic environment of the starter (Chu et al., 2023). It was likely associated with initial aroma development and may have contributed to the formation of volatile esters and higher alcohols after transfer to the tofu matrix (Chu et al., 2023; Xu et al., 2024). After inoculation onto the solid tofu matrix at M01, the system entered a transient colonization stage, marked by the temporary emergence of early colonizers such as Exophiala dermatitidis (LDA = 4.88). A clearer fungal transition was observed at M02, where the dominant biomarkers shifted to Diutina rugosa (LDA = 5.60) and Dipodascaceae sp. (LDA = 5.20). Based on the reported extracellular lipolytic and proteolytic capacities of these taxa in food fermentations, their proliferation is theoretically interconnected with the matrix degradation process, which would continuously supply free amino acids and fatty acids as essential precursors for subsequent flavor formation (Lotti et al., 1998). By M03, the fungal community became more heterogeneous, with Trichosporon asteroides (LDA = 5.24) and Clavispora lusitaniae (LDA = 4.74) emerging as representative biomarkers. In the mature stage (M04), the Mao-tofu fermentation niche was characterized by late-stage basidiomycetous yeasts, represented by Trichosporon coremiiforme (LDA = 5.64) and Cutaneotrichosporon lineages (LDA = 5.01). Their enrichment in the final stage suggests a close association with the sensory maturation of Mao-tofu, potentially contributing to the accumulation of mature volatile compounds, including aldehydes, ketones, and aromatic esters.

3.3. Dynamic evolution of volatile flavor profiles

3.3.1. Multivariate statistical analysis of metabolites

To evaluate the overall differentiation of volatile profiles during fermentation, a PLS-DA model was established based on the detected volatile metabolites. As illustrated in Fig. 3A, samples from different fermentation stages occupied distinct metabolic spaces, with Component 1 and Component 2 explaining 80.7% and 14.0% of the total variance, respectively. Notably, a clear directional trajectory was observed along Component 1. SW and M01 were located on the positive side of Component 1, whereas the M02, M03, and M04 progressively shifted towards the negative side. This clear spatial separation indicates that the volatile profile was continuously and systematically reshaped across successive fermentation phases rather than fluctuating randomly.

Fig. 3.

Partial least squares discriminant analysis (PLS-DA) of volatile metabolite profiles during traditional Huizhou Mao-tofu fermentation. The PLS-DA score plot (A) shows the discrimination of metabolite profiles among SW and Mao-tofu samples collected at different fermentation stages, and the permutation test (B) validates the robustness and predictive ability of the PLS-DA model.

To assess the predictive performance of the model, a permutation test with 200 iterations was executed (Fig. 3B). The resulting R² and Q² intercepts were 0.3318 and − 0.6837, respectively. The clearly negative intercept of the Q² regression line indicates that the PLS-DA model was reliable and free from overfitting. This overall metabolic differentiation provides a sound statistical basis for the subsequent analysis of volatile metabolite classes and key aroma compounds.

3.3.2. Succession of volatile metabolite classes

The volatile profile of traditional Huizhou Mao-tofu underwent a clear progressive reconstruction across successive fermentation stages, reflecting a transition from the initial background to a chemically more diverse and mature aroma system. This process was first evaluated at the level of volatile metabolite classes based on compound diversity and relative abundance (Fig. 4A). A total of 16 classes of volatile metabolites were detected across all fermentation stages. In SW, the volatile profile was dominated by alcohols (37.94%), acids (14.99%), aldehydes (12.88%), and hydrocarbons (9.45%). After inoculation onto tofu, the volatile classes were rapidly restructured. In the early stage (M01), the relative abundance of organoheterocyclic compounds increased markedly from 5.62% to 22.25%, whereas esters increased from 5.63% to 14.11%. At the same time, acids decreased to 5.20%. These changes suggest that microbial carbohydrate metabolism became more active than lipid metabolism at this stage, potentially promoting ester formation, while amino acid catabolism contributed to the accumulation of organoheterocyclic compounds (Guan et al., 2024; Wang et al., 2025). As fermentation progressed, the volatile profile gradually shifted towards a more mature composition enriched in higher alcohols, aldehydes, ketones, and phenols. In the final maturation stage (M04), alcohols accounted for 41.83% of the total volatile profile, whereas aldehydes, ketones, and phenols accounted for 14.80%, 7.12%, and 4.89%, respectively.

