Deciphering Spatiotemporal Dynamics of Fermented Grains in the Jiangxiangxing Baijiu Production Process: Insights From a Transformer‐Based Deep Learning Model

ABSTRACT

The solid‐state fermentation of Jiangxiangxing Baijiu exhibits marked spatiotemporal heterogeneity in microbial communities and physicochemical parameters. We characterized the microbial succession and physicochemical dynamics of Zaopei (fermented grains) across Da‐hui rounds (Rounds 3–5) and applied LimiX, a Transformer‐based deep learning framework, for fermentation state monitoring. Spatial stratification explained 99% of microbial community variance (PERMANOVA, P < 0.001), with inner layers harboring greater bacterial diversity than surface layers. Staphylococcus and Weissella correlated negatively with starch and reducing sugar, while Oceanobacillus and Acetobacter strongly discriminated among rounds. LimiX achieved an AUC of 1.000 for round classification and outperformed GLM and Random Forest (RF) in discriminating fermentation layers. External validation yielded 100% accuracy for round differentiation and 92.86% for temporal stage prediction. SHAP analysis identified Klebsiella and Thermoascus as the primary drivers of spatial and temporal stratification, respectively. These results demonstrate the capability of deep learning to decode complex fermentation data and provide a theoretical and technical basis for the intelligent digitalization of Baijiu production.

Practical Applications

A Transformer‐based deep learning model was developed to integrate amplicon sequencing data and physicochemical indices for decoding the spatiotemporal dynamics of fermented grains in Jiangxiangxing Baijiu fermentation. This framework supports intelligent process monitoring and quality control, with potential applicability to other complex food fermentation systems.

1. Introduction

Jiangxiangxing Baijiu is renowned for its distinctive sauce‐like aroma, full‐bodied taste, and lingering fragrance. This unique flavor profile results from the complex interplay of raw materials, environmental conditions, production techniques, and aging processes (Guo et al. 2021; Wang 2022). The brewing process for jiangxiangxing baijiu is extremely complex, involving multiple rounds of steaming, pile fermentation, and long‐term aging; the entire production cycle typically spans more than a year (Figure 1).

FIGURE 1.

The production process of Jiangxiangxing Baijiu.

Pile fermentation of fermented grains (Zaopei) is a critical stage in Jiangxiangxing Baijiu production, shaping the liquorliquorring the liquorction, shaping the liquorliquorof raw materials, environmental conditions (Niu et al. 2025). Multi‐omics research integrating micro biome and flavor metabolomics has confirmed the significant contribution of pile fermentation to flavor formation (Chen et al. 2024). Microbial dynamics during piling facilitate exchange between fermented grains and the environment, driving continuous community succession that progressively enriches beneficial microorganisms (Niu et al. 2025).

Both microbial consortia and physicochemical parameters within the fermentation pile are highly dynamic. The solid‐state fermentation process inherently generates spatial heterogeneity in microbial community structure and fermentation status at different pile locations (Chen et al. 2024), and considerable research has characterized this spatiotemporal heterogeneity (Wu et al. 2023; Wu et al. 2024; Yang et al. 2023; Yang et al. 2025; Li et al. 2025; Lu et al. 2023).

However, these studies have largely remained descriptive and have not addressed the challenge of predicting or interpreting such complex heterogeneity from high‐dimensional, sparse omics data. Conventional models such as generalized linear models (GLM) and Random Forest (RF) often struggle with inherent data sparsity and are susceptible to over fitting when incorporating numerous microbial abundance features (Karwowska et al. 2025). Moving beyond descriptive analysis requires computational approaches capable of learning the underlying causal structure of the ecosystem.

To this end, we applied LimiX—a Transformer‐based deep learning framework—to model the spatiotemporal heterogeneity of fermented grains from the Da‐hui rounds (rounds 3–5). Unlike traditional models that seek patterns directly from sparse observed data, LimiX employs causal synthetic data training and conditional mask modeling, enabling it to learn universal dependencies and impute missing information with minimal preprocessing and lower over fitting risk (Belkadi et al. 2025; Kuo et al. 2025; Zhang et al. 2025).

This study aimed to:(1) characterize the spatiotemporal dynamics of microbial abundance and key physicochemical indices (acidity, moisture, starch, and reducing sugar) across rounds; (2) evaluate LimiXat predictive performance in classifying spatial layers and fermentation rounds against GLM and RF; and (3) identify key microbial biomarkers of fermentation status. This framework provides a theoretical foundation and technical pathway for precise prediction and intelligent regulation of pile fermentation status in Jiangxiangxing Baijiu production.

