Cancer prediction system and method based on metabonomics mass spectrum data parameterization

Abstract

The invention further provides a cancer prediction system based on metabonomics mass spectrum data parameterization. The cancer prediction system comprises a sample collection module used for collecting detection samples, a mass spectrum detection module used for conducting mass spectrum scanning, a mass spectrum data processing module and a data classification module. The mass spectrum data processing module is used for performing signal selective addition on the mass spectrum signals, and fitting and classifying the added mass spectrum signals to obtain metabolite pathway characteristics of a detection sample; and the data classification module adopts a machine learning algorithm to construct a prediction model according to the metabolite pathway characteristics to obtain a final prediction result. The prediction system can improve the prediction accuracy, and solves the problem that the accuracy is low when serum metabolome is used for predicting cancers.

Description

Technical Field

[0001] This invention relates to the field of medical model construction technology, specifically to a cancer prediction system and method based on metabolomics mass spectrometry data parameterization. Background Technology

[0002] Malignant tumors are common medical diseases. Current imaging examinations (CT / MRI) have insufficient sensitivity (<50%) for sub-centimeter lesions, and tissue biopsies are invasive and difficult to perform dynamic monitoring. Developing highly sensitive, non-invasive early diagnostic technologies has become a core requirement for precision oncology.

[0003] CN118604104A discloses a mass spectrometry analysis system for endometrial cancer detection based on multi-module machine learning. It utilizes liquid chromatography-tandem mass spectrometry (LC-MS) to acquire metabolic and peptide molecular fingerprint information of the sample, and combines multiple machine learning methods to discover cancer-related features. Based on these features and the machine learning methods, a classifier is constructed for the identification of endometrial cancer. Finally, the judgment results of multiple independent models are integrated to obtain the final identification result. Specifically, it includes a sample acquisition module, a sample preprocessing module, a mass spectrometer, a mass spectrometry data processing and conversion module, a machine learning algorithm module, and a comprehensive analysis module. This invention breaks through the traditional approach of being limited to a single biomarker, a single sample, or a single machine learning module, effectively improving the diagnostic performance for endometrial cancer, colorectal cancer, lung cancer, and other cancers.

[0004] CN118280558A discloses a cancer diagnostic analysis method and system based on mass spectrometry data. This addresses the problem that when using black-box models as diagnostic systems, doctors often struggle to understand the reasons behind the model's predictions. Specifically, the method involves: using a file preprocessing algorithm to read mass spectrometry scan files from the same directory folder to obtain mass spectrometry data; then organizing the parameterized mass spectrometry data into a .csv format for use; pre-training an extreme gradient boosting algorithm on the mass spectrometry data using a Bayesian optimization algorithm; optimizing and selecting the optimal parameters based on multiple pre-training results; adjusting the extreme gradient boosting algorithm parameters based on the retained optimal parameters; using the pre-trained model to perform diagnostic classification using mass spectrometry data from healthy and cancer patients, and providing diagnostic results; and using the SHAP algorithm to analyze the model's diagnostic process, identifying the features in the mass spectrometry data upon which the model's diagnostic results are based, and outputting thirty features that significantly influence the diagnostic results.

[0005] Metabolomics analyzes small molecule metabolites (≤1500 Da) in biological fluids (serum / urine) to directly reflect the pathophysiological state of the body. Compared with genomics / proteomics, it has advantages such as real-time dynamic monitoring, easy sample acquisition, and controllable cost.

[0006] However, three major technical bottlenecks remain in this field: the challenge of biomarker screening, with over 10,000 types of body fluid metabolites and concentrations spanning nine orders of magnitude, making it difficult for traditional univariate analysis to distinguish between disease-specific fluctuations and individual physiological variations; model generalization defects, with existing machine learning models (such as support vector machines (SVM) and random forests) generally showing a decrease in area under the curve (AUC) of 0.15-0.2 during cross-center validation, due to insufficient batch effect correction and lack of modeling of nonlinear metabolic interactions; and barriers to clinical translation, with most studies limited to a single cancer type, lacking the ability to differentiate between multiple cancer types, and without establishing a quantitative association model with TNM staging (tumor node metastasis classification). Summary of the Invention

[0007] This invention addresses the systematic errors introduced by the need for metabolite annotation in traditional metabolomics by providing a cancer diagnostic system based on metabolomics mass spectrometry data parameterization. This prediction system can improve prediction accuracy and solve the problem of low accuracy in cancer prediction using serum metabolomics.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A cancer prediction system based on metabolomics mass spectrometry data parameterization includes a sample acquisition module, a mass spectrometry detection module, a mass spectrometry data processing module, and a data classification module; The sample acquisition module is used to collect test samples; The mass spectrometry detection module is used to perform mass spectrometry scanning on the test sample to obtain a mass spectrometry signal; The mass spectrometry data processing module is used to perform signal selection summation on the mass spectrometry signal, and after summation, the mass spectrometry signal is fitted and classified to obtain the metabolite pathway characteristics of the detected sample; The data classification module uses machine learning algorithms to construct a prediction model based on the characteristics of the metabolite pathways, and obtains the final prediction result.