Fig. 4.

Dynamic changes in volatile metabolite classes and potential aroma-active compounds during traditional Huizhou Mao-tofu fermentation. Relative abundance of different volatile metabolite classes across the five stages (A). Venn diagram and bar chart illustrating the number of differentially accumulated volatile metabolites among specific comparison groups (B). Stage-specific relative odor activity values (ROAVs) of 14 compounds exhibiting an ROAV ≥1 in at least one fermentation stage (C). ROAV-based relative contributions of the 14 selected compounds across fermentation stages (D).

To further evaluate the extent of volatile metabolite reorganization, differential metabolite analysis was performed across the selected comparison stages (Fig. 4B). The greatest metabolic divergence was observed immediately after inoculation, with 578 differential metabolites identified between SW and M01, indicating profound early metabolic reprogramming following transfer to the tofu matrix. The numbers of differential metabolites between adjacent subsequent stages also remained high, reaching 220 for M02 versus M01, 302 for M03 versus M02, and 323 for M04 versus M03. In addition, the Venn diagram revealed a core set of 58 overlapping differential metabolites shared across all five comparisons. Together, these results indicate that flavor formation in Mao-tofu was a continuous and directional process rather than a series of isolated random changes.

3.3.3. Identification of potential aroma-active compounds based on ROAV analysis

Relative odor activity value (ROAV) analysis was used to identify potential aroma-active compounds during Huizhou Mao-tofu fermentation (Fig. 4C and D). Experimental and library retention indices (RIs) for the 14 identified potential key aroma-active compounds are provided in Table S2. A total of 14 compounds showed ROAV values greater than 1 in at least one fermentation stage and were therefore considered potential key aroma-active compounds (Table S3). The number of these compounds increased from 3 in SW to 10 in M01, decreased to 7 in M02, and then increased to 10 and 11 in M03 and M04, respectively. This pattern indicates that the aroma-active profile was rapidly reconstructed after transfer to the tofu matrix and became progressively more complex during fermentation maturation.

In SW, damascenone, 1-octen-3-one, and 2,4-decadienal were the main compounds with ROAV ≥1. Damascenone showed the highest ROAV value and was assigned as the reference compound (ROAV = 100), accounting for 86.44% of the ROAV-based relative contribution (RC). As a highly potent ketone odorant derived from the degradation of carotenoids catalyzed by acids or enzymes, damascenone possesses a low odor threshold and imparts characteristic aromatic notes resembling rose, apple, tobacco, and honey (Zhang et al., 2024). Furthermore, minor active aroma contributors in SW included 1-octen-3-one (ROAV 10.10 and RC 8.72%) and 2,4-decadienal (ROAV 1.34 and RC 1.16%), reflecting the baseline presence of volatiles derived from lipid oxidation and phenols in the SW background.

Following inoculation onto the tofu matrix, the aroma-active profile underwent marked restructuring (Fig. 4D). The RC of damascenone decreased from 86.44% in SW to 2.35% at M01, suggesting that its floral-fruity contribution was diluted or masked after entry into the tofu matrix. At M01, 2,4-decadienal and 1-octen-3-one were the two largest potential contributors, accounting for 47.93% and 42.00% of the total RC, respectively. These two compounds represent volatiles originating from lipid oxidation and, within the tofu matrix, 2,4-decadienal is likely associated with the oxidation and oxidative cleavage of soybean lipids rich in linoleic acid (Ludwig et al., 2021). Sensory evaluations typically link 2,4-decadienal to fatty, fried, and green notes alongside aromas resembling cooked meat, whereas 1-octen-3-one is characterized by metallic and mushroom-like attributes (Wang et al., 2021; Zhang et al., 2024). Other compounds with ROAV ≥1 at M01 included 2-pentylfuran (ROAV = 2.44), ethyl hexanoate (ROAV = 2.13), 1-octen-3-ol (ROAV = 1.97), 2-nonenal (ROAV = 1.68), among others. Together, these compounds broadened the potential aroma spectrum during early fermentation, contributing fatty, green, mushroom-like, fruity, sulfur-containing and alcoholic notes.

