A probiotic phenotype prediction method based on multi-dimensional genomic feature fusion and stacking ensemble learning and application thereof

Abstract

The application discloses a probiotic phenotype prediction method based on multi-dimensional genomic feature fusion and stacked ensemble learning and application thereof. The construction method of the probiotic phenotype prediction model comprises the following steps: obtaining the whole genome sequence of the probiotic with known probiotic phenotype label, constructing a training data set and an independent test set; extracting k-mer sequence composition features and pan-genome homology features at the whole genome level; matching and fusing the k-mer features and the pan-genome features to construct a k-mer-pan-genome fusion feature matrix corresponding to different phenotypes; taking the fusion feature matrix and the corresponding phenotype label as input, training a machine learning model by using a stacked ensemble learning framework, and obtaining a probiotic phenotype prediction model for determining a specific phenotype. The application innovatively deeply fuses the ensemble learning algorithm and the SHAP explainability framework, accurately removes invalid redundant noise of high-dimensional data, and accurately quantifies and extracts core specific k-mers and pan-genome features that determine a specific probiotic phenotype.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, specifically to a method for predicting probiotic phenotypes based on multidimensional genomic feature fusion and stacked ensemble learning, and its application. Background Technology

[0002] Probiotics are live microorganisms that, when ingested in adequate amounts, provide health benefits to the host. In recent years, the global probiotic market has experienced rapid and sustained expansion as strains with clearly defined health-promoting effects have been continuously developed. However, a significant translational gap has long existed between basic research and industrial commercial applications of probiotics: academia typically focuses on specific biological activities of strains (such as antibacterial properties and hypoglycemic effects) to screen for highly functional strains; while industry, in actual production, urgently needs strains with excellent industrial production performance (such as growth rates in high-density fermentation) and ecological host adaptability (such as acid and bile salt tolerance for intestinal colonization). This disconnect between laboratory and factory evaluation results in many functional strains that perform well in the laboratory facing bottlenecks such as low production efficiency and poor survival rates during productization, significantly increasing the industry's trial-and-error costs and severely hindering the overall development of the probiotic industry.

[0003] In recent years, with the popularization of high-throughput whole genome sequencing (WGS) technology and the development of computational biology, using machine learning algorithms to directly predict the macroscopic phenotype of probiotics based on genomic features (such as k-mer sequence characteristics) has become a highly promising cutting-edge alternative in this field. For example, some studies have developed the "iProbiotics" tool using k-mer (a substring of DNA sequence of length k) feature extraction technology and Support Vector Machine (SVM) algorithms to assess the probiotic potential of bacterial strains. Other studies have combined k-mer profiling with phenotypic features, establishing the "VLCPredictor" framework based on the Random Forest (RF) algorithm to evaluate the functional potential of *Lactobacillus curvatureensis*. CN 202610582226.X also discloses a method for extracting nucleotide subsequence fragments from genomic sequence data using sliding windows of multiple preset lengths, obtaining sets of nucleotide subsequence fragments corresponding to each preset length; mapping each nucleotide subsequence fragment to integer words based on its frequency of occurrence in the genomic sequence data, obtaining word sequences; and inputting the word sequences corresponding to each preset length into a tolerance trait prediction model to obtain tolerance trait prediction results. These advances demonstrate the feasibility of machine learning at the genomic prediction level.

[0004] However, existing genome-based machine learning prediction technologies still face two major technical bottlenecks that urgently need to be overcome: First, the "curse of dimensionality" and computational bottleneck caused by high-dimensional features. The whole genome sequence of probiotics contains massive amounts of genetic information. When converted into machine learning input features, it often forms a high-dimensional and sparse feature matrix of tens or even hundreds of thousands of elements. Traditional single machine learning models are prone to overfitting when processing such high-dimensional data, resulting in poor generalization ability. Simultaneously, the input of massive redundant features and background noise inevitably leads to extremely high computational costs, resulting in slow prediction response and making it difficult to meet the needs of real-time evaluation of large-scale whole genome data in industry. Second, the "black box" nature of complex models and the lack of biological interpretability. To improve prediction accuracy, existing technologies tend to use complex high-order nonlinear models (such as deep neural networks and ensemble learning). Although these models can output high prediction accuracy, their internal decision-making mechanisms are highly complex, resembling a "black box," making it impossible for researchers to know which specific genomic features truly drive the model's prediction results. In the actual development of the probiotic industry, the industry not only needs to know "which strain has a superior phenotype," but also urgently needs to know "why it has that superior phenotype" (i.e., what are the specific genes or regulatory elements that determine acid resistance, bile salt tolerance, or high growth rate). This lack of biological interpretability seriously hinders subsequent confirmation of the probiotic mechanism of strains, discovery of molecular targets, and targeted genetic improvement.

[0005] Therefore, there is an urgent need in this field for an innovative screening method that can both guarantee high-precision prediction and break the "black box" barrier, achieving computational lightweighting and mechanism interpretability. Summary of the Invention

[0006] The technical problem this invention aims to solve is that the existing evaluation system for probiotic industrial and ecological phenotypes (such as growth rate, acid resistance, and bile salt tolerance) suffers from a disconnect between laboratory and factory transformation, resulting in low industrial transformation efficiency and high trial-and-error costs for many functional strains. Furthermore, existing single machine learning prediction models are prone to the "curse of dimensionality" when processing high-dimensional, sparse genomic data, leading to enormous computational overhead and extremely low prediction efficiency. In addition, existing complex prediction models generally exhibit "black box" characteristics, lacking interpretability of underlying biological mechanisms and making it difficult to directly correlate macroscopic phenotype prediction results with specific molecular targets. This invention provides a probiotic phenotype prediction method based on multidimensional genomic feature fusion and stacked ensemble learning. Specifically, it further provides a lightweight optimized prediction model construction method that breaks through the "black box" and significantly reduces dimensionality and noise, enabling rapid and accurate prediction of macroscopic phenotypes such as probiotic production performance and host adaptability, as well as the analysis of underlying mechanisms.

[0007] This invention is proposed against this backdrop, not only constructing a high-precision basic prediction framework based on multidimensional feature fusion and stacking ensemble learning, but also innovatively introducing the game theory-based SHapley Additive Interpretation (SHAP) interpretable machine learning algorithm. This method can accurately quantify the global marginal contribution of massive high-dimensional genomic features to the prediction of specific phenotypes, successfully remove redundant noise, reduce the dimensionality of tens of thousands of features to at least a few core genetic markers, and thereby reconstruct a highly lightweight probiotic function prediction system (such as the ProBio-Fast system in this embodiment).

[0008] This technological breakthrough not only effectively overcomes the hardware bottlenecks of slow computation and demanding computing power requirements of traditional high-dimensional models, achieving "second-level" accurate predictions for high-throughput genomic data; more significantly, it endows artificial intelligence models with strong biological interpretability, directly anchoring macroscopic phenotypic predictions to the microscopic molecular target level. This not only fills the gap between macroscopic phenotypic identification and underlying gene mechanism analysis, but also provides a highly advanced and translational computational biology tool for exploring the mechanisms of probiotic function, targeting and modifying industrial strains, and high-throughput mining of novel probiotic resources with independent intellectual property rights.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows: In a first aspect, the present invention provides a method for constructing a probiotic phenotype prediction model based on multidimensional genomic feature fusion and stacked ensemble learning, characterized by comprising the following steps: (1) Obtain the whole genome sequence of probiotics with known probiotic phenotypic tags, and construct a training dataset and an independent test set; (2) Extract k-mer sequence composition features and pan-genome homology features at the whole genome level, respectively; (3) Match and fuse the k-mer sequence composition features with the pan-genome homology features to construct the k-mer-pan-genome fusion feature matrix corresponding to different phenotypes; (4) Using the fused feature matrix and the corresponding phenotypic label as input, a machine learning model is trained using a stacked ensemble learning framework to obtain a probiotic phenotypic prediction model for specific phenotypic determination.

[0010] Preferably, the extraction of k-mer sequence composition features at the whole genome level in step (2) specifically involves: extracting sequence composition features at the whole genome level using k-mer counting analysis to obtain a k-mer feature matrix, where k values ​​are one or more of 6, 7, 8, 9, or 10; the extraction of pan-genome homology features in step (2) specifically involves: performing pan-genome analysis on the genome annotation file to obtain a pan-genome reference sequence, and using the TBLASTX method to perform homology comparison between the genome sequences of each strain and the pan-genome reference sequence to obtain a pan-genome feature matrix.