[0009] This invention pioneers a mass spectrometry data parameterization method, utilizing all the information contained in the mass spectrometry scan data, thus providing a superior replacement for the peak alignment and metabolite annotation steps required in traditional metabolomics. Through machine learning, a model is built to achieve early differential diagnosis and prediction of various cancers, including prostate cancer, endometrial cancer, colon cancer, and kidney cancer. The innovative application of reverse feature engineering reveals novel disease pathways in cancer.

[0010] The test sample includes one or more of blood samples, tissue samples, or urine samples; preferably, the test sample includes a tumor sample; the tumor sample includes one or more of colon cancer, cervical cancer, ovarian cancer, prostate cancer, renal cell carcinoma, benign tumor, and clear renal cell carcinoma.

[0011] Preferably, the mass spectrometry data processing module performs selective signal summation on the mass spectrometry signal, and the summed mass spectrometry signal is fitted and classified to obtain the metabolite pathway characteristics of the detected sample, specifically including: S1-1, calculate the mass-to-charge ratio m / z of the sum of all metabolites in the biological pathway; S1-2, with the mass-to-charge ratio m / z as the center and 0-50ppm as the radius, the mass spectrometry signals obtained by the mass spectrometry detection module are summed to obtain the summed mass spectrometry signal; S1-3, the modified expectation-maximum algorithm is used to fit the summed mass spectrometry signal with radial basis function (RBF) to obtain the fitted metabolite characteristic parameters; S1-4: Based on the biological pathway, the fitted metabolite characteristic parameters are classified to obtain metabolite pathway characteristics.

[0012] Preferably, before performing signal selection summation on the mass spectrometry signal, the mass spectrometry data processing module normalizes the retention time (rt) of the mass spectrometry signal to 0-30 minutes; that is, it uses the readMSData function of the MSnbase package in R to obtain the rt list of the mass spectrometry scan file, and sets rt=(rt-rt[1]) / rt[length(rt)]*1800 to 0-30 minutes. Where rt[1] is the start time of the mass spectrometry scan, and rt[length(rt)] represents the end time of the mass spectrometry scan.

[0013] Preferably, the fitted metabolite characteristic parameters are obtained by performing radial basis function fitting on the summed mass spectrometry signals using a modified expectation-maximum algorithm, specifically including the following steps: S2-1, sampling the core parameter c of the radial basis function of each metabolite mass spectrum signal based on the summed mass spectrum signal intensity; S2-2, Initialize the width parameter of the radial basis function; S2-3 uses multiple linear regression to obtain the weight parameters of the radial basis function, and uses gradient descent to optimize the width parameter initialized in S2-2, so that the width parameter is limited to the range (5, 60). Repeat steps S2-2 and S2-3 1-5 times to obtain the width parameter of the radial basis function of each metabolite mass spectrometry signal. Fit the mass spectrometry signal of each metabolite according to the weight parameters obtained in S2-3 to obtain the fitted metabolite characteristic parameters.

[0014] This invention improves the Expectation-Maximization (EM) algorithm by splitting the M-step (Maximization Step). The original M-step maximum likelihood method estimates all values ​​of the implicit function, but its convergence requires a large number of computational steps, resulting in slow computation. This new method splits the M-step into three steps, S2-1 to S2-3, to calculate the center, width, and height parameters respectively, thus accelerating the computation. Furthermore, the number of computational steps required for convergence can be specified, achieving efficient and accurate calculation of metabolite characteristic parameters.

[0015] Preferably, when the mass spectrometry data processing module adds the mass spectrometry signal, for metabolites with a mass-to-charge ratio (m / z) greater than 400, the ppm of the metabolite is calculated as 1ppm = m / z / 1000000; for metabolites with a mass-to-charge ratio less than or equal to 400, the ppm of the metabolite is calculated as 1ppm = 400 / 1000000, in order to increase the acquisition window for metabolite molecules with a smaller mass-to-charge ratio.

