Method for screening key active bacteria in Luzhou-flavor liquor pit mud and readable storage medium

Abstract

The application discloses a screening method for key active bacteria in Luzhou-flavor liquor pit mud and a readable storage medium, and belongs to the technical field of liquor brewing and fermentation microbial screening. The method first selects pit pools with significant quality differences and collects pit mud samples, removes dead bacteria interference through PMA treatment, extracts active bacteria DNA and performs 16S rRNA high-throughput sequencing to construct an active bacteria abundance matrix; then, candidate bacteria genera are coarsely screened through a random forest model, combined with LASSO stability selection for fine screening of key features, and finally, a secondary random forest modeling is performed for verification and output of a key active bacteria list. The application can accurately identify key active bacteria in pit mud, eliminate dead bacteria interference, and has stable feature selection and strong interpretability, and is suitable for pit mud systems in different production areas, and provides accurate technical support for high-quality pit mud cultivation and liquor flavor regulation.

Description

Technical Field

[0001] This invention relates to the field of microbial screening technology in baijiu brewing and fermentation, specifically to a method for screening key live bacteria in the cellar mud of strong-aroma baijiu and a readable storage medium. Background Technology

[0002] Strong-aroma baijiu is the mainstream aroma type of baijiu in my country. It adopts a continuous fermentation process in mud pits. The anaerobic bacteria in the pit mud can degrade the raw materials and generate key flavor precursors such as hexanoic acid and butyric acid, which play a decisive role in the formation of baijiu flavor and raw material yield. Its unique "pit aroma" is a core sensory attribute that directly determines the product's sensory characteristics and market competitiveness. The formation of this characteristic flavor is closely related to the pit mud. As the core carrier of the brewing system, the quality of the pit mud directly affects the flavor and quality stability of baijiu.

[0003] In the prior art, CN120161006 discloses a method, device, equipment, and medium for detecting the quality of cellar mud based on multimodal analysis. This method collects near-infrared (NIR) and mid-infrared (MIR) spectral data of cellar mud samples, performs spectral preprocessing, constructs an NIR-MIR fusion model from the standardized spectral data using a multivariate statistical analysis algorithm, and finally extracts key component information from the cellar mud samples based on the NIR-MIR fusion model and performs a comprehensive evaluation to generate a quality assessment result. Although this method can achieve cellar mud quality evaluation, it relies on professional spectral acquisition equipment, requiring high expertise and incurring high testing costs. Furthermore, the microbial community of cellar mud is complex, containing both actively metabolizing live bacteria and quiescent / dead microbial components. The function of cellar mud depends on the metabolism of active microorganisms; live bacteria can produce volatile organic acids, which then undergo esterification reactions with alcohols to generate aroma esters, represented by ethyl hexanoate, which together constitute the rich and complex cellar aroma flavor of strong-aroma baijiu. Relying solely on spectral and physicochemical indicators to characterize the overall state of pit mud makes it difficult to accurately identify key microbial components and their contribution to pit mud quality.

[0004] In the brewing of strong-aroma baijiu, accurately identifying key microorganisms in high-quality cellar mud is an important theoretical basis and technical prerequisite for optimizing cellar mud cultivation process, directional regulation of microbial community structure, and improving the flavor and quality stability of baijiu. The identification method of key microorganisms in cellar mud has always been a research focus in the industry.

[0005] Currently, a mature system for the precise identification of microorganisms in fermentation pit mud has not yet been established, and related research still needs to be carried out and improved. In the field of baijiu brewing, CN120796068 discloses a method for screening key microorganisms in the fermentation process of soy sauce-flavored baijiu. The method involves constructing a "fermentation tripartite" of pre-fermentation mash, daqu (fermentation starter), and post-fermentation mash. Based on the microbial abundance data in the fermentation tripartite samples as input features, a Lasso linear model is constructed. Then, based on the constructed LASSO model, potential microbial species are predicted, and key microorganisms in the fermentation process of soy sauce-flavored baijiu are screened based on the prediction results. This method is applicable to the fermentation system of soy sauce-flavored baijiu, but it only detects total microorganisms and does not distinguish between live and dead bacteria. It cannot accurately identify key live bacteria with metabolic functions, making it difficult to support the targeted regulation of pit mud microbial communities and the artificial cultivation of high-quality pit mud, thus restricting the upgrading of strong-aroma baijiu brewing technology and the stability of quality. Meanwhile, this method uses a single Lasso regression model, which is susceptible to sample perturbation, has insufficient stability in key feature selection, and does not combine feature selection with model generalization ability evaluation, making it difficult to output a reproducible and interpretable list of key live bacteria in industrial applications.