[0011] Preferably, the construction of the k-mer-pan-genome fusion feature matrix corresponding to different phenotypes in step (3) specifically involves: reading the k-mer feature matrix with different k values ​​under the corresponding phenotype and matching and merging it with the pan-genome feature matrix according to the strain number; the training of the machine learning model using the stacked ensemble learning framework in step (4) includes a first-layer base model and a second-layer meta-model; wherein, the first-layer base model learns the association between genomic features and probiotic phenotypes from different perspectives such as linear sparse features, random forest ensemble tree features, neural network nonlinear mapping features and gradient boosting ensemble tree features, and splices the predicted probabilities output by the first-layer base model to construct the meta-feature matrix; the second-layer meta-model uses a standardized L2 regularized logistic regression model to learn the complementary relationship between the prediction results of different base models and outputs the final binary classification prediction probability.

[0012] Preferably, the first base model includes resilient network logistic regression, random forest, multilayer perceptron, extreme gradient boosting, lightweight gradient boosting machine, and categorical gradient boosting algorithm for model construction.

[0013] Preferably, a lightweight optimized prediction model is further constructed from the probiotic phenotype prediction model obtained above, i.e., the primary prediction model, as follows: (1) Introduce the SHapley additive interpretation method to calculate the global SHAP absolute value of the prediction result of each input feature for a specific phenotype in the primary prediction model; (2) Sort the data in descending order according to the absolute value of the global SHAP, and screen the top N core genomic features. The core genomic features include specific k-mer sequence composition features and pan-genome homology features. Preferably, N is 20. (3) Using the extracted core genomic features as input data after dimensionality reduction, the model is retrained using the same stacked ensemble learning framework to construct a lightweight optimized prediction model based on the core genomic features.

[0014] Preferably, the probiotic phenotype is one or more of growth rate, acid resistance and bile salt resistance, and the probiotic is Lactobacillus plantarum. The core genomic features are: Core genomic features of growth performance include genes or sequences and feature types: group_3839, pan-genetic; dkgB, pan-genetic; group_2329, pan-genetic; group_753, pan-genetic; CATACTA, k-mer; group_1600, pan-genetic; CATGAGC, k-mer; group_2964, pan-genetic; ACTATAT, k-mer; CGGAAGA, k-mer; AGCAATG, k-mer; AGCTCGG, k-mer; group_1312, pan-genetic; group_2950, ​​pan-genetic; CGATCAA, k-mer; CCATAAC, k-mer; CCGTGAC, k-mer; murJ, pan-genetic; CTCCTTA, k-mer; group_3633, pan-genetic; the Genbak accession number for dkgB is YP_006750892.1, and the Genbak accession number for murJ is WP_089535142.1. The core genetic characteristics of acid tolerance include genes or sequences and characteristic types: CTAATTTTTA (SEQ ID NO. 14), k-mer; CGGCAAACTC (SEQ ID NO. 15), k-mer; AACCGGGGCA (SEQ ID NO. 16), k-mer; TGTTTAGCAA (SEQ ID NO. 17), k-mer; CGTTCAAAGC (SEQ ID NO. 18), k-mer; GGTCTGACTA (SEQ ID NO. 19), k-mer; GGAGTTTGCC (SEQ ID NO. 20), k-mer; TGAAAAGAAA (SEQ ID NO. 21), k-mer; ACATTCAACA (SEQ ID NO. 22), k-mer; GGTATATAA (SEQ ID NO. 23), k-mer; GAATATGGCA (SEQ ID NO. 24), k-mer; ATTTGGCATG (SEQ ID NO. 25), k-mer; GGCACTAATA (SEQ ID NO. 14), k-mer; GATATGCA (SEQ ID NO. 24), k-mer; ATTTGGCATG (SEQ ID NO. 25), k-mer; GGCACTAATA (SEQ ID NO. 16), k-mer; GATATATGA (SEQ ID NO. 17), k-mer; GGTTCAAAGC (SEQ ID NO. 18), k-mer; GGTCTGACTA (SEQ ID NO. 19), k-mer; GGAGTTTGCC (SEQ ID NO. 20), k-mer; GTGAAAGA (SEQ ID NO. 21), k-mer; GATATATAA (SEQ ID NO. 22), k-mer; GGTATATAA (SEQ ID NO. 23), k-mer; GATATGGCA (SEQ ID NO. 24), k-mer; ATTTGGCATG (SEQ ID NO. 25), k-mer; GGCACTAATA (SEQ ID NO. 16), k-mer; GATATATGA (SEQ ID NO. 17), k NO.26), k-mer; group_2792, pan-gene; CGCAAGACAC (SEQ ID NO.27), k-mer; CAGATTAGCC (SEQ ID NO.28), k-mer; CCATCTTTTA (SEQ ID NO.29), k-mer; thiM, pan-gene; group_3633, pan-gene; ACAACTGAAT (SEQ ID NO.30), k-mer; The Genbak accession number for thiM is WP_054682236.1; The core genomic features of bile salt tolerance include genes or sequences and feature types: TAGGCATATA (SEQ ID NO. 31), k-mer; ACCAAATAGT (SEQ ID NO. 32), k-mer; ATTGCCAATG (SEQ ID NO. 33), k-mer; ATTTGAACTA (SEQ ID NO. 34), k-mer; group_3452, pan-genome; patB, pan-genome; ACTAGCATTA (SEQ ID NO. 35), k-mer; group_2289, pan-genome; ACGTTTGACT (SEQ ID NO. 36), k-mer; GGTGTTGTCA (SEQ ID NO. 37), k-mer; TAAATCAGGA (SEQ ID NO. 38), k-mer; AATAAACAGC (SEQ ID NO. 39), k-mer; AAATCAGGAG (SEQ ID NO. 40), k-mer; ACCTGCCACA (SEQ ID NO. 31), k-mer; ACCTGCCACA ...1), k-mer; ACCTGCCACA (SEQ ID NO. 31), k- NO.41), k-mer; group_3741, pan-genome; mutS2, pan-genome; group_3007, pan-genome; ATGTCGGTGG (SEQ ID NO.42), k-mer; ATCAGGAGAG (SEQ ID NO.43), k-mer; group_3633, pan-genome; patB's Genbak accession number is YP_007988004.1; mutS2's Genbak accession number is YP_007414940.1; The nucleotide sequence of group_753 is shown in SEQ ID NO.1; The nucleotide sequence of group_1312 is shown in SEQ ID NO.2; The nucleotide sequence of group_1600 is shown in SEQ ID NO.3; The nucleotide sequence of group_2289 is shown in SEQ ID NO.4; The nucleotide sequence of group_2329 is shown in SEQ ID NO.5; The nucleotide sequence of group_2792 is shown in SEQ ID NO.6; The nucleotide sequence of group_2950 is shown in SEQ ID NO.7; The nucleotide sequence of group_2964 is shown in SEQ ID NO.8; The nucleotide sequence of group_3007 is shown in SEQ ID NO.9; The nucleotide sequence of group_3452 is shown in SEQ ID NO.10; The nucleotide sequence of group_3633 is shown in SEQ ID NO.11; The nucleotide sequence of group_3741 is shown in SEQ ID NO.12; The nucleotide sequence of group_3839 is shown in SEQ ID NO.13.

[0015] A second aspect of the present invention provides a probiotic phenotype prediction model constructed by the above-described construction method.

[0016] A third aspect of the present invention provides a lightweight optimized prediction model constructed according to the above-described construction method.

[0017] A fourth aspect of the present invention provides a method for screening candidate functional strains of probiotics, comprising the following steps: (1) Obtain the whole genome sequence file of the probiotic to be tested and annotate it; (2) Extract k-mer sequence composition features and generate corresponding pan-genome homology features through homology comparison; (3) Match and fuse the k-mer sequence composition features with the pan-genome homology features to form a fusion feature vector of the strain to be tested; or, extract the core genomic features determined by the method to form a dimensionality-reduced core feature vector. (4) Input the fused feature vector into the probiotic phenotype prediction model or input the core feature vector into the lightweight optimized prediction model; (5) Output the predicted probability and / or classification result of the probiotic to be tested belonging to the positive category under a specific phenotype; (6) Compare the predicted probability with the preset classification threshold, and classify, screen and output the candidate functional strains.