[0016] The machine learning algorithms include any one or more of the following constructed using Python's sklearn package: Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Random Forest, and LightGBM; or any one or more of the following constructed using Python's pytorch package: Graph Attention Network (GAT), Graph Convolutional Network (GCN), and Transformer. Preferably, the random forest model is constructed using the Random Forest Classifier function under the Ensemble Learning module of the scikit-learn machine learning library in Python.

[0017] Preferably, the transformer model and the MLP model are written and constructed using the torch framework in Python.

[0018] This invention also provides a cancer prediction method based on metabolomics mass spectrometry data parameterization, comprising the following steps: Step 1: Align the mass spectrometry scan time with the reference and perform a mass spectrometry scan on the sample to obtain the mass spectrometry signal; Step 2: Calculate the mass-to-charge ratio m / z of the sum of all metabolites in the biological pathway; Step 3: Sum the mass spectrometry signals obtained in Step 1 with the mass-to-charge ratio m / z in Step 2 as the center and 0-50ppm as the radius to obtain the summed mass spectrometry signal. Step 4: Use the improved expectation-maximum algorithm to perform radial basis function fitting on the summed mass spectrometry signal to obtain the fitted metabolite characteristic parameters; Step 5: Classify the fitted metabolite characteristic parameters according to the biological pathway to obtain the pathway characteristics of the metabolites; Step 6: Based on the pathway characteristics of metabolites, a prediction model is constructed using machine learning algorithms to obtain the final prediction results.

[0019] Preferably, in step 2, all metabolites of the biological pathway include all metabolites with mass-to-charge ratios recorded in any one of the databases such as KEGG, hmdb, and BIGG, including [M + H] + [M] + Na] + [M] + K] + [M] - H] - [M] + CH3COO - ]wait.

[0020] When the biological pathway is all metabolites with mass-to-charge ratio recorded in the KEGG database, preferably, in step 2 all metabolites refer to human metabolic pathway information obtained from the KEGG database through the KEGGRREST package in R language, and all metabolites whose actual molecular weight is not recorded are deleted. The actual molecular weight of hydrogen is 1.007825.

[0021] The sample data used in this invention comprises clinical cancer classification datasets such as MTBLS6039, MTBLS3444, MTBLS1122, MTBLS3838, ST001705, ST002521, and MTBLS4294. These datasets include mass spectrometry data from serum or tissue samples of one or more cancers, such as colon cancer, cervical cancer, ovarian cancer, prostate cancer, renal cell carcinoma, benign tumors, and clear renal cell carcinoma, as well as serum or tissue samples from healthy individuals. Sample features are extracted through steps 2-5, and the above eight types are distinguished using the method described in step 6. The overall accuracy rate reaches over 98%. This invention can identify the progression stage of cancer, providing an accurate tool for early detection and diagnosis of cancer.

[0022] Compared with the prior art, the present invention has the following beneficial effects: (1) In this invention, the mass spectrometry data parameterization method is used to construct a cancer prediction system by utilizing all the information contained in the mass spectrometry scan data, avoiding systematic errors introduced by metabolite annotation, and significantly improving the accuracy of model prediction.

[0023] (2) This invention utilizes the metabolic characteristics of biological metabolic networks by taking the metabolite characteristics of pathways as units, which is beneficial for the analysis and research of cancer mechanisms. Attached Figure Description

[0024] Figure 1This is a schematic diagram illustrating the process of using the cancer prediction system of the present invention for prediction in Examples 1-6.

[0025] Figure 2 This is the classification verification result of the MTBLS1122 dataset in Example 1.

[0026] Figure 3 This is the classification verification result of the MTBLS3444 dataset in Example 2.

[0027] Figure 4 This is the classification verification result of the MTBLS3838 dataset in Example 3.

[0028] Figure 5 This is the classification verification result of dataset ST002521 in Example 4.

[0029] Figure 6 This is the classification verification result of dataset ST001705 in Example 5.

[0030] Figure 7 This is the classification verification result of the MTBLS6039 dataset in Example 6.

[0031] Figure 8 This is the classification verification result of multiple datasets in Example 7. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Modifications or equivalent substitutions made by those skilled in the art based on their understanding of the technical solutions of this invention, without departing from the spirit and scope of the invention, should be covered within the protection scope of this invention.

[0033] All raw materials used in the following specific implementation methods were purchased from the market.

[0034] Example 1 A cancer prediction system based on metabolomics mass spectrometry data parameterization includes a sample acquisition module, a mass spectrometry detection module, a mass spectrometry data processing module, and a data classification module; the system is used to make predictions on the MTBLS1122 dataset.

