Screening method of tumor metabolism markers

Abstract

The invention belongs to the technical field of biological medicine, and particularly relates to a screening method of tumor metabolism markers. According to the screening method, wide targeted metabonomics data and targeted metabonomics data are utilized and combined with a biomarker recognition and optimization algorithm (BIO-FIRE) based on function and feature importance recursion enhancement to screen tumor metabolic markers. According to the screening method, a stable metabolic marker combination can be identified in large-scale and multi-center research data, so that a set of efficient, accurate and generalizable gastric cancer metabonomics diagnosis scheme is constructed.

Description

Technical Field

[0001] The present invention relates to the field of biomedical technologies, and particularly to a method for screening a metabolite marker combination and its application in tumor diagnosis. Background Art

[0002] Metabolomics is generally divided into two research strategies: non-targeted and targeted. Targeted metabolomics focuses on the quantitative analysis of specific metabolites, with high sensitivity and specificity, and can be used for the quantitative verification of known markers and the translation of clinical applications. As a new generation of targeted metabolomics technology, widely targeted metabolomics combines the advantages of targeted metabolomics and non-targeted metabolomics detection technologies, and can achieve the quantification of hundreds of metabolites (see the non-patent literature A Novel Integrated Method for Large-Scale Detection, Identification, and Quantification of Widely Targeted Metabolites: Application in the Study of Rice Metabolomics. [J] Volume 6, Issue 6, November 2013, Pages 1769-1780).

[0003] In recent years, non-invasive cancer diagnosis methods based on plasma metabolomics have received extensive attention. Metabolites are the final products of the body's biochemical activities, and their dynamic changes can directly reflect the metabolic reprogramming of cancer cells, host-tumor interactions, and the occurrence and development of diseases. Compared with genes and proteins, the fluctuations of metabolites are more significant, and plasma metabolome analysis has the advantages of simple operation, relatively low cost, and good repeatability, so it has become an important research direction for non-invasive cancer diagnosis. However, existing metabolomics research still faces many challenges in tumor diagnosis applications. First, many studies are based on small patient cohorts and lack multi-center, large-sample validation, which affects the generalization ability of the model. Second, the detection technologies and data analysis processes used in different studies lack unified standards, and metabolomics data are easily affected by multiple factors such as sampling, detection platforms, and analysis methods, resulting in poor comparability between different studies. In addition, traditional biomarker screening strategies mainly focus on the statistical significance of individual metabolites, while ignoring the systematic role of metabolic functional modules, and the biomarkers selected have certain limitations in biological interpretation. Summary of the Invention

[0004] To solve the problems of the existing technology, this application combines unsupervised functional module analysis, feature importance screening, and recursive optimization strategies to provide a biomarker identification and optimization algorithm (Biomarker Identification and Optimization Algorithm via Functional and Importance-based Recursive Enhancement, BIO-FIRE). The algorithm includes three key steps: 1) functional module clustering and interpretability enhancement; 2) potential biomarker identification based on feature importance; 3) biomarker determination based on the principle of recursive optimization. Based on the combination of biological interpretability, statistical importance, and model simplicity, this algorithm forms an efficient and optimized biomarker screening process, which not only improves the biological significance of the screened biomarkers but also ensures their superiority in terms of statistics and model simplicity.

[0005] This application also establishes a large-scale, multi-center clinical research system, integrates broad-target and targeted metabolomics technologies, and develops a new algorithm (Biomarker Identification and Optimization Algorithm via Functional and Importance-based Recursive Enhancement, BIO-FIRE) based on biological functional modules and recursive optimization strategies to obtain a robust method for screening tumor metabolic biomarkers.

[0006] The specific solution is as follows:

[0007] In the first aspect of the present invention, a method for screening tumor metabolic biomarkers is provided, and the screening method includes:

[0008] S1) Screening and obtaining preliminary screening metabolites from the samples of the tumor patient group and the non-tumor control group, and performing qualitative and quantitative analysis;

[0009] S2) Using the biomarker identification and optimization algorithm (BIO-FIRE) based on functional and feature importance recursive enhancement to perform feature extraction, and further determining tumor metabolic biomarkers from the preliminary screening metabolites.

[0010] The BIO-FIRE algorithm includes:

[0011] A) Using the non-negative matrix factorization method (NMF) to perform functional module clustering on the preliminary screening metabolites;

[0012] B) Using the Boruta algorithm to perform feature screening by calculating importance scores;

[0013] C) Based on the principle of recursive optimization, sort the importance scores of each biomarker calculated in B) from high to low, and sequentially select and include them in the metabolite biomarker subset starting from the first one until a metabolite subset that can maximize the coverage of the most functional modules is selected, which is the tumor metabolite biomarker.

[0014] Preferably, the S1) includes constructing a tumor-specific metabolite database using wide-targeted metabolomics and determining differential metabolites, and further screening and quantifying the pre-screened metabolites using targeted metabolomics.

[0015] Preferably, the S1) includes:

[0016] 1) Constructing a tumor-specific metabolite database using wide-targeted metabolomics and determining differential metabolites; preferably, performing full-scan mode detection on the sample using a quadrupole time-of-flight mass spectrometer (QTOF) to obtain metabolite ion pair signals, and further transferring the ion pair signals of the metabolites to a tandem quadrupole composite linear ion trap mass spectrometer (QTRAP6500) to obtain a tumor-specific metabolite database and determine differential metabolites;

[0017] 2) Screening and quantifying the pre-screened metabolites using targeted metabolomics; preferably, the detection range of the targeted metabolomics includes the differential metabolites in step 1), and further preferably also includes clinically and literature-recorded tumor metabolite biomarkers; preferably, the literature-recorded tumor metabolite biomarkers include those with a literature reporting frequency greater than 5 or recognized tumor metabolite biomarkers. Preferably, liquid chromatography is used to separate the differential metabolites and mass spectrometry is used to quantify the pre-screened metabolites. The liquid chromatography uses a T3 column, and the mass spectrometry uses a T3 column.

[0018] Further preferably, the S1) includes: performing full-scan mode detection on the sample using a quadrupole time-of-flight mass spectrometer (QTOF) to obtain metabolite ion pair signals, and further transferring the ion pair signals of the metabolites to a tandem quadrupole composite linear ion trap mass spectrometer (QTRAP6500) to obtain a tumor-specific metabolite database and determine differential metabolites; then using liquid chromatography to separate the differential metabolites and mass spectrometry to quantify the pre-screened metabolites.

[0019] Preferably, the pre-screened metabolites include various substances such as amino acids and their derivatives, bile acids and their conjugates, purine / pyrimidine metabolites, steroid hormones, lipids (fatty acids, phospholipids, sphingolipids), biogenic amines, organic acids, sugars, flavonoids / alkaloids, etc., involving amino acid metabolism, bile acid metabolism, nucleotide metabolism, lipid metabolism, hormone metabolism, drug metabolism, and microbial metabolism.

[0020] In a specific embodiment of the present invention, the liquid chromatography and mass spectrometry conditions used in the wide-targeted metabolomics are:

[0021] Chromatographic column: Waters ACQUITY UPLC HSS T3 C18, column temperature was 40 °C;

[0022] Mass spectrometry conditions: Electrospray ionization (ESI) source temperature 500 °C, mass spectrometry voltage 5500 V (positive) or -4500 V (negative), ion source gas I (GS I) 55 psi, gas II (GS II) 60 psi, curtain gas (CUR) 25 psi, collision-activated dissociation (CAD) parameter was set to high.

[0023] In the triple quadrupole (Qtrap), each ion pair was scanned and detected in MRM mode according to the optimized declustering potential (DP) and collision energy (CE).

[0024] In a specific embodiment of the present invention, the liquid chromatography and mass spectrometry conditions used in the targeted metabolomics were as follows:

[0025] Chromatographic column: Waters ACQUITY UPLC HSS T3 C18, column temperature was 40 °C;.

[0026] Mass spectrometry conditions: Electrospray ionization (ESI) source was set at a temperature of 500 °C, mass spectrometry voltage 5500 V (positive), -4500 V (negative), ion source gas I (GS I) 55 psi, gas II (GS II) 60 psi, curtain gas (CUR) 25 psi, collision-activated dissociation (CAD) parameter was set to high. In the triple quadrupole (Qtrap), each ion pair was scanned and detected in MRM mode according to the optimized declustering potential (DP) and collision energy (CE).

[0027] In the screening method, before targeted metabolomics detection, it is preferably further included to perform linear discriminant analysis (LDA) grouping on the ion pair signals in the constructed tumor-specific metabolite database, and use the P value and FC value of the statistical model and univariate statistical T test to screen for differential metabolic signals. Preferably, FDR < 0.05, FC > 1.2 or FC < 0.80, VIP > 1 are set as the differential significance criteria.

[0028] Preferably, the grouping is to divide into a tumor patient group and a non-tumor control group.

[0029] More preferably, the tumor group includes different subtypes or different development stages of the same tumor.

[0030] Preferably, the BIO-FIRE algorithm includes:

[0031] A) Using NMF to perform functional module clustering on the initially screened metabolites, and naming the obtained modules according to the KEGG annotation information;

[0032] B) The Boruta algorithm performs feature screening on the functional module data extracted in A) by calculating the importance score to obtain potential metabolic markers;

[0033] C) Based on the principle of recursive optimization, sort the importance scores of each marker calculated in B) from high to low, and sequentially select and include them in the metabolic marker subset starting from the first one until a metabolite subset that can maximize the coverage of all functional modules covered by the potential metabolic markers in B) is selected, which is the tumor metabolic marker.

[0034] Preferably, the NMF introduces the KL divergence optimization strategy to strengthen the sparse association between metabolites and functional modules. Preferably, the objective function is:

[0035]

[0036] Among them: X represents the original metabolite data matrix, with dimensions of n×m (n: number of metabolites, m: number of samples); W represents the weight matrix, with dimensions of n×r (r: number of functional modules), indicating the contribution of metabolites to functional modules; H represents the activation matrix, with dimensions of r×m, indicating the distribution of functional modules in samples; represents the square of the Frobenius norm, which is used to measure the matrix reconstruction error λ1, λ2 represent the L1 regularization parameters, which control the sparsity of W and H; λ3 represents the KL divergence regularization parameter, which adjusts the strength of the sparsity constraint; represents the sparsity prior matrix (such as generated by thresholding or low-rank templates), which guides the sparse distribution of the weight matrix W; represents the Kullback-Leibler divergence, which measures the difference between the weight matrix W and the prior sparse matrix ; i and j represent the row and column indices of the matrix (i: metabolite index, j: functional module or sample index).

