Mass spectrometry classification method, biomarker discovery method and device based on deep learning

Abstract

The present application discloses a mass spectrometry classification method based on deep learning, a biomarker discovery method and device, and relates to the field of mass spectrometry analysis technology. Only using MS1 information, by combining the model architecture of CNN and Transformer, it is possible to simultaneously capture local features and global associations in mass spectrometry data, thereby improving the ability to accurately classify mass spectrometry samples; using a local linear model to interpret the classification results, it is possible to identify MS1 features that play a key role in classification decisions, and a protein identification module based on key MS1 features can associate mass spectrometry features with corresponding proteins, providing a new strategy for early diagnosis of diseases and helping researchers to more accurately understand the biological significance and mechanism of action of biomarkers.

Description

Technical Field

[0001] The present application relates to the technical field of mass spectrometry analysis, and in particular to a mass spectrometry classification method based on deep learning, a biomarker discovery method and a device. Background Art

[0002] Mass spectrometry is an efficient bioanalytical method that can provide detailed information about proteins and metabolites in biological samples, and therefore has great potential in early diagnosis of cancer diseases and biomarker discovery.

[0003] Deep learning technology has shown great potential in mass spectrometry data analysis with its powerful feature learning and pattern recognition capabilities. In particular, models such as Convolutional Neural Networks (CNN) and Transformer are outstanding in processing high-dimensional data and learning complex features. However, previous deep learning mass spectrometry classification methods mostly rely on secondary mass spectrometry MS2 information, do not deeply explore the potential of primary mass spectrometry MS1 information, and do not provide interpretable biomarkers. These defects have hindered the practical application of deep learning-based mass spectrometry classification methods. Summary of the invention

[0004] In order to solve at least one of the technical problems mentioned in the background technology, the purpose of the present application is to provide a mass spectrometry classification method based on deep learning, a biomarker discovery method and device, which can classify mass spectrometry data relying solely on MS1 ​​information and provide interpretable biomarkers.

[0005] To achieve the above objectives, this application provides the following technical solutions:

[0006] In a first aspect, the present application provides a method for constructing a mass spectrometry classification model based on deep learning, comprising:

[0007] S1, acquiring a plurality of primary mass spectrometry data sets, including a primary mass spectrometry data set classified as a first category and a primary mass spectrometry data set classified as a second category; the primary mass spectrometry data set includes a mass-to-charge ratio, a peak value, and a retention time;

[0008] S2, after performing noise reduction processing on the primary mass spectrometry data set, performing a peak alignment operation, and discretizing the primary mass spectrometry data set after the peak alignment to obtain an intensity matrix; the number of rows of the intensity matrix is ​​equal to the number of time points of the retention time, and the number of columns is equal to the number of mass-to-charge ratio windows after the discretization processing; according to the maximum peak value of each row of the intensity matrix, each row element of the intensity matrix is ​​standardized to obtain a standardized intensity matrix;

[0009] S3, constructing a mass spectrometry classification model using the standardized intensity matrix and the classification information corresponding to the primary mass spectrometry data set.

[0010] Furthermore, the peak alignment operation adopts an alignment method based on retention time or an alignment method based on peak shape features.

[0011] Furthermore, the discretization process includes:

[0012] Dividing the mass-to-charge ratio range into a plurality of mass-to-charge ratio windows of equal width;

[0013] The peak values ​​of multiple peaks located in the same mass-to-charge ratio window are summed, and the sum of the peak values ​​is taken as the peak value corresponding to the mass-to-charge ratio window.

[0014] Furthermore, the mass spectrometry classification model includes a CNN module and a Transformer module. The performance of the mass spectrometry classification model is measured using a cross entropy loss function. The model parameters are optimized using an Adam optimizer until the training is stopped when the loss function converges.

[0015] In the second aspect, an embodiment of the present application provides a mass spectrometry classification method based on deep learning, which uses the above-mentioned step S2 to construct a standardized intensity matrix corresponding to the primary mass spectrometry data set to be classified, and inputs it into the mass spectrometry classification model constructed above, thereby obtaining classification information of the mass spectrometry data set to be classified, and the classification information includes a first category and a second category.

