A machine learning-assisted non-targeted metabolomics method for screening characteristic metabolites that inhibit vomitoxin synthesis

Abstract

This invention discloses a machine learning-assisted non-targeted metabolomics method for screening characteristic metabolites that inhibit vomitoxin synthesis. Starting from the perspective of machine learning-assisted non-targeted metabolomics screening of biomarkers, this invention utilizes a combined model of random forest and neural networks to overcome the "black box" problem of deep learning. In five-fold cross-validation, the model with the highest average classification accuracy is selected to identify metabolites that appear more than four times in the top 12 most important classification features. Then, through metabolic pathway enrichment analysis, key metabolites in the pathways are identified, narrowing down the range of biomarkers. Finally, exogenous addition validation is performed to screen metabolites that can inhibit vomitoxin synthesis. This invention can be applied to the efficient identification and control of pathogenic microorganisms and their toxins in the food industry, animal husbandry, and clinical settings.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of analytical chemistry, food engineering technology and bioinformatics, and in particular relates to a novel method for screening characteristic metabolites that inhibit the synthesis of vomitoxin by integrating machine learning and metabolic pathway information for non-targeted metabolomics. This method can be applied to the efficient identification and control of emetic Bacillus cereus and its toxins in the food industry, aquaculture, and clinical settings. Background Technology

[0002] Bacillus cereus is an important foodborne pathogen whose spores are highly resilient. Commonly found in soil, dust, and sediment, Bacillus cereus frequently contaminates a variety of foods, particularly improperly stored grains, rice, and pasta. In the European Union, a significant portion of food and waterborne outbreaks are related to bacterial toxins, particularly those produced by Clostridium, Staphylococcus, and Bacillus cereus. The vomiting syndrome caused by Bacillus cereus is triggered by vomitoxin, a highly bioactive toxin pre-synthesized in the food matrix that is very stable to heat, acid, and digestive enzymes and is not neutralized by heat treatment or exposure to gastric acid. Clinically, this syndrome is characterized by nausea and vomiting, typically appearing 0.5 to 6 hours after ingestion of contaminated food. In severe cases of poisoning, there is a risk of acute liver failure, which can be fatal. Therefore, the control of emetic Bacillus cereus and its toxins is crucial.

[0003] The identification of Bacillus cereus requires not only efficient and accurate identification at the species level but also rapid assessment of the strain's toxin-producing capacity, highlighting the necessity of innovative strategies. Several taxonomic models specifically designed to assess potential risks have been established, enabling the prediction of aflatoxin contamination risk 35-47 days in advance using early biomarkers. Therefore, screening for highly specific and reliable biomarkers is crucial for the early prediction and assessment of pathogenic microbial contamination and its toxins. Currently, biomarker screening methods for non-targeted metabolomics data mainly focus on univariate analysis and multivariate statistical learning methods. With the development of machine learning, deep learning has become the preferred method due to its advantages in handling large-scale and complex datasets, especially when extracting features directly from raw data. This invention proposes a novel biomarker screening method using a combined model of random forests and neural networks, and provides a new approach to controlling pathogenic microorganisms and their toxins by regulating metabolic flux through functional biomarkers. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method for screening characteristic metabolites that inhibit the synthesis of vomitoxin by integrating machine learning and metabolic pathway information, and to apply it to the identification and control of pathogenic microorganisms and their toxins in the food industry, aquaculture, and animal husbandry.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method integrating machine learning and metabolic pathway information to assist in non-targeted metabolomics screening for characteristic metabolites that inhibit vomitoxin synthesis includes the following steps:

[0007] (1) Sample processing and acquisition of non-targeted metabolomics data

[0008] The bacterial cultures of Bacillus cereus and Bacillus cereus grown to the mid-log phase were centrifuged, the supernatant was discarded and the bacterial cells were collected. The bacterial cells were washed with PBS and the metabolic reaction was quickly quenched with liquid nitrogen. Cold extraction solution was added, and the cells were ground or sonicated. The supernatant was collected for UPLC-MS / MS detection for data acquisition. The quality control sample was prepared by mixing the supernatants obtained from Bacillus cereus and Bacillus cereus in equal volumes.

