Biomarker screening method based on broad-target metabolomics and machine learning and its selected chronic kidney disease biomarker panel and application

By combining broad-target metabolomics and machine learning methods, the problem of difficulty in screening biomarkers in traditional technologies is solved, and 7 biomarkers of chronic kidney disease are screened out, which can achieve more efficient disease diagnosis and simplify the data analysis process. Improves the accuracy and adaptability of the diagnosis.

CN117825480BActive Publication Date: 2024-11-08WUHAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311825227.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-27
Publication Date
2024-11-08
Estimated Expiration
2043-12-27

AI Technical Summary

Technical Problem

Traditional metabolomics technology is difficult to accurately screen out biomarkers from high-throughput, large-scale metabolites data, resulting in limited application value for early diagnosis of diseases, especially in the field of chronic kidney disease.

Method used

Using biomarker screening methods based on broad-target metabolomics and machine learning, machine learning processes are combined with feature reconstruction, feature screening and model construction by building a local library of standards, non-targeted and targeted metabolomics technologies, Biomarkers of chronic kidney disease were screened out.

Benefits of technology

This method can automatically and concisely extract key information from massive data, quickly screen out biomarkers, build a model with high predictive ability, strong adaptability and generalization ability, and is suitable for biomarker screening for chronic kidney disease, reducing the number of biomarkers that have been found in chronic kidney disease. Manual intervention improves the accuracy and flexibility of diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117825480B_ABST
    Figure CN117825480B_ABST
Patent Text Reader

Abstract

The present invention discloses a biomarker screening method based on broad-target metabolomics and machine learning, and its selected chronic kidney disease biomarker group and application. The present invention is used to screen biomarkers and construct an optimal prediction model. It uses machine learning methods to process high-throughput, high-dimensional, and high-correlation metabolite abundance matrices determined by metabolomics. Through the steps of feature reconstruction, feature screening, model construction, and model screening, efficient biomarker screening and construction of an optimal prediction model can be achieved. Using this method, 7 metabolites were screened out from the plasma of patients with chronic kidney disease as biomarkers, and a prediction model was constructed to distinguish patients with chronic kidney disease in different stages. The average area under the curve of the prediction model can reach 0.99. The present invention provides a streamlined operation tool for processing complex broad-target metabolomics data, which can efficiently screen out biomarkers and build a prediction model with excellent prediction performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of biomedical technology, and in particular to a biomarker screening method based on broad-target metabolomics and machine learning, and a selected chronic kidney disease biomarker group and application thereof. Background Art

[0002] The metabolome is a complete collection of small molecules in an organism, reflecting downstream changes in the genome, transcriptome, and proteome. Metabolomics analysis can identify metabolites that regulate biological processes and cell physiology and determine the overall state of an organism. Compared with genes and proteins, metabolites can reflect disease states and physiological characteristics of diseases in a more real-time and dynamic manner, which makes the metabolome have good applications in areas such as early disease diagnosis.

[0003] Metabolomics technology includes two strategies: non-targeted analysis and targeted analysis. Non-targeted metabolomics technology can detect all metabolites in a sample without bias and has a wide range of metabolite coverage. Targeted metabolomics technology has good specificity, high sensitivity and accurate quantitative capabilities. Broad-targeted metabolomics methods combine the advantages of non-targeted and targeted technologies, and can achieve large-scale qualitative and quantitative analysis of metabolites in biological samples.

[0004] However, metabolomics offline data has the characteristics of high throughput, huge types and numbers of metabolites, multicollinearity between metabolites, unclear distribution characteristics of original data, and orders of magnitude differences in metabolite abundance values. These characteristics make it difficult for traditional statistical methods based on data distribution characteristics and probability inference to extract valuable information and accurately screen biomarkers, which greatly limits the promotion and application value of metabolomics technology. Machine learning has great advantages in processing such complex data sets. It can quickly identify the rules and patterns implicit in the data distribution, extract important information, and build a prediction model with better prediction performance.

[0005] The present invention discloses a biomarker screening method based on broad-target metabolomics and machine learning, and uses the method to screen out biomarkers of chronic kidney disease. Summary of the invention

[0006] In order to solve the technical problem, the present invention provides a biomarker screening method based on broad-target metabolomics and machine learning and a selected chronic kidney disease biomarker group, that is, a method for screening biomarkers using broad-target metabolomics and machine learning and a chronic kidney disease biomarker group screened by the method. The machine learning process in the present invention is used to screen biomarkers from broad-target metabolomics raw data, and the operation process is simple; in addition, the present invention uses the above process to screen 7 metabolites as biomarkers of chronic kidney disease.