At stage M02, 1-octen-3-one (ROAV 100 and RC 58.28%) and 2,4-decadienal (ROAV 50.58 and RC 29.08%) jointly accounted for 87.36% of the total RC. Their combined contribution remained high at M03, reaching 79.76%. Meanwhile, the RC of damascenone increased to 3.99%, and maturation-related compounds such as ethyl 2-methylbutyrate (RC 2.11%), skatole (RC 1.52%), and nonanal (RC 0.77%) became more prominent. The increase in ethyl 2-methylbutyrate suggests that esterification reactions between alcohols and organic acids became more evident during middle-to-late fermentation, contributing fruity and ester-like notes (Zheng et al., 2024). The emergence of skatole may be associated with microbial transformation of tryptophan or indole-related intermediates (Fu et al., 2022).

At the final maturation stage M04, 1-octen-3-one remained the dominant compound (ROAV 100 and RC 53.46%), followed by 2,4-decadienal (ROAV 38.62 and RC 20.58%). This indicates that lipid oxidation-derived carbonyls still formed the major aroma-active backbone of mature Mao-tofu. Damascenone was also retained among the potential key aroma-active compounds at M04, consistent with the continued presence of floral and honey-like aroma potential in the mature product. The occurrence of ethyl hexanoate, ethyl 2-methylbutyrate, and ethyl isovalerate among the M04 potential aroma-active compounds was consistent with ester formation from alcohols and organic acids, whereas nonanal and 2-nonenal were associated with fatty acid oxidation.

Overall, ROAV analysis revealed a transition from the damascenone-dominated aroma background of SW to a Mao-tofu profile dominated by lipid oxidation-derived compounds, particularly 1-octen-3-one and 2,4-decadienal. During maturation, esters, furans, aldehydes, alcohols, and nitrogen-containing compounds further increased aroma complexity. Because ROAV values were calculated from internal standard-normalized semi-quantitative data, the identified compounds should be regarded as potential contributors to the overall aroma profile.

3.4. Rate-based multi-omics associations between microbial succession and volatile aroma formation

Based on the preceding analyses, 10 bacterial genera, 8 fungal genera, and 14 ROAV-defined key aroma-active compounds were selected for rate-based multi-omics analysis. The O2PLS joint score plots showed that the four adjacent-stage transitions occupied largely distinct regions of the joint bacteria–volatile (Fig. 5A) and fungi–volatile (Fig. 5B) latent space, with microbial and volatile data points co-localizing within each transition. In both models, the M01–SW transition separated clearly along the negative tu[1] axis, consistent with the pronounced community restructuring that followed transfer from SW to the tofu matrix. The M02–M01 and M03–M02 transitions overlapped substantially in the tu[1]–tu[2] space, suggesting a more continuous, lower-amplitude covariation pattern during intermediate fermentation. By contrast, the M04–M03 transition formed a distinct cluster along tu[2], indicating a different rate-coupling pattern during final maturation.

Fig. 5.

Rate-based integration of microbial succession and changes in apparent volatile activity during traditional Huizhou Mao-tofu fermentation. O2PLS joint-score plots integrating the rates of change in 10 bacterial genera and 14 volatile compounds (A) or eight fungal genera and 14 volatile compounds (B). MFA individual score plot (C). MFA variable factor map (D) shows the distribution of bacterial genera, fungal genera, and potential key aroma-active compounds in the MFA space. Spearman correlation matrices between the rates of change in the 14 volatile compounds and bacterial (E) or fungal (F) genera. p values were adjusted separately for the bacterial and fungal matrices using the Benjamini–Hochberg procedure. *, q < 0.05; **, q < 0.01; ***, q < 0.001.

A three-block MFA was subsequently performed to integrate the selected bacterial genera, fungal genera, and key aroma-active compounds (Fig. 5C), in which Dim1 and Dim2 explained 33.82% and 21.85% of the total rate variance, respectively. M01–SW and M02–M01 occupied the negative region of Dim1, whereas M03–M02 occupied the positive region (Fig. 5C). M04–M03 was further separated along positive direction of Dim2.