[0009] (2) Preprocessing of non-targeted metabolomics data

[0010] The acquired UPLC-MS / MS data underwent preprocessing steps including deconvolution, peak alignment, redundancy removal, standardization, and normalization. Metabolites were annotated based on secondary mass spectrometry characteristic ions. Metabolite identification was performed using external and local database retrieval tools. Metabolites that could be qualitatively compared in more than three external databases or scored greater than 70 points in the local database were selected for training the deep learning model and screening biomarkers.

[0011] (3) Training of convolutional neural network models and screening of differential metabolites

[0012] The metabolites selected in step (2) were trained using a combined random forest and convolutional neural network model. The model ranked the metabolites according to the importance scores of all samples using the random forest algorithm. Then, metabolites with different counts were introduced into the convolutional neural network model according to the ranking for the classification of Bacillus cereus and Bacillus cereus vomiting. Five-fold cross-validation was used to evaluate the prediction accuracy of the convolutional neural network models with different metabolite counts, and the model with the highest average classification prediction accuracy was selected. Finally, the top 12 metabolites in terms of importance and repeated more than four times in the five-fold cross-validation were selected as biomarker candidates.

[0013] (4) Metabolic pathway enrichment analysis

[0014] The selected biomarker candidates were subjected to metabolic pathway enrichment analysis to identify key metabolites in the enrichment pathways and obtain biomarkers to be validated.

[0015] (5) Exogenous addition to validate the function of biomarkers

[0016] A certain amount of the biomarker to be validated was added to the culture medium and cultured with emetic Bacillus cereus. Then, the vomitoxin was measured, with the group without the added biomarker as a control. Finally, the biomarker that can inhibit the synthesis of vomitoxin was obtained.

[0017] Furthermore, in step (1), the cold extraction solution is 100% methanol, a mixture of methanol / water at a volume ratio of 8:2, or a mixture of methanol / acetonitrile / water at a volume ratio of 4:4:2.

[0018] Furthermore, in step (1), the UPLC-MS / MS detection conditions are as follows:

[0019] Chromatographic conditions: The column was a Thermo Syncronis HILIC column, 100 × 2.1 mm, 1.7 μm; mobile phase A was purified water containing 5 mM ammonium formate and 0.1% formic acid; mobile phase B was acetonitrile; the flow rate was 0.30 mL / min; the injection volume was 4 μL; and the column temperature was 30 °C. Elution method: gradient elution, with the following program: 0–5 min, mobile phase A volume fraction 5%; 5–18 min, mobile phase A volume fraction changing from 5% to 50%; 18–19 min, mobile phase A volume fraction changing from 50% to 5%; 19–22 min, mobile phase A volume fraction 5%.

[0020] Mass spectrometry conditions: The spray voltage of the heated electrospray ion source was set to 3 kV, and the capillary temperature was 320℃; positive ion mode was selected for mass spectrometry scanning, the primary mass spectrometry resolution was 35000, the mass scan range was 70-1000 m / z, and the secondary mass spectrometry was obtained by NCE gradient energy collision with the parent ion with collision energies of 20, 40, and 60 eV, the secondary mass spectrometry resolution was 17500, and the automatic gain control was set to 50000.

[0021] Furthermore, the data preprocessing software in step (2) includes XCMS, MS-DIAL, MZmine 2, and CompoundDiscorvery.

[0022] Furthermore, in step (2), the external database includes mzCloud. TM Chemspider TM KEGG and BioCyc; the local database retrieval tools include mzVault.TM Spectral library and quality list.

[0023] Furthermore, in steps (2) and (3), to overcome the "black box" problem of deep learning models, this invention utilizes a combined model of random forest (RF) and deep learning to select the features that contribute the most to classification. In practical applications, depending on the composition and differences of the data, the deep learning model can be a convolutional neural network (CNN), a residual network (ResNet), a recurrent neural network (RNN), a transformer, etc. Preferably, a combined model of random forest and convolutional neural network is used.