[0018] A fourth aspect of this invention provides a probiotic probiotic function prediction system based on multidimensional feature fusion and stacked ensemble learning, comprising: (1) Input module: used to receive the whole genome sequence file of the probiotic to be tested and to perform genome sequence annotation; (2) Feature acquisition module: used to extract k-mer sequence composition features based on the whole genome sequence of the probiotic to be tested, generate corresponding pan-genome homology features, and match and fuse the k-mer sequence composition features with the pan-genome homology features to generate a fused feature vector; or in the lightweight prediction system (ProBio-Fast), the feature acquisition module only extracts the core genome features (preferably the Top 20 key features) determined based on SHAP parsing, generates a dimensionality-reduced feature vector, and avoids massive redundant feature calculations; (3) Model calling module: used to call the probiotic phenotype prediction model obtained after training, or to call the lightweight optimized prediction model constructed based on the core genomic features; (4) Result output module: used to output the predicted probability and classification result of phenotypic category.

[0019] The fifth aspect of the present invention provides a computer-readable storage medium or computer device: The storage medium stores a computer program, or the device includes a processor and a computer program, which, when executed, implements the method for constructing a prediction model as described in any of the preceding claims, or the method for screening candidate functional strains as described above.

[0020] Compared with the prior art, the advantages of this invention are as follows: (1) Breaking the "black box" characteristic of the model and achieving a leap from macroscopic prediction to underlying mechanism analysis (key innovation): This invention innovatively integrates advanced ensemble learning algorithms with the SHAP interpretability framework, accurately removing invalid and redundant noise from high-dimensional data, and precisely quantifying and extracting the core specific k-mers and pan-genome features (such as the top 20 core features) that determine specific probiotic phenotypes (such as acid resistance, bile salt resistance, etc.). This not only endows artificial intelligence with strong biological interpretability, but also provides clear molecular target basis for the subsequent targeted genetic improvement of industrial strains and in-depth confirmation of probiotic mechanisms.

[0021] (2) Significantly reduced computational cost and second-level high-precision prediction (key innovation): The lightweight prediction system (ProBio-Fast) based on the reconstruction of extracted core target genes successfully compressed the model input features from tens of thousands of dimensions to tens of dimensions (e.g., 20 dimensions). This system completely overcomes the bottlenecks of traditional high-dimensional models, such as slow computation, susceptibility to the "curse of dimensionality," and stringent hardware requirements. Evaluation confirms that while significantly reducing computational overhead and achieving "second-level" response at the whole genome level, it still maintains good prediction accuracy (e.g., a bile salt tolerance prediction AUC of up to 96.80%), perfectly meeting the industry's demand for high-throughput, low-cost real-time screening.

[0022] (3) Innovative multidimensional genomic feature fusion to fully characterize probiotic potential: In response to the problem of incomplete characterization of single genomic information, this invention creatively combines k-mer features that reflect the underlying sequence pattern with pan-genome features that reflect the high-level function, effectively making up for the lack of intuitive functional annotation of k-mer and the difficulty of pan-genome in reflecting the variation of non-coding and regulatory regions, and comprehensively characterizing the genetic underlying logic of strain phenotype.

[0023] (4) Dual-layer stacked integrated architecture to overcome the generalization bottleneck of a single model: The Stacking dual-layer prediction network (base learner + meta learner) is adopted, which fully absorbs the complementary advantages of various advanced algorithms such as decision tree, multilayer perceptron, and LightGBM in capturing nonlinear biological interactions. It effectively overcomes the blind spot of a single model in specific complex phenotypes and significantly improves the model's cross-phenotype generalization ability and robustness.

[0024] (5) Rigorous model evaluation and industrial transformation value: In the model construction stage, a 7:3 independent test set partitioning and a 10-fold stratified cross-validation strategy were adopted to strictly ensure class balance and eliminate the risk of data leakage and overfitting. The constructed automated prediction system effectively bridges the gap between probiotics in laboratory biological research and macro-industrial production evaluation in factories, greatly reducing the cost of trial and error in the industry, and providing a disruptive technological paradigm for accelerating the discovery and commercialization of new probiotic strains. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall construction process of the probiotic phenotype prediction model described in the embodiments of the present invention.

[0026] Figure 2 This is a schematic diagram of the probiotic function prediction process of the ProBio system according to an embodiment of the present invention. Detailed Implementation

[0027] To more clearly illustrate the technical solution of the present invention, the following detailed description is provided in conjunction with specific embodiments, but these descriptions should not be construed as limiting the present invention; they are merely some embodiments of the present invention.

[0028] The culture media involved in the following examples are as follows: MRS agar plates (g / L): Casein digest 10.0 g / L, beef extract 5.0 g / L, yeast extract 4.0 g / L, glucose 20.0 g / L, Tween 80 1.0 mL / L, K2PO4·3H2O 2.0 g / L, sodium acetate 5.0 g / L, triammonium citrate 2.0 g / L, MgSO4·7H2O 0.2 g / L, MnSO4·4H2O 0.05 g / L, agar 20 g / L, with water as the solvent. The preparation method is to mix all the components evenly and sterilize them.

[0029] MRS broth medium (g / L): Casein digest 10.0 g / L, beef extract 10.0 g / L, yeast extract 4.0 g / L, glucose 20.0 g / L, Tween 80 1.0 mL / L, K2PO4·3H2O 2.0 g / L, sodium acetate 5.0 g / L, triammonium citrate 2.0 g / L, MgSO4·7H2O 0.2 g / L, MnSO4·4H2O 0.05 g / L, with water as the solvent. It is prepared by mixing all the components evenly and then sterilizing.

[0030] Example 1: Isolation and Identification of Probiotics 1.1. Isolation of probiotics Traditional fermented foods and animal feces from Xinjiang and Tibet were collected as samples. Under aseptic conditions, 0.1 g of sample was added to 10 mL of MRS broth medium, vortexed, and then cultured anaerobically at 37 °C for 24 h. 0.5 mL of the bacterial solution was then serially diluted. Physiological saline was added to prepare 10... -1 Up to 10 -8 Dilute gradient bacterial suspensions, select 10 -6 10 -7 10 -8 Three gradient bacterial suspensions were each pipetteed at 100 μL onto MRS agar plates, spread evenly using a spreader, and then incubated anaerobically at 37 °C for 48 h. Morphologically typical colonies from the plates were picked and streaked onto MRS agar plates for purification. After purification, single colonies were inoculated into MRS liquid medium and anaerobically cultured at 37 °C for 48 h. The culture was then stored in a -80 °C cryogenic freezer with 30% glycerol to obtain potential probiotic strains.

[0031] 1.2. Identification of Probiotics (1) Sample preparation: 0.2 g of fresh colonies obtained in 1.1 were resuspended in 300 μL of ultrapure water, 900 μL of anhydrous ethanol was added, the mixture was vortexed and centrifuged to remove the supernatant, 50 μL each of 70% formic acid and acetonitrile were added and mixed thoroughly, the mixture was centrifuged at 13,000×g at room temperature for 2 min and the supernatant was collected to obtain the sample.

[0032] (2) MALDI-TOF MS detection: 1 μL of the prepared sample was added to the target plate and allowed to dry. Then, 1 μL of α-cyano-4-hydroxycinnamic acid matrix solution (10 mg / mL) was added, and after drying, the sample was analyzed. MALDI-TOF MS detection was performed on a Microflex instrument. ® The experiment was conducted on the smart (Bruker, Leipzig, Germany) platform: calibration was performed using the Bacterial Test Standard (Bruker) before each experiment; subsequently, 20... 0h The z Smartbeam solid-state laser bombarded the target 240 times in cationic linear mode, acquiring spectra in the 2-20 kDa range to form the protein fingerprint of the test strain. A file package containing the spectra file (.fid) was saved. The spectra file was then imported into Bruker Biotyper. ® The software obtains the species information of the strain to be tested.

[0033] (3) Determination of probiotic species in this study: A total of 317 strains of Lactobacillus plantarum were obtained from the samples in this study. All of them were identified as Lactobacillus plantarum in the Bruker Biotyper system, with scores between 2.0 and 2.2, indicating that the identification results are reliable.

[0034] Example 2: Construction of a probiotic phenotype database 2.1. Activation of probiotics The probiotics to be tested were inoculated from glycerol tubes onto MRS agar plates and anaerobically cultured at 37 °C for 48 h. Colony morphology and color were observed. For strains with uniform colony morphology and color, single colonies were picked and inoculated into MRS broth medium and anaerobically cultured at 37 °C for 48 h; then, they were inoculated into fresh MRS broth medium and anaerobically cultured at 37 °C for 48 h to obtain activated probiotics. The activated probiotics were then used to detect probiotic phenotypes, including growth rate, acid tolerance, and bile salt tolerance.