[0035] The sample collection module is used to collect test samples; the data source for this embodiment is the dataset MTBLS1122 from the MetaboLights website. The dataset includes 197 colon tumor samples, including ascending colon cancer and descending colon cancer, and 39 normal tissue samples are added to the total dataset.

[0036] The mass spectrometry detection module is used to perform mass spectrometry scanning on the sample to obtain the mass spectrometry signal. The specific conditions for mass spectrometry scanning are: solvent flow rate of 0.5 mL / min and sample injection volume of 1 μL. For HPLC-MS analysis, mobile phase A is an aqueous solution of 25 mM NH4OH and 25 mM NH4OAc, while mobile phase B is acetonitrile in electrospray ionization (ESI) positive and negative ion modes, respectively. The linear gradient settings are as follows: 0-0.5 min: 95% B; 0.5-7 min: 95%-65% B; 7-8 min: 65%-40% B; 8-9 min: 40% B; 9-9.1 min: 40%-95% B; 9.1-12 min: 95% B, and the obtained mass spectrometry signal is shown.

[0037] The mass spectrometry data processing module performs selective summation on the mass spectrometry signals, and then fits and classifies the summed mass spectrometry signals to obtain the metabolite pathway characteristics of the detected samples. Specific steps include: Step 1: The mass spectrometry processing module first aligns the retention time from 12 minutes to the baseline of 30 minutes; Step 2: Calculate the mass-to-charge ratio m / z of the summation ion [M+H]+ of all metabolites recorded in the KEGG database, and sum the mass spectrometry signals aligned in Step 1 with a radius of 0-50 ppm to obtain the summation mass spectrometry signal; Step 3: Use the improved expectation-maximum algorithm to fit the radial basis function of the neutralized mass spectrometry signal to obtain the fitted metabolite characteristic parameters; Step 4: Based on the biological pathways recorded in the KEGG database, classify the fitted metabolite characteristic parameters to obtain metabolite pathway characteristics.

[0038] Step 5: Based on the pathway characteristics of metabolites, the data classification module uses the sklearn package in Python to construct Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting Tree (XGB), and LightGBM respectively to predict and classify cancer and normal human samples. The classification ability of each model is verified using ten-fold cross-validation. The analysis process is as follows: Figure 1 As shown, the results are as follows Figure 2 As shown, the highest accuracy of 98% can be achieved using the XGB and LightGBM classifiers, the highest area under the curve of 0.98 is achieved using the LightGBM classifier, the highest F1 score of 0.98 is achieved using the XGB and LightGBM classifiers, and the highest recall of 0.98 is achieved using the XGB and LightGBM classifiers.

[0039] Example 2 The same prediction system as in Example 1 was used for prediction. The data source for this example is the dataset MTBLS3444 from the Metabolic Lamp website. The dataset includes 396 endometrial cancer samples and 225 healthy volunteer samples added to the total dataset.

[0040] The mass spectrometry detection module is used to perform mass spectrometry scanning on the test sample to obtain the mass spectrometry signal. The specific conditions for mass spectrometry scanning are as follows: gradient elution of the analyte is performed using a 0.1% formic acid aqueous solution (A2) and a 0.1% formic acid acetonitrile solution (B2), or a 5 mM ammonium formate aqueous solution (A3) and acetonitrile (B3) at a flow rate of 0.25 ml / min. After equilibration, 2 μL of each sample is injected. The linear gradient increase of solvent B2 / B3 (v / v) is as follows: 0-1 min, 2% B2 / B3; 1-9 min, 2-50% B2 / B3; 9-12 min, 50-98% B2 / B3; 12-13.5 min, 98% B2 / B3; 13.5-14 min, 98-2% B2 / B3; 14-20 min, 2% B2 positive model (14-17 min, 2% B3 negative model), and the obtained mass spectrometry signal is obtained.

[0041] The mass spectrometry data processing module performs selective summation on the mass spectrometry signals, and then fits and classifies the summed mass spectrometry signals to obtain the metabolite pathway characteristics of the detected samples. Specific steps include: Step 1: The mass spectrometry processing module first aligns the retention time from 17 minutes to the baseline of 30 minutes; Step 2: Calculate the mass-to-charge ratio m / z of the summation ion [M+H]+ of all metabolites recorded in the KEGG database, and sum the mass spectrometry signals aligned in Step 1 with a radius of 0-50 ppm to obtain the summation mass spectrometry signal; Step 3: Use the improved expectation-maximum algorithm to fit the radial basis function to the summed mass spectrometry signal to obtain the fitted metabolite characteristic parameters; Step 4: Based on the biological pathways recorded in the KEGG database, classify the fitted metabolite characteristic parameters to obtain the pathway characteristics of the metabolites.