[0016] In a third aspect, the present application embodiment provides a deep learning mass spectrometry classification system, comprising:

[0017] A data input module, used to obtain a primary mass spectrometry data set to be classified;

[0018] A data storage module, used for storing standardized intensity matrices and classification information corresponding to a plurality of primary mass spectrometry data sets;

[0019] a classification module, connected to the data input module and the data storage module respectively, for first constructing a mass spectrum classification model according to the standardized intensity matrices corresponding to the multiple mass spectrum data sets and their classification information, and then inputting the primary mass spectrum data set to be classified into the model to obtain the classification information of the primary mass spectrum data set to be classified;

[0020] A result output module, connected to the classification module, for outputting the classification result;

[0021] The standardized intensity matrix is ​​constructed according to the above step S2.

[0022] Furthermore, the result output module is further connected to the data storage module, and is used to input the standardized intensity matrix corresponding to the primary mass spectrometry data set to be classified and the classification result into the data storage module.

[0023] In a fourth aspect, the present application provides a method for biomarker discovery based on deep learning, comprising:

[0024] Acquiring a primary mass spectrometry data set, and performing noise reduction processing, peak alignment operation, discretization processing, and standardization processing on the primary mass spectrometry data set, wherein the primary mass spectrometry data set is split into a plurality of mass-to-charge ratio windows;

[0025] For each mass-to-charge ratio window, a set of perturbation samples is generated respectively; in each of the perturbation samples, only the peak value corresponding to the mass-to-charge ratio window is changed;

[0026] Inputting each group of perturbation samples corresponding to all mass-to-charge ratio windows into the mass spectrum classification model constructed above to obtain corresponding output values; constructing a local linear model using all the perturbation samples and their corresponding output values, wherein the coefficient of the local linear model reflects the influence of the peak value of each mass-to-charge ratio window on the prediction result; wherein the absolute value of the coefficient indicates the influence, and the sign of the coefficient indicates the direction of the influence;

[0027] According to the coefficients of the local linear model, the mass-to-charge ratio windows that are ranked before a preset value are selected from large to small, as key mass spectrometry features; and the biomarkers are determined based on the key mass spectrometry features.

[0028] Furthermore, the step of determining the biomarker based on the key mass spectrometry features is determined using the open source mass spectrometry data analysis tool MaxQuant.

[0029] In a fifth aspect, the present application embodiment provides a biomarker discovery device based on deep learning, comprising:

[0030] an acquisition unit, used for acquiring a primary mass spectrum data set, and performing noise reduction, peak alignment, discretization and standardization on the primary mass spectrum data set, wherein the primary mass spectrum data set is split into a plurality of mass-to-charge ratio windows;

[0031] A generating unit, for generating a set of disturbance samples for each mass-to-charge ratio window; in each of the disturbance samples, only the peak value corresponding to the mass-to-charge ratio window is changed;

[0032] A construction unit is used to input each group of disturbance samples corresponding to all mass-to-charge ratio windows into the mass spectrum classification model constructed above to obtain corresponding output values; a local linear model is constructed using all the disturbance samples and their corresponding output values, wherein the coefficient of the local linear model reflects the influence of the peak value of each mass-to-charge ratio window on the prediction result; wherein the absolute value of the coefficient indicates the influence, and the sign of the coefficient indicates the direction of the influence;

[0033] A determination unit is used to select, according to the coefficients of the local linear model, the mass-to-charge ratio windows that are ranked before a preset value in descending order as key mass spectrometry features; and determine the biomarker based on the key mass spectrometry features.

[0034] Compared with the prior art, the beneficial effects of this application are:

[0035] Using only MS1 information, by combining the CNN and Transformer model architecture, it is possible to simultaneously capture local features and global associations in mass spectrometry data, thus improving the ability to accurately classify mass spectrometry samples.

[0036] The local linear model is used to interpret the classification results, which can identify the MS1 features that play a key role in classification decisions. The protein identification module based on the key MS1 features can associate the mass spectrometry features with the corresponding proteins, providing a new strategy for the early diagnosis of the disease and helping researchers to more accurately understand the biological significance and mechanism of action of biomarkers. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A flowchart of a method for constructing a mass spectrometry classification model based on deep learning provided in Example 1 of the present application;

[0038] Figure 2 A structural diagram of a deep learning mass spectrometry classification system provided in Example 3 of the present application;

[0039] Figure 3 A flowchart of a biomarker discovery method based on deep learning provided in Example 4 of the present application;

[0040] Figure 4 A volcano diagram visualization display diagram of protein identification using original mass spectrometry files provided in Example 4 of the present application;

[0041] Figure 5 A volcano diagram visualization display diagram of protein identification using key mass spectrometry features provided in Example 4 of the present application;

[0042] Figure 6 This is a structural diagram of a deep learning-based biomarker discovery device provided in Example 5 of the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0044] In order to solve the above-mentioned problems existing in the prior art, the embodiments of the present application provide a mass spectrometry classification method based on deep learning, a biomarker discovery method and a device, which are described in detail below.