[0037] The NMF described above includes: determining the optimal rank through multi-rank synchronous calculation combined with the inflection point analysis of the sum of squared residuals; preferably, multi-rank synchronous calculation (the rank range can be set from 3 to 8) combined with the inflection point analysis of the sum of squared residuals, and finally determining the optimal decomposition rank (preferably, rank = 5).

[0038] Preferably, for the candidate rank r ∈ {3, 4, …, 8}, calculate the NMF decomposition in parallel:

[0039]

[0040] where r represents the candidate decomposition rank (number of functional modules), and the value range is from 3 to 8; W r， H r represent the weight matrix and activation matrix when the rank is r; X is the original metabolite data matrix (rows represent metabolites, columns represent samples); represents the Frobenius norm, which is used to measure the error between X and WH; RSS represents the sum of squared residuals (Residual Sum of Squares), that is used to evaluate the fitting effects of different ranks; the inflection point is the point with the largest slope change in the RSS curve, which is used to automatically select the optimal rank (such as r = 5).

[0041] Preferably, screen the top 20% of the features with the highest interpretability for the functional modules (preferably using the principle of maximum contribution).

[0042] Preferably, control the functional module overlap rate below 15%. More preferably, use the Jaccard similarity to control the functional module overlap rate below 15%.

[0043] More preferably, the NMF described above includes screening the top 20% of the features with the highest interpretability for the functional modules using the principle of maximum contribution, and / or realizing the biological annotation of the functional modules through temporal trajectory analysis and pathway enrichment analysis, and using the Jaccard similarity to control the functional module overlap rate below 15%.

[0044] Preferably, the function for screening features using the principle of maximum contribution:

[0045] Selected j ={i|W i,j ≥Quantile(W :,j ,0.8)}

[0046] Among them, Selected j represents the set of metabolites retained in functional module j (the top 20%); W i,j represents the weight value of metabolite i in functional module j; W:,j represents the j-th column of the weight matrix (the weights of all metabolites in functional module j); Quantile(W:,j,0.8) represents the 80th percentile of the weights in functional module j, and only metabolites higher than this value are retained.

[0047] Preferably, the function used when controlling the functional module overlap rate below 15% is:

[0048]

[0049] Among them, W j , W k represent the weight vectors of functional module j and functional module k; Supp(W j ) represents the set of metabolites with non-zero weights in functional module j (i.e., {i∣W i,j >0}); Supp(W k ) represents the set of metabolites with non-zero weights in functional module k; ∩, ∪ represent the intersection and union operators of sets; 0.15 represents the maximum allowable overlap rate between functional modules (Jaccard similarity threshold).

[0050] In the Boruta algorithm, the splitting criterion for feature Xj at decision tree node t is:

[0051]

[0052] Among them, j represents the feature index (such as the number of genes or modules); t represents the current node in the decision tree; represents the Gini impurity of node t; C represents the total number of categories of the target variable (in classification tasks); p(c∣t) represents the proportion of samples in node t belonging to category c; N represents the total number of samples in node t; N L , N R represent the number of samples in the left and right child nodes after splitting; t L , t R represent the left and right child nodes; ΔGini(j,t): the reduction in Gini impurity brought by feature j at node t.

[0053] The tumor described is a hematological tumor or a solid tumor, preferably gastric cancer.

[0054] In a specific embodiment of the present invention, the tumor is gastric cancer, and the functional modules include a glutamine metabolism functional module, a lipid metabolism functional module, a host-microbe co-metabolism functional module, a metabolic network bridge functional module, and a branched-chain amino acid metabolism functional module.

[0055] The sample is a body fluid. Preferably, the body fluid is selected from blood, plasma or serum.

[0056] In a specific embodiment of the present invention, the sample is plasma.

[0057] The number of the tumor metabolic markers is 10 - 15, preferably 12.

[0058] In a specific embodiment of the present invention, the tumor metabolic markers include nicotinamide, taurine, mono-2-ethylhexyl phthalate, sphingosine, 9,12,13-trihydroxy-octadecenoic acid, glycyl-phenylalanine, N-phenylacetyl-L-glutamine, (±)12-hydroxy-5Z,8Z,10E,14Z-eicosatetraenoic acid ((±)12-HETE), 5'-deoxy-5'-methylthioadenosine, traumatic acid, hypoxanthine, and azelaic acid.

[0059] In a second aspect of the present invention, there is provided a tumor metabolic marker obtained by the screening method.

[0060] In a third aspect of the present invention, there is provided a computer-readable storage medium or a device comprising a computer-readable storage medium, on which a computer program is stored, and the computer program includes a BIO-FIRE algorithm, and the BIO-FIRE algorithm includes:

[0061] A) Using the non-negative matrix factorization method (NMF) to perform functional module clustering on the preliminarily screened metabolites;

[0062] B) The Boruta algorithm performs feature screening by calculating importance scores;

[0063] C) Based on the principle of recursive optimization, sort the importance scores of each marker calculated in B) from high to low, and sequentially select and include them in the metabolite subset of the metabolic markers starting from the first one until a metabolite subset that can maximize the coverage of the most functional modules is selected, which is the tumor metabolic marker.

[0064] The tumor is a hematological tumor or a solid tumor; preferably, the tumor is gastric cancer.

[0065] Preferably, the functional modules include a glutamine metabolism functional module, a lipid metabolism functional module, a host-microbe co-metabolism functional module, a metabolic network bridge functional module, and a branched-chain amino acid metabolism functional module.

[0066] In a fourth aspect of the present invention, there is provided an application of the computer-readable storage medium or a device comprising the computer-readable storage medium in screening for tumor markers.

[0067] In a fifth aspect of the present invention, there is provided a method for constructing a tumor diagnosis model, the construction method comprising:

[0068] i) Collecting the concentration detection results of metabolic markers of a tumor patient group and a non-tumor control group;

[0069] ii) Constructing a tumor diagnosis model based on the information collected in step i);

[0070] Alternatively, the construction method comprises:

[0071] I) Collecting samples of subjects and detecting the concentrations of metabolic markers;

[0072] II) Clinically diagnosing the subjects and dividing the subjects into a tumor patient group and a non-tumor control group;

[0073] III) Constructing a tumor diagnosis model based on the detection results of step I) and the diagnosis results of step II).

[0074] The metabolic markers are the metabolic markers screened by the screening method described in the first aspect. Preferably, they include nicotinamide, taurine, mono-2-ethylhexyl phthalate, sphingosine, 9,12,13-trihydroxy-octadecenoic acid, glycyl-phenylalanine, N-phenylacetyl-L-glutamine, (±)12-hydroxy-5Z,8Z,10E,14Z-eicosatetraenoic acid ((±)12-HETE), 5'-deoxy-5'-methylthioadenosine, traumatic acid, hypoxanthine, and azelaic acid.

[0075] Preferably, the tumor is a hematological tumor or a solid tumor. More preferably, the tumor is gastric cancer. The methods for clinical diagnosis include, but are not limited to, electronic gastroscopy examination.

[0076] The sample is a body fluid; preferably, the body fluid is selected from blood, plasma or serum.

[0077] The methods for detecting the concentrations of metabolic markers in the test samples include one or more of nuclear magnetic resonance spectroscopy, mass spectrometry, chromatography, high performance liquid chromatography-tandem mass spectrometry, Fourier transform ion cyclotron resonance, ion mobility spectrometry, electrochemical detection, Raman spectroscopy, or radioactive labeling.

[0078] The algorithm used for constructing the diagnostic model is a machine learning model, preferably including one or more of logistic regression (LR), k-nearest neighbor algorithm (KNN), naive Bayes (NB), support vector machine (SVM Radial), random forest (RF), XGBoost, TreeNet, gradient boosting machine (GBM), LASSO, or neural network (NNET).

[0079] In the sixth aspect of the present invention, there is provided a diagnostic model obtained by the construction method as described in the fifth aspect.

[0080] In the seventh aspect of the present invention, there is provided an application of a metabolite biomarker in the preparation of a product for diagnosing gastric cancer, and the metabolite biomarker includes nicotinamide, taurine, mono-ethylhexyl phthalate, sphingosine, 9,12,13-trihydroxy-octadecenoic acid, glycyl-phenylalanine, N-phenylacetyl-L-glutamine, (±)12-hydroxy-5Z,8Z,10E,14Z-eicosatetraenoic acid ((±)12-HETE), 5'-deoxy-5'-methylthioadenosine, traumatic acid, hypoxanthine, and azelaic acid.

[0081] The tumor is a hematological tumor or a solid tumor. Preferably, the tumor is gastric cancer, such as early gastric cancer, gastric cancer negative for traditional tumor markers, gastric cancer with a tumor major axis less than 4 cm, gastroesophageal junction cancer, etc.

[0082] The metabolite biomarker is a metabolite biomarker in body fluids. Preferably, the body fluid is selected from blood, plasma, or serum; more preferably, it is a metabolite biomarker in plasma.

[0083] The product includes a reagent for detecting the metabolite biomarker. Preferably, the reagent detects the concentration of the metabolite biomarker.

[0084] Preferably, the methods used for detecting the metabolite biomarker include one or more of nuclear magnetic resonance spectroscopy, mass spectrometry, chromatography, high performance liquid chromatography-tandem mass spectrometry, Fourier transform ion cyclotron resonance, ion mobility spectrometry, electrochemical detection, Raman spectroscopy, or radioactive labeling.

[0085] In a specific embodiment of the present invention, the method used for detecting the metabolite biomarker is high performance liquid chromatography-tandem mass spectrometry.

[0086] Preferably, the product includes a diagnostic model, a kit, a test strip, a chip, or a device.

[0087] Preferably, the product further includes a reagent for sample processing and / or a standard of the metabolite biomarker.

[0088] In the eighth aspect of the present invention, a metabolic marker for diagnosing gastric cancer is provided. The metabolic marker includes nicotinamide, taurine, mono-2-ethylhexyl phthalate, sphingosine, 9,12,13-trihydroxy-octadecenoic acid, glycyl-phenylalanine, N-phenylacetyl-L-glutamine, (±)12-hydroxy-5Z,8Z,10E,14Z-eicosatetraenoic acid ((±)12-HETE), 5'-deoxy-5'-methylthioadenosine, traumatic acid, hypoxanthine, and azelaic acid.