[0042] The data classification module uses the sklearn package in Python to build support vector machines, random forests, extreme gradient boosting trees, and LightGBM classifiers to classify cancer and normal human samples. The classification ability of the classifiers is verified by ten-fold cross-validation.

[0043] The results are as follows Figure 3As shown, the highest accuracy of 94% can be achieved using the XGB and LightGBM classifiers, the highest area under the curve of 0.81 is achieved using the XGB classifier, the highest F1 score of 0.93 is achieved using the XGB classifier, and the highest recall of 0.94 is achieved using the XGB and LightGBM classifiers.

[0044] Example 3 The same prediction system as in Example 1 was used for prediction. The data source for this example is the dataset MTBLS3838 from the Metabolic Lamp website. The dataset includes 112 clear renal cell carcinoma samples and 52 healthy volunteer samples added to the total dataset.

[0045] The mass spectrometry detection module is used to perform mass spectrometry scanning on the sample to obtain the mass spectrometry signal. The specific conditions for mass spectrometry scanning are as follows: Acetonitrile:water (60:40 v / v) is mixed with 10 mM ammonium acetate and 0.1% acetic acid (mobile phase A), and isopropanol:acetonitrile (90:10 v / v) is mixed with 10 mM ammonium acetate and 0.1% acetic acid (mobile phase B). The program is as follows: 0.0-2.0 min, 40-43% B; 2.0-2.1 min, 43-50% B; 2.1-12.0 min, 50-54% B; 12.0-12.1 min, 54-70% B; 12.1-18.0 min, 70-99% B; 18.0-19.0 min, 99% B. The flow rate is kept constant at 0.4 mL / min for 12 min, and then reduced to 0.30 mL / min from 12 to 19 min. The gradient was then restored to its initial conditions within 11 minutes. The injection volume was 2 μL. The temperatures of the column and autosampler tray were set to 55 °C and 5 °C, respectively, to obtain the mass spectrometry signal.

[0046] The mass spectrometry data processing module performs selective summation on the mass spectrometry signals, and then fits and classifies the summed mass spectrometry signals to obtain the metabolite pathway characteristics of the detected samples. Specific steps include: Step 1: The mass spectrometry processing module first aligns the retention time from 19 minutes to the baseline of 30 minutes; Step 2: Calculate the mass-to-charge ratio m / z of the summation ion [M+H]+ of all metabolites recorded in the KEGG database, and sum the mass spectrometry signals aligned in Step 1 with a radius of 0-50 ppm to obtain the summation mass spectrometry signal; Step 3: Use the improved expectation-maximum algorithm to fit the radial basis function to the summed mass spectrometry signal to obtain the fitted metabolite characteristic parameters; Step 4: Based on the biological pathways recorded in the KEGG database, classify the fitted metabolite characteristic parameters to obtain the pathway characteristics of the metabolites.

[0047] The classification module uses the sklearn package in Python to build support vector machines, random forests, extreme gradient boosting trees, and LightGBM classifiers to classify cancer and normal human samples. The classification ability of the classifiers is verified by ten-fold cross-validation.

[0048] The results are as follows Figure 4 As shown, the highest accuracy of 83% can be achieved using the LightGBM classifier, the highest area under the curve of 0.7 is achieved using the XGB and LightGBM classifiers, the highest F1 score of 0.80 is achieved using the XGB and LightGBM classifiers, and the highest recall of 0.83 is achieved using the LightGBM classifier.

[0049] Example 4 The same prediction system as in Example 1 was used for prediction. The data source for this example is dataset ST002521 from the Metabolomics Workbench website. The dataset includes 208 ovarian cancer samples and 137 healthy volunteer samples added to the total dataset.

[0050] The mass spectrometry detection module is used to perform mass spectrometry scanning on the sample to obtain the mass spectrometry signal. The specific conditions for mass spectrometry scanning are as follows: 10 mM ammonium acetate, water / acetonitrile (40:60 v / v) (mobile phase A), 10 mM ammonium acetate and 2-isopropanol / acetonitrile (90:10 v / v). 0.1% formic acid (mobile phase B). Flow gradient: 0-1 min 80-40% A; 1-5 min 40-30% A; 5-5.5 min 30-15% A; 5.5-8 min 15-10% A; 8-8.2 min 10-0% A; 8.2-10.5 min 0% A; 10.5-10.7 min 0-80% A; 10.7-12.0 min 80% A. Column temperature 50℃.