The MFA variable factor map (Fig. 5D) resolved the overall covariance structure into three coherent and largely non-overlapping groups of vectors. The first group was oriented towards positive Dim1 and comprised Enterococcus, Streptococcus, Cutaneotrichosporon, Lactobacillus, Limosilactobacillus, and Pichia. These microbial vectors aligned with the ΔV/Δt vectors of damascenone, nonanal, ethyl 2-methylbutyrate and ethyl isovalerate. This alignment suggested that changes in the relative abundance of these bacterial and fungal taxa covaried with changes in the apparent aroma activity of the selected esters, aldehyde and damascenone.

The second group was oriented towards negative Dim1 and comprised Comamonas, Diutina, Dipodascaceae, and Mucor. These taxa aligned with 1-octen-3-one, 2,4-decadienal, 1-octen-3-ol, 2-nonenal, 2,4-nonadienal, 3-methyl-1-pentanol, and dimethyl trisulfide, which contributed predominantly lipid-oxidation- and sulfur-derived apparent aroma activity. This rate-coupling pattern was consistent with the extracellular lipolytic capacity reported for Diutina and related Dipodascaceae yeasts (Lotti et al., 1998). It was also compatible with the role of surface Mucor mycelia in lipid hydrolysis during Mao-tofu ripening, which may increase the availability of polyunsaturated fatty acid precursors for subsequent oxidative aroma formation (Hiol et al., 1999). The third group comprised Acinetobacter, Kurthia, Weissella, Corynebacterium, and Clavispora, and was aligned with skatole. This axis was distinct from the lipid- and ester-associated axes and was instead related to microbial tryptophan catabolism.

Spearman correlation analyses of the ΔM/Δt and ΔV/Δt values were visualized as heatmaps to resolve the pairwise microbial–volatile relationships underlying the MFA groupings (Fig. 5E, F). In the bacterial network (Fig. 5E), Comamonas displayed the broadest range of positive correlations, significantly tracking 1-octen-3-one, 2,4-decadienal, 2-pentylfuran, ethyl hexanoate, 2-nonenal, 1-octen-3-ol, 2,4-nonadienal, and 3-methyl-1-pentanol (ρ = 0.555–0.851, q < 0.05). This pattern was consistent with the lipid-oxidation-related direction identified by MFA and with the predominance of 1-octen-3-one and 2,4-decadienal described in Section 3.3.3. However, these coordinated changes may reflect shared responses to fatty acid precursor availability and oxygen exposure at the tofu surface rather than the direct production of these compounds by Comamonas. Enterococcus and Lactococcus were positively correlated with damascenone, ethyl isovalerate, and nonanal (Enterococcus: ρ = 0.541–0.875; Lactococcus: ρ = 0.618–0.636; q < 0.05), but negatively correlated with 2-pentylfuran and ethyl hexanoate (Enterococcus: ρ = −0.681; Lactococcus: ρ = −0.704; q < 0.01). Their positive associations with ethyl isovalerate and nonanal broadly corresponded to the positive-Dim1 ester–aldehyde direction in the MFA map. Streptococcus was positively correlated with damascenone (ρ = 0.612, q < 0.05) but negatively correlated with skatole (ρ = −0.824, q < 0.001), whereas Weissella and Corynebacterium were positively correlated with skatole (ρ = 0.832 and 0.770, respectively; q < 0.01).

Significant associations in the fungal network were sparser and more compound-specific (Fig. 5F). Trichosporon was positively correlated with 1-octen-3-one (ρ = 0.722, q < 0.05), whereas Pichia and Exophiala were negatively correlated with this compound (ρ = −0.668 and − 0.605, respectively; q < 0.05). Pichia was additionally positively correlated with nonanal (ρ = 0.611, q < 0.05). Because fungal C8 volatiles can arise from linoleic acid oxidation, these contrasting associations may reflect differences in lipid-associated metabolism, colonization timing and access to fatty acid precursors rather than a uniform fungal contribution (Karrer et al., 2021). Diutina was negatively correlated with ethyl 2-methylbutyrate, ethyl isovalerate, and nonanal (ρ = −0.656 to −0.627, q < 0.05), whereas Cutaneotrichosporon was positively correlated with ethyl 2-methylbutyrate and nonanal (ρ = 0.598–0.620, q < 0.05). These opposing relationships were consistent with the separation of the two genera in the MFA map. Clavispora displayed the largest number of significant relationships among the eight fungal genera, comprising negative correlations with 1-octen-3-one, 2-nonenal, 1-octen-3-ol, 2,4-nonadienal, and 3-methyl-1-pentanol (ρ = −0.716 to −0.588, q < 0.05), together with a positive correlation with skatole (ρ = 0.714, q < 0.05). Despite their notable MFA loadings, Mucor and the unclassified Dipodascaceae genus showed no significant genus-level associations with any of the 14 target compounds after BH–FDR correction.