[0045] The embodiments of the present application are described using mass spectrometry data of hepatocellular carcinoma HCC as an example. It should be noted that the method provided in the present application can also be used for mass spectrometry data classification of other cancers, without limitation thereto.

[0046] Embodiment 1:

[0047] like Figure 1 As shown, Figure 1 A flowchart of a method for constructing a deep learning mass spectrometry classification model provided in an embodiment of the present application includes the following steps:

[0048] Step S1, acquiring multiple primary mass spectrometry data sets, including primary mass spectrometry data sets classified into the first category and primary mass spectrometry data sets classified into the second category, the primary mass spectrometry data sets including mass-to-charge ratio, peak value and retention time.

[0049] Primary mass spectrometry data files were obtained from clinically collected HCC samples and non-HCC samples, and the files were converted into mzML file format to obtain HCC primary mass spectrometry data sets and non-HCC primary mass spectrometry data sets. The primary mass spectrometry data sets contained information such as mass-to-charge ratio m / z, peak value, and retention time.

[0050] Step S2, preprocessing the primary mass spectrometry data set, including noise reduction, peak identification and peak alignment, window division discretization and standardization.

[0051] During the mass spectrometry data acquisition process, noise will be present in the mass spectrometry data due to the noise of the instrument itself, environmental factors, and various interferences during sample preparation. These noises will affect the accuracy of subsequent data analysis and interpretation, so it is necessary to denoise the mass spectrometry data.

[0052] The mass spectrometry data was denoised using the Gaussian filter in the open source mass spectrometry data analysis tool pyopenms. GaussFilter is a linear filter that performs weighted averaging on mass-to-charge ratio and peak data, and the coefficients in its weighted window template are calculated using Gaussian distribution.

[0053] Specifically, the coefficients of the Gaussian window are By length The Gaussian distribution is calculated as:

[0054]

[0055] in , x represents the window area, Represents the standard deviation of the Gaussian distribution.

[0056] Next, the mass spectrometry data is subjected to peak recognition and peak alignment. Peak recognition is the process of accurately detecting and locating peaks representing different compounds or molecules from mass spectrometry data. The accuracy of peak recognition is crucial for subsequent molecular identification and quantitative analysis.

[0057] Mass spectrometry data has two-dimensional patterns in mass-to-charge ratio and retention time dimensions in a series of MS1 scans. These patterns are called features, which show chromatographic elution curves in the retention time dimension and isotope patterns in the mass-to-charge ratio dimension. Use the Feature Finder tool in pyopenms to identify these features from two-dimensional data, and treat peaks with similar mass-to-charge ratios and similar retention times as the same feature.

[0058] After peak identification, the next step is to perform peak alignment. In mass spectrometry data, mass spectra obtained from different samples or under different conditions may have peak position drift, which will affect subsequent comparison and analysis. Therefore, it is necessary to align the peaks to ensure that the peak positions of the same compound in different samples are consistent. Peak alignment methods include alignment based on retention time, alignment based on peak shape and features, etc. Peak alignment can eliminate the differences in peak positions between different samples and improve the comparability and repeatability of data. Use the Map Alignment Algorithm PoseClustering tool in pyopenms to align feature peaks and save the mzML file after peak alignment. This is a graph alignment algorithm based on pose clustering. It selects some of the highest intensity peak features for each spectrum and finds the most likely transition between RTs by analyzing the peak pair distance, where a peak pair refers to the pairing between a peak in one spectrum and a corresponding (or similar) peak in another spectrum.

[0059] The primary mass spectrometry data set after peak alignment was discretized and normalized.

[0060] In most cases, the range of ion mass-to-charge ratio scans and the number of peaks in each spectrum change dynamically, which is not suitable for analysis using deep learning models. Therefore, these data should first be uniformly discretized.