[0024] Furthermore, the key steps in data preprocessing in step (2) include data cleaning, removal of confounding factors, QC normalization, and internal standard normalization. By combining data preprocessing and noise reduction algorithms with various data filtering strategies or considering the loss of outliers, data cleaning is performed to remove confounding factors such as pollutants in the MS source, mobile phase impurities, and differences in sample matrix complexity, effectively solving the noise problem and improving the overall data quality.

[0025] Furthermore, in step (5), the culture medium includes, but is not limited to, R2A culture medium, as well as nutrient-rich culture media such as LB and TSB, with R2A being preferred, and the culture time being 12-74h, preferably 24h.

[0026] Furthermore, in step (5), biomarkers that can inhibit the synthesis of vomitoxin are screened to obtain amino acids and vitamins. Preferably, the vitamin is biotin and the amino acid is L-leucine. The substance that inhibits the synthesis of vomitoxin can be extended to other key metabolites of the same metabolic pathway.

[0027] The present invention also provides the application of the above-described method in the efficient identification and control of pathogenic microorganisms and their toxins in the food industry, aquaculture, and clinical settings.

[0028] Furthermore, the pathogenic microorganism and its toxin are emetic Bacillus cereus and its vomitoxin.

[0029] The present invention has the following beneficial effects:

[0030] (1) The screening model provided in this invention overcomes the "black box" problem of deep learning models in biological screening and can provide biomarker candidates with higher specificity than those obtained by traditional univariate analysis and multivariate statistical analysis.

[0031] (2) The screening method in this invention integrates biological pathway information, improves the predictive ability of biomarkers, and can be used to reveal complex biological relationships, thereby improving the efficiency of identification and verification of reliable biomarkers.

[0032] (3) The exogenous addition of biomarkers in this invention can significantly inhibit the synthesis of vomitoxin in emetic Bacillus cereus by changing the metabolic flow, providing a new method for controlling vomitoxin in emetic Bacillus cereus in the food industry, aquaculture or clinical practice.

[0033] (4) The biomarkers verified in this invention can also distinguish between Bacillus cereus and vomiting Bacillus cereus, and are used for the identification of vomiting Bacillus cereus and its vomiting toxin. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the screening method of the present invention.

[0035] Figure 2 These are the biomarker candidates screened by this invention.

[0036] Figure 3 This is a pathway enrichment analysis of the biomarker candidates screened in this invention.

[0037] Figure 4 Functional verification of exogenous biomarkers in this invention. Detailed Implementation

[0038] To more concisely and clearly demonstrate the technical solution, purpose, and advantages of the present invention, the technical solution of the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0039] The *Bacillus cereus* and *Bacillus cereus* var. *emeticus* used in this invention to verify the feasibility of the technical solution were isolated by our laboratory in previous work. The UPLC-Q Exactive Focus MS used in this invention comes from the omics analysis platform of the State Key Laboratory of Applied Microbiology, South China. The deep learning models in this invention are all trained using Python, and the deep learning platforms involved include the two mainstream learning frameworks, Tensorflow and Keras.

[0040] The LB medium, R2A medium, sample vials, inner tubes, syringes, and 0.22 μm microporous membranes used in this invention were purchased from Huankai Company; LC-MS grade acetonitrile and methanol reagents were purchased from Merck (Darmstadt, Germany); HPLC grade formic acid (FA) was purchased from Sigma-Aldrich (USA); PBS buffer was purchased from Beijing Solarbio Science & Technology Co., Ltd.; and the internal standard L-2-chlorophenylalanine (Yuanye Biotechnology) was purchased from Baiyingli Chuang Biotechnology (Beijing) Co., Ltd.

[0041] Example 1: Non-targeted metabolomics analysis of Bacillus cereus and emetic Bacillus cereus

[0042] All bacterial strains were transferred from glycerol tubes to LB agar and cultured at 37°C for 20 h. Single colonies were then picked and cultured in LB liquid medium at 37°C and 200 rpm for 16 h. 100 μL of the bacterial culture was then transferred to 15 mL of fresh LB liquid medium and cultured at 37°C and 200 rpm until mid-log phase (OD). 600 When the concentration reaches 0.6±0.05, take 10 mL of bacterial culture, centrifuge at 4000g for 10 min, discard the supernatant, wash the bacterial cells once with PBS, discard the supernatant, and then quickly quench with liquid nitrogen and store at -80℃.