2. Materials and Methods

2.1. Sample Collection

Zaopei samples were obtained from a Jiangxiangxing Baijiu production enterprise during Rounds 3–5 (Da‐hui rounds), yielding 182 samples in total.

2.2. Experimental Design

In each of three consecutive Da‐hui rounds, grains were piled into a conical heap (∼1.5 m height, ∼3 m base diameter) and fermented for ∼96 h. Samples were collected from two depths: the surface (5–10 cm) and the inner layer (40–50 cm). At each depth, samples were taken at three time points: 0 h (immediately after piling), ∼48 h (mid‐fermentation), and ∼96 h (before pit transfer), with three equidistant subsamples serving as technical replicates. In total, 182 samples were collected, of which 140 were used for model training and validation, and 42 were held out as an independent test set. All samples were immediately placed into sterile containers, transported on ice, and stored at −80°C until analysis.

2.3. Instruments and Equipment

The instruments and equipment used in this study included a refrigerated ultracentrifuge (Model CS150NX, Eppendorf, Germany), a Qubit 4 Fluorometer (Thermo Fisher Scientific, USA), a QuantStudio Real‑Time PCR System (Thermo Fisher Scientific, USA), and a Near‑Infrared (NIR) Spectroscopy Analyzer (Thermo Fisher Scientific, USA).

2.4. Methods

2.4.1. Determination of Physicochemical Parameters

Moisture content was determined gravimetrically. Approximately 10 g of Zaopei sample was dried in an evaporating dish at 130°C for 90 min until constant weight. Moisture percentage was calculated based on weight loss.

For acidity measurement, 10 g of sample was extracted with 100 mL distilled water at room temperature for 30 min with intermittent stirring every 10 min. The extract was filtered, and 10 mL of filtrate was titrated with 0.1 mol/L NaOH using phenolphthalein as an indicator. The acidity was expressed as the volume (mL) of 0.1 mol/L NaOH consumed per gram of Zaopei.

Reducing sugar content was determined using Fehling's method. The same filtrate prepared for acidity was used. An aliquot of 10 mL filtrate was added to 5 mL each of Fehlingded to e filtrate prepared for ac30 mL of water. The mixture was heated to boiling and titrated with 2 g/L glucose standard solution with methylene blue as an indicator. A blank titration was performed. The reducing sugar content was calculated as glucose equivalent per 100 g of Zaopei.

Starch content was measured after acid hydrolysis. A 2.00 g ground sample was hydrolyzed with 100 mL of 1:4 HCl in a boiling water bath under reflux for 35 min. After cooling, the hydrolysate was neutralized to weak acidity with 20% NaOH, filtered, and diluted to 500 mL. The filtrate was then titrated against Fehlingtrate was then5 mL A + 5 mL B) with glucose standard solution. Starch content was calculated by multiplying the glucose equivalent by 0.9 and the dilution factor.

2.4.2. Amplicon Sequencing of Microbial Communities

2.4.2.1. DNA Extraction

Fermented grain samples (50 g each) were stored at −20°C in sterile 50 mL centrifuge tubes and subsequently shipped to Sangon Biotech (Shanghai) Co., Ltd., for genomic DNA extraction. DNA was extracted using the E.Z.N.A. Mag‐Bind Soil DNA Kit (Omega Bio‐tek, USA). The concentration of the extracted genomic DNA was accurately quantified using the Qubit 3.0 DNA Assay Kit (Thermo Fisher Scientific, USA).

2.4.2.2. PCR Amplification

Bacterial 16S V3–V4 and fungal ITS regions were amplified using a two‐step PCR strategy. In step 1, the target regions were amplified in 30 µL reactions containing 1 µL forward primer, 1 µL reverse primer, 10–20 ng template DNA, and nuclease‐free water. The primer pairs were as follows: for bacteria, 16S forward 5′‐CCTACGGGNGGCWGCAG‐3′ and reverse 5′‐GACTACHVGGGTATCTAATCC‐3′; for fungi, ITS forward 5′‐GCATCGATGAAGAACGCAGC‐3′ and reverse 5′‐TCCTCCGCTTATTGATATGC‐3′. After mixing and brief centrifugation, thermal cycling was performed as follows: 94°C for 3 min; 5 cycles of 94°C for 30 s, 45°C for 20 s, and 65°C for 30 s; 20 cycles of 94°C for 20 s, 55°C for 20 s, and 7°C for 30 s; 72°C for 5 min; and a 10°C hold. Step 2, the indexing PCR for Illumina compatibility, was carried out in 30 µL reactions containing 15 µL of 2× Hieff Robust PCR Master Mix, 1 µL Index‐Primer F, 1 µL Index‐PCR Reverse primer, 20–30 ng of purified step‐1 product, and nuclease‐free water. After mixing and brief centrifugation, the thermal program was as follows: 95°C for 3 min; 5 cycles of 94°C for 20 s, 55°C for 20 s, and 72°C for 30 s; 72°C for 5 min; and a 10°C hold.