[0089] The metabolic marker is a metabolic marker in body fluid. Preferably, the body fluid is selected from blood, plasma or serum; more preferably, it is a metabolic marker in plasma.

[0090] In the ninth aspect of the present invention, a method for diagnosing a tumor is provided. The method includes detecting the concentration of a metabolic marker in a sample of a subject.

[0091] The metabolic marker includes nicotinamide, taurine, mono-2-ethylhexyl phthalate, sphingosine, 9,12,13-trihydroxy-octadecenoic acid, glycyl-phenylalanine, N-phenylacetyl-L-glutamine, (±)12-hydroxy-5Z,8Z,10E,14Z-eicosatetraenoic acid ((±)12-HETE), 5'-deoxy-5'-methylthioadenosine, traumatic acid, hypoxanthine, and azelaic acid.

[0092] Preferably, the sample is body fluid. More preferably, the body fluid is selected from blood, plasma or serum.

[0093] The tumor is a hematological tumor or a solid tumor. Preferably, the tumor is gastric cancer.

[0094] The method includes one or more of nuclear magnetic resonance spectroscopy, mass spectrometry, chromatography, high performance liquid chromatography-tandem mass spectrometry, Fourier transform ion cyclotron resonance, ion mobility spectrometry, electrochemical detection, Raman spectroscopy, or radioactive labeling to detect the concentration of the metabolic marker.

[0095] Preferably, the method includes inputting the detected concentration of the metabolic marker into a constructed machine learning model to output a probability value, and comparing the output probability value with a threshold to determine whether the subject has a tumor.

[0096] The threshold is obtained from previous experiments, that is, by experiments and data analysis of the difference degree of metabolic markers between the gastric cancer patient group and the non-gastric cancer control group, training a machine learning model to determine the threshold.

[0097] In the tenth aspect of the present invention, a tumor diagnosis system is provided. The diagnosis system includes a device, unit or module for determining whether a subject has a tumor according to the concentration of a metabolic marker.

[0098] Preferably, the diagnostic system includes:

[0099] A data detection device, unit or module for detecting the concentration of metabolic markers in a sample;

[0100] A data input device, unit or module for inputting the concentration data of metabolic markers;

[0101] A data analysis device, unit or module for analyzing the tumor risk based on the concentration data of metabolic markers;

[0102] A data output device, unit or module for outputting the analysis result of whether an individual has a tumor.

[0103] The analysis uses a machine learning model, preferably one or more of logistic regression (LR), k-nearest neighbor algorithm (KNN), naive Bayes (NB), support vector machine (SVM Radial), random forest (RF), XGBoost, TreeNet, gradient boosting machine (GBM), LASSO or neural network (NNET).

[0104] The metabolic markers include nicotinamide, taurine, mono-2-ethylhexyl phthalate, sphingosine, 9,12,13-trihydroxy-octadecenoic acid, glycyl-phenylalanine, N-phenylacetyl-L-glutamine, (±)12-hydroxy-5Z,8Z,10E,14Z-eicosatetraenoic acid ((±)12-HETE), 5'-deoxy-5'-methylthioadenosine, traumatic acid, hypoxanthine and azelaic acid.

[0105] In the eleventh aspect of the present invention, a device is provided, and the device includes the diagnostic system described in the tenth aspect.

[0106] The term "subject" can be a human or non-human animal, including "patient", "suspected patient" and "healthy individual", etc. This term does not indicate a specific age or gender, covering adult or neonatal subjects as well as fetuses. The non-human animal can be a wild animal, zoo animal, economic animal, pet, experimental animal, etc. Specifically, the non-human animal includes but is not limited to pigs, cows, sheep, horses, donkeys, foxes, raccoons, minks, camels, dogs, cats, rabbits, mice (such as rats, mice, guinea pigs, hamsters, gerbils, chinchillas, squirrels) or monkeys, etc.

[0107] The term "diagnosis" refers to finding out whether a patient has a disease or disorder in the past, at the time of diagnosis or in the future, or finding out the progress of the disease or its possible future progress.

[0108] The present invention has the following beneficial effects:

[0109] 1. Large-scale, multi-center data support: This application integrates large-scale clinical samples from multiple authoritative medical centers to ensure the robustness and clinical applicability of research results;

[0110] 2. Combined analysis of untargeted and targeted metabolomics: Untargeted metabolomics is used to screen potential candidates, and then targeted metabolomics is combined for quantitative verification to ensure the specificity and repeatability of metabolites;

[0111] 3. Systematic metabolic biomarker screening strategy: This application has developed the BIO-FIRE algorithm, which uses non-negative matrix factorization (NMF) to mine functional modules of metabolites, and combines the Boruta feature selection algorithm and recursive optimization principle to determine the smallest biomarker combination covering the most functional modules, improving the comprehensiveness and efficiency of biomarker screening. In particular, the inventors of this application used targeted metabolomics combined with the RFE-RF machine learning algorithm for feature extraction in previous research (see patent document 2025103171683), without using a screening strategy covering functional modules. Although the AUC value of the screened biomarker combination is high, the number is large and the cost increases.

[0112] 4. Applicable to multiple clinical application scenarios: The screening method of this application is not only applicable to routine gastric cancer screening, but also can provide an efficient diagnostic plan in complex clinical scenarios such as early gastric cancer, gastric cancer with negative traditional tumor markers, and gastroesophageal junction cancer.

[0113] 5. This application breaks through the limitations of existing metabolomics research, constructs an efficient, stable and popularizable metabolomics biomarker screening and diagnosis process for non-invasive screening and precise diagnosis of gastric cancer, and provides an extensible technical platform for the field of metabolic research. Brief Description of the Drawings

[0114] Hereinafter, the embodiments of the present invention will be described in detail with reference to the drawings, wherein:

[0115] Figure 1 The results of untargeted metabolomics detection are shown, wherein, Figure 1 A is a heat map covering 3327 metabolic signals, wherein, HC represents the healthy group, BGD represents the benign disease group, EGC represents the early gastric cancer group, and AGC represents the advanced gastric cancer group. Figure 1 B is a linear discriminant analysis (LDA) graph, wherein, HC represents the healthy group, BGD represents the benign disease group, EGC represents the early gastric cancer group, and AGC represents the advanced gastric cancer group. Figure 1 C shows the process of screening differential metabolic signals taking the gastric cancer (GC) vs non-gastric cancer (NGC) comparison group as an example, wherein VIP represents the importance of variables in the projection, FC represents the fold change, and FDR represents the false discovery rate. Figure 1D represents the differential metabolic signals and their union of 10 comparison groups, where HC represents the healthy group, GC represents the gastric cancer group, BGD represents the benign disease group, EGC represents the early gastric cancer group, and AGC represents the advanced gastric cancer group;

[0116] Figure 2 The screening and identification process of differential metabolic ion pairs is shown;

[0117] Figure 3 The various evaluation indexes of NMF under different decomposition ranks are shown;

[0118] Figure 4 The heatmaps of different clustering methods are shown, where Figure 4 A is the heatmap of hclust clustering, Figure 4 B is the heatmap of Kmean clustering, Figure 4 C is the heatmap of NMF clustering;

[0119] Figure 5 Partial results during the operation of the Boruta algorithm are shown, where Figure 5 A reflects the rigor of the Boruta algorithm for feature screening, Figure 5 B reflects the comprehensiveness of the coverage of important metabolites screened by the Boruta algorithm on functional modules;

[0120] Figure 6 The detailed process of determining markers using the recursive optimization principle is shown;

[0121] Figure 7 The importance score ranking (A) of 12 modeling metabolic markers and the difference in their expression levels between the gastric cancer (GC) group and the non-gastric cancer (NGC) group (B) are shown;

[0122] Figure 8 The bar charts of the coverage rate of functional modules of metabolic markers screened by the Bortua algorithm, LASSO algorithm, and RF algorithm and the proportion of newly deconvoluted substances are shown;

[0123] Figure 9 The performance of the sum of P-values after FDR correction and the sum of absolute values of logFC of metabolic markers screened by the Bortua algorithm, LASSO algorithm, and RF algorithm are shown;

[0124] Figure 10 The modeling robustness of metabolic markers screened by the Bortua algorithm, LASSO algorithm, and RF algorithm is shown;

[0125] Figure 11Shown are the ROC curves and corresponding AUC values of the gastric cancer metabolic marker diagnostic model constructed using 10 machine learning algorithms in the test set (A); the radar chart shows the performance comparison of the metabolic marker diagnostic models constructed by 10 machine learning algorithms and 3 traditional tumor marker diagnostic models in the test set (B), and the evaluation indicators include accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV); the ROC curve (C) of the random forest (RF) model of PMB-P12, CEA, CA199, and CA724, as well as the bar charts (D) of the AUC value, accuracy, sensitivity, and specificity.

[0126] Figure 12 Shown is the diagnostic efficacy of metabolite combinations covering 4 functional modules and those covering a single functional module.

[0127] Figure 13 Shown is the discrimination ability of 10 diagnostic models of 12 metabolic markers in complex clinical scenarios. Among them, Figure A shows gastric cancer with negative traditional tumor markers, Figure B shows gastric cancer with a diameter less than 4 cm, Figure C shows extremely early-stage (stage IA) gastric cancer, and Figure D shows gastroesophageal junction cancer. Detailed implementation manners

[0128] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0129] It should be noted that the methods used in the present invention are all conventional methods unless otherwise specified, and the reagents used in the present invention are all commercially available products unless otherwise specified.

[0130] Example 1: Construction of the gastric cancer metabolic map and identification of differential metabolites based on the widely targeted metabolomics technology

[0131] 1 Collection of plasma samples in the discovery cohort

[0132] After obtaining the consent of the patients, peripheral venous blood plasma samples were collected from 267 non-gastric cancer subjects (including 100 healthy individuals and 167 patients with benign gastric diseases) and 202 gastric cancer patients (including 134 early gastric cancer patients and 68 advanced gastric cancer patients) from two research centers (Beijing Friendship Hospital and Hubei Provincial People's Hospital) as the discovery cohort. Among them, the diagnosis criteria for gastric cancer patients were confirmed by postoperative pathology; the non-gastric cancer group samples included healthy individuals without gastric diseases after physical examination and patients with benign gastric diseases such as chronic gastritis, erosive gastritis, gastric ulcer, and gastric polyps after hospital examination. All gastric cancer patients and non-gastric cancer group subjects had no history of other malignant tumors, no other major systemic diseases, and no chronic medical history of long-term medication. Samples of healthy individuals (HC) and patients with benign diseases (BGD) were included as the non-gastric cancer group (NGC), and samples of early (EGC) and advanced gastric cancer (AGC) patients were included as the gastric cancer group (GC). All blood samples were collected in the early morning fasting state of the subjects, and then centrifuged to separate plasma, which was immediately stored in a -80 °C refrigerator. During the research process, plasma samples were taken out and thawed according to experimental requirements for subsequent analysis.