[0051] The mass spectrometry data processing module performs selective summation on the mass spectrometry signals, and then fits and classifies the summed mass spectrometry signals to obtain the metabolite pathway characteristics of the detected samples. Specific steps include: Step 1: The mass spectrometry processing module first aligns the retention time from 12 minutes to the baseline of 30 minutes; Step 2: Calculate the mass-to-charge ratio m / z of the summation ion [M+H]+ of all metabolites recorded in the KEGG database, and sum the mass spectrometry signals aligned in Step 1 with a radius of 0-50 ppm to obtain the summation mass spectrometry signal; Step 3: Use the improved expectation-maximum algorithm to fit the radial basis function of the summed mass spectrometry signal in Step 3 to obtain the fitted metabolite characteristic parameters; Step 4: Based on the biological pathways recorded in the KEGG database, classify the fitted metabolite characteristic parameters to obtain the pathway characteristics of the metabolites.

[0052] The data classification module uses the sklearn package in Python to build support vector machines, random forests, extreme gradient boosting trees, and LightGBM classifiers to classify cancer and normal human samples. The classification ability of the classifiers is verified by ten-fold cross-validation.

[0053] The results are as follows Figure 5 As shown, the LightGBM classifier can achieve the highest accuracy of 80%, the highest area under the curve of 0.53, the highest F1 score of 0.72, and the highest recall of 0.80.

[0054] Example 5 The same prediction system as in Example 1 was used for prediction. The data source for this example is dataset ST001705 from the Metabolomics Workbench website. The dataset includes 82 renal cell carcinoma samples and 174 healthy volunteer samples added to the total dataset.

[0055] The mass spectrometry detection module is used to perform mass spectrometry scanning on the sample to obtain the mass spectrometry signal. The specific conditions for mass spectrometry scanning are as follows: 10 mM ammonium acetate, water / acetonitrile (40:60 v / v) (mobile phase A), 10 mM ammonium acetate and 2-isopropanol / acetonitrile (90:10 v / v). 0.1% formic acid (mobile phase B). Flow gradient: 0-1 min 80-40% A; 1-5 min 40-30% A; 5-5.5 min 30-15% A; 5.5-8 min 15-10% A; 8-8.2 min 10-0% A; 8.2-10.5 min 0% A; 10.5-10.7 min 0-80% A; 10.7-12.0 min 80% A. Column temperature 50℃.

[0056] The mass spectrometry data processing module performs selective summation on the mass spectrometry signals, and then fits and classifies the summed mass spectrometry signals to obtain the metabolite pathway characteristics of the detected samples. Specific steps include: Step 1: The mass spectrometry processing module first aligns the retention time from 12 minutes to the baseline of 30 minutes; Step 2: Calculate the mass-to-charge ratio m / z of the summation ion [M+H]+ of all metabolites recorded in the KEGG database, and sum the mass spectrometry signals aligned in Step 1 with a radius of 0-50 ppm to obtain the summation mass spectrometry signal; Step 3: Use the improved expectation-maximum algorithm to fit the radial basis function to the summed mass spectrometry signal to obtain the fitted metabolite characteristic parameters; Step 4: Based on the biological pathways recorded in the KEGG database, classify the fitted metabolite characteristic parameters to obtain the pathway characteristics of the metabolites.

[0057] The data classification module uses the sklearn package in Python to build support vector machines, random forests, extreme gradient boosting trees, and LightGBM classifiers to classify cancer and normal human samples. The classification ability of the classifiers is verified by ten-fold cross-validation.

[0058] The results are as follows Figure 6 As shown, the highest accuracy of 93% can be achieved using the XGB classifier, the highest area under the curve of 0.90 using the XGB and LightGBM classifiers, the highest F1 score of 0.92 using the XGB classifier, and the highest recall of 0.93 using the XGB classifier.

[0059] Example 6 The same prediction system as in Example 1 was used for prediction. The data source for this example is the MTBLS6039 dataset from the Metabolic Lamp website. The dataset includes 20 prostate cancer samples and 60 healthy volunteer samples added to the total dataset.

[0060] The mass spectrometry detection module is used to perform mass spectrometry scanning on the sample to obtain the mass spectrometry signal. The specific conditions for mass spectrometry scanning are as follows: the chromatographic column is maintained at 40℃, the temperature of the autosampler is 8℃, and the analyte is eluted with gradient elution using 0.1% formic acid aqueous solution (A2) and 0.1% formic acid acetonitrile solution (B2) or 5 mM ammonium formate aqueous solution (A3) and acetonitrile solution (B3) at a flow rate of 0.25 ml / min. After equilibration, 2 μl of the sample is injected. The linear gradient of solvent B2 / B3 (v / v) is as follows: 0-1 min, 2% B2 / B3; 1-9 min, 2-50% B2 / B3; 9-12 min, 50-98% B2 / B3; 12-13.5 min, 98% B2 / B3; 13.5-14 min, 98-2% B2 / B3; 14-20 min, 2% B2 positive model (14-17 min, 2% B3 negative model).