[0035] 2.2. Detection of probiotic growth rate Activated probiotic strains were used, and the cell density was adjusted to OD. 600= 1.00 ± 0.05 as the inoculum. 10 µL of the inoculum was inoculated into 190 µL of MRS broth and anaerobically cultured at 37 ℃ for 48 h. At 0, 12, 24, 36, and 48 h of culture, the bacterial density (OD) of the strain was determined using a multi-functional microplate reader (Biotek, USA). 600 (Value). The growth rate of each strain was measured in 5 biological replicates.

[0036] Based on the OD values ​​of each strain at 0, 12, 24, 36, and 48 hours 600 Dynamic changes (OD) 0h OD 12h OD 24h OD 36h OD 48h ), calculate the relative growth rate (RGR) and maximum specific growth rate (GR) of the strain. μmax RGR is calculated as the area under the growth rate curve (GR-AUC) and the area under the curve of growth rate. The formula for calculating RGR is: RGR (%) = (OD) / (Growth Rate - Area Under the Curve) / (Growth Rate - Area Under the Curve) = G / AUC. 48h - OD 0h ) / OD 0h × 100%; GR μmax Based on the OD between two adjacent detection time points 600 The changes in the values ​​are calculated, and the maximum value is taken in each time interval; GR-AUC is based on the OD at each time point. 600 Dynamic changes were calculated using the trapezoidal integral method. Simultaneously, the OD values ​​for the same strain were... 0h and OD 48h Paired t-tests were performed to determine whether the bacterial density increased significantly during the detection period. Strains with significantly increased bacterial density were included in subsequent analyses.

[0037] Furthermore, ①OD is adopted 48h ②RGR, ③GR μmax The growth rate of the strain was comprehensively scored using four indicators: ①OD, ②GR-AUC, and ③GR-AUC. 48h The scoring criteria are: when OD 48h > 1.10 hours = 2 minutes, 0.90 < OD 48h ≤ 1.10 hours is counted as 1 minute, OD 48h ① ≤ 0.90, score 0 points; ② The scoring criteria for RGR are: 2 points when RGR > 400%, 1 point when 200% < RGR ≤ 400%, and 0 points when RGR ≤ 200%; ③ GR μmaxThe scoring criteria are: when GR μmax > 0.15 h -1 2 minutes, 0.09 h -1 < GR μmax ≤ 0.15 h -1 1 minute, GR μmax ≤ 0.09 h -1 ④ The scoring criteria for GR-AUC are as follows: 2 points are awarded when GR-AUC > 55, 1 point is awarded when 45 < GR-AUC ≤ 55, and 0 points are awarded when GR-AUC ≤ 45. The sum of the four indicators is used as the comprehensive score of the strain. The classification method for the growth rate of the strain is as follows: when the comprehensive score of the strain is ≥ 7 points, the strain is classified as a strain with a fast growth rate (marked as "1" in the "Growth Rate" field of the database); otherwise, it is judged as a strain with a slow growth rate (marked as "0" in the "Growth Rate" field of the database).

[0038] 2.3. Acid resistance test of probiotics The pH of the MRS broth medium was adjusted to 3.5 using HCl. The activated probiotic strain was then used, and the cell density was adjusted to OD0.05. 600 = 1.00 ± 0.05 as the inoculum. 10 µL of the inoculum was inoculated into 190 µL of MRS broth medium at pH = 3.5 and anaerobically cultured at 37 ℃ for 48 h. The bacterial density (OD) of the strain was measured using a multi-functional microplate reader at 0, 12, 24, 36, and 48 h of the start of culture. 600 (Value). The acid tolerance test for each strain was performed in 5 biological replicates.

[0039] Based on the OD values ​​of each strain at 0, 12, 24, 36, and 48 h in an acidic environment 600 Dynamic changes (aOD) 0h aOD 12h aOD 24h aOD 36h aOD 48h ), calculate the relative growth rate (RGR in acidic environment, aRGR) and maximum specific growth rate (aGR) of the strain in acidic environment. μmax ) and the area under the growth rate curve (aGR-AUC). aRGR, aGR μmax The calculation method for aGR-AUC is the same as that in the growth rate detection section.

[0040] Furthermore, using ①aOD 48h 、②aRGR、③aGR μmaxThe acid resistance of the strain was comprehensively scored using four indicators: ①aOD, ②aGR-AUC, and ③aGR-AUC. 48h The scoring criteria are: when aOD 48h > 1.00 hours = 2 minutes, 0.60 < aOD 48h 1 point is awarded for a score ≤1.00, aOD 48h ① ≤ 0.60, score 0 points; ② The scoring criteria for aRGR are: 2 points when aRGR > 300%, 1 point when 100% < aRGR ≤ 300%, and 0 points when aRGR ≤ 100%; ③ aGR μmax The scoring criteria are: when aGR μmax >0.08 h -1 2 minutes, 0.04 h -1 < aGR μmax ≤ 0.08 h -1 1 minute on the clock, aGR μmax ≤ 0.04 h -1 ④ The scoring criteria for aGR-AUC are as follows: 2 points are awarded when aGR-AUC > 40, 1 point is awarded when 20 < aGR-AUC ≤ 40, and 0 points are awarded when aGR-AUC ≤ 20. The sum of the four indicators is used as the comprehensive score of the strain. The classification method for acid tolerance of strains is as follows: when the comprehensive score of a strain is ≥ 6 points, the strain is classified as a strain with strong acid tolerance (marked as "1" in the "Acid Tolerance" field of the database); otherwise, it is judged as a strain with weak acid tolerance (marked as "0" in the "Acid Tolerance" field of the database).

[0041] 2.4. Detection of bile salt tolerance in probiotics Bile salts were added to MRS broth medium to prepare an MRS broth medium containing 0.01% bile salts. Activated probiotic strains were then used, and the cell density was adjusted to OD0.01. 600 = 1.00 ± 0.05 as the inoculum. 10 µL of the inoculum was inoculated into 190 µL of MRS broth containing 0.01% bile salts and anaerobically cultured at 37 ℃ for 48 h. The bacterial density (OD) of the strain was measured using a multi-functional microplate reader at 0, 12, 24, 36, and 48 h of the start of culture. 600 (Value). The bile salt tolerance test for each strain was performed in 5 biological replicates.

[0042] Based on the OD values ​​of each strain at 0, 12, 24, 36, and 48 h under bile salt stress... 600 Dynamic changes (bOD) 0h bOD 12h bOD 24h bOD 36h bOD48h The relative growth rate (RGR) under bilesalt stress environment (bRGR) and maximum specific growth rate (bGR) of the strain were calculated. μmax ) and the area under the growth rate curve (bGR-AUC). bRGR, bGR μmax The calculation method for bGR-AUC is the same as that for the growth rate detection section.

[0043] Furthermore, using ①bOD 48h , ②bRGR, ③bGR μmax The strain's bile salt tolerance was comprehensively scored using four indicators: bOD, bGR-AUC, and bGR-AUC. 48h The scoring criteria are: when bOD 48h > 0.45 hours = 2 minutes, 0.35 hours = 2 minutes <bOD 48h ≤ 0.45 seconds counts as 1 minute, bOD 48h ① ≤ 0.35, score 0 points; ② The scoring criteria for bRGR are: 2 points when bRGR > 60%, 1 point when 30% < bRGR ≤ 60%, and 0 points when bRGR ≤ 30%; ③ bGR μmax The scoring criteria are: when bGR μmax > 0.02 h -1 2 minutes, 0.01 h -1 < bGR μmax ≤ 0.02 h -1 1 minute on the clock, bGR μmax ≤ 0.01 h -1 ④ The scoring criteria for bGR-AUC are as follows: 2 points when bGR-AUC > 15, 1 point when 6 < bGR-AUC ≤ 15, and 0 points when bGR-AUC ≤ 6. The sum of the four indicators is used as the comprehensive score of the strain. The classification method for the bile salt tolerance of the strain is as follows: when the comprehensive score of the strain is ≥ 5 points, the strain is classified as a strain with strong bile salt tolerance (marked as "1" in "Bile Salt Tolerance" in the database); otherwise, it is judged as a strain with weak bile salt tolerance (marked as "0" in "Bile Salt Tolerance" in the database).