Collectively, the genus-level correlation analysis identified three principal co-variation patterns. The first was a lipid-oxidation-associated group represented primarily by Comamonas and, more selectively, Trichosporon. The second was an ester- and amino-acid-associated group in which Enterococcus, Lactococcus and Cutaneotrichosporon tracked selected esters and aldehydes. The third was a tryptophan-related group characterized by contrasting associations of Weissella, Corynebacterium, Streptococcus and Clavispora with skatole. These patterns broadly corresponded to the multivariate structure identified by MFA (Fig. 5D). However, several genera with notable MFA loadings, including Kurthia, Mucor and the unclassified Dipodascaceae genus, showed no significant pairwise correlations. This discrepancy was expected because MFA captures shared covariance across complete data blocks, whereas Spearman analysis evaluates individual genus–compound pairs. Moreover, the associations were derived from rates of co-change in semi-quantitative apparent aroma activity rather than externally calibrated production rates. They should therefore be interpreted as exploratory evidence of co-variation, rather than proof that an individual genus produced or consumed a specific volatile compound.

Future studies should further characterize the viable and metabolically active microbial populations involved in Mao-tofu fermentation. Because amplicon sequencing provides DNA-based profiles of relative community composition, subsequent investigations should combine culture-based enumeration and viability-selective quantitative methods with targeted single-strain fermentation, enzyme activity measurements and pathway-level validation of the priority genera identified here.

4. Conclusion

This study provides a stage-resolved framework for understanding microbial succession and volatile flavor evolution during traditional Huizhou Mao-tofu fermentation. From the SW to the subsequent tofu fermentation stages, both bacterial (ANOSIM R = 0.9837) and fungal (R = 0.7474) communities underwent clear succession, accompanied by marked restructuring of the volatile flavor profile. The SW was dominated by Lactobacillus and Pichia and exhibited an aroma background strongly characterized by damascenone (RC 86.44%). After transfer to the tofu matrix, this initial aroma background was rapidly displaced by a surge of volatiles derived from lipid oxidation, principally 2,4-decadienal and 1-octen-3-one, which jointly accounted for 74.0–87.4% of the ROAV-based relative contribution from M02 to M04. These volatiles dominated the early and intermediate stages and were likely associated with substrate-dependent oxidation processes and broader community-level metabolism. As fermentation progressed, a specialized microbial guild characteristic of the maturation phase, primarily comprising Enterococcus, Trichosporon, and Cutaneotrichosporon, became enriched alongside a more complex aroma profile incorporating additional esters, aldehydes, and nitrogen-containing compounds, together with a partial recovery of damascenone. Rate-based multi-omics analysis of adjacent-stage change rates resolved three temporally distinct co-variation patterns: a lipid-oxidation-associated group represented by Comamonas and Trichosporon; an ester- and amino-acid-associated group comprising Enterococcus, Lactococcus, and Cutaneotrichosporon; and a tryptophan-catabolism-related group involving Weissella, Corynebacterium, Streptococcus, and Clavispora. Overall, these findings advance the understanding of microbiota succession and flavor evolution during traditional Huizhou Mao-tofu fermentation, providing a baseline for future studies on strain-directed flavor modulation in this traditional fermented soybean product.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

During the preparation of this work, the authors used ChatGPT (OpenAI) and Claude (Anthropic) to improve the readability, clarity, and linguistic quality of the manuscript. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

CRediT authorship contribution statement

Mingming Qi: Writing – original draft, Methodology, Investigation, Formal analysis, Conceptualization. Tingting Jia: Software, Methodology. Fengyao Pan: Software, Methodology. Mengying Ren: Methodology, Formal analysis. Junshun Zhang: Validation, Data curation. Shoubao Yan: Supervision, Project administration, Funding acquisition, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

Supplementary material

Additional data and supporting information for this study are provided in Appendix A.

Data availability

Data will be made available on request.