A mass spectrometry detection method based on deep learning

By analyzing the mass spectrometry data, the deviation degree parameters PC are obtained, and the concentrated data is selected to train the deep learning network model, and its accuracy is evaluated, which solves the problem of model overfitting in the existing technology, and improves the accuracy of mass spectrometry detection and model optimization. ability.

CN118412062BActive Publication Date: 2024-12-06LANZHOU BAIYUAN GENE TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410578474.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-11
Publication Date
2024-12-06
Estimated Expiration
2044-05-11

AI Technical Summary

Technical Problem

The existing deep learning-based mass spectrometry detection methods lack analysis of the degree of the training set dataset, resulting in an increased risk of model overfitting and the accuracy of mass spectrometry detection of detected substances cannot be guaranteed.

Method used

By analyzing the mass spectrometry data, the deviation degree parameter PC is obtained, the relatively concentrated data is selected as the training set, the deep learning network model is trained, and the test data set is used to evaluate the model accuracy, and the reasons for the low accuracy are determined for optimization and regulation.

Benefits of technology

It reduces the risk of overfitting the deep learning network model, improves the mass spectrometry detection accuracy of the detected substances, and realizes the detection and optimization regulation of the model accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118412062B_ABST
    Figure CN118412062B_ABST
Patent Text Reader

Abstract

The present invention belongs to the technical field of mass spectrometry detection. The present invention provides a mass spectrometry detection method based on deep learning, including: obtaining mass spectrometry data of a substance to be detected, extracting information features in the mass spectrometry data, analyzing the information features of the mass spectrometry data, processing and outputting a deviation degree parameter PC of the mass spectrometry data according to a mean value ratio of ion quantity deviation and a mass-to-charge ratio deviation ratio, marking the mass spectrometry data according to the deviation degree parameter PC of the mass spectrometry data, marking it as training mass spectrometry data if the deviation degree parameter PC is less than a deviation degree parameter threshold PCY, integrating all the training mass spectrometry data into a training data set, and using the training data set to train a convolutional neural network to obtain a trained deep learning network model. The present invention is conducive to reducing the risk of overfitting of the deep learning network model and improving the performance of the deep learning network model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of mass spectrometry detection, and specifically to a mass spectrometry detection method based on deep learning. Background Art

[0002] Mass spectrometry is an important biological sample analysis technology, which is widely used in scenarios such as peptide sequence determination and protein post-translational modification identification. However, traditional mass spectrometry data processing methods are mainly based on feature matching and statistical models, which have limited analytical capabilities for complex biological samples and are difficult to meet the needs of high-precision and high-throughput analysis. Therefore, it is of great significance to develop a mass spectrometry detection method based on deep learning.

[0003] A Chinese patent application with publication number CN112505133A discloses a mass spectrometry detection method based on deep learning, including: based on a mass spectrometer, by changing the ionization voltage, detecting the substance to be detected, and obtaining a number of mass spectra corresponding to different ionization voltages; based on the mass spectra, selecting part of the mass spectra to construct an initial data set, and based on the initial data set obtained corresponding to different ionization voltages, constructing an initial mass spectrometry deep learning network model; selecting the remaining mass spectra to construct a training set, training the mass spectrometry deep learning network model, and obtaining a target mass spectrometry deep learning network model; based on the target mass spectrometry deep learning network model, analyzing the mass spectrometry, and resolving the substance to be detected through the distribution of fragment peaks.

[0004] In the prior art, the deep learning network model is trained by the acquired mass spectrometry data, but there is a lack of analysis of the selected training set, and the concentration of the training set data cannot be guaranteed, which increases the risk of overfitting of the model and thus cannot guarantee the accuracy of the model. At the same time, there is a lack of analysis of the accuracy of the model after training, and the accuracy of mass spectrometry detection of the detected object cannot be guaranteed.

[0005] To this end, the present invention provides a mass spectrometry detection method based on deep learning. Summary of the invention

[0006] In order to make up for the deficiencies of the prior art, at least one technical problem raised in the background technology is solved.