[0061] Here, we first divide the entire m / z range into windows of equal width. Then, in each mass spectrum, all peaks in the same window are summed in the m / z dimension to obtain the intensity matrix (IM).

[0062] ;

[0063] in, Indicates that the scan index is In the primary mass spectrum, Peaks in the window The number of rows i of the intensity matrix is ​​equal to the number of time points of the retention time, and the number of columns j is equal to the number of mass-to-charge ratio windows m / z_bin after discretization.

[0064] Finally, since the scale of the peak intensity in each window is inconsistent, the peak value needs to be divided by the maximum peak value of each scan for normalization, that is, each row element in IM is divided by the maximum value of the row element to obtain the Normalized Intensity Matrix (NIM).

[0065] The standardized intensity matrix NIM obtained through the above operation is used as the input of the subsequent mass spectrometry classification model.

[0066] Step S3, constructing an HCC mass spectrometry classification model using the normalized intensity matrix NIM and the classification information corresponding to the primary mass spectrometry data set.

[0067] In the process of model construction, CNN and Transformer were used as the basic structure of the classification model, the cross entropy loss function was used, and the Adam optimizer was used for model parameter learning.

[0068] After the raw mass spectrometry data has passed through the previous data preprocessing module, each mass spectrometry sample is converted into a tensor with a dimension of [RT_SIZE, 15400], where RT_SIZE represents the number of primary mass spectra in the mass spectrometry sample. Represents the number of m / z_bins divided in the mass spectrometry data, and the size of each mz_bin is 0.1. For the tensor representation of each sample, first sum it in the RT_SIZE dimension, and the dimension of the result data is [1, 15400]. Next, a one-dimensional CNN is used to extract local features. By performing convolution operations on the local features of the mass spectrometry data in the m / z dimension, the important information in the mass spectrometry data can be effectively captured and the representation ability of the features can be improved. Next, the multi-head attention mechanism (Multi Head Attention) in Transformer is introduced to further integrate the local features. Multi Head Attention can consider information at different positions at the same time and learn context-related feature representations, which is conducive to improving the model's understanding and representation ability of mass spectrometry data. Finally, the extracted features are layer-normalized to normalize the feature distribution, and a multi-layer fully connected layer is used for the final classification. Through this step, the model can map the mass spectrometry data to the category label to achieve accurate classification of the sample.

[0069] On the test set, the classification indicators of the mass spectrometry classification model are shown in Table 1.

[0070] Table 1 Model classification indicators on the test set

[0071]

[0072] As can be seen from Table 1, the mass spectrometry classification model performs well in terms of accuracy, precision, recall and F1 score on the test set, with an AUC greater than 0.9. The mass spectrometry classification model has strong classification capabilities.

[0073] The classification report of the mass spectrometry classification model on the test set is shown in Table 2.

[0074] Table 2 Classification report of the model on the test set

[0075]

[0076] As can be seen from Table 2, the mass spectrometry classification model provided in this application can achieve a classification accuracy of more than 90% by using only MS1 information.

[0077] Embodiment 2:

[0078] The embodiment of the present application also provides a mass spectrometry classification method based on deep learning, which uses the above step S2 to construct a standardized intensity matrix corresponding to the primary mass spectrometry data set to be classified, and inputs it into the above mass spectrometry classification model to obtain classification information of the mass spectrometry data set to be classified.

[0079] Taking HCC as an example, when the output value of the mass spectrometry classification model is greater than 0.5, it means that the input mass spectrometry data is judged to be HCC; when the output value of the mass spectrometry classification model is not greater than 0.5, it means that the input mass spectrometry data is judged to be non-HCC.

[0080] Embodiment three:

[0081] like Figure 2 As shown, Figure 2 A structural diagram of a deep learning mass spectrometry classification system provided in Example 3 of the present application, the system comprising:

[0082] A data input module, used to obtain a primary mass spectrometry data set to be classified;

[0083] A data storage module, used for storing standardized intensity matrices and classification information corresponding to a plurality of primary mass spectrometry data sets;

[0084] The classification module is connected to the data input module and the data storage module respectively, and is used to first classify the X init and its classification information to construct a mass spectrum classification model, and then input the primary mass spectrum data set to be classified into the model to obtain the classification information of the primary mass spectrum data set to be classified;

[0085] A result output module, connected to the classification module, for outputting classification results;

[0086] The normalized intensity matrix is ​​constructed according to step S2.