[0006] Studies have shown that the microbial community structure of cellar mud in strong-aroma baijiu exhibits significant differences between production areas and distilleries. Influenced by factors such as raw materials, processing techniques, cellar age, and geographical climate, the dominant microbial communities in different cellar muds vary considerably, and key functional microorganisms lack universality across production areas and distilleries. Existing evaluation methods based on total microbial communities or fixed functional microbial lists are difficult to adapt to cellar mud systems with different ecological niches, lacking sufficient transferability and practicality. Summary of the Invention

[0007] In view of the above-mentioned shortcomings in the prior art, the purpose of this invention is to provide a method for screening key live bacteria in the cellar mud of strong-aroma baijiu and a readable storage medium, so as to solve the problems of the prior art being unable to distinguish between live and dead bacteria, unstable feature screening, and poor adaptability across production areas.

[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: On the one hand, a method for screening key live bacteria in the cellar mud of strong-aroma baijiu is provided, including the following steps: Step 1: Select target cellars based on the quality differences between the cellar mud and the raw liquor produced, and conduct sensory evaluation of the cellar mud and raw liquor to classify their quality grades. Step 2: During the fermentation cycle, samples of different quality grades of pits are taken without opening the pits to obtain parallel pit mud samples. Step 3: Homogenize the pit mud sample and add buffer solution to mix well to prepare pit mud suspension; Step 4: Perform PMA treatment on the pit mud suspension to selectively inhibit the amplification of dead bacterial DNA, and collect live bacterial cells by centrifugation; Step 5: Extract the genomic DNA of the live bacteria, amplify the 16S rRNA gene fragment and perform high-throughput sequencing to obtain the active microbial sequence and abundance information; Step 6: Based on the abundance information and combined with the aforementioned classification results of pit mud quality grades, construct an abundance matrix of active bacteria with bacterial genera as characteristics, samples as samples, and pit mud quality as classification labels. Step 7: Perform dataset partitioning and data standardization preprocessing on the viable bacterial abundance matrix; Step 8: Construct the first random forest model using the preprocessed data, and obtain a set of candidate key fungal genera related to the quality grade of pit mud based on the importance of fungal genera characteristics. Step 9: Based on bootstrap resampling and LASSO regression, the stability selection of the candidate key bacterial genera set is performed to refine the core key bacterial genera that are highly associated with high-quality pit mud; Step 10: Construct a second random forest model based on the core key bacterial genera and verify it to obtain a classification model that can determine the quality grade of pit mud based on microbial characteristics; Step 11: Output the names, feature importance, and selection frequency of key live bacteria that are strongly correlated with the quality grade of the pit mud according to the classification model.

[0009] Furthermore, in step 1, the method for evaluating the quality of the fermentation pit includes: Step 101: Establish a quality evaluation team. The team members are professional researchers from the winery, with rich experience in baijiu tasting and cellar mud evaluation, and are able to accurately identify the sensory characteristics of baijiu and cellar mud. Step 102: Evaluate the color, aroma, taste, and style of the raw wine produced in cellars of different quality according to the detailed rules and scoring criteria specified in Table 1. Step 103: According to the standards specified in the artificial sensory evaluation table for pit mud quality in Table 2, the two types of pit mud are scored by sensory evaluation of color, texture and odor.

[0010] Furthermore, in step 2, the method for collecting the pit mud samples includes: Step 201, Sampler embedding method as follows Figure 1 As shown, a sampler was buried within a 30 cm × 20 cm area in the center of the cellar wall; Step 202: When taking samples, the cellar is not opened. Instead, three samplers are directly taken as three samples at the same time point. Step 203: After sampling, transfer the sample to a sterile sampling bag, record the sample information, and store at 4°C.