[0061] The mass spectrometry data processing module performs selective summation on the mass spectrometry signals, and then fits and classifies the summed mass spectrometry signals to obtain the metabolite pathway characteristics of the detected samples. Specific steps include: Step 1: The mass spectrometry processing module first aligns the retention time from 12 minutes to the baseline of 30 minutes; Step 2: Calculate the mass-to-charge ratio m / z of the summation ion [M+H]+ of all metabolites recorded in the KEGG database, and sum the mass spectrometry signals aligned in Step 1 with a radius of 0-50 ppm to obtain the summation mass spectrometry signal; Step 3: Use the improved expectation-maximum algorithm to fit the radial basis function to the summed mass spectrometry signal to obtain the fitted metabolite characteristic parameters; Step 4: Based on the biological pathways recorded in the KEGG database, classify the fitted metabolite characteristic parameters to obtain the pathway characteristics of the metabolites.

[0062] The data classification module uses the sklearn package in Python to build support vector machines, random forests, extreme gradient boosting trees, and LightGBM classifiers to classify cancer and normal human samples. The classification ability of the classifiers is verified by ten-fold cross-validation.

[0063] The results are as follows Figure 7 As shown, the LightGBM classifier can achieve the highest accuracy of 95%, the highest area under the curve of 0.94, the highest F1 score of 0.95, and the highest recall of 0.95.

[0064] Example 7 Prediction was performed using the same prediction system as in Example 1. The data source for this example was the datasets MTBLS1122, MTBLS3444, MTBLS6039, MTBLS3838, ST001705, and MTBLS2521 from the Metabolic Lamp website. A total of 1722 samples were collected, including 196 colon cancer samples, 82 renal cell carcinoma samples, 73 endometrial cancer samples, 60 cervical cancer samples, 20 prostate cancer samples, 153 clear renal cell carcinoma samples, 273 ovarian cancer samples, 368 benign tumors or inflammation samples, and 497 healthy controls. Examples 1-6 were individual validations performed on each dataset; this example validates the data by combining all examples into a single dataset.

[0065] The specific conditions for mass spectrometry scanning performed by the mass spectrometry acquisition module are the same as those described in Examples 1-6.

[0066] The mass spectrometry data processing module performs selective summation on the mass spectrometry signals, and then fits and classifies the summed mass spectrometry signals to obtain the metabolite pathway characteristics of the detected samples. Specific steps include: Step 1: The mass spectrometry processing module first aligns the retention time from 17 minutes to the baseline of 30 minutes; Step 2: Calculate the mass-to-charge ratio m / z of the summation ion [M+H]+ of all metabolites recorded in the KEGG database, and sum the mass spectrometry signals aligned in Step 1 with a radius of 0-50 ppm to obtain the summation mass spectrometry signal; Step 3: Use the improved expectation-maximum algorithm to fit the radial basis function of the summed mass spectrometry signal in Step 3 to obtain the fitted metabolite characteristic parameters; Step 4: Based on the biological pathways recorded in the KEGG database, classify the fitted metabolite characteristic parameters to obtain the pathway characteristics of the metabolites.

[0067] The classification module uses Python's sklearn to construct extreme gradient boosting trees to classify mixed cancer samples and normal human samples.

[0068] The results are as follows Figure 8 As shown, the XGB model correctly classified all 256 samples except for one healthy sample that was misclassified as renal cell carcinoma, one healthy sample that was misclassified as prostate cancer, one ovarian cancer sample that was misclassified as clear renal cell carcinoma, and one clear renal cell carcinoma sample that was misclassified as cervical cancer.