[0007] The present invention analyzes mass spectrometry data to obtain a deviation degree parameter PC, and selects relatively concentrated mass spectrometry data as a training data set to train the deep learning network model according to the deviation degree parameter PC, which is beneficial to reducing the risk of overfitting of the deep learning network model, improving the performance of the deep learning network model, and facilitating mass spectrometry detection of detected substances. The present invention also realizes the detection of the accuracy of the deep learning network model and determines the reason for the low accuracy of the deep learning network model, which is convenient for subsequent optimization and regulation of the deep learning network model.

[0008] The technical solution adopted by the present invention to solve the technical problem is: a mass spectrometry detection method based on deep learning, comprising:

[0009] Acquire mass spectrum data of the substance to be detected, and perform data preprocessing on the mass spectrum data;

[0010] Extracting information features from mass spectrometry data, wherein the information features include: mass-to-charge ratios of characteristic fragment ions and the number of ions;

[0011] Analyze the information characteristics of the mass spectrometry data to obtain deviation data, wherein the deviation data includes the mean ratio of ion number deviation and the mass-to-charge ratio deviation ratio. According to the mean ratio of ion number deviation and the mass-to-charge ratio deviation ratio, the deviation degree parameter PC of the mass spectrometry data is processed and outputted;

[0012] The mass spectrum data are marked according to the deviation degree parameter PC of the mass spectrum data;

[0013] If the deviation degree parameter PC is greater than or equal to the deviation degree parameter threshold PCY, the mass spectrum data is marked as non-training mass spectrum data;

[0014] If the deviation degree parameter PC is less than the deviation degree parameter threshold PCY, it is marked as training mass spectrometry data;

[0015] All training mass spectrometry data are integrated into a training data set, and the convolutional neural network is trained using the training data set to obtain a trained deep learning network model.

[0016] As a further technical solution of the present invention, the mass-to-charge ratio deviation ratio is obtained by:

[0017] The mass-to-charge ratios of the characteristic fragment ions of all mass spectrometry data are summed and averaged to obtain the average mass-to-charge ratio of the characteristic fragment ions, and the average mass-to-charge ratio is marked as the reference mass-to-charge ratio;

[0018] The mass-to-charge ratio of the characteristic fragment ion of the mass spectrometry data is subtracted from the reference mass-to-charge ratio, and the absolute value of the difference is taken to obtain the mass-to-charge ratio difference. The mass-to-charge ratio difference is then ratio-processed with the reference mass-to-charge ratio to obtain the mass-to-charge ratio deviation ratio.

[0019] As a further technical solution of the present invention, the method for obtaining the mean value ratio of the ion quantity deviation is as follows:

[0020] Obtain the type and quantity of each ion in the information features of the mass spectrometry data, sum and average the number of ions of the same type in the information features of all the mass spectrometry data to obtain the number mean of ions of the same type, perform difference processing on the number of each ion of the same type in the information features of the mass spectrometry data and the corresponding number mean of ions of the same type, take the absolute value of all the differences obtained to obtain the deviation number of each ion of the same type, perform ratio processing on the deviation number of each ion of the same type and the corresponding number mean of ions of the same type to obtain the deviation number ratio of each ion of the same type, sum and average the deviation number ratios of each ion of the same type to obtain the proportion of the mean deviation of the ion number.

[0021] As a further technical solution of the present invention, it also includes:

[0022] Use the test data set to test the trained deep learning network model, analyze the test output results, obtain the accurate evaluation value PG, and judge the accuracy of the deep learning network model based on the accurate evaluation value PG;

[0023] Based on the low accuracy of the deep learning network model, the mass spectrometry data of the abnormal detection substance is compared and analyzed with the training mass spectrometry data to obtain the data overlap value CH, and the data overlap value CH is compared with the data overlap threshold. According to the comparison result, it is determined whether the reason for the low accuracy of the deep learning network model is that the mass spectrometry data of the detection substance exceeds the range of the training mass spectrometry data.

[0024] As a further technical solution of the present invention, the test data set represents mass spectrometry data of several substances to be detected with known compound contents and types;

[0025] The test output results include the compound type and content of the substance to be detected.