[0011] Furthermore, in step 3, the method for preparing the pit mud suspension includes: Step 301: Place the sterile bag containing the pit mud into the beater homogenizer and homogenize it at a speed of 10 times / s for 5 minutes at room temperature. Step 302: Accurately weigh 1 g of the homogenized pit mud sample into 100 mL of PBS buffer (1 M, pH=7.4), mix at 150 rpm for 30 min to obtain pit mud suspension.

[0012] Furthermore, in step 4, the PMA treatment method for the pit mud suspension includes: Step 401: Under light-protected conditions, accurately pipette 1 mL of the pit mud suspension and two portions of the control group into 1.5 mL centrifuge tubes respectively; Step 402: Add 50 μM PMA reagent to the suspension and mix thoroughly; Step 403: Incubate in the dark for 10 minutes; Step 404: After incubation, expose to a 60W blue light for 15 minutes; Step 405: After the sample is treated with PMA, it is centrifuged at 5000 g for 10 min, and the supernatant is removed to collect the bacterial cells.

[0013] Furthermore, in step 5, the method for extracting bacterial DNA from the pit mud and collecting 16S rRNA data includes: Step 501: Extract genomic DNA from live bacteria in PMA-treated pit mud using a commercial soil genomic DNA extraction kit; Step 502: After testing the purity and integrity using a NanoDrop micro spectrophotometer, store at -80℃ for further experiments; Step 503: Amplify the 16S rDNA V3V4 variable region of live bacterial genomic DNA using universal primers 515F / 806R to construct a high-throughput sequencing library and perform Illumina MiSeq paired-end sequencing. Step 504: Perform quality control on the raw sequencing sequences using FASTP; Step 505: Use FLASH software to concatenate the sequences; Step 506: Use the DADA2 sequence denoising method to denoise the data and obtain ASV (Amplicon Sequence Variant) representative sequence and abundance information; Step 507: Randomly extract all sample sequences to a uniform data size according to the minimum number of sample sequences; Step 508: Use Qiime2 to generate species classification information for each ASV, and screen key microorganisms based on genus-level classification.

[0014] Furthermore, in step 6, the method for constructing the microbial abundance matrix includes: Step 601: Construct a viable abundance matrix with genera as rows and samples as columns, based on the number of ASVs for each genera; Step 602: Assign a high-quality or ordinary quality label to each pit mud sample and classify and code the sample labels.

[0015] Furthermore, in step 7, the method for preprocessing the viable bacterial abundance matrix data includes: Step 701: Divide all sample data into training and test sets in a 7:3 ratio; Step 702: On the training set, standardize the data to the 0-1 range.

[0016] Furthermore, in step 8, the first coarse screening method for random forest modeling includes: Step 801: Construct a random forest classification model on the training set, with 50 decision trees; Step 802: Evaluate the model convergence using out-of-bag error (OOB error) and calculate the classification accuracy on the test set; Step 803: Based on the prediction error of the external bag replacement, obtain the importance score of each bacterial genus feature, and screen out features with a contribution greater than 0 as candidate bacterial genera.

[0017] Furthermore, in step 9, the LASSO combined with stability selection method for screening key features includes: Step 901: Perform 50 bootstrap resampling cycles. In each round of resampling, perform bootstrap sampling on the training set to construct a binary LASSO logistic regression model.

[0018] Step 902: The model training process uses 5-fold cross-validation. Based on the 1-SE principle, that is, the sparsest model whose cross-validation error does not exceed the optimal error plus one standard error is selected and the coefficient vector of the model is extracted.

[0019] Step 903: Then count the number of times the coefficient of each feature is non-zero in all categories and all bootstrap repetitions, and calculate its stability frequency of being selected; Step 904: Based on the preset stability threshold, the final set of key bacterial genera is obtained; Step 905: When no feature meets the condition at the current threshold, the top certain proportion with the highest stability frequency can be automatically selected to ensure the robustness and adaptability of the screening process.

[0020] Furthermore, in step 10, the second random forest modeling and validation method includes: Step 1001: Using the selected key fungal genera features, re-standardize the training and test sets to construct a new random forest classification model; Step 1002: Evaluate model performance comprehensively through multiple methods, including test set accuracy, confusion matrix, learning curves composed of training and test errors under different training set sizes, and hierarchical 5-fold cross-validation; principles; Step 1003: Based on the out-of-bag importance of the final model, recalculate the contribution of key microbial communities to form a list of key viable microorganisms with a combination of random forest importance and stability frequency.