The final amplified products were purified, quantified, and pooled in equimolar concentrations for subsequent library construction. Paired‐end sequencing was performed on the MiSeq platform according to the manufacturer's instructions.

2.4.2.3. Purification and Quantification of Amplicons

Pooled PCR products were purified with the AxyPrep DNA Gel Extraction Kit and quantified by QuantiFluor‐ST. Sequencing libraries were prepared with the TruSeq DNA PCR‐Free Library Preparation Kit and sequenced on the Illumina MiSeq platform (2 × 300 bp paired‐end).

2.4.3. Data Processing

2.4.3.1. Amplicon Sequence Data Processing

Raw sequencing data were processed in QIIME2 (v2020.6). Demultiplexed paired‐end reads were quality‐filtered with the q2‐quality‐control plugin (Phred ≥ Q20 over a 10‐bp sliding window) and subsequently denoised and merged using DADA2 (qiime dada2 denoise‐paired; parameters: –p‐trim‐left‐f 13, –p‐trim‐left‐r 13, –p‐trunc‐len‐f 240, –p‐trunc‐len‐r 210) based on quality profiles. Chimeras were removed via DADA2's consensus method. Taxonomy was assigned to the resulting amplicon sequence variants (ASVs) with the q2‐feature‐classifier against the SILVA 138 reference database (99% OTU similarity). A phylogenetic tree was constructed using the q2‐phylogeny plugin with MAFFT alignment and FastTree. To normalize sequencing depth, the ASV table was rarefied (subsampled without replacement) to 35,000 sequences per sample, corresponding to the minimum depth after quality filtering. All subsequent alpha and beta diversity analyses were performed on this rarefied dataset. Raw sequence data are available at the NCBI Sequence Read Archive under BioProject accession PRJNA1436829 (http://www.ncbi.nlm.nih.gov/bioproject/PRJNA1436829).

2.4.3.2. Functional Prediction, Data Analysis and Visualization

Bacterial functional potentials were predicted from the rarefied ASV table using PICRUSt2 (v2.5.1) with default parameters against the KEGG database, and the resulting pathway abundances were filtered for carbohydrate and amino acid metabolism pathways. Fungal ecological guilds and trophic modes were assigned to fungal ASVs using FUNGuild (v1.1) based on taxonomic annotations. All subsequent data analysis was conducted in R (v4.2.3). The vegan package was used to compute weighted ASV counts, phylum and genus level relative abundances, Shannon index (alpha diversity), Bray‐Curtis dissimilarity (beta diversity), and principal coordinates analysis (PCoA). Differentially abundant taxa were identified by LEfSe (LDA score > 2). ASV abundances were Z‐score normalized prior to modeling, and visualizations were generated with ggplot2. For predictive modeling, the dataset of 140 samples was randomly split into stratified training (80%) and test (20%) sets. The LimiX model was initialized from the LimiX‐16 M checkpoint and run in inference‐only mode with PyTorch 2.8/CUDA 12.8 without fine‐tuning (https://github.com/limix‐ldm‐ai/LimiX); thus, conventional training loss curves were unavailable. Overfitting was assessed by computing cross‐entropy loss on both sets using the LimiX predictor with dynamic k‐NN context retrieval (samples_per_class = 15) and temperature scaling (T = 1.3), and the losses were visualized as bar charts (Figures S1–S3). A Generalized Linear Model (L2‐regularized logistic regression, cross‐entropy loss, ‘sag’ solver, max_iter = 50) and a RF classifier (100 trees, Gini impurity) were trained on the same split and optimized via 50‐trial grid search. Feature importance was assessed using model coefficients (GLM), Gini importance (RF), and SHAP values (LimiX).