[0007] To achieve the above purpose, the technical solution provided by the present invention is as follows:

[0008] In the first aspect, the present invention provides a biomarker screening method based on broad-target metabolomics and machine learning, wherein the broad-target metabolomics technology includes: construction of a local library of standard products, non-targeted metabolomics based on time-of-flight mass spectrometry, and targeted metabolomics based on triple quadrupole mass spectrometry; the machine learning technology includes: feature reconstruction, feature screening, model construction, and model selection. The model construction includes 6 supervised machine learning algorithms: random forest, XGBoost, multi-layer perceptron, kernel support vector machine, linear support vector machine, and Logistc regression.

[0009] The application of the broad-target metabolomics process in the above method, the specific analysis steps are as follows:

[0010] Collection and processing of blood samples: 2 ml of fasting venous blood was collected using an EDTA anticoagulation blood collection tube, and serum was extracted after centrifugation (4 °C, 10 min, 12000 rpm) and placed in a -80 °C refrigerator; after all samples were collected, all blood samples were taken out, thawed on ice, vortexed for 10 s, and the supernatant (50 μL) was extracted into a centrifuge tube; 300 μL of 20% acetonitrile methanol internal standard extract was added; after centrifugation (4 °C, 10 min, 12000 rpm), 200 μL of supernatant was transferred and allowed to stand (-20 °C, 30 min); centrifuged again (4 °C, 3 min, 12000 rpm) and 100 μL of supernatant was transferred into the liner tube of the injection bottle for analysis on the machine.

[0011] Qualitative non-targeted metabolomics: ① Establish a project database. The data-dependent acquisition mode in ultra-high performance liquid chromatography tandem high-resolution mass spectrometry is used to analyze the compound information in the QC of the mixed sample, and compare it with the standard information in the local library (mass transition and retention time under the multiple reaction detection mode MRM). After removing duplicate compounds, the final MRM mass transition information is obtained and summarized into a project database; ② Identify the metabolites in the blood samples. A quadrupole time-of-flight mass spectrometer is used to collect MS / MS spectra in information-dependent acquisition mode. Analyst1.6.3 software (SCIEX) is used to process the raw data obtained from the non-targeted metabolomics experiment for chromatographic peak detection, extraction and alignment. When identifying metabolites, first use the local database for identification. The library contains a list of chemical standards and selected compounds. The comparison content includes accurate mass ( m / z ), MS / MS spectra and retention times; subsequently, further metabolite identification was performed based on accurate mass, isotope patterns and MS / MS spectra according to public databases such as mzCloud, METLIN, HMDB and KEGG.

[0012] Targeted metabolomics quantification: Relative quantitative analysis of metabolites in plasma samples was performed using a triple quadrupole linear ion trap mass spectrometer. Metabolites in the established project database were used for MRM, and the analysis was performed in positive and negative ion modes, respectively. The instrument was controlled by Analyst 1.6.3 software (SCIEX). The raw data of the experiment were processed using MultiQuant3.0.3 software (SCIEX), including chromatographic peak integration and correction, and the resulting peak area was calculated as the relative abundance of each metabolite in the sample. The repeatability of metabolite determination was evaluated by relative standard deviation (RSD). Mass spectral peaks with RSD greater than 30% were excluded. The final data set included metabolite name, abundance, and sample ID.

[0013] In a second aspect, the present invention provides an application of the above-mentioned biomarker screening method, wherein the application is a data analysis process, comprising the following steps:

[0014] Step S1: Use broad-target metabolomics technology to conduct qualitative and quantitative analysis of metabolites in blood samples; specifically include: S1.1: Collect and process blood samples; S1.2: Establish a project analysis database: Use high-resolution mass spectrometry to analyze metabolites in mixed sample QC, retain ion information, and establish a project database after deduplication with the local standard library; S1.3: Use triple quadrupole mass spectrometry to determine the metabolite content in all samples; S1.4: Use MultiQuant 3.0.3 software to process raw mass spectrometry data and obtain the metabolite abundance matrix.