[0133] 2 Construction of a gastric cancer-specific metabolite ion pair database

[0134] Database construction process: To study the relationship between plasma metabolome and gastric cancer, we used 20% of the plasma samples in the above discovery cohort (with an equal number of samples in each group) and different mass spectrometry detection techniques (QTOF and QTRAP) and metabolite databases to widely collect plasma metabolite ion characteristics of gastric cancer and constructed a gastric cancer plasma-specific metabolome database. First, the gastric cancer plasma samples were detected using the full-scan mode with a high-resolution instrument QTOF to ensure that as many metabolite ion characteristics as possible were detected. However, due to the characteristics of the QTOF instrument, its detection sensitivity for metabolites with low concentrations is low, and it is difficult to capture the signals of low-content substances. Therefore, the detected ion pair signals were continued to be transferred to the QTRAP6500 instrument, and the AB SCIEX QTRAP6500 MIM-IDA-EPI scan mode was used to supplement the detection of low-concentration metabolites. Combining all the above metabolite ion characteristics, metabolite information in the local standard substance database, and literature metabolite information, after optimization and duplicate removal, a gastric cancer plasma-specific metabolome database was formed.

[0135] 3 Plasma-wide targeted metabolomics detection

[0136] 1) Sample pretreatment

[0137] Take out the sample collected in step 1 from the -80°C refrigerator and thaw it on ice until there are no ice cubes in the sample (all subsequent operations are required to be carried out on ice); after the sample is thawed, vortex it for 10 s to mix evenly, take 50 μL of the sample and add it to the centrifuge tube with the corresponding number; add 300 μL of pure methanol internal standard extraction solution (containing L-phenylalanine internal standard with a concentration of 100 ppm); vortex for 5 min, let it stand for 24 h, and then centrifuge at 12000 r / min at 4°C for 10 min; aspirate 270 μL of the supernatant and concentrate it for 24 h; then add it to 100 μL of the complex solution composed of acetonitrile and water in a volume ratio of 1:1 for LC-MS / MS analysis. Take 20 μL from each sample and mix them into a quality control sample (QC), which is collected once every 15 samples.

[0138] 2) Detection and analysis of sample metabolites

[0139] Determine the liquid chromatography conditions as follows:

[0140] Chromatographic column: Waters ACQUITY UPLC HSS T3 C18 1.8 μm, 2.1 mm * 100 mm; column temperature is 40°C; injection volume is 2 μL.

[0141] Mobile phase: Phase A is an aqueous solution containing 0.1% acetic acid, and phase B is an acetonitrile solution containing 0.1% acetic acid. The elution gradient program is as follows: at 0 min, the volume ratio of phase A to phase B is 95:5; at 11.0 min, the volume ratio of phase A to phase B is 10:90; at 12.0 min, the volume ratio of phase A to phase B is 10:90; at 12.1 min, the volume ratio of phase A to phase B is 95:5; at 14.0 min, the volume ratio of phase A to phase B is 95:5 V / V. Flow rate is 0.4 mL / min.

[0142] Determine the mass spectrometry conditions as follows:

[0143] The temperature of the electrospray ionization (ESI) source is 500°C, the mass spectrometry voltage is 5500 V (positive) or -4500 V (negative), the ion source gas I (GS I) is 55 psi, the gas II (GS II) is 60 psi, the curtain gas (CUR) is 25 psi, and the collision-activated dissociation (CAD) parameter is set to high.

[0144] In the triple quadrupole (Qtrap), each ion pair is scanned and detected in MRM mode according to the optimized declustering potential (DP) and collision energy (CE).

[0145] Analyze and detect the samples respectively according to the determined liquid chromatography conditions and mass spectrometry conditions: For all the samples in the above discovery cohort, adopt a metabolomics method combining enhanced ion scanning mass spectrometry (MIM-EPI) and time-of-flight mass spectrometry (TOF) with a multi-reaction monitoring acquisition mode, and integrate the local standard database to construct a gastric cancer plasma metabolite database to obtain the original mass spectrometry data of each plasma sample.

[0146] 3) Pretreatment and integration of chromatogram peak areas

[0147] Based on the constructed gastric cancer serum-specific metabolite database, perform qualitative and quantitative mass spectrometry analysis on the metabolites of the samples. Different molecular weight metabolites can be separated by liquid chromatography. Characteristic ions of each substance are screened through the multi-reaction monitoring mode (MRM) of a triple quadrupole, and the signal intensity (CPS) of the characteristic ions is obtained in the detector. Open the sample's off-line mass spectrometry file with MultiQuant 3.0.3 software to perform the integration and correction of chromatographic peaks. The peak area (Area) of each chromatographic peak represents the relative content of the corresponding substance. Set S / N>5, and retain the peaks with a retention time deviation of no more than 0.2 min. Finally, export all the chromatographic peak area integration data for storage.

[0148] 4) Experimental quality control

[0149] By performing an overlapping display analysis on the total ion current chromatograms of the mass spectrometry detection and analysis of different quality control (QC) samples, the repeatability of metabolite extraction and detection, that is, technical repeatability, can be judged. The high stability of the instrument provides an important guarantee for the repeatability and reliability of the data. The CV value, that is, the coefficient of variation, is the ratio of the standard deviation of the original data to the average of the original data, which can reflect the degree of data dispersion. The empirical cumulative distribution function (ECDF) can be used to analyze the frequency of the CV of substances less than the reference value. The higher the proportion of substances with a lower CV value in the QC samples, the more stable the experimental data: When the proportion of substances with a CV value less than 0.5 in the QC samples is higher than 85%, it indicates that the experimental data is relatively stable; when the proportion of substances with a CV value less than 0.3 in the QC samples is higher than 75%, it indicates that the experimental data is very stable. At the same time, monitor the change of the CV value of the L-phenylalanine internal standard during the detection process. If the change of the internal standard CV value is less than 20%, it indicates good instrument stability during the detection process.

[0150] 4 Construction of plasma metabolite profiles of gastric cancer patients and identification of differential metabolites

[0151] (1) Data preprocessing

[0152] A total of 3327 ion pairs were detected in the above step 1. Data preprocessing was performed on the results of normalizing the peak area integration data. According to the 80% rule, when a metabolite is detected in at least 4 / 5 of the samples in a certain group, the metabolite is considered to be detectable. According to this rule, 83 metabolites were robustly detected in all 928 samples. For a small number of undetected metabolites (<1 / 5 samples), they were filled with 1 / 2 of the lowest detection value in that group for subsequent statistical analysis.

[0153] (2) Analysis of the results of broad-target metabolomics detection

[0154] The results of the heat map analysis covering 3327 metabolic signals showed that there were significant differential expressions in different metabolic signals among the healthy control group (HC), benign gastric disease group (BGD), early gastric cancer group (EGC), and advanced gastric cancer group (AGC) Figure 1 A). To effectively distinguish the above four groups, in this embodiment, linear discriminant analysis (LDA) of supervised learning was further used to analyze the data. As Figure 1 shown in B, the overall differences among the groups can be distinguished by the linear combination of the characteristics of all compounds. Each point in the figure represents a sample, and the closer the distance between the sample points, the closer the contents of its various compounds. The results showed that the sample points of the HC, BGD, EGC, and AGC groups showed an obvious aggregation trend, while there was a good separation between the sample points of the HC group and those of the other three groups.

[0155] To more comprehensively explore the differences among the 4 groups to determine the most differentially expressed metabolic signals, pairwise comparisons and sum comparisons were further performed on the 4 groups. A total of 10 comparison groups were set up, and differential metabolite analysis was performed between the two groups using the peak area integration data. At the same time, the VIP value of the multivariate statistical model, the P value and FC value of the univariate statistical T test were used to screen for differential metabolic signals, and FDR<0.05, FC>1.2 or FC<0.80, VIP>1 were set as the differential significance criteria Figure 1 C), and the differential metabolic signals of the 10 comparison groups were screened out and their union was taken Figure 1 D), and finally 1097 differential metabolic signals were obtained for subsequent ion pair deconvolution and substance identification.

[0156] (3) Deconvolution and substance identification of differential metabolic signals

[0157] To improve the confidence in targeted detection of metabolites, 792 signals with high-resolution data were selected from the 1097 differential metabolic signals screened previously for accurate spectral deconvolution. The spectral deconvolution process mainly includes two parts: ① Molecular formula fitting and compound matching: First, for the unknown ion pairs, according to their accurate mass-to-charge ratios, through the software peakview2.2 (AB Sciex), molecular formula fitting was carried out, with the mass-to-charge ratio set within the range of less than 10 ppm. After comprehensive evaluation, a suitable molecular formula was selected. Using the Chemspider database (www.Chemspider.com) and Pubchem database, multiple possible compounds corresponding to this molecular formula were obtained. Combining biological and biochemical information, the most likely candidate compounds were screened out, and the fragment fitting in the fomular pane in peakview was used for secondary matching with the unknowns, with a matching rate > 70%. ② Standard substance verification: To confirm the identity of the compound, a standard substance was purchased for verification. The unknown metabolites were confirmed according to the following four criteria: column retention time, isotope pattern, MS fragmentation (collision-induced dissociation MS / MS), and accurate mass. The detailed results of this process are as Figure 2 , and finally, after deduplication, 84 differentially metabolized substances identified by standard substances (MSI Level1) were obtained. The specific information of the 84 differential metabolites is shown in Table 1.

[0158] Table 1

[0159]

[0160]

[0161]

[0162]

[0163] Example 2: Determination of initially screened metabolites based on targeted metabolomics technology

[0164] Collection of plasma samples in the modeling cohort

[0165] In this example, peripheral venous blood plasma from 460 non-gastric cancer subjects (including 216 healthy people and 244 patients with benign gastric diseases) and 468 gastric cancer patients (including 247 early gastric cancer patients and 221 advanced gastric cancer patients) from four research centers (Beijing Friendship Hospital, Cancer Hospital of the Chinese Academy of Medical Sciences, Hubei Provincial People's Hospital, Tongji Hospital Affiliated to Tongji Medical College of Huazhong University of Science and Technology) was collected as the modeling cohort. The patient inclusion and exclusion criteria were the same as those in Example 1. The blood collection time was in the early morning fasting state. After centrifugation, all plasma samples were stored in a -80°C refrigerator. When conducting the research, the plasma samples were taken out and thawed for subsequent analysis.