[0021] The key live bacteria mentioned in this invention refer to live bacteria genera that are significantly positively correlated with high-quality pit mud and can be stably and repeatedly selected by the model.

[0022] Furthermore, in step 11, the method for outputting and displaying the results includes: Step 1101: Output the names of each key bacterial genus, their importance scores in the final model, and their selection frequency; Step 1102: Generate visualization results such as importance bar chart, cumulative contribution line chart, OOB error curve, learning curve, and cross-validation confusion matrix.

[0023] On the other hand, a computer-readable storage medium is provided, storing a computer program, which, when executed by a processor, implements the steps of the screening method for key live bacteria in the cellar mud of strong-aroma baijiu as described above, and during the execution process, calls the cellar mud quality grade classification results as classification labels.

[0024] The beneficial effects of this invention are as follows: 1. Precisely target live bacteria and eliminate interference from dead bacteria: The combination of propidium bromide (PMA) treatment and high-throughput sequencing detects only functional bacteria with metabolic activity in the pit mud, effectively overcoming the shortcomings of conventional sequencing in distinguishing between live and dead bacteria, and better meeting the actual production needs such as artificial cultivation of pit mud and inoculation of functional bacteria. 2. Stable feature selection and strong resistance to sample disturbance: The dual selection mechanism combining random forest feature importance and LASSO stability selection significantly improves the robustness of key microbial features and solves the problems of single model being easily affected by sample fluctuations and the non-reproducibility of selection results. 3. Key features are highly simplified, facilitating practical verification: While ensuring and improving classification accuracy, the features of hundreds of original bacterial genera are compressed to less than ten, which greatly reduces the difficulty of subsequent microbial isolation and purification, functional verification and mechanism research.

[0025] 4. Excellent interpretability and practicality: The evaluation of key live bacteria is based on a combination of random forest importance score and LASSO selection frequency, which has both statistical significance and biological significance, making it easy to form a standardized and interpretable list of key bacteria in industrial scenarios. 5. Strong adaptability and cross-regional promotion: The method does not rely on a fixed list of strains and can automatically adapt and screen according to different production areas and different cellar ecosystems, solving the problems of poor universality and difficulty in migration and application of existing technologies. 6. Easy to software-based and industrially deploy: The entire process can be packaged into an algorithm module or cloud service, which can be integrated with the winery's existing cellar mud testing and production management system to achieve rapid identification of cellar mud quality and automatic output of key live bacteria lists. Attached Figure Description

[0026] Figure 1 Sensory radar chart of the raw spirit and scores for each sensor; Figure 2 A schematic diagram (a) of the sampling of pit mud samples during fermentation and a diagram of the actual burial (b); Figure 3 Comparison of Ct values ​​of pit mud samples before and after PMA treatment; TP, before PMA treatment, NP, after PMA treatment; Figure 4 Comparison of Ct values ​​before and after PMA treatment for pure culture samples of Escherichia coli; (a) Escherichia coli in logarithmic growth phase; (b) Escherichia coli that died from heat; VB, before PMA treatment; TB, after PMA treatment; Figure 5 The out-of-bag error curve for the first random forest (RF) model in Example 1; Figure 6 This is a frequency selection diagram for LASSO stability in Example 1; Figure 7 The learning curve of the final model in Example 1; Figure 8 This is the row-normalized confusion matrix of the final 5-fold hierarchical cross-validation model in Example 1; Figure 9 The following is a bar chart showing the importance of key features and the cumulative contribution curve of the final model in Example 1; Figure 10 The out-of-bag error curve for the first random forest model in Example 2; Figure 11 This is a frequency selection diagram for LASSO stability in Example 2; Figure 12 The learning curve of the final model in Example 2; Figure 13 This is the row-normalized confusion matrix of the final 5-fold hierarchical cross-validation model in Example 2; Figure 14The following is a bar chart showing the importance of key features and the cumulative contribution curve of the final model in Example 2; Figure 15 The importance bar chart and cumulative contribution curve for feature selection using RF alone. Detailed Implementation