[0015] Step S2: Use machine learning technology to analyze the metabolite abundance matrix, screen biomarkers and build a prediction model; specifically including: S2.1: Feature reconstruction, normal transformation and scaling of the metabolite abundance matrix; S2.2: Feature screening, using filtering and packaging methods to screen out the most important metabolites as biomarkers; S2.3: Model construction, using supervised learning methods, determining the best parameter combination of each model through grid search with cross-validation, and building 6 prediction models; S2.4: Model selection, determining the best model as the prediction model through the F1 score, AUC and DCA curve of the model on the test set or external validation set. The filtering method includes univariate analysis and variance analysis; the packaging method includes algorithms based on random forest models and recursive feature elimination; the supervised learning method includes 6 traditional machine learning algorithms: random forest, XGBoost, multi-layer perceptron, kernel support vector machine, linear support vector machine, and Logistc regression.

[0016] Step S3: Screening of biomarkers for chronic kidney disease. Specifically, it includes: S3.1: Collecting blood samples, selecting chronic kidney disease patients and healthy controls from the hospital as research subjects, using EDTA anticoagulation blood collection tubes to collect fasting venous blood from the research subjects, extracting the supernatant after centrifugation and freezing it for storage; S3.2: Processing plasma samples: After the samples are collected, take them out of the refrigerator, and process the blood samples according to the steps of "gradient thawing-adding extraction solution-centrifugation-extracting supernatant", and prepare blank controls and mixed samples for QC; S3.3: Conducting broad-target metabolomics analysis; S3.4: Using machine learning to screen biomarkers and build a prediction model.

[0017] In a third aspect, the present invention provides a chronic kidney disease biomarker panel, which is obtained based on the application of the above-mentioned screening method; the chronic kidney disease biomarker panel includes 7 blood metabolites: 5-aminoimidazole ribonucleotide, succinic acid, aldehyde lysine, 5-methylthioadenosine, 5-methoxyindoleacetic acid, deoxyribose-1-phosphate and 2-hydroxyethyl sulfonic acid.

[0018] In a fourth aspect, the present invention provides an artificial intelligence model training method, wherein the application of the machine learning process is an analysis method, and the analysis steps are as follows:

[0019] Dataset organization: First, exogenous compounds such as drugs and pollutants and their metabolites in the metabolite dataset were excluded, and only endogenous metabolites were retained; second, the endogenous metabolite abundance matrix was log-transformed to check for outliers.

[0020] Feature reconstruction: Minmax scaling is used to scale the abundance distribution value of each metabolite in the sample to between 0 and 1; each endogenous metabolite is defined as a feature, which is the feature column of the data set, and whether or not there is a disease or disease classification is defined as a label, which is the label column of the data set. Each row of the data set represents a sample.

[0021] Feature screening: The data set was randomly split into a training set and a test set at a ratio of 80% and 20%. First, feature screening based on univariate analysis was used, and the number of screened features was defined as 10; then feature screening based on the random forest model was used, and it was gradually iterated until the number of selected features was between 5 and 15; finally, the intersection of the two was taken as the biomarker group.

[0022] Model construction: With the biomarker group in the previous step as the feature and the disease condition as the label, the data set was randomly split into a training set and a test set at a ratio of 80% and 20%, and six supervised learning models were trained one by one using a grid search with cross-validation: Logistic regression, support vector machine, naive Bayes, integrated decision tree (random forest and XGBoost) and multi-layer perceptron. The optimal parameter combination of each model was obtained with accuracy as the indicator.

[0023] Model selection: Using R when label values ​​are continuous variables 2 Evaluate the predictive performance of each model; when the label value is a binary variable, use accuracy, AUC, and F1 score to compare the predictive ability of each model. When there is an external data set, perform external validation. When the label value is a multi-class variable, use macro-average and micro-average F1 scores, macro-average and micro-average AUC to compare the predictive ability of each model. Select the model with the best predictive ability as the final prediction model.

[0024] The advantages and beneficial effects of the present invention are as follows:

[0025] 1. The method of the present invention makes the process of broad-target metabolomics data analysis more concise. Traditional statistical methods are usually unable to handle high-throughput, large-scale, and complex metabolite abundance matrices. This method can automatically learn features and extract key information from massive data, especially in the capture and processing of nonlinear relationships, and can quickly and easily screen out biomarkers.

[0026] 2. The method of the present invention makes the process of broad-target metabolomics data analysis more automated. The method of the present invention does not require prior knowledge and does not need to process thousands of metabolite information one by one, reducing the need for manual intervention, which is very important for handling large-scale tasks and real-time applications. Traditional statistics need to rely on pre-selected metabolites and cannot simultaneously process complex data sets with high dimensions and high correlations.

[0027] 3. The biomarkers screened and the prediction models constructed by the method of the present invention have stronger predictive ability, adaptability and generalization ability. Compared with traditional statistical analysis methods, they can perform better when facing new data sets and are more flexible when dealing with new problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the analysis process of the screening method of the present invention.