[0043] Resuspend the bacterial cells in 200 μL of methanol / water = 8:2 (v / v) solution, add 0.1500 ± 0.0020 g glass beads and grind for 10 min. Centrifuge at 15000 g for 10 min and collect the supernatant. Add 200 μL of methanol / acetonitrile / water solution = 4:4:2 (v / v) (containing 2 μg / mL L-2-chlorophenylalanine) to the precipitate and grind for 5 min. Centrifuge at 15000 g for 10 min and collect the supernatant, combining it with the supernatant from the previous step. Dry the collected supernatant in a vacuum concentrator and store at -80℃. Before loading, reconstitute the supernatant with 200 μL of acetonitrile / water = 8:2 (v / v) solution, centrifuge at 15000 g for 10 min, and transfer 100 μL to a sample vial. Take 10 μL of the solution from each sample and mix to prepare the quality control (QC) sample. The blank sample is a sample obtained without bacterial inoculation, with other treatment steps consistent with the experimental group. Each sample group underwent four independent biological replicates.

[0044] The prepared samples were analyzed by UPLC-MS / MS. Chromatographic conditions: The column was a Thermo Syncronis HILIC column (100 × 2.1 mm, 1.7 μm); mobile phase A was purified water (containing 5 mM ammonium formate and 0.1% formic acid); mobile phase B was acetonitrile; flow rate was 0.30 mL / min; injection volume was 4 μL; column temperature was 30℃; gradient elution was used, and the program is shown in Table 1. Mass spectrometry conditions: The spray voltage of the heat electron spray ionization (HESI) source was set to 3 kV, and the capillary temperature was 320℃. Positive ion mode was selected for mass spectrometry. The primary mass spectrometry resolution was 35000, and the mass scan range was 70-1000 m / z. Secondary mass spectrometry was obtained by NCE gradient energy collisions with the precursor ion at collision energies of 20, 40, and 60 eV, with a secondary mass spectrometry resolution of 17500. Automatic gain control was set to 50000. During sample injection, the QC sample is first injected into the first 8 needles, and then the samples are injected randomly to avoid systematic bias. A QC sample is injected every 8 samples to correct for instrument fluctuations.

[0045] Table 1 Elution gradient of mobile phase

[0046]

[0047]

[0048] Example 2: Data Preprocessing and Training of Deep Learning Models

[0049] The acquired untargeted metabolomics data were processed using Compound Discoverer 3.1 for peak identification, extraction, alignment, and integration, followed by secondary mass spectrometry database matching and annotation. Multiple databases and spectral library search tools (including mzCloud) were also used. TM Chemspider TM KEGG and BioCyc) and local database retrieval tools (such as mzVault) TM Spectral libraries or quality lists are used for compound identification. Metabolites that can be qualitatively compared in more than three databases simultaneously or score greater than 70 in a local database are ultimately selected for training the deep learning model and screening biomarkers. The screening process is as follows: Figure 1 .

[0050] Eighteen samples from each of the two groups, *Bacillus cereus* and *Bacillus emeticus*, were analyzed, and 267 metabolites were detected. Eighty biofunctional metabolites were selected and trained using a combined random forest (RF) and convolutional neural network (CNN) model. This model ranked metabolites by importance score (default metric: average decrease in accuracy) across all samples using the RF algorithm. Subsequently, the top k (k = 12, 16, or 32) metabolites were introduced into the CNN model for *Bacillus cereus* and *Bacillus emeticus* classification. The predictive efficiency of the CNN model with different metabolite counts was evaluated using the area under the curve (AUC) of five-fold cross-validation. According to Table 2, the model with the top 12 metabolites achieved the highest average AUC. To enhance the specificity of biomarker identification, cross-analysis was performed on the top 12 metabolites ranked by importance in the five-fold cross-validation, revealing nine metabolites that appeared more than four times. Figure 2 A) lists the following metabolites: biotin, benzidine, L-leucine, histidine, 4-aminophenol, 2'-O-methylamide, 2-naphthaleneacetic acid, cis-4-hydroxy-D-proline, and 6-acetaminohexanoic acid, and provides the differences in these metabolites in univariate analysis. Figure 2 (B, 2C and 2D).