[0027] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0028] Example 1 S10. Based on the company's production experience, select fermentation pits with stable fermentation states and significant differences in the quality of the pit mud and the resulting raw liquor. An evaluation team composed of five professional researchers from the distillery, all with extensive experience in baijiu tasting and pit mud evaluation, is formed, capable of accurately identifying the sensory characteristics of baijiu and pit mud. Following the detailed rules and scoring standards specified in Table 1, the color, aroma, taste, and style of the raw liquor produced from pits of different qualities are evaluated. According to the standards specified in the artificial sensory evaluation table for pit mud quality in Table 2, sensory scores for color, texture, and aroma are given for the two different qualities of pit mud.

[0029] Table 1 Table 2 Table 3 Sensory scores of high-quality and ordinary pit mud In this embodiment, a sensory score of ≥50 for cellar mud is considered high-quality, while <50 is considered ordinary. High-quality cellar mud has a fine and soft texture, a uniform dark brown color, and a pure aroma with a rich cellar fragrance; ordinary cellar mud has an uneven texture with layering, a dull color, lacks aroma, and has a noticeable ammonia smell. The color, aroma, taste, flavor, and overall score of the raw liquor produced from cellar mud of different qualities vary significantly. The overall sensory score of raw liquor produced from high-quality cellar mud is ≥90, while the overall sensory score of raw liquor produced from ordinary cellar mud is <90.

[0030] S20. During a 60-day fermentation cycle in two different seasons, samples were taken at six time points: 0d, 5d, 15d, 25d, 45d, and 60d, to track changes in the fermentation mud. Samples were taken directly at 0d and 60d with the fermentation pit open, while samples were taken at the other four time points using a buried mud sampler. The sampler was buried as follows... Figure 2As shown, samplers were embedded within a 30 cm × 20 cm area in the center of the pit wall. The pit was not opened during sampling; instead, three samplers were directly extracted as three samples from the same time point, resulting in a total of 72 samples. After sampling, the samples were transferred to sterile sampling bags, and the sample information was recorded and stored at 4°C for PMA processing.

[0031] S30: Place the sterile bag containing the pit mud into a homogenizer and homogenize it at a speed of 10 times / s for 5 minutes at room temperature. Accurately weigh 1 g of the homogenized pit mud sample into 100 mL of PBS buffer (1 M, pH=7.4) and mix at 150 rpm for 30 minutes to obtain a pit mud suspension.

[0032] S40: Under light-protected conditions, accurately pipette 1 mL of the above-mentioned pit mud suspension into a 1.5 mL centrifuge tube, add 50 μM PMA reagent to the suspension, mix thoroughly, and incubate in the dark for 10 min. After incubation, expose to a 60 W blue light for 15 min. After PMA treatment, centrifuge the sample at 5000 g for 10 min, remove the supernatant, and collect the bacterial cells.

[0033] Depend on Figure 3 It can be seen that the Ct value of the pit mud sample (TP) before PMA treatment is significantly lower than that after PMA treatment (NP), indicating that the PMA treatment sample contained dead bacteria, resulting in a higher amplification efficiency. PMA treatment can effectively remove dead bacteria from the pit mud sample.

[0034] Depend on Figure 4 (a) It can be seen that the number of bacteria in the logarithmic growth phase of *E. coli* was higher before PMA treatment, but there was no significant difference before and after treatment. Therefore, it can be concluded that PMA treatment did not have a significant effect on the amplification of viable bacterial DNA. Figure 4 (b) It can be found that in the heat-killed group, the number of E. coli before treatment was significantly higher than that after treatment, which indicates that PMA treatment has a significant inhibitory effect on the amplification of DNA from dead bacteria.

[0035] S50: Genomic DNA of live bacteria from PMA-treated pit mud was extracted using a commercial soil genomic DNA extraction kit. After purity and integrity were tested using a NanoDrop micro-spectrophotometer, the DNA was stored at -80℃ for further experiments. The live bacterial genomic DNA was amplified using universal primers 515F / 806R to amplify the 16S rDNA V3V4 variable region to construct a high-throughput sequencing library, which was then subjected to Illumina MiSeq paired-end sequencing. The raw sequencing sequences were quality controlled using FastP, assembled using FLASH software, and denoised using the DADA2 sequence denoising method to obtain representative ASV (Amplicon Sequence Variant) sequences and abundance information. Sequences from all samples were randomly selected to a uniform data volume according to the minimum sample sequence quantity. Simultaneously, Qiime2 was used to generate species classification information for each ASV, and key microorganisms were screened based on genus-level classification.