[0166] Targeted metabolomics detection of plasma samples

[0167] (1) Determine the targeted detection range

[0168] Based on the 84 differential metabolites screened and identified in the early stage, through systematic retrieval of gastric cancer metabolome research and bibliometric analysis, metabolites reported in the literature more than 5 times or recognized tumor metabolic markers were summarized to obtain 32 literature metabolites. After removing duplicates, a total of 102 metabolites to be detected were finally obtained. These metabolites mainly cover various substances such as amino acids and their derivatives, bile acids and their conjugates, purine / pyrimidine metabolites, steroid hormones, lipids (fatty acids, phospholipids, sphingolipids), biogenic amines, organic acids, carbohydrates, flavonoids / alkaloids, etc., involving multiple physiological and pathological processes such as amino acid metabolism, bile acid metabolism, nucleotide metabolism, lipid metabolism, hormone metabolism, drug metabolism, and microbial metabolism, reflecting the metabolic status of the body.

[0169] (2) Sample pretreatment

[0170] Take out the samples collected in step 1 from the -80°C refrigerator and thaw them on ice until there are no ice crystals left; after the samples are thawed, vortex for 10 seconds to mix well. Take 50 μL of the sample, add 150 μL of the extraction solution (the extraction solution contains an isotope internal standard with a concentration of 100 ppm), then vortex and mix for 3 minutes, centrifuge at 12000 rpm and 4°C for 10 minutes, and leave it standing overnight in the -20°C refrigerator. The next day, centrifuge at 12000 rpm and 4°C for 5 minutes, take 170 μL of the supernatant, and transfer it to a 96-well plate in sequence. After completing the protein precipitation treatment, seal the well plate for LC-MS / MS analysis. Another 20 μL of each sample was mixed to prepare a quality control sample (QC), and the collection was carried out at intervals of every 15 samples.

[0171] (3) Determine the detection conditions for detection

[0172] Targeted quantitative analysis used a T3 column for metabolite separation, and the specific detection conditions are as follows:

[0173] T3 column liquid chromatography conditions: The chromatographic column is Waters ACQUITY UPLC HSS T3 C18 1.8 μm, 2.1 mm × 100 mm, the column temperature is set at 40°C, and the injection volume is 2 μL. Mobile phase A is an aqueous solution containing 0.1% acetic acid, and mobile phase B is an acetonitrile solution containing 0.1% acetic acid; the elution gradient is: 0 minutes, A:B = 95:5; 11.0 minutes, A:B = 10:90; 12.0 minutes, A:B = 10:90; 12.1 minutes, A:B = 95:5; 14.0 minutes, A:B = 95:5, and the flow rate is set at 0.4 mL / min.

[0174] Mass spectrometry conditions: The temperature of the electrospray ionization source (ESI) was set at 500 °C, and the mass spectrometry voltages were 5500 V (positive ion mode) and -4500 V (negative ion mode), respectively. The pressure of Ion Source Gas I (GS I) was 55 psi, the pressure of Gas II (GS II) was 60 psi, and the pressure of curtain gas (CUR) was 25 psi. Collisionally activated dissociation (CAD) was set to high. In the triple quadrupole (Qtrap), optimized declustering potential (DP) and collision energy (CE) were used for MRM mode scanning to detect the signals of each ion pair.

[0175] (4) Pretreatment and integration of chromatogram peak areas

[0176] The mass spectrometry data were processed by MultiQuant 3.0.3 software. According to the retention time and peak shape information of the standards, the mass spectrometry peaks of the target substances in each sample were integrated and corrected to ensure the accuracy of qualitative and quantitative analysis. All samples were subjected to qualitative and quantitative analysis, and the peak area of each chromatographic peak reflected the relative concentration of the substance. By substituting into the linear equation and calculation formula, the qualitative and quantitative analysis results of the target substances in all samples were finally obtained.

[0177] (5) Calculation of metabolite concentration

[0178] Standard solutions with different concentrations ranging from 0.01 ng / mL to 500 ng / mL were prepared, and the mass spectrometry peak intensity data corresponding to each concentration of the standard were obtained. By plotting the standard curve between the concentration ratio of external standard to internal standard (Concentration Ratio) and the peak area ratio (Area Ratio), the concentration of each substance was further analyzed. The peak area ratios detected in all samples were substituted into the linear equation of the standard curve, and the concentration calculation was performed using the calculation formula. The dilution factor was set to 3 in MultiQuant 3.0.3 software, and finally the concentration (ng / mL) of the substance in the sample was calculated, thereby obtaining the content data of the substance in the sample.

[0179] (6) Experimental quality control

[0180] By performing total ion current map overlay analysis on the mass spectrometry data of different quality control (QC) samples, the repeatability of metabolite extraction and detection, i.e., technical replicates, can be evaluated. The stability of the instrument is crucial for ensuring the reliability and repeatability of the data. The coefficient of variation (CV), which is the ratio of the standard deviation to the mean of the raw data, is used to reflect the discreteness of the data. The frequency of CV values of substances below the reference value can be analyzed through the empirical cumulative distribution function (ECDF). If the proportion of substances with low CV values in the QC samples is high, it indicates that the experimental data is relatively stable. When the CV values of all QC samples are less than 0.3, it shows that the data stability is good; if the proportion of substances with CV values less than 0.2 in the QC samples exceeds 90%, it indicates that the data is very stable. In addition, by monitoring the change in the CV value of the isotope internal standard, if the change in the internal standard CV is less than 20%, it indicates that the instrument maintains good stability during the detection process.

[0181] Perform data preprocessing on the qualitative and quantitative results obtained in step 2. According to the 80% rule, when a metabolite is detected in at least 4 / 5 of the samples in a certain group, the metabolite is considered to be detectable. According to this rule, 83 metabolites were robustly detected in all 928 samples. For a small number of metabolites that were not detected (<1 / 5 samples), they were filled with 1 / 2 of the lowest detected value in that group for subsequent statistical analysis.

[0182] Example 3 Screening of Gastric Cancer Metabolic Markers Based on the BIO-FIRE Algorithm

[0183] Algorithm Development: In this example, a biomarker identification and optimization algorithm based on functional and feature importance recursive enhancement (Biomarker Identification and Optimization Algorithm via Functional and Importance-based Recursive Enhancement, BIO-FIRE) was developed and applied to the screening of gastric cancer markers. Taking the data of 83 metabolites in 928 samples in Example 2 as an example, the implementation process of this algorithm mainly includes 3 steps:

[0184] I. Functional Module Clustering and Interpretability Enhancement

[0185] The main purpose of this step is to utilize the internal structure in the metabolic data, extract functional modules with biological significance, and provide a solid functional framework and data support for subsequent biomarker screening and optimization through functional module clustering and interpretability enhancement. By accurately dividing the functional units of different metabolic processes, this method aims to reveal the key regulatory factors in the metabolic pathway, thereby enhancing the overall biological interpretability and clinical application value.

[0186] Non-negative Matrix Factorization (NMF) is selected as the core algorithm for this step, mainly due to its natural adaptability to non-negative data and additive interpretability. Metabolomic data is essentially non-negative numerical values. NMF can naturally preserve the physical and biological meanings of the data when processing such data. The decomposition results not only reveal the potential functional modules in the data but also quantitatively reflect the contributions of each metabolite to different functional modules through the output weight matrix. Compared with traditional clustering methods, NMF can more effectively avoid information overlap and ensure the orthogonality between functional modules, thus enhancing the independence and accuracy of functional interpretation.

[0187] In the NMF algorithm, the objective function is usually a problem of minimizing the reconstruction error based on matrix factorization. NMF attempts to decompose a non-negative matrix X (usually the metabolite concentration matrix) into the product of two non-negative matrices W and H:

[0188] X≈W×H

[0189] Where: X is the original metabolic data matrix (rows represent metabolites, columns represent samples); W is the weight matrix from genes / metabolites to functional modules (indicating the contribution of each metabolite in different functional modules); H is the activation matrix of functional modules (indicating the weighted distribution of each sample on different functional modules).

[0190] NMF optimizes W and H by minimizing the reconstruction error:

[0191]

[0192] Where, denotes the Frobenius norm, which is used to measure the error between X and WH.

[0193] In addition, sparsity constraints or regularization terms can be introduced, for example:

[0194] L1 regularization (improving the sparsity of functional modules): Encourages the elements in W or H to approach zero, so that each metabolite mainly belongs to a few functional modules;

[0195] L2 regularization (preventing overfitting): Prevents individual metabolites from having too much contribution to a certain functional module and improves the simulation stability. Generally speaking, the objective function is usually written as:

[0196]

[0197] Where:

[0198] λ1, λ2 are regularization parameters; ||W||1 and ||H||1 control sparsity, ensure the orthogonality between functional modules, and reduce the overlap of metabolites in multiple functional modules.

[0199] In specific implementation, the present invention improves the traditional NMF by adopting a dynamic rank selection and consensus enhancement strategy. First, through multi-rank synchronous calculation (in the present invention, the rank range is set from 3 to 8) combined with the analysis of the inflection point of the sum of squared residuals, the optimal decomposition rank (in the present invention, rank = 5) is finally determined. The specific process is shown in Figure 3 . At the same time, each candidate rank is independently run 100 times to generate a consensus matrix to verify the stability of the functional modules. Further, the KL divergence (Kullback-Leibler Divergence) optimization strategy is adopted to strengthen the sparse association between metabolites and functional modules, and the top 20% of the features with the highest explanatory power for the functional modules are selected through the principle of maximum contribution to eliminate noise interference. Finally, the biological annotation of the functional modules is realized through temporal trajectory analysis and pathway enrichment analysis, and the overlap rate of the functional modules is controlled below 15% by using the Jaccard similarity to ensure the independence and accuracy of each functional module. The specific improvements are as follows:

[0200] 1. Dynamic rank selection and consensus enhancement strategy

[0201] The traditional NMF requires manual specification of the rank, while the present invention automatically determines the optimal rank through multi-rank synchronous calculation and stability verification.