[0087] The result output module is further connected to the data storage module, and is used to input the standardized intensity matrix and classification results corresponding to the primary mass spectrometry data set to be classified into the data storage module. In this way, the mass spectrometry classification model is continuously optimized during the application process, thereby further improving the accuracy of classification.

[0088] Embodiment 4:

[0089] like Figure 3 As shown, Figure 3 A flowchart of a deep learning-based biomarker discovery method provided in Example 4 of the present application, the method comprising the following steps:

[0090] A primary mass spectrometry data set is acquired, and noise reduction, peak alignment, discretization and standardization are performed on the primary mass spectrometry data set. The primary mass spectrometry data set is split into multiple mass-to-charge ratio windows m / z_bin.

[0091] This step is the same as the step of processing the primary mass spectrometry data set in Example 1, and will not be repeated here for the sake of brevity.

[0092] For each mass-to-charge ratio window, a set of perturbation samples is generated respectively; in each perturbation sample, only the peak value corresponding to the mass-to-charge ratio window is changed;

[0093] Input each group of perturbation samples corresponding to all mass-to-charge ratio windows into the mass spectrometry classification model constructed in the first embodiment to obtain corresponding output values; use all perturbation samples and their corresponding output values ​​to construct a local linear model, the coefficient of the local linear model reflects the influence of the peak value of each mass-to-charge ratio window on the prediction result; wherein the absolute value of the coefficient indicates the influence, and the sign of the coefficient indicates the direction of the influence.

[0094] For the trained mass spectrometry classification model, the interpretability tool LIME was used to interpret the model and find the mass spectrometry features that play a key role in the classification results of HCC samples. These key features can help understand the decision-making process of the model and provide biological explanations, providing clues for subsequent biomarker discovery.

[0095] Our trained deep learning mass spectrometry classification model is denoted as The LIME method aims to Find a local linear model near , so that and Close enough, and The complexity is low and easy to explain.

[0096]

[0097]

[0098]

[0099]

[0100] In the above formula, Representative sample points The best interpretable model nearby, Measure the distance between the local linear model g and the mass spectrometry classification model f, Representation sample The weight of represents the complexity of the local linear model, represents the model parameters of the local linear model, Representation sample Nearby perturbation data points, Representation sample A vector representation of nearby perturbed data points.

[0101] Specifically, the LIME method will be analyzed at the sample point A series of perturbed data Z are generated near the input space, which are slightly different but not too far from the original instance. For mass spectrometry data, perturbations are mainly generated by adding Gaussian noise to its continuous features. For each perturbed data point Z, LIME uses the original complex classification model f to make predictions, thereby obtaining a probability distribution of the classification model output, which reflects the uncertainty of the classification model on different inputs near the sample. LIME is based on the distance between the perturbed sample and the original sample. To calculate the weight of each perturbed sample. The closer the perturbed sample is to the original sample, the greater its weight. Next, use these weighted perturbation data z and its corresponding model to predict , LIME trains a simple local linear model g. This local model is designed to capture the relationship between input and output around a specific sample. Finally, LIME uses this simple local linear model g to explain the predictions of the original mass spectrometry classification model f. The coefficients of the local linear model g can be directly interpreted as the importance of the features, thereby providing an explanation for the original model's predictions. Through this process, LIME is able to provide interpretability for specific samples without modifying the original model, allowing users to understand why the model makes specific predictions, thereby increasing trust in the model's predictions.

[0102] In the LIME method, when generating perturbed data, usually only one m / z_bin value is adjusted at a time, while keeping the values ​​of other m / z_bins unchanged. This is to independently evaluate the impact of each m / z_bin feature on the model prediction results.

[0103] Specifically, the LIME workflow includes:

[0104] Select an instance (i.e. a mass spectrum sample) that you want to interpret.

[0105] For the selected instance, multiple perturbation samples are generated for each m / z_bin feature. In each perturbation sample, only the value of one m / z_bin is changed (increased or decreased), while keeping the values ​​of all other m / z_bins the same as the original instance. This generates a set of perturbation samples for each m / z_bin.

[0106] Each perturbation sample is input into the original model (i.e., the mass spectrometry classification model) to obtain the corresponding prediction result.