[0036] S60: Construct a viable abundance matrix based on the number of ASVs for each genus, with the genus as the row and the sample as the column. Assign a quality label of "high quality" or "normal quality" to each mud sample, resulting in an 856×72 matrix. The first row is the sample name, the second row is the sample label, and the first column is the genus name.

[0037] S70: Classify and encode the sample labels; divide all sample data into training and test sets in a 7:3 ratio; the training set contains 51 samples and the test set contains 21 samples; on the training set, standardize the data to the 0-1 range.

[0038] S80: Construct a random forest classification model on the training set with 50 decision trees; evaluate the model convergence using out-of-bag (OOB) error, and calculate the classification accuracy on the test set. Obtain the importance score of each genus feature based on the OOB permutation prediction error, and select features with a contribution greater than 0 as candidate genera.

[0039] like Figure 5 As shown, in this embodiment, the out-of-bag error after modeling using all live bacteria data gradually decreases with the number of trees, reaching a plateau at 50 trees, indicating a suitable number of trees. The accuracy of random forest modeling based on 50 trees is 95.24%. The importance score of each genus feature is obtained based on the out-of-bag permutation prediction error, and features with a contribution greater than 0 are selected as candidate genera. In this embodiment, a total of 88 genera with a contribution greater than 0 are obtained through coarse screening.

[0040] S90: Perform 50 bootstrap resampling iterations. In each resampling iteration, bootstrap sampling is performed on the training set. A binary LASSO logistic regression model is constructed for each category. The model training process employs 5-fold cross-validation, selecting the sparsest model whose cross-validation error does not exceed the optimal error plus one standard error, based on the 1-SE principle. The coefficient vector of this model is then extracted. Subsequently, the number of times each feature's coefficient is non-zero across all categories and all bootstrap repetitions is counted, calculating its selected stability frequency. The final set of key bacterial genera is obtained based on a preset stability threshold. When no feature meets the condition at the current threshold, the top proportion with the highest stability frequency is automatically selected to ensure the robustness and adaptability of the selection process.

[0041] like Figure 6 As shown, in this embodiment, among the 100 sets of coefficient vectors accumulated from 50 resampling and two class models, the stability frequency of most bacterial genera was below 0.20, with only a few genera showing significantly higher occurrence frequencies. Setting 0.45 as the stability threshold, nine bacterial genera were ultimately selected that contributed significantly to the quality of the pit mud and exhibited high stability during extensive resampling training.

[0042] S100: Using the selected key bacterial genus features, the training and test sets are re-standardized; a new random forest classification model is constructed, and the model performance is comprehensively evaluated through various methods such as test set accuracy, learning curves composed of training errors and test errors under different training set sizes, and hierarchical 5-fold cross-validation.

[0043] like Figure 7 As shown, in this embodiment, a random forest model is reconstructed based on the nine selected bacterial genera. With the increase in the training set size, both training and testing errors show a decreasing and regionally stable characteristic. The model does not exhibit significant overfitting or underfitting, and the learning curve shows stable performance. Figure 8 As shown, in this embodiment, after 5-fold stratified cross-validation, the final random forest model built based on these 9 genera exhibits high prediction accuracy across different categories with minimal classification bias. The contribution of key bacterial communities is recalculated based on the out-of-bag importance of the final model, forming a list of key viable bacteria based on a combination of random forest importance and stability frequency. Figure 9 As shown in this embodiment, in the final constructed random forest model, the cumulative contribution rate of the first 6 genera exceeds 80%, and the cumulative contribution rate of the first 7 genera exceeds 90%.

[0044] S110: Output the names of each key bacterial genera, their importance scores in the final model, and their selection frequency; generate visualization results such as importance bar charts, cumulative contribution line charts, OOB error curves, learning curves, and cross-validation confusion matrices. As shown in Table 3, in this embodiment, a total of 9 bacterial genera were selected.