[0026] As a further technical solution of the present invention, the accurate evaluation value PG is obtained by:

[0027] Count the number of abnormal detection substances and the number of all detected substances with known compound content and types, and perform ratio processing to obtain the ratio of the number of abnormal detection substances;

[0028] Perform difference processing on the abnormal category proportion characterization value ZB corresponding to the abnormal detection substance and the abnormal category proportion characterization threshold value ZBY, and take the absolute value of the difference to obtain the abnormal category proportion deviation value corresponding to the abnormal detection substance;

[0029] The proportion deviation values ​​of the abnormal types corresponding to all abnormal detection substances are summed up and averaged to obtain the proportion deviation mean of the abnormal types corresponding to the abnormal detection substances;

[0030] The mean deviation of the proportion of abnormal types corresponding to the abnormal detection substances is processed by ratio with the threshold value ZBY representing the proportion of abnormal types to obtain the mean deviation ratio of the proportion of abnormal types corresponding to the abnormal detection substances;

[0031] The accurate evaluation value PG is obtained by multiplying the proportion of the abnormal detection substances by the mean deviation ratio of the proportion of the abnormal types corresponding to the abnormal detection substances.

[0032] As a further technical solution of the present invention, the labeling process of the abnormal detection substance includes:

[0033] The compound type and content of the substance to be detected outputted by the test are marked as the detection type and the detection content, respectively, and the compound type and content of the substance to be detected with known compound content and type are marked as the actual type and the actual content, respectively;

[0034] Compare all detected species with all actual species, mark the species that do not exist in the actual species as abnormal species among the detected species, and count the number and content of abnormal species;

[0035] The number of abnormal species is processed by ratio with the number of detected species to obtain the number ratio of abnormal species, and the content of abnormal species is processed by ratio with the content of detected species to obtain the content ratio of abnormal species;

[0036] The proportion of the number of abnormal types is multiplied by the proportion of the content of the abnormal types to obtain the proportion representation value ZB of the abnormal types.

[0037] As a further technical solution of the present invention, the marking process of the abnormal detection substance further includes:

[0038] Compare the abnormal category proportion representation value ZB with the abnormal category proportion representation threshold ZBY,

[0039] If the abnormal category ratio characterization value ZB is greater than or equal to the abnormal category ratio characterization threshold value ZBY, the substance to be detected with known compound content and type is marked as an abnormal detection substance;

[0040] If the ratio characterization value ZB of the abnormal category is less than the ratio characterization threshold ZBY of the abnormal category, it is marked as a normal detection substance.

[0041] As a further technical solution of the present invention, the data coincidence value CH is obtained in the following manner:

[0042] Obtaining a mass spectrum in the mass spectrum data of the abnormal detection substance and a mass spectrum in the training mass spectrum data;

[0043] Convert each mass spectrum into a feature vector, wherein the feature vector includes a comparison feature vector and a reference feature vector, measure the cosine angle between the comparison feature vector and the reference feature vector, compare the cosine angle with a cosine angle threshold, and if the cosine angle is less than or equal to the cosine angle threshold, mark the comparison feature vector as a similar feature vector; if the cosine angle is greater than the cosine angle threshold, mark the comparison feature vector as a non-similar feature vector;

[0044] The mass spectra corresponding to the similar feature vectors are marked as coincident mass spectra, the number of coincident mass spectra and the number of mass spectra in the mass spectrum data of all abnormal detection substances are counted, and they are ratio-processed to obtain the data coincidence value CH.

[0045] As a further technical solution of the present invention, the data coincidence value CH is compared with the data coincidence threshold value. The specific comparison process is as follows:

[0046] The preset data overlap threshold is CHY;

[0047] If the data overlap value CH is greater than or equal to the data overlap threshold CHY, it means that the reason for the low accuracy of the deep learning network model is not that the mass spectrum data of the detected substance exceeds the range of the training mass spectrum data;

[0048] If the data overlap value CH is less than the data overlap threshold CHY, it means that the reason why the accuracy of the deep learning network model is low is that the mass spectrum data of the detected substance exceeds the range of the training mass spectrum data.

[0049] The beneficial effects of the present invention are as follows:

[0050] 1. The present invention obtains mass spectrum data of a substance to be detected, extracts information features in the mass spectrum data, obtains a deviation degree parameter PC of the mass spectrum data according to the information features of the mass spectrum data, marks the mass spectrum data according to the deviation degree parameter PC of the mass spectrum data, obtains a training data set and a non-training data set, trains a convolutional neural network using the training data set, and obtains a trained deep learning network model. The present invention obtains the deviation degree parameter PC by analyzing the mass spectrum data, selects relatively concentrated mass spectrum data as the training data set according to the deviation degree parameter PC to train the deep learning network model, which is beneficial to reducing the risk of overfitting of the deep learning network model, improving the performance of the deep learning network model, and facilitating mass spectrometry detection of the detected substance.