[0044] Example 3: Construction of a Probiotic Genome Database 3.1. DNA extraction from probiotics Whole-genome DNA was extracted from all bacterial strains using a bacterial DNA extraction kit (Huankai Biotechnology, China). 1–2 mL of bacterial culture in the logarithmic growth phase was transferred to a 1.5 mL centrifuge tube and centrifuged at 10,000 × g for 1 min, then the supernatant was discarded. Next, 120 µL of bacterial suspension and 4 µL of RNase A were added to the bacterial pellet, vortexed, and incubated at 37 °C for 30 min. Then, 220 µL of lysis buffer and 20 µL of proteinase K were added, vortexed, and incubated at 70 °C for 10–15 min. Following this, 250 µL of anhydrous ethanol was added, and the mixture was vortexed for 15 s. The entire mixture was transferred to a DNA silica membrane adsorption column, centrifuged at 10,000 × g for 1 min, and the filtrate was discarded. The cells were then washed with 500 µL of protein removal buffer and 500 µL of wash buffer, centrifuged at 10,000 × g for 1 min, and the filtrate was discarded. Finally, the adsorption column was placed in a new 1.5 mL sterile centrifuge tube, and 100 µL of elution buffer was added to the center of the adsorption column membrane. The tube was incubated at room temperature for 5 min, followed by centrifugation at 10,000 × g to elute the DNA. The extracted genomic DNA was stored at -20 °C for use in whole-genome sequencing.

[0045] 3.2. Whole genome sequencing of probiotics The concentration and quality of extracted genomic DNA were assessed using the Qubit dsDNA High Sensitivity Kit (Invitrogen, USA) and the Qubit 3.0 quantitative PCR instrument (Invitrogen, USA). Sequencing libraries were then constructed using the AMT RapidDNA-Seq Kit for Illumina (Cistro, USA). The steps included DNA fragmentation and end repair, adapter ligation, magnetic bead purification, fragment selection (target fragment 300-600 bp), and library PCR amplification. The fragment distribution and concentration of the library were assessed using the Agilent High Sensitivity DNA Kit (Agilent Technologies, USA) and the Agilent 2100 Bioanalyzer (Agilent Technologies, USA) to confirm quality. After mixing and denaturing the qualified libraries, 2×150 bp paired-end sequencing was performed using an Illumina NextSeq 550 sequencer (Inmena, USA) with the High Output v2.5 kit (Inmena, USA).

[0046] 3.3. Whole genome assembly and annotation of probiotics Trimmomatic software (v0.39) was used to perform quality control on the processed data, removing low-quality reads and adapter sequences. The resulting high-quality reads were then assembled de novo using SPAdes software (v3.13.1) to obtain genome contigs. After assembly, the assembly quality needed to be evaluated. Quast software (v5.0.2) was used to screen genomes that met the criteria (genome size 1.80–3.50 Mb, GC% 38.50%–50.00%, contig number < 150, contig N50 > 50 kb). Finally, Prokka software (v1.13) was used to annotate the screened genomes, predicting coding sequences, rRNA, tRNA, and other gene elements.

[0047] Example 4: Constructing a Probiotic Function Prediction System for Probiotics Using Machine Learning Combined with Whole Genome Analysis 4.1. Processing of Probiotic Phenotype Datasets For all probiotic phenotypic data, data cleaning and outlier handling were performed first: the validity of each data point was checked, and then statistical methods were used to identify and remove outliers that significantly deviated from the main distribution trend, ensuring reliable data quality. Subsequently, the phenotypic data were standardized according to the method in Example 2: the Z-score standardization method was used to transform the original phenotypic data into a standard normal distribution with a mean of 0 and a standard deviation of 1, to eliminate the influence of different units of measurement and to make the model converge faster and more stable. All data processing was performed in a Python 3.8.9 environment, primarily using packages such as pandas and scikit-learn.

[0048] 4.2. Dataset Partitioning To rigorously evaluate the model's generalization ability while mitigating the risk of overfitting, this study employs a model evaluation strategy combining a training set and an independent test set. First, the probiotic dataset, with completed phenotypic labeling and matching to genomic features, is randomly divided into a training set and an independent test set in a 7:3 ratio. A stratified sampling strategy is used during the partitioning process to ensure that the ratio of positive to negative samples in the training and test sets remains as consistent as possible, thereby reducing the impact of class distribution differences on the model evaluation results.

[0049] Model training, feature selection, hyperparameter optimization, and model fusion are all performed solely within the training set. The independent test set is not involved in model training, feature selection, or parameter tuning; it is only used for final model performance evaluation. To improve the stability of model training and fully utilize the training set samples, a 10-fold hierarchical cross-validation method is further employed within the training set for model training and validation. Specifically, the training set is divided into 10 mutually exclusive subsets according to the proportion of phenotypic labels. In each round, 9 subsets are used as the training subset, and the remaining subset is used as the validation subset. This process is repeated 10 times, ensuring that each subset is used as the validation set once. During each round of cross-validation, feature selection, hyperparameter optimization, and base learner training are all performed only on the corresponding training subset. The validation subset is used only for model performance evaluation and validation during the ensemble model construction process.

[0050] By employing the aforementioned data partitioning method, model parameters and feature combinations can be optimized during the training phase, while the final model can be unbiasedly evaluated using an independent test set. This study established corresponding datasets for three binary classification phenotypes: growth rate, acid tolerance, and bile salt tolerance, and used the same data partitioning strategy for model training and performance evaluation.

[0051] 4.3. Genomic Feature Extraction To transform probiotic whole-genome sequences of different lengths into fixed-dimensional numerical features suitable for training machine learning models, this embodiment constructs k-mer sequence composition features, pan-genome features, and k-mer-pan-genome fusion features, respectively. These features are used for machine learning model construction, feature combination comparison, and performance evaluation for three probiotic phenotypes: growth rate, acid tolerance, and bile salt tolerance.

[0052] First, k-mer features were extracted based on the probiotic whole-genome assembly sequence. A k-mer is a continuous DNA subsequence of length k, reflecting the sequence composition characteristics of the strain at the whole-genome level. In this embodiment, KMC2 software was used to perform k-mer counting analysis on all probiotic genome sequences, setting k values ​​to 6, 7, 8, 9, and 10 to obtain the k-mer abundance information for each strain at different k values. Subsequently, the k-mer counting results of each strain were merged according to strain number to construct a corresponding k-mer feature matrix. Each row represents a strain, each column represents a k-mer feature, and the matrix elements represent the abundance or normalized abundance of that k-mer in the corresponding strain's genome. For features that are completely missing or constant features that remain unchanged across all samples, filtering was performed based on the training set during the subsequent model training phase to reduce the impact of invalid features on model training.

[0053] Secondly, pan-genome features were constructed based on the annotation results of each probiotic genome. Specifically, Panaroo software was used to perform pan-genome analysis on all probiotic genome annotation files to obtain a non-redundant pan-genome gene set and the corresponding pan-genome reference sequence. In this embodiment, Panaroo employed a rigorous cleaning mode for gene cluster correction and erroneous annotation filtering to reduce the impact of low-quality annotations, gene fragmentation, and incorrect clustering on subsequent feature construction.

[0054] After obtaining the pan-genome reference sequence, the TBLASTX method was further used to perform homology comparison between the genomic sequences of each strain and the pan-genome reference sequence, and a strain-level pan-genome homology feature matrix was constructed based on the comparison results. Each row in this matrix represents a strain, and each column represents a pan-genome gene feature. The matrix elements are used to represent the homology level of the corresponding pan-genome gene in that strain. In this embodiment, for pan-genome genes detected in the strain, their corresponding alignment identity value is used as the feature value; for undetected pan-genome genes, the feature value is recorded as 0. Through the above method, the genomic annotation information and homology gene composition information of each strain are transformed into a fixed-dimensional pan-genome feature matrix, which is used to supplement functional gene-level information that is difficult to directly interpret using k-mer sequence composition features.

[0055] Finally, the k-mer feature matrix and the pan-genome homology feature matrix were matched and fused according to strain number to construct a k-mer-pan-genome fusion feature matrix. For the three phenotypes of growth rate, acid tolerance, and bile salt tolerance, k-mer feature matrices with different k values ​​for the corresponding phenotypes were read and merged with the pan-genome feature matrix according to strain number. The fused feature matrix contains both whole-genome sequence composition information and pan-genome functional gene information, which can be used to evaluate the complementary effects of the two types of genomic features in different probiotic phenotype prediction tasks.

[0056] Through the above processing, this embodiment obtains the k-mer feature matrix, the pan-genome feature matrix, and the k-mer-pan-genome fusion feature matrix, respectively. These three types of feature matrices serve as inputs for subsequent machine learning model training and performance comparison. During subsequent model training, all supervised feature selection, low-frequency feature filtering, and model parameter learning are completed only within the training set; the independent test set does not participate in feature selection or model training and is only used for final performance evaluation.