[0029] Figure 2 These are the 7 biomarkers of chronic kidney disease and their structural formulas in the present invention.

[0030] Figure 3 This is the process of using this method to screen chronic kidney disease biomarkers and the prediction model established in the embodiments of the present invention. DETAILED DESCRIPTION

[0031] The following will use specific examples in conjunction with the accompanying drawings to describe in detail the process of screening biomarkers for chronic kidney disease using the method of the present invention. Unless otherwise specified, the various raw materials, reagents, instruments and equipment used in the present invention can be purchased from the market or can be prepared by existing methods.

[0032] Example 1 Using a broad-target metabolomics approach to obtain plasma metabolomics data from patients with chronic kidney disease at different stages

[0033] First, blood samples were collected. A total of 62 patients with chronic kidney disease (22 patients with stage 1-3, 20 patients with stage 4, and 20 patients with stage 5) were selected from the Department of Nephrology of a university hospital, and 21 healthy controls were selected from the Department of Physical Examination. 2 ml of fasting venous blood was collected from the subjects using EDTA anticoagulant blood collection tubes, and the supernatant was extracted after centrifugation and frozen in a -80 °C refrigerator for storage.

[0034] Second, process the plasma samples. After all 83 samples were collected, they were taken out of the refrigerator and thawed on ice (all subsequent operations were performed on ice). After vortexing for 10 s, the supernatant (50 μL) was extracted into a centrifuge tube; 300 μL of 20% acetonitrile methanol internal standard extract was added; after centrifugation (4 °C, 10 min, 12000 rpm), 200 μL of supernatant was transferred to stand (-20 °C refrigerator, 30 min); centrifugation (4 °C, 3 min, 12000 rpm) was performed again, and 180 μL of supernatant was transferred to the inner liner of the injection bottle for analysis on the machine. Blank control and mixed sample quality control (QC) samples were also prepared. The first injection was the blank control, followed by the actual sample, and one injection of QC was injected every 10 injections of the actual sample.

[0035] Third, a broad-target metabolomics analysis was performed. The local library already had the mass transitions and retention times of 2,800 standard products under MRM. The data-dependent acquisition mode of ultra-high performance liquid chromatography tandem high-resolution mass spectrometry was used (chromatographic instrument: ExionLC AD, chromatographic column: Waters ACQUITY UPLC HSS T3 C18 column (2.1 mm×100 mm, 1.8μm), mass spectrometry: TripleTOF TM 6600, SCIEX,) analyzed the compound information in QC, compared it with the local library, removed duplicate compounds, and added MRM mass transition and retention time information of 315 compounds. A total of 3115 MRM mass transitions (1910 in positive ion mode and 1205 in negative ion mode) were obtained and summarized into the project database; ② Identify the metabolites in the blood samples. A quadrupole time-of-flight (QTOF) mass spectrometer was used to acquire MS / MS spectra in information-dependent acquisition mode (IDA). SCIEX's Analyst1.6.3 software was used to process the raw data obtained from the non-targeted metabolomics experiment for chromatographic peak detection, extraction and alignment. When identifying metabolites, the local database was first used for identification. The library contains chemical standards and a list of selected compounds. The comparison content includes accurate mass ( m / z), MS / MS spectra and retention time; then, further metabolite identification was performed based on accurate mass, isotope pattern and MS / MS spectrum according to public databases such as mzCloud, METLIN, HMDB and KEGG. Finally, the metabolite abundance matrix was obtained. A total of 1431 metabolite information was obtained, of which 401 metabolites were primary identifications and 174 metabolites were secondary identifications; among the 1431 metabolites, 539 compounds were endogenous metabolites, and their abundance information was retained as the final endogenous metabolite abundance matrix.

[0036] Example 2 Using machine learning to screen biomarkers and build a prediction model.

[0037] First, the dataset was organized. The abundance matrix of 539 endogenous metabolites was log-transformed, and no outliers were found after inspection.

[0038] Second, feature engineering. Minmax scaling was used to scale the abundance distribution values ​​of each metabolite in 83 samples to between 0 and 1. The data set was sorted, with a total of 541 columns, of which 539 metabolites were feature columns, chronic kidney disease stages were label columns, and blood sample IDs were sample ID columns. Each row of the data set represents a blood sample.