[0051] 2. The present invention uses a test data set to test the trained deep learning network model to determine the accuracy of its deep learning network model. Based on the low accuracy of the deep learning network model, the mass spectrum data of the abnormal detection substance is compared and analyzed with the training mass spectrum data to obtain the data overlap value CH, and the data overlap value CH is compared with the data overlap threshold value. According to the comparison result, it is determined whether the reason for the low accuracy of the deep learning network model is that the mass spectrum data of the detection substance exceeds the range of the training mass spectrum data, and a control signal is generated, wherein the control signal is a control start signal and a control analysis signal. Based on the control start signal, the mass spectrum data of the detection substance that exceeds the range of the training mass spectrum data is added to the training mass spectrum data, and the deep learning network model is retrained, so that the deep learning network model continues to learn and train. The present invention realizes the detection of the accuracy of the deep learning network model and determines the reason for the low accuracy of the deep learning network model, which is convenient for subsequent optimization and control of the deep learning network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The present invention will be further described below in conjunction with the accompanying drawings.

[0053] Figure 1 is a flow chart of a mass spectrometry detection method based on deep learning described in Example 1 of the present invention;

[0054] Figure 2 is a flow chart of a mass spectrometry detection method based on deep learning described in Example 2 of the present invention;

[0055] Figure 3 This is a flow chart of a method for obtaining an accurate evaluation value in a mass spectrometry detection method based on deep learning described in Example 2 of the present invention. DETAILED DESCRIPTION

[0056] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0057] Example 1

[0058] like Figure 1 As shown, a mass spectrometry detection method based on deep learning according to an embodiment of the present invention includes:

[0059] Step 1: Obtain mass spectrometry data of the substance to be detected and perform data preprocessing on the mass spectrometry data, including: denoising, baseline correction and peak detection;

[0060] In some embodiments, the mass spectrum data of the substance to be detected includes: a mass spectrum, an ion spectrum, etc., wherein the mass spectrum data is obtained by a mass spectrometer;

[0061] In some embodiments, the substance to be detected is present in several copies;

[0062] Step 2: extracting information features from the mass spectrum data, wherein the information features include: the mass-to-charge ratio of the characteristic fragment ions and the number of ions;

[0063] In some embodiments, the information features are extracted by a deep learning algorithm (convolutional neural network);

[0064] Step 3: According to the information characteristics of the mass spectrometry data, the deviation degree parameter PC of the mass spectrometry data is obtained, and the mass spectrometry data is marked according to the deviation degree parameter PC of the mass spectrometry data to obtain a training data set and a non-training data set, and the convolutional neural network is trained using the training data set to obtain a trained deep learning network model;

[0065] Exemplarily, the deviation degree parameter PC is obtained in the following manner:

[0066] The information features in the mass spectrometry data are analyzed to obtain deviation data, wherein the deviation data includes the mass-to-charge ratio deviation ratio and the ion number deviation mean ratio. The mass-to-charge ratio deviation ratio and the ion number deviation mean ratio are multiplied to obtain the deviation degree parameter PC.

[0067] Exemplarily, the mass-to-charge ratio deviation ratio is obtained as follows:

[0068] The mass-to-charge ratios of the characteristic fragment ions of all mass spectrometry data are summed and averaged to obtain the average mass-to-charge ratio of the characteristic fragment ions, and the average mass-to-charge ratio is marked as the reference mass-to-charge ratio;

[0069] Subtract the mass-to-charge ratio of the characteristic fragment ion of the mass spectrum data from the reference mass-to-charge ratio, take the absolute value of the difference to obtain the mass-to-charge ratio difference, perform ratio processing on the mass-to-charge ratio difference and the reference mass-to-charge ratio to obtain the mass-to-charge ratio deviation ratio;

[0070] Exemplarily, the method for obtaining the mean ratio of ion quantity deviation is as follows:

[0071] Obtain the type and quantity of each ion in the information feature of the mass spectrometry data, sum and average the number of ions of the same type in the information features of all the mass spectrometry data to obtain the number mean of ions of the same type, perform difference processing on the number of each ion of the same type in the information feature of the mass spectrometry data and the corresponding number mean of ions of the same type, and take the absolute value of all the differences obtained to obtain the deviation number of each ion of the same type, perform ratio processing on the deviation number of each ion of the same type and the corresponding number mean of ions of the same type to obtain the deviation number ratio of each ion of the same type, sum and average the deviation number ratios of each ion of the same type to obtain the proportion of the mean deviation of the number of ions;

[0072] The deviation degree parameter PC is compared with the deviation degree parameter threshold. The specific comparison process is as follows:

[0073] The preset deviation degree parameter threshold is PCY;

[0074] If the deviation degree parameter PC is greater than or equal to the deviation degree parameter threshold PCY, the mass spectrum data is marked as non-training mass spectrum data;

[0075] If the deviation degree parameter PC is less than the deviation degree parameter threshold PCY, it is marked as training mass spectrometry data;

[0076] Integrate all training mass spectrometry data to obtain a training data set;

[0077] It should be noted that the training data set includes all training mass spectrometry data;

[0078] Use the training data set to train the convolutional neural network to obtain a trained deep learning network model;

[0079] The technical solution of the embodiment of the present invention is: obtaining mass spectrum data of the substance to be detected, extracting information features in the mass spectrum data, obtaining a deviation degree parameter PC of the mass spectrum data according to the information features of the mass spectrum data, marking the mass spectrum data according to the deviation degree parameter PC of the mass spectrum data, obtaining a training data set and a non-training data set, using the training data set to train the convolutional neural network to obtain a trained deep learning network model. The present invention obtains the deviation degree parameter PC by analyzing the mass spectrum data, and selects relatively concentrated mass spectrum data as the training data set to train the deep learning network model according to the deviation degree parameter PC, which is beneficial to reducing the risk of overfitting of the deep learning network model, improving the performance of the deep learning network model, and facilitating mass spectrometry detection of the detected substance.

[0080] Example 2

[0081] like Figure 2 As shown, based on Example 1, a mass spectrometry detection method based on deep learning described in an embodiment of the present invention includes:

[0082] Step 4: Use the test data set to test the trained deep learning network model, analyze the test output results, obtain the accurate evaluation value PG, and judge the accuracy of the deep learning network model based on the accurate evaluation value PG;

[0083] Specifically, Figure 3 As shown, several sets of mass spectrometry data of substances to be detected with known compound contents and types are obtained, and the mass spectrometry data are input into the deep network learning model as a test data set to obtain the test output result of the deep network learning model, wherein the test output result includes the compound type and content of the substances to be detected;

[0084] The compound type and content of the substance to be detected outputted by the test are marked as the detection type and the detection content, respectively, and the compound type and content of the substance to be detected with known compound content and type are marked as the actual type and the actual content, respectively;

[0085] Compare all detected species with all actual species, mark the species that do not exist in the actual species as abnormal species among the detected species, and count the number and content of abnormal species;

[0086] The number of abnormal species is processed by ratio with the number of detected species to obtain the number ratio of abnormal species, and the content of abnormal species is processed by ratio with the content of detected species to obtain the content ratio of abnormal species;

[0087] The number ratio of abnormal species is multiplied by the content ratio of abnormal species to obtain the ratio representation value ZB of abnormal species;

[0088] Compare the abnormal category proportion characterization value ZB with the abnormal category proportion characterization threshold, and mark the substance to be detected with known compound content and category according to the comparison result;

[0089] Specifically, the threshold representing the proportion of abnormal types is preset to be ZBY;

[0090] If the abnormal category ratio characterization value ZB is greater than or equal to the abnormal category ratio characterization threshold value ZBY, the substance to be detected with known compound content and type is marked as an abnormal detection substance;

[0091] If the abnormal category proportion characterization value ZB is less than the abnormal category proportion characterization threshold value ZBY, it is marked as a normal detection substance;

[0092] Count the number of abnormal detection substances and the number of all detected substances with known compound content and types, and perform ratio processing to obtain the ratio of the number of abnormal detection substances;