[0057] 4.4. Establishment and Optimization of Machine Learning Models This embodiment, based on the constructed k-mer feature matrix, pan-genome feature matrix, and k-mer-pan-genome fusion feature matrix, establishes machine learning prediction models for three binary phenotypes of probiotics: growth rate, acid tolerance, and bile salt tolerance. To fully utilize the complementary advantages of different machine learning algorithms in high-dimensional genomic feature learning, this embodiment employs a stacking ensemble learning framework to construct the probiotic function prediction model. The overall construction process of the probiotic function prediction model is as follows: Figure 1 As shown.

[0058] like Figure 1 As shown, the model construction process in this embodiment mainly includes three parts: data and feature construction, model training, and probiotic function prediction. In the data and feature construction stage, the whole genome sequence data of probiotics is first quality-controlled and annotated. Then, k-mer features and pan-genome features are extracted, and strain-level k-mer feature matrices, pan-genome feature matrices, and k-mer-pan-genome fusion feature matrices are constructed. In the model training stage, the corresponding feature matrices and phenotypic labels are used as input. The first-layer base model generates prediction probabilities, and a meta-feature matrix is ​​further constructed. The second-layer meta-model outputs the final prediction probabilities. In the function prediction stage, the whole genome sequence of the strain to be predicted undergoes the same feature extraction and feature fusion process, and is then input into the trained model. The model outputs the prediction probabilities and classification results for the corresponding probiotic phenotypes.

[0059] During the model evaluation phase, the strain-level feature matrix corresponding to each phenotype was first divided into a training set and an independent test set in a 7:3 ratio. A stratified sampling strategy was employed to ensure that the ratio of positive to negative samples in the training set and the independent test set remained as consistent as possible. Model training, feature filtering, feature selection, base learner training, and meta-learner training were all completed only within the training set. The independent test set did not participate in model training, feature selection, or parameter tuning; it was used only for the final performance evaluation.

[0060] For k-mer features and k-mer-pan-genome fusion features, this embodiment employs supervised k-mer feature selection using XGBoost feature importance within the training set. Specifically, using only training set samples and their phenotypic labels, the XGBoost model is trained within the training set using hierarchical cross-validation. Candidate k-mer features are selected based on their importance stability and cumulative contribution across different trade-offs. The independent test set does not participate in the k-mer feature selection process; only the same feature columns are extracted according to the feature set determined by the training set for final testing and evaluation. For pan-genome features, low-frequency, very high-frequency, and constant features are filtered within the training set to reduce the impact of invalid features on model training.

[0061] Within the training set, a 10-fold hierarchical cross-validation process is further employed to generate the out-of-fold prediction results of the first-layer base learner. Specifically, the training set is divided into 10 mutually exclusive subsets according to the proportion of phenotypic labels. In each round, 9 subsets are selected as the training subset, and the remaining subset is used as the validation subset. Each base learner is trained on the training subset and outputs the positive class prediction probability for the validation subset. After 10 iterations, the first-layer out-of-fold prediction probability for each sample in the training set is obtained. For the independent test set, predictions are made by the base learners obtained during the 10-fold training process, and the average of the prediction probabilities is taken as the first-layer prediction representation of the test sample under that base learner.

[0062] The first base model utilizes algorithms including Resilient Logistic Regression, Random Forest, Multilayer Perceptron, Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting (LightGBM), and CatBoost to construct the model. These base models learn the association between genomic features and beneficial phenotypes from different perspectives, such as linear sparse features, random forest ensemble tree features, neural network nonlinear mapping features, and gradient boosting ensemble tree features. Specifically, Resilient Logistic Regression is suitable for linear discrimination under high-dimensional sparse features; Random Forest can reduce the variance of a single model by ensembling multiple decision trees; Multilayer Perceptron can learn (high-dimensional) nonlinear feature combinations; XGBoost improves model diversity and reduces overfitting risk through a stronger randomized splitting strategy; LightGBM effectively handles complex nonlinear relationships and feature interactions; and CatBoost excels at efficiently handling class features and effectively avoids target leakage through a ranking boosting mechanism, enhancing the model's generalization ability in complex feature spaces.

[0063] Subsequently, the predicted probabilities output by each base model in the first layer are concatenated to construct a meta-feature matrix. Each row of this meta-feature matrix corresponds to a bacterial strain, and each column corresponds to the positive prediction probability output by a base model. The second-layer meta-model employs a standardized L2 regularized logistic regression model to learn the complementary relationships between the prediction results of different base models and outputs the final binary classification prediction probability. When the prediction probability is greater than or equal to 0.5, the strain is determined to belong to the positive category; when the prediction probability is less than 0.5, the strain is determined to belong to the negative category.

[0064] For different k-mer lengths (k = 6, 7, 8, 9, and 10), and for k-mer features, pan-genome features, and k-mer-pan-genome fusion features, the aforementioned stacked ensemble learning framework was used for model training and comparison. By comparing the prediction performance of different k values, different feature inputs, and different model structures on independent test sets, the optimal combination of genomic features and model schemes suitable for predicting different probiotic phenotypes were determined.

[0065] 4.5. Machine Learning Model Evaluation and Performance Comparison To verify the effectiveness of the probiotic function prediction model constructed in this invention, this embodiment evaluates the prediction performance of different k values, different genomic feature inputs, and different model structures on an independent test set, in order to determine the preferred model scheme suitable for predicting probiotic growth rate, acid resistance, and bile salt tolerance.

[0066] In this embodiment, the area under the precision-recall curve (AUPRC), the area under the ROC curve (AUC), accuracy (ACC), F1 score, and Matthews correlation coefficient (MCC) are used as performance evaluation metrics for the model.

[0067] First, for each probiotic phenotype, this embodiment compares the predictive performance of the k-mer and pan-genome fusion model on independent test sets under k = 6, 7, 8, 9, and 10 conditions to evaluate the impact of different k-mer lengths on the fusion model performance. The comparison results of the fusion model's independent test set performance under different k-mer lengths are shown in Table 1. As shown in Table 1, the fusion model exhibits certain predictive ability under different k values, but the optimal k-mer length differs for different phenotypes, indicating that the k-mer length affects the model's ability to represent genomic information related to different probiotic phenotypes.

[0068] Furthermore, to compare the impact of different genomic feature inputs on the predictive performance of the stacked ensemble model, this embodiment constructs stacked ensemble models based on k-mer features, pan-genome features, and k-mer-pan-genome fusion features, respectively, and evaluates them on independent test sets. The results are shown in Table 1. Table 1 shows that the optimal feature inputs for different probiotic phenotypes are not entirely consistent. In the growth rate prediction task, the k7-mer and pan-genome fusion features achieve better AUPRC and AUC; in the acid tolerance prediction task, the k10-mer and pan-genome fusion features achieve the best overall performance; and in the bile salt tolerance prediction task, the k10-mer and pan-genome fusion features achieve the best overall performance. Compared with the k-mer or pan-genome models alone, the models constructed using the k-mer-pan-genome fusion features all show the best performance in AUPRC (P < 0.001). These results indicate that using k-mer-pan-genome fusion feature inputs helps improve model predictive performance.

[0069] After comparing different feature inputs, this embodiment further compares the predictive performance of different machine learning model structures based on the optimized feature inputs for each phenotype. Specifically, three probiotic phenotype prediction models were constructed using Decision Tree (DT), K-nearest neighbors (KNN), Ridge classifier, and stacked ensemble model, respectively. A comprehensive comparison was performed based on metrics such as AUPRC, AUC, ACC, F1, and MCC, and the results are shown in Table 2. As shown in Table 2, under the same feature inputs and the same independent test set partitioning, the stacked ensemble model exhibited better overall predictive performance in the three probiotic function prediction tasks. In the growth rate prediction task, the stacked ensemble model achieved a higher AUC, indicating that the model is well-suited for predicting growth rate-related phenotypes. In the acid resistance prediction task, the stacked ensemble model achieved higher AUC and AUPRC, indicating that it can effectively integrate the predictive information from different base learners and improve the model's ability to discriminate acid resistance phenotypes. In the task of predicting bile salt tolerance, the stacked ensemble model performed better on indicators such as AUC, AUPRC, ACC, F1 and MCC, demonstrating a stable predictive ability for bile salt tolerance phenotypes.