3. Results

3.1. Spatiotemporal Succession and Functional Dynamics of the Bacterial Community

Rarefaction curves for all samples reached a plateau (Figure 2a). Alpha‐diversity indices (Chao1 and Shannon) exhibited spatiotemporal variation, with the inner layer showing higher bacterial richness and diversity than the surface layer, especially in later fermentation stages (Figure 2b; Table S1). PCoA based on Bray‐Curtis distances revealed that sample clustering was driven primarily by fermentation round (PERMANOVA: R² = 0.99, P < 0.001) and secondarily by spatial layer (Figure 2c; Table S2). At the genus level, the relative abundances of Lactobacillus, Acetobacter, and Virgibacillus changed dynamically over fermentation (Figure 2d). Bacillota dominated the community structure across all rounds (Figure 2e). RDA identified starch, acidity, and reducing sugars as the primary drivers of community variation (Figure 2f). Spearman correlation analysis showed that Staphylococcus and Weissella were negatively correlated with starch and reducing sugars (P < 0.01), whereas Kroppenstedtia and Saccharopolyspora were negatively correlated with acidity and moisture (Figure 2g). LRT identified Oceanobacillus and Acetobacter as discriminative genera across rounds, while Sphingobacterium and Lactobacillus differed between spatial strata (Figure 2h). Functional prediction indicated “carbohydrate metabolism” and “amino acid metabolism” as the predominant pathways; pathways related to “development and regeneration” were enriched in Round 4 (Figure 2i; Table S3). Co‐occurrence network analysis revealed predominantly positive edges, with Bacillota and Pseudomonadota as central nodes (Figure 1, 2; Table S4).

FIGURE 2.

Characteristics of the microbial community structure in fermented grain mash.

3.2. Structural Reorganization and Trophic Shifts in the Fungal Mycobiome

Rarefaction curves plateaued (Figure 3a); fungal alpha diversity fluctuated across rounds, with the surface layer showing lower diversity than the inner layer in the final round (Figure 3b; Table S5). PCoA showed a major community shift distinguishing Round 3 from subsequent rounds (PERMANOVA: Rs = 0.99, P < 0.001) (Figure 3c; Table S6). Pichia, Paecilomyces, and Kazachstania were the prevalent genera (Figure 3d), and Pichia maintained high relative abundance throughout. FUNGuild analysis revealed significant shifts in trophic modes across rounds and layers (PERMANOVA, P < 0.05), with plant pathogen and saprotroph guilds varying in abundance (Figure 3e; Table S7), and the fungal co‐occurrence network showed a modular topology with Ascomycota hubs and a balance of positive and negative interactions (Figure 3f; Table S8).

FIGURE 3.

Characteristics of the fungal community structure in fermented grain mash.

3.3. Transformer‐Based Deep Learning for Precise Fermentation State Prediction and Marker Discovery

To establish a robust framework for real‐time monitoring of fermentation status, we benchmarked the Transformer framework LimiX against GLM and RF. LimiX consistently outperformed traditional methods across all classification tasks, demonstrating the advantage of the Transformer architecture in capturing high dimensional nonlinear interactions in multi‐omics data (Figure 4a). For FRound″ classification, LimiX achieved perfect accuracy (AUC = 1.000), matching RF and surpassing GLM (AUC = 0.655) (Figure 4b); even in the more challenging “Layer” classification task, with boundary ambiguity, LimiX maintained robust accuracy (0.690), comparable to RF. Model generalization was assessed via cross‐entropy loss (Figures S1–S3): minimal loss discrepancy was observed for “Round”, whereas “Site” and “Time” tasks showed substantial increases, reflecting inherent spatial and temporal distribution shifts. Confusion matrices further confirmed LimiX's stability, with minimal misclassification in distinguishing fermentation stages (Figure 4c, Table S9).

FIGURE 4.

Performance comparison of spatiotemporal heterogeneity models for fermented grain.

To interpret model predictions, we applied SHapley Additive exPlanations (SHAP), revealing that decisions relied on keystone taxa and physicochemical parameters (Figure 4d, e). For “Round”, the fungal genus Zygosaccharomyces and starch content were the top features; for “Layer”, the bacterial genus Klebsiella was the primary discriminator; and Thermoascus played a pivotal role in predicting “Time”. These findings validate that the Transformer model attends to biologically relevant features, establishing specific taxa as diagnostic biomarkers for process monitoring.

3.4. Generalizability and Real‐World Applicability of the Deep Learning Framework

To rigorously assess the generalization capability of the LimiX framework beyond the training cohort, we conducted an external validation using an independent dataset of 42 Zaopei samples, integrating high‐dimensional microbial abundance profiles with physicochemical indices to mimic real‐world production scenarios. The model exhibited exceptional robustness on this unseen data (Table 1, Table S10): Strikingly, LimiX achieved perfect predictive fidelity for distinguishing fermentation rounds, yielding 100% classification accuracy, and maintained high performance in resolving spatiotemporal dynamics, attaining 92.86% accuracy for stacking fermentation stages (“Time”) and 83.33% for spatial layers (“Layer”). These results underscore the potential of LimiX as a reliable, data‐driven tool for standardized quality monitoring and process control in industrial bio‐fermentation systems.

TABLE 1.

42 test set samples validated the performance of the LimiX model.