[0093] Perform difference processing on the abnormal category proportion characterization value ZB corresponding to the abnormal detection substance and the abnormal category proportion characterization threshold value ZBY, and take the absolute value of the difference to obtain the abnormal category proportion deviation value corresponding to the abnormal detection substance;

[0094] The proportion deviation values ​​of the abnormal types corresponding to all abnormal detection substances are summed up and averaged to obtain the proportion deviation mean of the abnormal types corresponding to the abnormal detection substances;

[0095] The mean deviation of the proportion of abnormal types corresponding to the abnormal detection substances is processed by ratio with the threshold value ZBY representing the proportion of abnormal types to obtain the mean deviation ratio of the proportion of abnormal types corresponding to the abnormal detection substances;

[0096] The proportion of the number of abnormal detection substances and the mean deviation ratio of the proportion of abnormal types corresponding to the abnormal detection substances are multiplied to obtain the accurate evaluation value PG;

[0097] Compare the accurate evaluation value PG with the accurate evaluation threshold PGY, and judge the accuracy of its deep learning network model according to the comparison result;

[0098] Specifically, if the accurate evaluation value PG is less than the accurate evaluation threshold PGY, it means that the accuracy of the deep learning network model is high;

[0099] If the accurate evaluation value PG is greater than or equal to the accurate evaluation threshold PGY, it means that the accuracy of the deep learning network model is low;

[0100] Step 5: Based on the low accuracy of the deep learning network model, the mass spectrum data of the abnormal detection substance is compared and analyzed with the training mass spectrum data to obtain a data overlap value CH, and the data overlap value CH is compared with the data overlap threshold value. According to the comparison result, it is determined whether the reason for the low accuracy of the deep learning network model is that the mass spectrum data of the detection substance exceeds the range of the training mass spectrum data, and a control signal is generated, wherein the control signal is a control start signal and a control analysis signal;

[0101] Specifically, the mass spectrum in the mass spectrum data of the abnormal detection substance and the mass spectrum in the training mass spectrum data are obtained and compared. The specific comparison process is as follows:

[0102] Perform feature extraction on the mass spectrum, convert each mass spectrum into a feature vector, if the mass spectrum is a mass spectrum in the mass spectrum data of the abnormal detection substance, then mark the converted feature vector as a comparison feature vector, if the mass spectrum is a mass spectrum in the training mass spectrum data, then mark the converted feature vector as a reference feature vector;

[0103] It should be noted that the mass spectrum is subjected to feature extraction to convert each mass spectrum into a feature vector, wherein the extraction method includes but is not limited to: peak identification, peak area calculation, mass-to-charge ratio distribution;

[0104] All comparison feature vectors are compared and analyzed with all reference feature vectors respectively. Specifically, the cosine angle between the comparison feature vector and the reference feature vector is measured, and the cosine angle is compared with the cosine angle threshold. If the cosine angle is less than or equal to the cosine angle threshold, the comparison feature vector is marked as a similar feature vector. If the cosine angle is greater than the cosine angle threshold, the comparison feature vector is marked as a non-similar feature vector.

[0105] The mass spectra corresponding to the similar feature vectors are marked as coincident mass spectra, the number of coincident mass spectra and the number of mass spectra in the mass spectrum data of all abnormal detection substances are counted, and the ratios are processed to obtain the data coincidence value CH;

[0106] Compare the data coincidence value CH with the data coincidence threshold. The specific comparison process is as follows:

[0107] The preset data overlap threshold is CHY;

[0108] If the data overlap value CH is greater than or equal to the data overlap threshold CHY, it means that the reason for the low accuracy of the deep learning network model is not that the mass spectrum data of the detected substance exceeds the range of the training mass spectrum data, and a regulation analysis signal is generated;

[0109] If the data overlap value CH is less than the data overlap threshold CHY, it means that the reason for the low accuracy of the deep learning network model is that the mass spectrum data of the detected substance exceeds the range of the training mass spectrum data, and a control start signal is generated;

[0110] It should be noted that when the data overlap value is less than the data overlap threshold, it means that the mass spectrometry data input to the model and the mass spectrometry data used for model training have low similarity, and the model cannot recognize the new mass spectrometry data, resulting in inaccurate output results of the model;

[0111] In some embodiments, based on the regulation analysis signal, further analysis is performed on the reasons for low model accuracy, including but not limited to: data quality check, overfitting analysis, and underfitting analysis;