[0070] The results in Tables 1 and 2 show that k-mer length, genomic feature input type, and machine learning model structure all affect the probiotic function prediction performance. Based on the results of independent test sets for different phenotypes, this embodiment uses phenotype-specific feature inputs and a stacked ensemble learning framework to construct a probiotic function prediction model, thereby achieving effective prediction of probiotic functional phenotypes such as acid tolerance, bile salt tolerance, and growth rate.

[0071] Table 1. Performance comparison of stacked ensemble models on independent test sets under different genomic feature inputs. ;

[0072] Table 2. Performance comparison of different models on independent test sets for three probiotic function prediction tasks.

[0073] Example 5: Application of the Probiotic Probiotic Function Prediction System 5.1. ProBio System Flowchart for Predicting the Probiotic Function of Strains Traditional fermented food samples were collected, and probiotics were isolated, purified, and identified according to Example 1. For candidate strains identified as probiotics, whole-genome sequencing was performed according to the method in Example 3 to obtain the whole-genome sequence file of the strain to be predicted. Subsequently, the whole-genome sequence of the strain to be predicted was input into the ProBio system to predict the probiotic functions of the strain, such as acid tolerance, bile salt tolerance, and growth rate.

[0074] The ProBio system's probiotic function prediction process is as follows: Figure 2 As shown, firstly, the whole genome sequence of the probiotic to be predicted is input into the system, and the genome sequence is annotated. Secondly, k-mer features are extracted based on the genome sequence of the strain to be predicted, and corresponding pan-genome features are generated through TBLASTX alignment. Subsequently, the k-mer features and pan-genome features are matched and fused to form a fused feature vector of the strain to be predicted. Then, the fused feature vector is input into a pre-trained stacked ensemble learning model to obtain the predicted probability of the strain belonging to the positive category under the three phenotypes of growth rate, acid tolerance, and bile salt tolerance. Finally, the predicted probability is compared with a preset classification threshold. If the predicted probability is greater than the preset threshold, the strain is marked as the positive category of the corresponding probiotic function. If the predicted probability is not greater than the preset threshold, it is marked as the negative category, and the final prediction result and classification label are output.

[0075] Through the above process, the ProBio system can automatically complete genomic feature extraction, pan-genome feature construction, feature fusion, model prediction, and classification result output after obtaining the whole genome sequence of the test strain, enabling rapid assessment of the acid resistance, bile salt tolerance, and growth rate of candidate probiotics. This system can be used for the initial screening and functional evaluation of probiotics from traditional fermented food sources, thereby reducing the workload of traditional phenotypic screening experiments and improving the efficiency of probiotic functional screening.

[0076] 5.2. Performance analysis of the ProBio system for predicting the probiotic phenotypes of 25 strains of *Lactobacillus plantarum*. This study selected 25 newly isolated *Lactobacillus plantarum* strains for ProBio efficacy validation. The whole genome data of the strains were submitted according to the method described above, and after inputting into ProBio, rapid assessments of the strains' growth rate, acid tolerance, and bile salt tolerance were obtained. Simultaneously, the actual probiotic phenotypes of the 25 strains were detected and phenotypically classified according to the protocol in Example 2. The results obtained by the two methods are shown in Table 3.

[0077] The results showed that, regarding growth rate, 24 out of 25 *Lactobacillus plantarum* strains exhibited the same predicted and measured phenotypes, with an accuracy rate of 96.00%; regarding acid tolerance, 23 out of 25 strains exhibited the same predicted and measured phenotypes, with an accuracy rate of 92.00%; and regarding bile salt tolerance, 23 out of 25 strains exhibited the same predicted and measured phenotypes, with an accuracy rate of 92.00%. Therefore, ProBio demonstrates good predictive efficacy for the probiotic phenotypes of *Lactobacillus plantarum*.

[0078] Table 3 Comparison of ProBio system prediction and experimental results for the probiotic phenotypes of 25 newly isolated *Lactobacillus lactis* strains.

[0079] Example 6: Construction and efficacy evaluation of a probiotic function prediction system based on lactic acid bacteria-specific k-mers and genes 6.1. Key Gene Feature Mining Influencing the Probiotic Phenotype of Lactic Acid Bacteria In Example 4, the stacked ensemble model constructed based on high-dimensional k-mer and pan-genome fusion features demonstrated excellent predictive performance. However, the input of tens of thousands of high-dimensional features inevitably led to high computational overhead, and the "black box" nature of complex machine learning models limited their application in explaining biological mechanisms. To achieve dimensionality reduction and noise reduction, and to further improve the model's generalization ability and robustness, as well as enhance its biological interpretability, this example aims to break the "black box" by directly anchoring the model's prediction results to key genetic markers that determine specific probiotic phenotypes of lactic acid bacteria. This provides a clear molecular target basis for subsequent research on the probiotic functional mechanisms of strains and targeted improvement.

[0080] Specifically, this embodiment introduces the Shapley Additive Explanations (SHAP) method to perform in-depth feature importance analysis on the stacked ensemble model that performed best on the independent test set in Example 4. Based on game theory principles, SHAP can accurately quantify the average marginal contribution of each feature value in all datasets to the final output of the model. By calculating the global SHAP absolute value of the features (Mean |SHAP|), the influence weight of each genomic feature (including specific k-mer sequences and pan-genome functional genes) on the ability to discriminate specific probiotic phenotypes can be intuitively and objectively evaluated.

[0081] In this embodiment, the global |SHAP| values ​​of all features in each phenotypic prediction model are sorted in descending order, and the top 20 features are strictly used as the criteria for selecting core features, eliminating a large number of redundant and low-contribution background noise features. After SHAP analysis and screening, the 20 key genomic features with the largest contributions to the three probiotic phenotypes of growth rate, acid tolerance, and bile salt tolerance are shown in Table 4.

[0082] Table 4. Twenty key gene characteristics affecting the probiotic phenotype of lactic acid bacteria

[0083] As shown in Table 4, the key features influencing the probiotic phenotype of lactic acid bacteria exhibit a multi-dimensional fusion, encompassing both k-mer fragments representing specific DNA sequence composition patterns (such as the specific 10-mer sequence TAGGCATATA affecting acid tolerance) and pan-genome gene clusters representing specific metabolic pathways or structural proteins (such as murJ affecting growth performance and dkgB affecting bile salt tolerance). This result further confirms that the probiotic function of lactic acid bacteria is a complex polygenic trait resulting from the combined effects of low-level sequence features and high-level functional genes.

[0084] 6.2. Construction and Performance Evaluation of a Lightweight Prediction System Based on Core Genomic Features (ProBio-Fast) After accurately identifying the Top 20 core genetic features shown in Table 5, to verify the representational power and practical application value of the core feature set, this invention discards the original high-dimensional whole-genome feature matrix and extracts only these 20 high-dimensional converged features as input data. Based on this, this embodiment refers to the model construction process in Embodiment 4.4, again using decision trees, K-nearest neighbors, and Ridge classifiers as base learners, and combining them with a logistic regression meta-learner, to retrain and construct a stacked ensemble optimization prediction model based on core features. This invention names this model the Lightweight Probiotic Function Prediction System—ProBio-Fast.

[0085] To rigorously evaluate the actual predictive performance of the ProBio-Fast system, this embodiment strictly follows the evaluation scheme of Embodiment 4.5. On the same independent test set, five core evaluation indicators, namely AUC, AUPRC, ACC, F1-score and MCC, are used to comprehensively verify the performance of ProBio-Fast. The evaluation results are shown in Table 5.

[0086] Table 5 ProBio-Fast System Performance Evaluation

[0087] As shown in Table 5, despite the significant compression of the input feature dimensions of the ProBio-Fast system (from tens of thousands to only 20 dimensions), the system still maintained extremely high and stable discriminative ability in the three probiotic function prediction tasks. These results fully demonstrate that the key k-mers and gene features extracted by the SHAP method in this invention have extremely high biological representativeness and information entropy. The constructed ProBio-Fast system not only successfully overcomes the hardware bottlenecks of slow computation and demanding computing power requirements of traditional high-dimensional models, achieving "second-level" prediction response at the whole genome level, but also maintains or even surpasses the benchmark prediction accuracy while making the model decision-making process transparent. This system provides a highly innovative and practical computational biology tool for future industrial applications to conduct rapid, low-cost, high-throughput probiotic targeted screening and in-depth basic research on probiotic mechanisms.