[0107] Use all perturbation samples (one sample for each change in m / z_bin) and their predictions to train a simple local linear model (such as linear regression or logistic regression). This local linear model will try to fit the relationship between the perturbation samples and the predictions.

[0108] The coefficients of the local linear model will reflect the degree of influence of each m / z_bin feature on the model prediction results. The size of the coefficient indicates the magnitude of the influence, and the sign of the coefficient (positive or negative) indicates the direction of the influence.

[0109] In this way, LIME is able to evaluate the independent contribution of each m / z_bin feature to the model prediction results, thereby providing interpretability of the model predictions.

[0110] Through the above method, the importance metric corresponding to each m / z_bin in the mass spectrum MS1 is obtained, and the 50 most important m / z_bins are selected as key mass spectrum features for subsequent analysis. The m / z_bin classification importance results of HCC mass spectrum samples on one of the test sets are shown in Table 3.

[0111] Table 3 Results of mz_bin classification importance analysis of mass spectrometry samples

[0112]

[0113] Based on the above key mass spectrometry features, we used the open source mass spectrometry data analysis tool MaxQuant to associate these features with the corresponding proteins. This step can help identify proteins associated with HCC.

[0114] For each sample, we obtained 50 characteristic m / z_bins that contributed the most to the determination of hepatocellular carcinoma and 50 characteristic m / z_bins that contributed the most to the distinction of normal tissues. Next, we filtered the original mzML files of each mass spectrometry sample by secondary mass spectrometry (MS2), retaining only the MS2s that overlapped with these m / z_bins with high contributions, and re-saved these mass spectrometry data as mzML files. We re-identified proteins in these re-saved mzML files using the tool MaxQuant, and the protein library to be searched was the Homo sapiens proteome FASTA library from https: / / www.uniprot.org / proteomes / UP000005640.

[0115] For the protein identification results of the mzML file, we used the volcano plot visualization method to display the significance of differentially expressed proteins. See the volcano plot visualization of the original mzML file using MaxQuant for protein identification. Figure 4 , the key features found by deep learning model and LIME method are shown in the mzML file using MaxQuant to visualize the protein identification results using the volcano map. Figure 5 .

[0116] From the comparison, we can find that the mzML file found by the LIME method retains most of the significantly different proteins, while the number of non-significantly different proteins is significantly reduced. This shows that our method has indeed found key differential proteins, namely biomarkers.

[0117] The deep learning-based biomarker discovery method proposed in this application can discover HCC biomarkers based on key mass spectrometry features, providing new technical means and methods for the early diagnosis and treatment of HCC, and helping researchers to more accurately understand the biological significance and mechanism of action of biomarkers.

[0118] Embodiment five:

[0119] like Figure 6 As shown, Figure 6 A structural diagram of a deep learning-based biomarker discovery device provided in Example 5 of the present application, the device comprising:

[0120] An acquisition unit is used to acquire a primary mass spectrum data set, and perform noise reduction, peak alignment, discretization and standardization on the primary mass spectrum data set, where the primary mass spectrum data set is split into multiple mass-to-charge ratio windows;

[0121] A generating unit is used to generate a set of perturbation samples for each mass-to-charge ratio window; in each perturbation sample, only the peak value corresponding to the mass-to-charge ratio window is changed;

[0122] A construction unit is used to input each group of disturbance samples corresponding to all mass-to-charge ratio windows into the mass spectrum classification model constructed in the first embodiment to obtain corresponding output values; a local linear model is constructed using all disturbance samples and their corresponding output values, wherein the coefficient of the local linear model reflects the influence of the peak value of each mass-to-charge ratio window on the prediction result; wherein the absolute value of the coefficient indicates the influence, and the sign of the coefficient indicates the direction of the influence;

[0123] The determination unit is used to select the mass-to-charge ratio window before the preset value according to the coefficients of the local linear model, sorted from large to small, as the key mass spectrometry feature; and determine the biomarker based on the key mass spectrometry feature.

[0124] The deep learning-based biomarker discovery device provided in Example 5 of the present application has the same implementation principle and technical effects as those of the aforementioned method Example 4. For the sake of brief description, for matters not mentioned in the device embodiment, reference can be made to the corresponding contents in the aforementioned method Example 4.