Claims

1. A cancer prediction system based on metabolomics mass spectrometry data parameterization, characterized in that, It includes a sample acquisition module, a mass spectrometry detection module, a mass spectrometry data processing module, and a data classification module; The sample acquisition module is used to collect test samples; The mass spectrometry detection module is used to perform mass spectrometry scanning on the test sample to obtain a mass spectrometry signal; The mass spectrometry data processing module is used to perform selective summation on the mass spectrometry signal, and after summation, the mass spectrometry signal is fitted and classified to obtain the metabolite pathway characteristics of the detected sample, specifically including: S1-1, calculate the mass-to-charge ratio m / z of the sum of all metabolites in the biological pathway; S1-2, with the mass-to-charge ratio m / z as the center and 0-50ppm as the radius, the mass spectrometry signals obtained by the mass spectrometry detection module are summed to obtain the summed mass spectrometry signal; S1-3, the improved expectation-maximum algorithm is used to fit the summed mass spectrometry signal with radial basis functions to obtain the fitted metabolite characteristic parameters; S1-4, classify the fitted metabolite characteristic parameters according to biological pathways to obtain metabolite pathway characteristics; The data classification module uses machine learning algorithms to construct a prediction model based on the characteristics of the metabolite pathways, and obtains the final prediction result.

2. The cancer prediction system based on metabolomics mass spectrometry data parameterization according to claim 1, characterized in that, The test samples include one or more of the following: blood samples, tissue samples, or urine samples.

3. The cancer prediction system based on metabolomics mass spectrometry data parameterization according to claim 1, characterized in that, The test samples include tumor samples; the tumor samples include one or more of the following: colon cancer, cervical cancer, ovarian cancer, prostate cancer, renal cell carcinoma, benign tumors, and clear renal cell carcinoma.

4. The cancer prediction system based on metabolomics mass spectrometry data parameterization according to claim 1, characterized in that, Before performing selective summation on the mass spectrometry signal, the mass spectrometry data processing module normalizes the retention time of the mass spectrometry signal to 0-30 minutes.

5. The cancer prediction system based on metabolomics mass spectrometry data parameterization according to claim 1, characterized in that, The modified expectation-maximum algorithm is used to fit the summed mass spectrometry signals using radial basis functions to obtain the fitted metabolite characteristic parameters. The specific steps include: S2-1, sampling the core parameter c of the radial basis function of each metabolite mass spectrum signal based on the summed mass spectrum signal intensity; S2-2, Initialize the width parameter of the radial basis function; S2-3 uses multiple linear regression to obtain the weight parameters of the radial basis function, and uses gradient descent to optimize the width parameter initialized in S2-2, so that the width parameter is limited to the range (5, 60). Repeat steps S2-2 and S2-3 1-5 times to obtain the width parameter of the radial basis function of each metabolite mass spectrometry signal. Fit the mass spectrometry signal of each metabolite according to the weight parameters obtained in S2-3 to obtain the fitted metabolite characteristic parameters.

6. The cancer prediction system based on metabolomics mass spectrometry data parameterization according to claim 1, characterized in that, When the mass spectrometry data processing module adds the mass spectrometry signal, for metabolites with a mass-to-charge ratio (m / z) greater than 400, the ppm calculation method is 1ppm = m / z / 1000000; for metabolites with a mass-to-charge ratio less than or equal to 400, the ppm calculation method is 1ppm = 400 / 1000000.

7. The cancer prediction system based on metabolomics mass spectrometry data parameterization according to claim 1, characterized in that, The machine learning algorithms used in the data classification module include one or more of the following: extreme gradient boosting, support vector machine, random forest, and LightGBM model constructed using the sklearn package in Python; or one or more of the following of the following: graph attention network, graph convolution, and Transformer model constructed using the pytorch package in Python.

8. A cancer prediction method based on metabolomics mass spectrometry data parameterization, characterized in that, Including the following steps: Step 1: Align the mass spectrometry scan time with the reference and perform a mass spectrometry scan on the sample to obtain the mass spectrometry signal; Step 2: Calculate the mass-to-charge ratio m / z of the sum of all metabolites in the biological pathway; Step 3: Sum the mass spectrometry signals obtained in Step 1 with the mass-to-charge ratio m / z in Step 2 as the center and 0-50ppm as the radius to obtain the summed mass spectrometry signal. Step 4: Use the improved expectation-maximum algorithm to perform radial basis function fitting on the summed mass spectrometry signal to obtain the fitted metabolite characteristic parameters; Step 5: Classify the fitted metabolite characteristic parameters according to the biological pathway to obtain the pathway characteristics of the metabolites; Step 6: Based on the pathway characteristics of metabolites, a prediction model is constructed using machine learning algorithms to obtain the final prediction results.

9. The cancer prediction method based on metabolomics mass spectrometry data parameterization according to claim 8, characterized in that, In step 1, the mass spectrometry signal is derived from mass spectrometry data of one or more datasets, including MTBLS6039, MTBLS3444, MTBLS1122, MTBLS3838, ST001705, ST002521, and MTBLS4294.

Images