[0051] Table 2. Model evaluation for different numbers of features.

[0052]

[0053] Example 3: Metabolic pathway enrichment analysis and exogenous addition functional verification

[0054] Using paired box plots ( Figure 3 A) To assess changes in nine biomarkers (i.e., the nine metabolites obtained in Example 2) between *Bacillus cereus* and *Bacillus emeticus*. Although the distribution of these biomarkers was comparable between the two groups, their metabolite expression was significantly opposite, suggesting that these metabolites may be associated with the production of vomitoxin. Metabolic pathway enrichment analysis of these nine metabolites was performed, and the results are as follows: Figure 3 As shown in B, these metabolites are enriched in four key metabolic pathways, including histidine metabolism, biotin metabolism, biosynthesis of valine, leucine, and isoleucine, and degradation of valine, leucine, and isoleucine. Metabolic pathway enrichment analysis narrowed the biomarker pool to L-leucine, histidine, and biotin.

[0055] To further verify the effects of L-leucine, histidine, and biotin on the synthesis of vomitoxin, 14 sets of exogenous addition validation experiments were conducted, including the addition of L-leucine (0.1mM, 0.3mM, 1mM, 3mM), histidine (0.1mM, 0.3mM, 1mM, 3mM), biotin (0.4μM, 4μM), leucine (0.3mM) + biotin (4μM), leucine (3mM) + biotin (4μM), histidine (0.3mM) + biotin (4μM), and histidine (3mM) + biotin (4μM). A 100-fold stock solution of the target metabolite was prepared. 1 mL of the stock solution was mixed with 99 mL of R2A medium, and *Bacillus cereus* var. *emeticus* was cultured for 24 h, followed by extraction and purification. Vomitoxin was detected by LC-MS / MS, with three biological replicates for each sample group.

[0056] Quantitative analysis of vomitoxin was performed using a C18 column (Acquity). Peptide BEH (1.7 μm, 2.1 × 100 mm, Waters). Ion source parameters included a capillary voltage of 5.5 kV, an ion source temperature of 550 °C, and a collision gas pressure of 9 psi. The column temperature was 40 °C, and the injection volume was 5 μL. Analysis was performed in MRM mode with argon as the collision gas. The mobile phase gradient consisted of a 10 mM ammonium formate aqueous solution containing 0.1% formic acid (solution A) and acetonitrile containing 0.1% formic acid (solution B), starting at 70% in solution B; linearly increasing to 90% within 3 min, remaining stable for 5 min; then decreasing to 70% within 0.1 min, and finally stabilizing after a 2-min settling period. Accurate identification and quantification were achieved using ion pairs 1170.70→172.30 and 1128.60→343.50, with the precursor ion detected in MS1 ​​being [M+NH4+]. + Vomitoxin standard solutions (acetonitrile solutions with concentrations ranging from 10 to 10,000 ng / mL, each solution containing 100 ng / mL valamicin as an internal standard) were used to establish the calibration curve.

[0057] The final verification results showed that biotin and L-leucine can inhibit the synthesis of vomiting toxins. Figure 4 ).