[0125] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and that the present application can be implemented in other specific forms without departing from the spirit or essential features of the present application. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present application is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present application.

Claims

1. A biomarker discovery method based on deep learning, characterized in that: include: Acquiring a primary mass spectrometry data set, and performing noise reduction processing, peak alignment operation, discretization processing, and standardization processing on the primary mass spectrometry data set, wherein the primary mass spectrometry data set is split into a plurality of mass-to-charge ratio windows; For each mass-to-charge ratio window, a set of perturbation samples is generated respectively; in each of the perturbation samples, only the peak value corresponding to the mass-to-charge ratio window is changed; Inputting each group of perturbation samples corresponding to all mass-to-charge ratio windows into the mass spectrum classification model to obtain corresponding output values; constructing a local linear model using all the perturbation samples and their corresponding output values, wherein the coefficient of the local linear model reflects the influence of the peak value of each mass-to-charge ratio window on the prediction result; wherein the absolute value of the coefficient indicates the influence, and the sign of the coefficient indicates the direction of the influence; According to the coefficients of the local linear model, the mass-to-charge ratio windows that are ranked before a preset value are selected from large to small order as key mass spectrometry features; and a biomarker is determined based on the key mass spectrometry features; The steps of constructing the mass spectrometry classification model include: S1, acquiring a plurality of primary mass spectrometry data sets, including a primary mass spectrometry data set classified as a first category and a primary mass spectrometry data set classified as a second category; the primary mass spectrometry data set includes a mass-to-charge ratio, a peak value, and a retention time; S2, after performing noise reduction processing on the primary mass spectrometry data set, performing a peak alignment operation, and discretizing the primary mass spectrometry data set after the peak alignment to obtain an intensity matrix; the number of rows of the intensity matrix is ​​equal to the number of time points of the retention time, and the number of columns is equal to the number of mass-to-charge ratio windows after the discretization processing; according to the maximum peak value of each row of the intensity matrix, each row element of the intensity matrix is ​​standardized to obtain a standardized intensity matrix; S3, constructing a mass spectrometry classification model using the standardized intensity matrix and the classification information corresponding to the primary mass spectrometry data set.

2. The method for discovering biomarkers based on deep learning according to claim 1, characterized in that: The step of determining the biomarkers based on the key mass spectrometry features is determined using the open source mass spectrometry data analysis tool MaxQuant.

3. The method for discovering biomarkers based on deep learning according to claim 1, characterized in that: The peak alignment operation adopts an alignment method based on retention time or an alignment method based on peak shape features.

4. The method for discovering biomarkers based on deep learning according to claim 1, characterized in that: The discretization process comprises: Dividing the mass-to-charge ratio range into a plurality of mass-to-charge ratio windows of equal width; The peak values ​​of multiple peaks located in the same mass-to-charge ratio window are summed, and the sum of the peak values ​​is taken as the peak value corresponding to the mass-to-charge ratio window.

5. The method for discovering biomarkers based on deep learning according to claim 1, characterized in that: The mass spectrometry classification model includes a CNN module and a Transformer module. The performance of the mass spectrometry classification model is measured using a cross entropy loss function. The model parameters are optimized using an Adam optimizer until the training is stopped when the loss function converges.

6. A biomarker discovery device based on deep learning, characterized in that: include: an acquisition unit, used for acquiring a primary mass spectrum data set, and performing noise reduction, peak alignment, discretization and standardization on the primary mass spectrum data set, wherein the primary mass spectrum data set is split into a plurality of mass-to-charge ratio windows; A generating unit, for generating a set of disturbance samples for each mass-to-charge ratio window; in each of the disturbance samples, only the peak value corresponding to the mass-to-charge ratio window is changed; A construction unit is used to input each group of disturbance samples corresponding to all mass-to-charge ratio windows into the mass spectrum classification model constructed in claim 1 to obtain corresponding output values; a local linear model is constructed using all the disturbance samples and their corresponding output values, wherein the coefficient of the local linear model reflects the influence of the peak value of each mass-to-charge ratio window on the prediction result; wherein the absolute value of the coefficient indicates the influence, and the sign of the coefficient indicates the direction of the influence; A determination unit is used to select, according to the coefficients of the local linear model, the mass-to-charge ratio windows that are ranked before a preset value in descending order as key mass spectrometry features; and determine the biomarker based on the key mass spectrometry features.

Images