[0039] Third, feature screening. The data set was randomly split into a training set and a test set at a ratio of 80% and 20%. Feature screening based on univariate analysis was used, and the number of screened features was defined as 10; then feature screening based on the random forest model was used, iterated 3 times, and the number of selected features was 12; the intersection of the univariate screening results and the random forest screening results was taken as the biomarker group. The seven plasma biomarkers screened were: 5-Aminoimidazole ribonucleotide (C8H14N3O7P), sciadonic acid (C20H34O2), allysine (C6H11NO3), 5'-Methylthioadenosine (C11H15N5O3S), 5-Methoxyindole-acetate (C11H11NO3), deoxyribose 1-phosphate (C5H11O7P), and 2-Hydroxyethanesulfonate (C2H6O4S).

[0040] Fourth, model construction. Using 7 metabolites as features and chronic kidney disease stages as labels, the data set was randomly split into training and test sets at a ratio of 80% and 20%, and six supervised learning models were trained one by one using grid search with cross-validation: Logistic regression, support vector machine, naive Bayes, integrated decision tree (random forest and XGBoost), and multi-layer perceptron. Using accuracy as an indicator, the optimal parameter combination of each model was obtained. Fifth, model selection. The predictive ability of each model was compared by F1 score, among which random forest had the best predictive efficiency (see Table 1 and Figure 3 ), the model can accurately predict the healthy control group and patients with chronic kidney disease stage 5, with an F1 score of 1.00, and can correctly distinguish all the control groups and patients with chronic kidney disease stage 5 in the test set. The model made one error in distinguishing between chronic kidney disease stage 1-3 and chronic kidney disease stage 4, predicting patients with chronic kidney disease stage 1-3 as chronic kidney disease stage 4. The F1 score for chronic kidney disease stage 4 was low, at 0.89, but the overall accuracy of the model was the highest, at 0.95. Therefore, it was determined as the final prediction model.

[0041] Table 1. F1 scores of random forest models based on 7 biomarkers.

[0042]

Claims

1. A chronic kidney disease biomarker panel, characterized in that: The chronic kidney disease biomarker panel includes seven blood metabolites: 5-aminoimidazole ribonucleotide, succinic acid, aldehyde lysine, 5-methylthioadenosine, 5-methoxyindoleacetic acid, deoxyribose-1-phosphate, and 2-hydroxyethylsulfonic acid; The chronic kidney disease biomarker panel is obtained based on the following screening method application; the application is a data analysis process, comprising the following steps: Step S1: Use broad-target metabolomics technology to conduct qualitative and quantitative analysis of metabolites in blood samples, specifically: S1.1: Collect and process blood samples; S1.2: Establish a project analysis database: Use high-resolution mass spectrometry to analyze metabolites in mixed sample QC, retain ion information, and establish a project database after deduplication with the local standard library; S1.3: Use triple quadrupole mass spectrometry to determine the metabolite content in all samples; S1.4: Use MultiQuant 3.0.3 software to process raw mass spectrometry data and obtain the metabolite abundance matrix; Step S2: Use machine learning technology to analyze the metabolite abundance matrix, screen biomarkers and build a prediction model, specifically: S2.1: Feature reconstruction, normal transformation and scaling of metabolite abundance matrix; S2.2: Feature screening, filtering and packaging methods are used to screen out the most important metabolites as biomarkers; S2.3: Model construction, supervised learning method is used to determine the best parameter combination of each model through grid search with cross-validation, and 6 prediction models are constructed; S2.4: Model selection, the best model is determined as the prediction model through the F1 score, AUC and DCA curve of the model on the test set or external validation set; Step S3: Screening chronic kidney disease biomarkers, specifically: S3.1: Collect blood samples. Select chronic kidney disease patients and healthy controls from the hospital as research subjects. Use EDTA anticoagulant blood collection tubes to collect fasting venous blood from the research subjects. Extract the supernatant after centrifugation and freeze it for storage; S3.2: Process plasma samples: After the samples are collected, take them out of the refrigerator and process the blood samples according to the steps of "gradient thawing-adding extraction solution-centrifugation-extracting supernatant". Prepare blank controls and mixed samples for QC; S3.3: Conduct broad-target metabolomics analysis; S3.4: Use machine learning to screen biomarkers and build prediction models.

2. The marker panel according to claim 1, characterized in that: In step S2: the filtering method includes univariate analysis and variance analysis; the packaging method includes an algorithm based on a random forest model and recursive feature elimination; the supervised learning method includes six traditional machine learning algorithms: random forest, XGBoost, multilayer perceptron, kernel support vector machine, linear support vector machine, and Logistc regression.