[0202] (1) Multi-rank synchronous calculation and inflection point analysis

[0203] Function form: For candidate ranks r ∈ {3, 4,..., 8}, the NMF decomposition is calculated in parallel:

[0204]

[0205] where r represents the candidate decomposition rank (the number of functional modules), and the value range is from 3 to 8; W r， H r represent the weight matrix and activation matrix when the rank is r; X is the original metabolite data matrix (rows represent metabolites, and columns represent samples); represents the Frobenius norm, which is used to measure the error between X and WH; RSS represents the residual sum of squares, that is used to evaluate the fitting effects of different ranks; the inflection point is the point with the largest slope change in the RSS curve, which is used to automatically select the optimal rank (such as r = 5).

[0206] (2) Consensus matrix to verify the stability of functional modules

[0207] Function form: For each candidate rank r, it is independently run 100 times to generate a consensus matrix Cr:

[0208]

[0209] Among them, C r represents the consensus matrix when the rank is r, with a dimension of n×n, indicating the frequency of metabolite co-occurrence in the same functional module; N represents the number of independent runs (N = 100); W r (k) represents the weight matrix obtained from the k-th independent run; Cluster k (·) represents the clustering function for the k-th run, which converts W r (k) into a binary association matrix (1 indicates that metabolites belong to the same functional module).

[0210] Stability evaluation: Select the optimal rank through the uniformity and consistency of Cr.

[0211] 2. Improvement of the objective function: Introduce the KL divergence optimization strategy

[0212] In the present invention, by introducing the KL divergence as a regularization term, the sparse association between metabolites and functional modules is strengthened. The improved objective function is as follows:

[0213]

[0214] Among them: X represents the original metabolite data matrix, with a dimension of n×m (n: number of metabolites, m: number of samples); W represents the weight matrix, with a dimension of n×r (r: number of functional modules), indicating the contribution of metabolites to functional modules; H represents the activation matrix, with a dimension of r×m, indicating the distribution of functional modules in samples; represents the square of the Frobenius norm, which is used to measure the matrix reconstruction error λ1, λ2 represent the L1 regularization parameters, which control the sparsity of W and H; λ3 represents the KL divergence regularization parameter, which adjusts the strength of the sparsity constraint; represents the sparsity prior matrix (such as generated by thresholding or low-rank templates), which guides the sparse distribution of the weight matrix W; represents the Kullback-Leibler divergence, which measures the difference between the weight matrix W and the prior sparse matrix ; i, j represent the row and column indices of the matrix (i: metabolite index, j: functional module or sample index).

[0215] 3. Feature screening and independence constraint of functional modules

[0216] (1) Screen features according to the maximum contribution principle

[0217] Functional form: After normalizing the weight matrix W column by column, retain the top 20% of the metabolites in each column (functional module), and eliminate

[0218] Remove low - contribution metabolites to improve the interpretability of functional modules:

[0219] Selected j ={i|W i,j ≥Quantile(W :,j ,0.8)}

[0220] Among them, Selected j represents the set of metabolites retained in functional module j (the top 20%); W i,j represents the weight value of metabolite i in functional module j; W:,j represents the j - th column of the weight matrix (the weights of all metabolites in functional module j); Quantile(W:,j,0.8) represents the 80th percentile of the weights in functional module j, and only metabolites higher than this value are retained.

[0221] (2) Control of functional module overlap rate

[0222] Functional form: Constrain the independence of functional modules through Jaccard similarity:

[0223]

[0224] Among them, W j ,W k represent the weight vectors of functional module j and functional module k; Supp(W j ) represents the set of metabolites with non - zero weights in functional module j (i.e., {i∣W i,j >0}); Supp(W k ) represents the set of metabolites with non - zero weights in functional module k; ∩, ∪ represent the intersection and union operators of sets; 0.15 represents the maximum allowable overlap rate between functional modules (Jaccard similarity threshold).

[0225] Results: 83 metabolites were classified by unsupervised NMF clustering analysis, and the obtained clustering modules were functionally named according to KEGG annotation information. Finally, five metabolic functional modules were obtained, namely: glutamine metabolism functional module (NMF cluster 1), lipid metabolism functional module (NMF cluster 2), host - microbe co - metabolism functional module (NMF cluster 3), metabolic network bridge functional module (NMF cluster 4), and branched - chain amino acid metabolism functional module (NMF cluster 5).

[0226] In addition, the non - negative matrix factorization (NMF) algorithm adopted was compared with two common unsupervised clustering methods, hierarchical clustering (hclust) and K - means clustering, and the clustering performance of each was systematically evaluated. The specific results are as follows: (1) Hierarchical clustering (hclust): The data distribution between different categories is severely unbalanced, and the number of metabolites within some categories is scarce, unable to provide sufficient statistical support, thus restricting the effective screening of subsequent functional modules ( Figure 4 A). (2) K - means clustering: The discrimination between clusters is relatively low, and there are obvious data overlaps within some clusters ( Figure 4 B). (3) NMF algorithm: In contrast, the NMF algorithm can efficiently divide all metabolites into five functional modules. There are significant differences between the functional modules and the metabolite quantity distribution is relatively balanced, thus providing a more reliable statistical basis for subsequent analysis ( Figure 4 C). II. Identification of potential biomarkers based on feature importance

[0227] The main objective of this step is to accurately identify potential biomarkers based on feature importance, so as to ensure that the subsequent constructed biomarker subset has high interpretability and predictive ability. By screening out features that contribute significantly to the response variable, this method can eliminate noise and redundant information, providing solid data support for establishing a robust biomarker combination.

[0228] The reason for choosing the Boruta algorithm lies in its unique feature selection mechanism and non - linear modeling ability. Boruta constructs a null distribution model by generating shadow features after random perturbation. Its iterative decision - making process (i.e., the feature importance score needs to continuously exceed the shadow features) can avoid false - positive selections in a single random forest calculation. At the same time, its strategy of retaining all relevant features can capture those features with low individual importance but significant joint effects, which is more suitable for revealing the synergistic effects between metabolites compared with methods such as LASSO, providing a more comprehensive perspective for biomarker research.

[0229] In terms of specific implementation, the present invention performs feature screening on the functional module data pre - extracted by NMF by calling the Boruta::Boruta() function in the Boruta package of R language, with parameter settings ntree = 500 and pValue = 0.001 to ensure high robustness and stability of the algorithm during the iterative process. The following is the specific training method and mathematical implementation process of the Boruta algorithm:

[0230] 1. Feature importance measurement (Gini importance)

[0231] Let the splitting criterion of feature Xj at decision tree node t be:

[0232]

[0233] Among them, j represents the feature index (such as the number of genes or functional modules); t represents the current node in the decision tree; represents the Gini impurity of node t; C represents the total number of categories of the target variable (in a classification task); p(c∣t) represents the proportion of samples in node t belonging to category c; N represents the total number of samples in node t; t L , t R represent the left and right child nodes; N L , N R represent the number of samples in the left and right child nodes after splitting; ΔGini(j,t): the reduction in Gini impurity brought by feature j at node t.

[0234] The cumulative importance of feature Xj in the entire tree is:

[0235]

[0236] Among them, represents the cumulative importance of feature Xj in a single decision tree; t represents a certain node (Node) in the decision tree; Splits(j) represents the set of all nodes in the entire tree that use feature Xj for splitting; ΔGini(j,t) represents the reduction in Gini impurity brought by feature Xj when splitting at node t.

[0237] Finally, the importance score of the random forest is obtained by averaging over all trees.

[0238] 2. Shadow Feature Generation Mechanism

[0239] For the original feature matrix X∈R n×p , generate the shadow feature matrix

[0240]

[0241] Among them, represents the generated shadow feature matrix; X represents the original feature matrix (the result after NMF extraction); :,k represents the k-th column (all rows) of the matrix; Permute(·) represents the operation of randomly permuting the column vector; k represents the feature index (from 1 to p).

[0242] 3. Wilcoxon Signed-Rank Test

[0243] For each feature X j , calculate the importance difference between it and the maximum shadow feature:

[0244]

[0245] Construct a statistic:

[0246]

[0247] where Importance j represents the importance score of the original feature j; represents the importance score of the shadow feature k; d j represents the importance difference between feature j and the best shadow feature; p represents the total number of variables or features in the dataset; R represents the number of iterations (the total number of rounds the algorithm runs); d j (i) represents the difference value of feature j in the i-th iteration; sign(·) represents the sign function (the positive or negative sign of the value); rank(|·|) represents the rank order of the absolute value; W represents the Wilcoxon signed-rank statistic.

[0248] 4. Significance threshold for Bonferroni correction

[0249] For m features to be tested, the adjusted significance level is:

[0250]

[0251] where α represents the original significance level (user-specified value, here it is 0.001); m represents the number of hypothesis tests conducted simultaneously (equal to the total number of features p); p represents the number of original features (the dimension after NMF extraction); α adj represents the corrected significance threshold. It is used to ensure that the overall error rate (Family-Wise Error Rate, FWER) does not exceed 0.001.

[0252] 5. Iteration termination condition

[0253] Define the feature status set where S represents the status set of all features; S j represents the status of the j-th feature; j represents the feature index; p represents the total number of features.

[0254] When the following condition is satisfied for t consecutive iterations:

[0255]

[0256] Then terminate the algorithm.

[0257] where s j (t)represents the state of feature j at the tth iteration (Confirmed / Rejected / Tentative); p represents the total number of features; II(·) represents the indicator function (1 when the condition is met, otherwise 0); t represents the threshold number of iterations without continuous state change (default is 20).

[0258] Results: This parameter combination not only significantly reduced the impact of random errors, but also screened out features closely related to the response variables through statistical significance tests, laying the foundation for the subsequent construction of the optimal marker subset and its clinical application. Figure 5 This is a partial display of the results during the operation of the Boruta algorithm. Figure 5 A reflects the rigor of Boruta algorithm in feature screening (when the importance of the original feature is significantly higher than that of all shadow features, it is judged as "important", when it is significantly lower than that of all shadow features, it is judged as "irrelevant", and when it cannot be clearly judged, it is classified as "pending", and the confidence is continuously improved through iteration). Figure 5 B reflects the comprehensive coverage of functional modules by important metabolites screened by the Boruta algorithm.

[0259] The Boruta feature selection algorithm was used to further screen key metabolite markers, and finally 57 potential metabolite markers (covering NMF clusters 2-5) were obtained and ranked according to the feature importance score.