Claims

1. A method for constructing a probiotic phenotypic prediction model based on multidimensional genomic feature fusion and stacked ensemble learning, characterized in that, Includes the following steps: (1) Obtain the whole genome sequence of probiotics with known probiotic phenotypic tags, and construct a training dataset and an independent test set; (2) Extracting k-mer sequence composition features at the whole genome level and extracting pan-genome homology features; (3) Match and fuse the k-mer sequence composition features with the pan-genome homology features to construct the k-mer-pan-genome fusion feature matrix corresponding to different phenotypes; (4) Using the fused feature matrix and the corresponding phenotypic label as input, a machine learning model is trained using a stacked ensemble learning framework to obtain a probiotic phenotypic prediction model for specific phenotypic determination.

2. The method for constructing a probiotic phenotype prediction model according to claim 1, characterized in that, The extraction of k-mer sequence composition features at the whole genome level in step (2) specifically involves: extracting sequence composition features at the whole genome level using k-mer counting analysis to obtain a k-mer feature matrix, where k values ​​are one or more of 6, 7, 8, 9, or 10; the extraction of pan-genome homology features in step (2) specifically involves: performing pan-genome analysis on the genome annotation file to obtain a pan-genome reference sequence, and using the TBLASTX method to perform homology comparison between the genome sequences of each strain and the pan-genome reference sequence to obtain a pan-genome feature matrix.

3. The method for constructing a probiotic phenotype prediction model according to claim 2, characterized in that, The construction of the k-mer-pan-genome fusion feature matrix corresponding to different phenotypes in step (3) is specifically as follows: read the k-mer feature matrix with different k values ​​under the corresponding phenotype, and match and merge it with the pan-genome feature matrix according to the strain number; the training of the machine learning model using the stacked ensemble learning framework in step (4) includes a first-layer base model and a second-layer meta-model; wherein, the first-layer base model learns the association between genomic features and probiotic phenotypes from different perspectives such as linear sparse features, random forest ensemble tree features, neural network nonlinear mapping features and gradient boosting ensemble tree features, and splices the predicted probabilities output by the first-layer base model to construct the meta-feature matrix; the second-layer meta-model uses a standardized L2 regularized logistic regression model to learn the complementary relationship between the prediction results of different base models and outputs the final binary classification prediction probability.

4. The method for constructing a probiotic phenotype prediction model according to claim 3, characterized in that, The first base model includes models constructed using resilient network logistic regression, random forest, multilayer perceptron, extreme gradient boosting, lightweight gradient boosting machine, and categorical gradient boosting algorithms.

5. The method for constructing a probiotic phenotype prediction model according to any one of claims 1 to 4, characterized in that, For the probiotic phenotype prediction model obtained by any one of claims 1-4, i.e., the primary prediction model, a lightweight optimized prediction model is further constructed as follows: (1) Introduce the SHapley additive interpretation method to calculate the global SHAP absolute value of the prediction result of each input feature for a specific phenotype in the primary prediction model; (2) Sort the data in descending order according to the absolute value of the global SHAP, and screen the top N core genomic features. The core genomic features include specific k-mer sequence composition features and pan-genome homology features. (3) Using the extracted core genomic features as input data after dimensionality reduction, the model is retrained using the same stacked ensemble learning framework to construct a lightweight optimized prediction model based on the core genomic features.

6. The method for constructing a probiotic phenotype prediction model according to claim 5, characterized in that, The probiotic phenotype is one or more of the following: growth rate, acid resistance, and bile salt tolerance; and the probiotic is Lactobacillus plantarum. The core genomic features are: Core genomic features of growth performance include genes or sequences and feature types: group_3839, pan-genetic; dkgB, pan-genetic; group_2329, pan-genetic; group_753, pan-genetic; CATACTA, k-mer; group_1600, pan-genetic; CATGAGC, k-mer; group_2964, pan-genetic; ACTATAT, k-mer; CGGAAGA, k-mer; AGCAATG, k-mer; AGCTCGG, k-mer; group_1312, pan-genetic; group_2950, ​​pan-genetic; CGATCAA, k-mer; CCATAAC, k-mer; CCGTGAC, k-mer; murJ, pan-genetic; CTCCTTA, k-mer; group_3633, pan-genetic; the Genbak accession number for dkgB is YP_006750892.1, and the Genbak accession number for murJ is WP_089535142.

1. The core genetic characteristics of acid tolerance include genes or sequences and characteristic types: SEQ ID NO.14, k-mer; SEQ ID NO.15, k-mer; SEQ ID NO.16, k-mer; SEQ ID NO.17, k-mer; SEQ ID NO.18, k-mer; SEQ ID NO.19, k-mer; SEQ ID NO.20, k-mer; SEQ ID NO.21, k-mer; SEQ ID NO.22, k-mer; SEQ ID NO.23, k-mer; SEQ ID NO.24, k-mer; SEQ ID NO.25, k-mer; SEQ ID NO.26, k-mer; group_2792, pan-gene; SEQ ID NO.27, k-mer; SEQ ID NO.28, k-mer; SEQ ID NO.29, k-mer; thiM, pan-gene; group_3633, pan-gene; SEQ ID NO.30, k-mer; thiM's Genbak login number is WP_054682236.1; Core genomic features of bile salt tolerance include genes or sequences and feature types: SEQ ID NO.31, k-mer; SEQ ID NO.32, k-mer; SEQ ID NO.33, k-mer; SEQ ID NO.34, k-mer; group_3452, pangenome; patB, pangenome; SEQ ID NO.35, k-mer; group_2289, pangenome; SEQ ID NO.36, k-mer; SEQ ID NO.37, k-mer; SEQ ID NO.38, k-mer; SEQ ID NO.39, k-mer; SEQ ID NO.40, k-mer; SEQ ID NO.41, k-mer; group_3741, pangenome; mutS2, pangenome; group_3007, pangenome; SEQ ID NO.42, k-mer; SEQ ID NO.43, k-mer; group_3633, pan-genome; patB's Genbak accession number is YP_007988004.1; mutS2's Genbak accession number is YP_007414940.1; The nucleotide sequence of group_753 is shown in SEQ ID NO.1; The nucleotide sequence of group_1312 is shown in SEQ ID NO.2; The nucleotide sequence of group_1600 is shown in SEQ ID NO.3; The nucleotide sequence of group_2289 is shown in SEQ ID NO.4; The nucleotide sequence of group_2329 is shown in SEQ ID NO.5; The nucleotide sequence of group_2792 is shown in SEQ ID NO.6; The nucleotide sequence of group_2950 is shown in SEQ ID NO.7; The nucleotide sequence of group_2964 is shown in SEQ ID NO.8; The nucleotide sequence of group_3007 is shown in SEQ ID NO.9; The nucleotide sequence of group_3452 is shown in SEQ ID NO.10; The nucleotide sequence of group_3633 is shown in SEQ ID NO.11; The nucleotide sequence of group_3741 is shown in SEQ ID NO.12; The nucleotide sequence of group_3839 is shown in SEQ ID NO.

13.

7. A probiotic phenotype prediction model constructed according to the construction method of claim 1.

8. A lightweight optimized prediction model constructed according to the construction method of claim 5.

9. A method for screening candidate functional strains of probiotics, characterized in that, Includes the following steps: (1) Obtain the whole genome sequence file of the probiotic to be tested and annotate it; (2) Extract k-mer sequence composition features and generate corresponding pan-genome homology features through homology comparison; (3) Match and fuse the k-mer sequence composition features with the pan-genome homology features to form a fusion feature vector of the strain to be tested; or, extract the core genomic features determined by the method described in claim 5 to form a dimensionality-reduced core feature vector. (4) Input the fused feature vector into the probiotic phenotype prediction model of claim 1 or input the core feature vector into the lightweight optimized prediction model of claim 5; (5) Output the predicted probability and / or classification result of the probiotic to be tested belonging to the positive category under a specific phenotype; (6) Compare the predicted probability with the preset classification threshold, and classify, screen and output the candidate functional strains.

10. A probiotic probiotic function prediction system based on multidimensional feature fusion and stacked ensemble learning, characterized in that, include: (1) Input module: used to receive the whole genome sequence file of the probiotic to be tested and to perform genome sequence annotation; (2) Feature acquisition module: used to extract k-mer sequence composition features based on the whole genome sequence of the probiotic to be tested, generate corresponding pan-genome homology features, and match and fuse the k-mer sequence composition features with the pan-genome homology features to generate a fused feature vector; (3) Model calling module: used to call the probiotic phenotype prediction model of claim 1 obtained by training, or to call the lightweight optimized prediction model of claim 5 constructed based on the core genomic features; (4) Result output module: used to output the predicted probability and classification result of phenotypic category.

Images