[0112] Based on the control start signal, the mass spectrum data of the detection substance that exceeds the training mass spectrum data range is added to the training mass spectrum data, and the deep learning network model is retrained, so that the deep learning network model continues to learn and train;

[0113] The technical solution of the embodiment of the present invention is: using a test data set to test the trained deep learning network model to determine the accuracy of the deep learning network model; based on the low accuracy of the deep learning network model, the mass spectrum data of the abnormal detection substance is compared and analyzed with the training mass spectrum data to obtain a data overlap value CH; the data overlap value CH is compared with the data overlap threshold; and according to the comparison result, it is determined whether the reason for the low accuracy of the deep learning network model is that the mass spectrum data of the detection substance exceeds the range of the training mass spectrum data, and a control signal is generated, wherein the control signal is a control start signal and a control analysis signal; based on the control start signal, the mass spectrum data of the detection substance that exceeds the range of the training mass spectrum data is added to the training mass spectrum data, and the deep learning network model is retrained, so that the deep learning network model continues to learn and train. The present invention realizes the detection of the accuracy of the deep learning network model and determines the reason for the low accuracy of the deep learning network model, which is convenient for optimizing and controlling the deep learning network model later.

[0114] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A mass spectrometry detection method based on deep learning, Features: include: Acquire mass spectrum data of the substance to be detected, and perform data preprocessing on the mass spectrum data; Extracting information features from mass spectrometry data, wherein the information features include: mass-to-charge ratios of characteristic fragment ions and the number of ions; Analyze the information characteristics of the mass spectrometry data to obtain deviation data, wherein the deviation data includes the mean ratio of ion number deviation and the mass-to-charge ratio deviation ratio. According to the mean ratio of ion number deviation and the mass-to-charge ratio deviation ratio, the deviation degree parameter PC of the mass spectrometry data is processed and outputted; The deviation degree parameter PC is obtained in the following manner: The mass-to-charge ratio deviation ratio and the ion number deviation mean ratio are multiplied to obtain the deviation degree parameter PC; The mass-to-charge ratio deviation ratio is obtained as follows: The mass-to-charge ratios of the characteristic fragment ions of all mass spectrometry data are summed and averaged to obtain the average mass-to-charge ratio of the characteristic fragment ions, and the average mass-to-charge ratio is marked as the reference mass-to-charge ratio; Subtract the mass-to-charge ratio of the characteristic fragment ion of the mass spectrum data from the reference mass-to-charge ratio, take the absolute value of the difference to obtain the mass-to-charge ratio difference, perform ratio processing on the mass-to-charge ratio difference and the reference mass-to-charge ratio to obtain the mass-to-charge ratio deviation ratio; The method for obtaining the mean value ratio of the ion quantity deviation is as follows: Obtain the type and quantity of each ion in the information feature of the mass spectrometry data, sum and average the number of ions of the same type in the information features of all the mass spectrometry data to obtain the number mean of ions of the same type, perform difference processing on the number of each ion of the same type in the information feature of the mass spectrometry data and the corresponding number mean of ions of the same type, and take the absolute value of all the differences obtained to obtain the deviation number of each ion of the same type, perform ratio processing on the deviation number of each ion of the same type and the corresponding number mean of ions of the same type to obtain the deviation number ratio of each ion of the same type, sum and average the deviation number ratios of each ion of the same type to obtain the proportion of the mean deviation of the number of ions; The mass spectrum data are marked according to the deviation degree parameter PC of the mass spectrum data; If the deviation degree parameter PC is greater than or equal to the deviation degree parameter threshold PCY, the mass spectrum data is marked as non-training mass spectrum data; If the deviation degree parameter PC is less than the deviation degree parameter threshold PCY, it is marked as training mass spectrometry data; All training mass spectrometry data are integrated into a training data set, and the convolutional neural network is trained using the training data set to obtain a trained deep learning network model.