[0045] Table 4 Example 2 In the method steps of Example 2, the similarities with Example 1 will not be repeated here; only the differences from Example 1 will be described. The difference between Example 2 and Example 1 is that 200 decision trees, 30 training samples, and 12 test samples are used; the rest are the same as described in Example 1.

[0046] like Figure 10 As shown, in this embodiment, the out-of-bag error after modeling using all live bacteria data gradually decreases with the number of trees, reaching a plateau at 200 trees, indicating a suitable number of trees. The accuracy of random forest modeling based on 200 trees is 83.33%. The importance score of each genus feature is obtained based on the out-of-bag permutation prediction error, and features with a contribution greater than 0 are selected as candidate genera. In this embodiment, a total of 80 genera with a contribution greater than 0 are obtained through coarse screening.

[0047] like Figure 11 As shown in this embodiment, in the 50 resampling cycles and the cumulative 100 sets of coefficient vectors from the two class models, the stability frequency of most bacterial genera was below 0.20, with only a few genera exhibiting significantly higher occurrence frequencies. Setting 0.4 as the stability threshold, a total of 7 bacterial genera were ultimately selected that contributed significantly to the quality of the pit mud and showed high stability during extensive resampling training.

[0048] like Figure 12 As shown, in this embodiment, a random forest model is reconstructed based on the seven selected bacterial genera. With the increase in the training set size, both training and testing errors show a decreasing and regionally stable characteristic, indicating no obvious overfitting or underfitting, and the learning curve is stable. Figure 13 As shown, in this embodiment, after 5-fold stratified cross-validation, the final random forest model built based on these 7 genera exhibits high prediction accuracy across different categories with minimal classification bias. The contribution of key bacterial communities is recalculated based on the out-of-bag importance of the final model, forming a list of key viable bacteria based on a combination of random forest importance and stability frequency. Figure 14 As shown in this embodiment, in the final constructed random forest model, the cumulative contribution rate of the first 5 genera exceeds 80%, the cumulative contribution rate of the first 6 genera exceeds 90%, and the final test set prediction accuracy is 91.67%.

[0049] As shown in Table 4, a total of 7 genera were screened in this embodiment.

[0050] Table 5 Example 3 A computer program product includes a computer program / instructions that, when executed by one or more processors, implement the steps of the screening method for key live bacteria in the cellar mud of strong-aroma baijiu as described in any embodiment of this application.

[0051] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, an optical disk, or a read-only memory (ROM). Computer-readable storage media such as ROM (Only Memory), or Random Access Memory (RAM).

[0052] Comparative Example To demonstrate that a dual screening mechanism combining random forest feature importance and LASSO stability selection can significantly improve the robustness of key microbial features and address the issues of single models being susceptible to sample fluctuations and the non-reproducibility of screening results, a comparative study was conducted, comparing the use of random forest alone for feature selection and the use of LASSO alone for stability selection.

[0053] Table 6 Comparison of various screening results From Table 6 and Figure 15 It can be seen that the RF model has the best overall performance in classification, with accuracy, recall, precision, and F1 score reaching 95.46%, 100%, 91.67%, and 95.65%, respectively. However, the selected features show a dispersed distribution in importance, with the cumulative contribution of the top 20 important features far below 80%, indicating poor concentration of key features. Using LASSO alone, its accuracy, precision, recall, and F1 score are all low, with none exceeding 90%, indicating that the LASSO model has limited ability to handle different sludge quality classification models. The RF combined with LASSO model achieves accuracy, precision, recall, and F1 score of 86.36%, 90.91%, 83.33%, and 86.96%, respectively, and the selected microorganisms all have high importance and selection frequency, indicating that this model achieves the dual advantages of accurate prediction and concentration of key features.

[0054] In summary, this invention overcomes the limitations of conventional high-throughput sequencing technology in distinguishing the active state of microorganisms and eliminating interference from dead bacteria. Furthermore, it overcomes the significant differences in microbial communities in cellar mud from different production areas and distilleries of strong-aroma baijiu, and the lack of universality of key microorganisms. It provides a method and storage medium for screening key microorganisms during the fermentation process of strong-aroma cellar mud based on propidium bromide azide combined with diversity sequencing.