[0260] 3. Determination of markers based on the principle of recursive optimization

[0261] The main goal of this step is to reasonably streamline and optimize the number of metabolic marker combinations, and gradually select markers by combining the NMF clustering modules obtained in step 1 with the importance scores obtained in step 2. The specific process is: sort according to the importance score of each marker, and include markers one by one in order from high to low until the selected marker set can maximize the coverage of all functional modules. In this process, each new marker added will further improve the coverage of the functional modules through its synergy with the existing markers, ensuring that each selected marker is biologically explanatory and can complement the previously selected markers.

[0262] Once the combination of selected markers can achieve the predetermined maximum functional module coverage, that is, meet the maximum number of clusters, the marker selection process stops. The core logic of this step is to maximize the explanatory coverage of functional modules by minimizing the number of markers, thereby improving the simplicity and diagnostic performance of the model. The detailed process of determining markers using the recursive optimization principle is shown in Figure 6 .

[0263] The results showed that when the top 5 metabolic markers were included, 1 NMF cluster could be covered; when the top 6 markers were included, 2 clusters could be covered; when the top 7 markers were included, 3 clusters could be covered; and when the top 12 markers were included, 4 clusters could be covered. Therefore, in this example, the top 12 potential metabolic markers were finally selected as the final modeling markers, including: nicotinamide, 9,12,13-trihydroxy-octadecenoic acid, mono-2-ethylhexyl phthalate, taurine, sphingosine, traumatic acid, 5'-deoxy-5'-methylthioadenosine, (±)12-hydroxy-5Z,8Z,10E,14Z-eicosatetraenoic acid ((±)12-HETE), hypoxanthine, glycyl-phenylalanine, azelaic acid, and N-phenylacetyl-L-glutamine. The specific information of the above 12 metabolic markers is shown in Table 2.

[0264] Table 2

[0265] Serial number Chinese name Molecular formula CAS number 1 Nicotinamide <![CDATA[C6H6N2O]]> 98-92-0 2 Taurine <![CDATA[C2H7NO3S]]> 107-35-7 3 Monooctyl phthalate <![CDATA[C 16 H 22 O4]]> 4376-20-9 4 Sphingosine <![CDATA[C 18 H 37 NO2]]> 123-78-4 5 9,12,13-Trihydroxy-octadecenoic acid <![CDATA[C 18 H 34 O5]]> 97134-11-7 6 Glycyl-phenylalanine <![CDATA[C 11 H 14 N2O3]]> 3321-03-7 7 N-Phenylacetyl-L-glutamine <![CDATA[C 13 H 16 N2O4]]> 28047-15-6 8 (+ / -)12-HETE <![CDATA[C 20 H 32 O3]]> 71030-37-0 9 5’-Deoxy-5’-methylthioadenosine <![CDATA[C 11 H 15 N5O3S]]> 2457-80-9 10 Traumatic acid <![CDATA[C 12 H 20 O4]]> 6402-36-4 11 Hypoxanthine <![CDATA[C5H4N4O]]> 68-94-0 12 Azelaic acid <![CDATA[C9H 16 O4]]> 123-99-9

[0266] In addition, the expression levels of 12 metabolites in the GC group (gastric cancer group) and the NGC group (non-gastric cancer group) were compared through statistical analysis. The results showed that the expression differences of all metabolites were statistically significant (P<0.001). Among them, 5'-deoxy-5'-methylthioadenosine and N-phenylacetyl-L-glutamine were significantly upregulated in the plasma of the GC group, while the other 10 metabolic markers were significantly downregulated in the plasma of the GC group ( Figure 7 ). The above results indicate that this combination of 12 metabolites can be used as potential metabolic markers for GC diagnosis, providing a new biological basis for the early detection and accurate diagnosis of gastric cancer.

[0267] Overall, through the deep integration of algorithm parameter innovation and biological interpretation, the present invention has promoted NMF and Boruta from tool-level applications to a biology-oriented analysis framework, forming an efficient and optimized marker screening process, which not only improves the biological interpretability of the screened markers but also ensures their statistical importance and model simplicity.

[0268] Example 4: Result comparison of feature importance screening algorithms

[0269] A comparison was made between the Boruta algorithm used in Example 3 and two common feature importance ranking algorithms, LASSO and RF. First, taking the 83 metabolites of 928 samples in Example 2 as an example, according to the importance scores calculated by each algorithm, the top 30 metabolites ranked by importance scores were respectively taken as candidate metabolic markers, and comparative analysis was carried out from the following multiple dimensions:

[0270] (1) Comprehensiveness and novelty of candidate markers

[0271] ① NMF functional module coverage: NMF divides all metabolites into 5 functional modules. The metabolites selected by LASSO only cover 3 functional modules (60%); while the metabolites selected by Boruta and RF algorithms both cover 4 functional modules (80%), indicating that the latter two have higher sensitivity and accuracy in reflecting the complexity and diversity of each functional module in the underlying metabolic network( Figure 8 ).

[0272] ② Ratio of newly identified metabolites: Among the top 30 metabolites selected by each algorithm, all the metabolites selected by the Boruta algorithm are differential metabolites identified by the previous wide-target metabolomics spectrum analysis (the ratio of newly identified metabolites is 100%), while 3 metabolites selected by the LASSO algorithm and the RF algorithm come from the later literature supplement (the ratio of newly identified metabolites is 90%), indicating that the Boruta algorithm can more accurately extract the core metabolites obtained from experimental data and ensure the objectivity and reproducibility of the screening results( Figure 8 ).

[0273] (2) Differential significance of candidate biomarkers

[0274] By comparing the performance of the metabolites selected by each algorithm in statistical indicators (such as the total sum of P-values after FDR correction and the total sum of absolute values of logFC), the results show that the overall difference of the metabolites selected by Boruta is more significant, manifested as the smallest total sum of FDR-P values and the largest total sum of absolute values of logFC, which statistically supports its advantage in group differentiation( Figure 9 ).

[0275] (3) Distribution balance and modeling robustness of biomarker combinations

[0276] Based on the 30 candidate metabolic biomarkers selected by each algorithm, the biomarker combinations for final modeling are determined by the principle of recursive optimization, and the following comparisons are made for these three biomarker combinations:

[0277] ① Distribution balance: In terms of the up- and down-regulation ratios of metabolites, the up- and down-regulation ratio of metabolites in the screening results of the LASSO algorithm is 0:6 (no up-regulated metabolites), the ratio of the screening results of the RF algorithm is 1:8 (including 1 up-regulated metabolite), and the ratio of the screening results of Boruta is 1:5 (including 2 up-regulated metabolites). It can be seen that the Boruta algorithm has a relative advantage in the distribution balance of up- and down-regulation.

[0278] In this example, peripheral venous blood plasma samples were collected from 153 non-gastric cancer subjects and 156 gastric cancer patients from four research centers (Beijing Friendship Hospital, Cancer Hospital of the Chinese Academy of Medical Sciences, Hubei Provincial People's Hospital, Tongji Hospital Affiliated to Tongji Medical College of Huazhong University of Science and Technology) as an independent validation cohort. The inclusion and exclusion criteria for all subjects, the targeted detection method for plasma samples, and the data preprocessing process were all consistent with those in Example 1 to ensure data comparability and analysis reliability.

[0279] ② Model robustness: The modeling cohort in Example 2 was divided into a training set and a test set at a ratio of 3:1. The RF model was used to perform diagnostic modeling on the above three metabolic marker combinations in the training set and was verified on the independent validation cohort. The results showed that, compared with the other two algorithms, the combination of 12 markers screened by the Boruta algorithm had the highest AUC value, accuracy, and specificity in the test set, indicating that this marker combination had obvious advantages in modeling robustness ( Figure 10 ).

[0280] Example 5: Establishment and verification of a gastric cancer diagnosis model based on multiple machine learning algorithms

[0281] 1 Model establishment and performance comparison based on 10 machine learning algorithms

[0282] To improve the early diagnosis efficiency of gastric cancer and optimize the modeling method of multi-marker combinations, in this example, 10 advanced machine learning algorithms (RF, TreeNet, GBM, XGBoost, SVM Radial, LASSO, LR, KNN, NNET, NB) were used to model the 12 metabolic markers (Table 2) screened by the above BIO-FIRE algorithm.

[0283] During the data analysis process, the train function in the caret package was used for model training, and the method parameter was set to select different machine learning algorithms. Hyperparameter optimization was performed using 10-fold cross-validation with trainControl(method = "cv", number = 10, classProbs = TRUE) to avoid model overfitting. When dividing the data set, to ensure the generalization ability of the model, in this example, the modeling cohort in Example 2 was divided into a training set and a test set at a ratio of 3:1. The training set was used for modeling, and the model performance was evaluated in the test set.

[0284] Model performance was evaluated using the receiver operating characteristic (ROC) curve and the area under the curve (AUC). The results showed that the gastric cancer diagnosis models established by 10 machine learning algorithms performed well in differentiating gastric cancer patients from non-gastric cancer patients. The AUC of the test set ranged from 0.877 to 0.921, the sensitivity ranged from 0.783 to 0.879, and the specificity ranged from 0.795 to 0.923, indicating that the models had good stability and generalization ability. Figure 11 A). To select the optimal machine learning algorithm, a radar chart was used to comprehensively compare the sensitivity, specificity, accuracy, F1 value, positive predictive value, and negative predictive value of the models constructed by 10 machine learning algorithms. The results showed that the RF model had the best comprehensive performance. Figure 11 B), so it was selected as the final gastric cancer diagnosis model.

[0285] 2 Independent verification of the optimal diagnosis model based on 12 metabolic markers

[0286] To further evaluate the stability and generalization ability of the optimal diagnosis model based on 12 metabolic markers constructed in this example, the independent verification cohort in Example 4 was used for verification.

[0287] In the independent verification cohort, the optimal random forest (RF) model was externally verified. The results showed that the AUC (area under the curve) of this model could reach 0.95, the diagnostic accuracy was 90.3%, and both the sensitivity and specificity exceeded 90%. It had a significant advantage over the diagnostic performance of traditional tumor diagnostic markers. Figure 11 C-D), indicating that the model still had excellent discriminatory ability and good clinical application potential in the independent population, and could provide reliable auxiliary decision-making support for the early screening and precise diagnosis of gastric cancer.