2. A mass spectrometry detection method based on deep learning according to claim 1, Features: Also includes: Use the test data set to test the trained deep learning network model, analyze the test output results, obtain the accurate evaluation value PG, and judge the accuracy of the deep learning network model based on the accurate evaluation value PG; The accurate evaluation value PG is obtained as follows: Count the number of abnormal detection substances and the number of all detected substances with known compound content and types, and perform ratio processing to obtain the ratio of the number of abnormal detection substances; Perform difference processing on the abnormal category proportion characterization value ZB corresponding to the abnormal detection substance and the abnormal category proportion characterization threshold value ZBY, and take the absolute value of the difference to obtain the abnormal category proportion deviation value corresponding to the abnormal detection substance; The proportion deviation values ​​of the abnormal types corresponding to all abnormal detection substances are summed up and averaged to obtain the proportion deviation mean of the abnormal types corresponding to the abnormal detection substances; The mean deviation of the proportion of abnormal types corresponding to the abnormal detection substances is processed by ratio with the threshold value ZBY representing the proportion of abnormal types to obtain the mean deviation ratio of the proportion of abnormal types corresponding to the abnormal detection substances; The proportion of the number of abnormal detection substances and the mean deviation ratio of the proportion of abnormal types corresponding to the abnormal detection substances are multiplied to obtain the accurate evaluation value PG; The marking process of the abnormal detection substance includes: The compound type and content of the substance to be detected outputted by the test are marked as the detection type and the detection content, respectively, and the compound type and content of the substance to be detected with known compound content and type are marked as the actual type and the actual content, respectively; Compare all detected species with all actual species, mark the species that do not exist in the actual species as abnormal species among the detected species, and count the number and content of abnormal species; The number of abnormal species is processed by ratio with the number of detected species to obtain the number ratio of abnormal species, and the content of abnormal species is processed by ratio with the content of detected species to obtain the content ratio of abnormal species; The number ratio of abnormal species is multiplied by the content ratio of abnormal species to obtain the ratio representation value ZB of abnormal species; Based on the low accuracy of the deep learning network model, the mass spectrum data of the abnormal detection substance is compared and analyzed with the training mass spectrum data to obtain the data overlap value CH, and the data overlap value CH is compared with the data overlap threshold value, and it is determined according to the comparison result whether the reason for the low accuracy of the deep learning network model is that the mass spectrum data of the detection substance exceeds the range of the training mass spectrum data; The data coincidence value CH is obtained in the following manner: Obtaining a mass spectrum in the mass spectrum data of the abnormal detection substance and a mass spectrum in the training mass spectrum data; Convert each mass spectrum into a feature vector, wherein the feature vector includes a comparison feature vector and a reference feature vector, measure the cosine angle between the comparison feature vector and the reference feature vector, compare the cosine angle with a cosine angle threshold, and if the cosine angle is less than or equal to the cosine angle threshold, mark the comparison feature vector as a similar feature vector; if the cosine angle is greater than the cosine angle threshold, mark the comparison feature vector as a non-similar feature vector; The mass spectra corresponding to the similar feature vectors are marked as coincident mass spectra, the number of coincident mass spectra and the number of mass spectra in the mass spectrum data of all abnormal detection substances are counted, and they are ratio-processed to obtain the data coincidence value CH.

3. A mass spectrometry detection method based on deep learning according to claim 2, Features: The test data set represents mass spectrometry data of several substances to be detected with known compound contents and types; The test output results include the compound type and content of the substance to be detected.

4. The mass spectrometry detection method based on deep learning according to claim 2, Features: The marking process of the abnormal detection substance further includes: Compare the abnormal category proportion representation value ZB with the abnormal category proportion representation threshold ZBY, If the abnormal category ratio characterization value ZB is greater than or equal to the abnormal category ratio characterization threshold value ZBY, the substance to be detected with known compound content and type is marked as an abnormal detection substance; If the ratio characterization value ZB of the abnormal category is less than the ratio characterization threshold value ZBY of the abnormal category, it is marked as a normal detection substance.

5. The mass spectrometry detection method based on deep learning according to claim 2, Features: Compare the data coincidence value CH with the data coincidence threshold. The specific comparison process is as follows: The preset data overlap threshold is CHY; If the data overlap value CH is greater than or equal to the data overlap threshold CHY, it means that the reason for the low accuracy of the deep learning network model is not that the mass spectrum data of the detected substance exceeds the range of the training mass spectrum data; If the data overlap value CH is less than the data overlap threshold CHY, it means that the reason why the accuracy of the deep learning network model is low is that the mass spectrum data of the detected substance exceeds the range of the training mass spectrum data.