[0055] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

[0056] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for screening key viable bacteria in the cellar mud of strong-aroma baijiu (Chinese liquor), characterized in that, Includes the following steps: Step 1: Select target cellars based on the quality differences between the cellar mud and the raw liquor produced, and conduct sensory evaluation of the cellar mud and raw liquor to classify their quality grades. Step 2: During the fermentation cycle, samples of different quality grades of pits are taken without opening the pits to obtain parallel pit mud samples. Step 3: Homogenize the pit mud sample and add buffer solution to mix well to prepare pit mud suspension; Step 4: Perform PMA treatment on the pit mud suspension to selectively inhibit the amplification of dead bacterial DNA, and collect live bacterial cells by centrifugation; Step 5: Extract the genomic DNA of the live bacteria, amplify the 16S rRNA gene fragment and perform high-throughput sequencing to obtain the active microbial sequence and abundance information; Step 6: Based on the abundance information and combined with the aforementioned classification results of pit mud quality grades, construct an abundance matrix of active bacteria with bacterial genera as characteristics, samples as samples, and pit mud quality as classification labels. Step 7: Perform dataset partitioning and data standardization preprocessing on the viable bacterial abundance matrix; Step 8: Construct the first random forest model using the preprocessed data, and obtain a set of candidate key fungal genera related to the quality grade of pit mud based on the importance of fungal genera characteristics. Step 9: Based on bootstrap resampling and LASSO regression, the stability selection of the candidate key bacterial genera set is performed to refine the core key bacterial genera that are highly associated with high-quality pit mud; Step 10: Construct a second random forest model based on the core key bacterial genera and verify it to obtain a classification model that can determine the quality grade of pit mud based on microbial characteristics; Step 11: Output the names, feature importance, and selection frequency of key live bacteria that are strongly correlated with the quality grade of the pit mud according to the classification model.

2. The screening method according to claim 1, characterized in that, In step 1, the sensory evaluation includes scoring four indicators for the color, aroma, taste, and style of the raw wine, as well as three indicators for the color, texture, and aroma of the cellar mud. Based on the scoring results, the quality grade classification of the cellar mud is determined as either high-quality or ordinary.

3. The screening method according to claim 1, characterized in that, In step 2, a buried sampler is used for sampling. The sampling location is a 30cm×20cm area in the center of the pit wall. The pit is not opened during sampling. Three samplers are taken at the same time point as parallel samples. All samples correspond to a clear pit mud quality grade.

4. The screening method according to claim 1, characterized in that, In step 3, the parameters for preparing the pit mud suspension are as follows: homogenize at a speed of 10 times / s for 5 min, take 1 g of pit mud and add 100 mL of 1M PBS buffer with pH=7.4, and mix at 150 rpm for 30 min.

5. The screening method according to claim 1, characterized in that, In step 4, the PMA treatment conditions were as follows: PMA concentration 50 μM, incubation in the dark for 10 min, exposure to 60 W blue light for 15 min, and centrifugation at 5000 g for 10 min to collect the bacterial cells.

6. The screening method according to claim 1, characterized in that, In step 5, the sequencing primers were 515F / 806R, and the ASV sequence was obtained by DADA2 noise reduction. Species classification annotation was performed at the genus level, and the obtained sequence and abundance information were assigned to the corresponding quality grade samples.

7. The screening method according to claim 1, characterized in that, In step 7, the training set and the test set are divided into layers at a ratio of 7:3, and the data are standardized to the 0-1 range. All datasets carry labels for the quality grade of the pit mud.

8. The screening method according to claim 1, characterized in that, In step 8, the number of decision trees in the first random forest model is 50. Based on the prediction error of the external bag permutation, bacterial genera with a contribution greater than 0 are selected as candidates. All candidate bacterial genera are correlated with the quality grade of the pit mud.

9. The screening method according to claim 1, characterized in that, In step 9, 50 bootstrap resampling operations were performed, and 5-fold cross-validation was used. Based on the 1-SE principle, the final key bacterial genera were screened according to the stability frequency threshold.

10. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, implements the steps of the screening method for key live bacteria in the cellar mud of strong-aroma baijiu as described in any one of claims 1-9, and during execution, calls the cellar mud quality grade classification results as classification labels.

Images