[0058] The above are merely preferred embodiments of the present invention. It should be noted that the above preferred embodiments should not be considered as limitations on the present invention, and the scope of protection of the present invention should be determined by the scope defined in the claims. For those skilled in the art, several improvements and modifications can be made without departing from the spirit and scope of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method of integrating machine learning and metabolic pathway information to aid non-targeted metabolomics screening for characteristic metabolites that inhibit the synthesis of emetic toxins, characterized in that, Comprising the following steps: (1) Sample processing and non-targeted metabolomics data acquisition Centrifuge the bacterial liquid of B. cereus and B. cereus emetic grown to the mid-log phase, discard the supernatant and collect the bacterial cells, wash the bacterial cells with PBS, then rapidly quench the metabolic reaction with liquid nitrogen, add cold extraction solution, grind or ultrasonically break, and then collect the supernatant for UPLC-MS / MS detection for data acquisition; The quality control sample is prepared by mixing the supernatant obtained from B. cereus and the supernatant obtained from B. cereus emetic in equal volumes; (2) Pretreatment of non-targeted metabolomics data The obtained UPLC-MS / MS data is pretreated by steps of deconvolution, peak alignment, redundancy removal, standardization, and normalization, and the metabolites are annotated according to the characteristic ions of the secondary mass spectrum; metabolite identification is performed by applying external database and local database retrieval tools, and metabolites that can be simultaneously aligned in more than three external databases or have a score greater than 70 in the local database are selected for training of a deep learning model and screening of biomarkers; (3) Training of convolutional neural network model and screening of differential metabolites The metabolites selected in step (2) are trained by a combined model of random forest and convolutional neural network, the model sorts the metabolites according to the importance scores of all samples by random forest algorithm, then different counts of metabolites are introduced into the convolutional neural network model according to the sorting for classification of B. cereus and B. cereus emetic; the prediction accuracy of the convolutional neural network model with different metabolite counts is evaluated using five-fold cross-validation, and the model with the highest average classification prediction accuracy is selected; finally, the metabolites that rank first in importance and appear more than four times in the five-fold cross-validation are selected as biomarker candidates; (4) Metabolic pathway enrichment analysis The biomarker candidates obtained by screening are subjected to metabolic pathway enrichment analysis to lock the key metabolites of the enriched pathways and obtain the biomarkers to be verified; (5) Exogenous addition verification of biomarker function A certain amount of biomarker to be verified is added to the culture medium and cultured with B. cereus emetic, then emetic toxin is determined, and the group without adding the biomarker to be verified is used as a control, and finally the biomarker that can inhibit the synthesis of emetic toxin is obtained.

2. The method of claim 1, wherein, In step (1), the cold extraction solution is 100% methanol, a mixture of methanol / water at a volume ratio of 8:2, or a mixture of methanol / acetonitrile / water at a volume ratio of 4:4:

2.

3. The method of claim 1, wherein, In step (1), the UPLC-MS / MS detection conditions are: Chromatographic conditions: the chromatographic column was Thermo Syncronis HILIC chromatographic column 100 × 2.1 mm, 1.7 μm, mobile phase A was purified water containing 5 mM ammonium formate and 0.1% formic acid, mobile phase B was acetonitrile, the flow rate was 0.30 mL / min, the injection volume was 4 μL, and the column temperature was 30℃; the elution mode was gradient elution, and the program was as follows: 0-5 min, the volume fraction of mobile phase A was 5%; 5-18 min, the volume fraction of mobile phase A changed from 5% to 50%; 18-19 min, the volume fraction of mobile phase A changed from 50% to 5%; 19-22 min, the volume fraction of mobile phase A was 5%. Mass spectrometric conditions: the spray voltage of the heated electrospray ion source was set to 3 kV, and the capillary temperature was 320℃; the mass spectrometric scanning was selected in positive ion mode, the first mass spectrometric resolution was 35000, the mass scanning range was 70-1000 m / z, the secondary mass spectrometric collision energy was obtained by using NCE gradient energy collision of parent ions, the collision energy was 20, 40, 60 eV, the secondary mass spectrometric resolution was 17500, and the automatic gain control was set to 50000.

4. The method of claim 1, wherein, The data preprocessing software in step (2) includes XCMS, MS-DIAL, MZmine 2 and Compound Discorvery.

5. The method of claim 1, wherein, In step (2), the external database includes mzCloud™, Chemspider™, KEGG and BioCyc; and the local database retrieval tool includes mzVault™ spectrum library and mass list.

6. The method of claim 1, wherein, In step (2), the deep learning model includes convolutional neural network CNN, residual network Resnet, recurrent neural network RNN and transformer Transformer.

7. The method of claim 1, wherein, In step (5), the biomarkers capable of inhibiting the synthesis of vomitoxin include amino acids and vitamins.

8. The method according to any one of claims 1-7, for use in the food industry, the breeding industry, and the efficient identification and prevention of C. emerosus and its vomitoxin.

Images