Group	Accuracy (%)
Round	100
Layer	83.33
Time	92.86

4. Discussion

The cussi times steaming, eight times fermentation, and seven times extraction”xtractionn namples validated the performance of the LimiX modeZaopei (Wu et al. 2023; Wu et al. 2024). Our data confirm that microbial assembly is deterministically governed by fermentation round and spatial layer. Consistent with Yang et al. (2025) and Li et al. (2025), the inner layer maintained higher bacterial richness. This stratification arises from environmental gradients: the surface remains aerobic and dry, selecting for Bacillus and Oceanobacillus (Lu et al. 2023; Ji et al. 2025), whereas the core turns anaerobic and accumulates heat, favoring Lactobacillus and Sphingobacterium (Shen et al. 2025). Across rounds, community succession was shaped by increasing acidity, rising reducing sugars, and declining starch (Zeng et al. 2025), which in turn drove negative correlations of Staphylococcus with reducing sugars and of Kroppenstedtia with acidity. Fungal dynamics paralleled these responsesrPichia persisted as a core yeast, while trophic guilds shifted from plant pathogens to saprotrophs, mirroring the nutrient‐ and stress‐driven patterns reported by Wang et al. (2021). Rare taxa also play key roles: conditionally rare taxa such as Lentibacillus and Pseudomonas exhibit distinct niche partitioning and contribute to yield variation (Li et al. 2026). The fungal network11″ partitioning and contribute to yield variation (d by ing reducing sugars, and declining starch (rface remains aerobic and dry, selecting for (“Lative bacterial network. Thus, spatiotemporal heterogeneity of flavor metabolites previously reported (Yang et al. 2023; Huang et al. 2024; Shi et al. 2025) is a downstream consequence of this spatially structured microbial metabolism.

Traditional linear models struggle with high‐dimensional, zero‐inflated micro biome data (Lee and Kim 2025). The Transformer based LimiX model learns latent dependencies from synthetic data and transfers them to real observations (Zhang et al. 2025), outperforming GLM and matching or exceeding RF in spatial discrimination. LimiX identified Thermoascus and Klebsiella as critical spatiotemporal predictors missed by standard analyses due to their higher‐order interactions. SHAP confirmed interpretability (Lee et al., 2025), consistent with studies linking low lactic acid and acidity to high yield rounds (Li et al. 2026) and identifying quality markers in sauce‐flavor baijiu (Li et al. 2024). External validation showed high accuracy, but the perfect round classification (100%, n = 42) requires cautious interpretation due to limited sample size. Cross‐entropy loss comparison indicated slight over fitting with the current sample size (Figures S1–S3). Overall, deep learning extracts structural information overlooked by conventional statistics, supporting process control (Rahman et al. 2025; Xiao et al. 2025).

Nomenclature

CS

Immediately after pile construction (0 h)

EC

Prior to transferring the grains to the fermentation pit (approximately 96 h)

GLM

Generalized Linear Model

Jiangxiangxing Baijiu

A type of Chinese Baijiu characterized by a sauce‐like aroma

LimiX

A Transformer‐based deep learning model designed for structured data modeling

MsS

The mid‐point of pile fermentation (approximately 48 h)

RF

Random Forest

SHAP

SHapley Additive exPlanations, a method for interpreting model predictions

Zaopei

Fermented grains used in the production of Jiangxiangxing Baijiu

Author Contributions

YingYu Huo: data curation, writing – original draft. Fei Zhang: investigation. GangReng Yang: supervision. XuJian Wu: resources. Xing Zhao: formal analysis, visualization. Cheng Wu: project administration. JunQiao Long: validation. DaChao Zhu: methodology. Xu Li: conceptualization, funding acquisition.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supplementary Figures: jfds71316‐sup‐0001‐FigureS1‐S3.docx

Supplementary Tables: jfds71316‐sup‐0002‐TableS1‐S10.docx

Supplementary Data: jfds71316‐sup‐0003‐SuppMatt.zip

Huo, Y. , Zhang F., Yang G., et al. 2026. “Deciphering Spatiotemporal Dynamics of Fermented Grains in the Jiangxiangxing Baijiu Production Process: Insights From a Transformer‐Based Deep Learning Model.” Journal of Food Science 91, no. 8: e71316. 10.1111/1750-3841.71316

Data Availability Statement

All fermentation grain sampling was conducted under the ethical guidelines of Guizhou Xijiu Co., Ltd., with written consent from the production management team. Raw sequence data and fully annotated metadata (including spatial coordinates, physicochemical measurements, and sampling timestamps) are publicly available in the NCBI SRA under BioProject PRJNA1436829.