[0288] Example 6: Influence of the number of covered functional modules on the diagnostic efficacy of gastric cancer

[0289] To further confirm the advantages of the metabolic marker combinations covering functional modules determined in this application, the diagnostic efficacy of the metabolic marker combinations covering multiple functional modules and those covering a single functional module was compared in this example. The metabolic marker combination covering 4 functional modules (nicotinamide + traumatic acid + 5'-deoxy-5'-methylthioadenosine + N-phenylacetyl-L-glutamine) and the metabolic marker combination covering a single functional module (9,12,13-trihydroxy-octadecenoic acid + 2-aminoethanesulfonic acid / taurine + hypoxanthine + azelaic acid) were modeled using the RF algorithm (modeled according to the method in Example 5). The samples were selected from the independent verification cohort described in Example 4. The results are as Figure 12 shown.

[0290] Example 7: Application of a combination of 12 metabolic markers in complex clinical scenarios

[0291] In clinical practice, the early diagnosis and subtype identification of gastric cancer still face many challenges, especially in the following complex or special diagnostic scenarios: (1) Ultra-early gastric cancer (stage IA gastric cancer): Due to the low tumor burden, it is often difficult for traditional imaging and serum markers to accurately identify. At the same time, these patients are absolute indications for endoscopic resection of early gastric cancer. Early diagnosis and treatment are of great significance for improving the survival rate of gastric cancer. (2) Gastric cancer with negative traditional tumor markers (CA19-9, CA72-4, CEA): These patients cannot be effectively screened by existing serological methods and are easily missed. (3) Gastric cancer with a tumor long diameter < 4 cm: Smaller tumors may not be obvious in imaging or traditional biomarker detection and are easily missed. (4) Gastroesophageal junction cancer: This subtype has intersections with both gastric cancer and esophageal cancer in anatomical and biological characteristics, increasing the diagnostic difficulty.

[0292] To verify the diagnostic value of the 12-metabolic marker combination (Table 2) constructed in this example in the above complex clinical scenarios, this example also collected the clinical information of the patients in the internal training set, internal test set in Example 2, and the independent validation cohort in Example 4, and respectively used the patients with ultra-early gastric cancer (stage IA gastric cancer) in the internal training set, internal test set, and independent validation cohort and the non-gastric cancer group as the cohort for modeling or validating ultra-early gastric cancer (stage IA gastric cancer), the patients with gastric cancer with negative traditional tumor markers and the non-gastric cancer group as the cohort for modeling or validating gastric cancer with negative traditional tumor markers, the patients with gastric cancer with a tumor long diameter < 4 cm and the non-gastric cancer group as the cohort for modeling or validating gastric cancer with a tumor long diameter < 4 cm, and the patients with gastroesophageal junction cancer and the non-gastric cancer group as the cohort for modeling or validating gastroesophageal junction cancer, and used 10 machine learning algorithms to build models in the internal training set and validate in the internal test set and independent validation cohort to ensure the comparability and stability of the results.

[0293] In the independent validation cohort, the RF diagnostic models based on 12 metabolic markers all showed high discrimination ability. The specific AUC results of the independent validation cohort are shown in Figure 13 .

Claims

1. A method for screening tumor metabolic markers, characterized in that: The screening method comprises: S1) Screening metabolites from samples of tumor patients and non-tumor control group, and performing qualitative and quantitative analysis; S2) Using the Biomarker Identification and Optimization algorithm based on Function and Feature Importance Recursive Enhancement (BIO-FIRE) for feature extraction, tumor metabolism markers were further identified from the initial screening metabolites; The BIO-FIRE algorithm includes: A) Non-negative matrix factorization (NMF) was used to cluster the functional modules of the initial screening metabolites; B) Boruta algorithm performs feature screening by calculating importance scores; C) Based on the recursive optimization principle, the importance score of each marker calculated in B) is ranked from high to low, and the metabolite marker subsets are selected in sequence starting from the first one, until the metabolite subset that can maximize the coverage of the most functional modules is selected, which is the tumor metabolic marker.

2. The screening method according to claim 1, characterized in that Said S1) comprises: 1) Using broad targeted metabolomics to construct a tumor-specific metabolite database and identify differential metabolites; preferably, using a quadrupole time-of-flight mass spectrometer (QTOF) in full scan mode to detect samples to obtain metabolite ion pair signals, and further transferring the metabolite ion pair signals to a tandem quadrupole composite linear ion trap mass spectrometer (QTRAP6500) to obtain a tumor-specific metabolite database and identify differential metabolites; 2) Targeted metabolomics screening to obtain preliminary screening metabolites and quantify them; preferably, the detection range of the targeted metabolomics includes the differential metabolites in step 1); Preferably, liquid chromatography is used to separate differential metabolites and mass spectrometry is used to quantify the initial screening metabolites, wherein the liquid chromatography is performed using a T3 column and the mass spectrometry is performed using a T3 column.

3. The screening method according to claim 1, characterized in that The NMF introduces KL divergence optimization strategy to strengthen the sparse association between metabolites and functional modules; Preferably, the objective function is: Where: X represents the original metabolic data matrix, with a dimension of n×m (n: number of metabolites, m: number of samples); W represents the weight matrix, with the dimension of n × r (r: number of functional modules), which indicates the contribution of metabolites to the functional modules; H represents the activation matrix, with a dimension of r×m, which represents the distribution of functional modules in the sample; Represents the square of the Frobenius norm, which is used to measure the matrix reconstruction error λ1, λ2 represent L1 regularization parameters, controlling the sparsity of W and H; λ3 represents the KL divergence regularization parameter, which adjusts the strength of the sparsity constraint; Represents a sparse prior matrix (such as generated by thresholding or low-rank templates), guiding the sparse distribution of the weight matrix W; Represents the Kullback-Leibler divergence, which measures the weight matrix W and the prior sparse matrix differences; i, j represent the row and column indices of the matrix (i: metabolite index, j: functional module or sample index).

4. The screening method according to any one of claims 1 to 3, characterized in that: The NMF comprises: determining the optimal rank by multi-rank synchronous calculation combined with residual square sum inflection point analysis; Preferably, for candidate rank r∈{3,4,…,8}, the NMF decomposition is computed in parallel: Among them, r represents the candidate decomposition rank (number of functional modules), ranging from 3 to 8; W r， H r Represents the weight matrix and activation matrix when the rank is r; X is the original metabolic data matrix (rows represent metabolites and columns represent samples); represents the Frobenius norm, which is used to measure the error between X and WH; RSS stands for Residual Sum of Squares, that is Used to evaluate the fitting effects of different ranks; The inflection point is the point in the RSS curve where the slope changes the most and is used to automatically select the optimal rank (eg, r=5).

5. The screening method according to any one of claims 1 to 4, characterized in that: In the Boruta algorithm, the splitting criterion of feature Xj at decision tree node t is: Where j represents the feature index (such as the number of a gene or module); t represents the current node in the decision tree; represents the Gini impurity of node t; C represents the total number of categories of the target variable (in classification tasks); p(c|t) represents the proportion of samples in node t belonging to category c; N represents the total number of samples of node t; N L , N R Indicates the number of samples of the left and right child nodes after splitting; t L , t R Indicates left and right child nodes; ΔGini(j,t): The reduction in Gini impurity brought by feature j at node t.

6. The screening method according to any one of claims 1 to 5, characterized in that: The tumor is a hematological tumor or a solid tumor; Preferably, the tumor is gastric cancer, and the functional modules include a glutamine metabolism functional module, a lipid metabolism functional module, a host-microorganism co-metabolism functional module, a metabolic network bridge functional module and a branched-chain amino acid metabolism functional module.

7. The screening method according to any one of claims 1 to 6, characterized in that: The sample is a body fluid; Preferably, the body fluid is selected from blood, plasma or serum; Preferably, the number of tumor metabolism markers is 10-15, preferably 12, and further preferably includes nicotinamide, taurine, monoethylhexyl phthalate, sphingosine, 9,12,13-trihydroxy-octadecenoic acid, glycyl-phenylalanine, N-phenylacetyl-L-glutamine, (±)12-hydroxy-5Z,8Z,10E,14Z-eicosatetraenoic acid ((±)12-HETE), 5'-deoxy-5'-methylthioadenosine, callous acid, hypoxanthine and azelaic acid.

8. A computer-readable storage medium or a device comprising a computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program includes a BIO-FIRE algorithm, and the BIO-FIRE algorithm includes: A) Non-negative matrix factorization (NMF) was used to cluster the functional modules of the initial screening metabolites; B) Boruta algorithm performs feature screening by calculating importance scores; C) Based on the recursive optimization principle, the importance score of each marker calculated in B) is ranked from high to low, and the metabolite marker subsets are selected in sequence starting from the first one, until the metabolite subset that can maximize the coverage of the most functional modules is selected, which is the tumor metabolic marker.

9. A method for constructing a tumor diagnostic model, characterized in that: The construction method comprises: i) collecting the concentration test results of metabolic markers in the tumor patient group and the non-tumor control group; ii) constructing a tumor diagnosis model based on the information collected in step i); Alternatively, the construction method comprises: 1) Collecting samples from subjects and detecting the concentration of metabolic markers; II) clinically diagnosing the subjects and dividing them into a tumor patient group and a non-tumor control group; III) constructing a tumor diagnosis model according to the detection results of step I) and the diagnosis results of step II); The metabolite marker is a metabolite marker screened by the screening method according to any one of claims 1 to 7; Preferably, the tumor is a blood tumor or a solid tumor; further preferably, the tumor is gastric cancer; Preferably, the sample is a body fluid; further preferably, the body fluid is selected from blood, plasma or serum; Preferably, the algorithm used to construct the diagnostic model is a machine learning model, preferably including one or more of logistic regression (LR), k-nearest neighbor algorithm (KNN), naive Bayes (NB), support vector machine (SVM Radial), random forest (RF), XGBoost, TreeNet, gradient boosting machine (GBM), LASSO or neural network (NNET).

10. Use of a metabolic marker in the preparation of a product for diagnosing gastric cancer, characterized in that: The metabolite markers include nicotinamide, taurine, monoethylhexyl phthalate, sphingosine, 9,12,13-trihydroxy-octadecenoic acid, glycyl-phenylalanine, N-phenylacetyl-L-glutamine, (±)12-hydroxy-5Z,8Z,10E,14Z-eicosatetraenoic acid ((±)12-HETE), 5'-deoxy-5'-methylthioadenosine, callous acid, hypoxanthine and azelaic acid; Preferably, the metabolite marker is a metabolite marker in plasma.

Images