Mass spectrometry peak detection analysis method and apparatus

Abstract

The present application relates to the technical field of computer algorithm field, especially to a kind of mass spectrum peak detection analysis method and its device.The mass spectrum peak detection analysis method specifically includes the following steps: step S1: mass spectrum data is regarded as three-dimensional point cloud, and mass spectrum data is divided into window;Step S2: point cloud data extraction is carried out in each window area, and point cloud data is pretreated;Step S3: the point cloud data after pretreatment is analyzed based on the end-to-end multi-task network of deep learning.By the present application, 3D analysis of mass spectrum data can be realized, and the accuracy of mass spectrum peak quantitative analysis can be effectively guaranteed.

Description

Technical Field

[0001] This invention relates to the technical field of computer algorithms, and in particular to a method and apparatus for mass spectrometry peak detection and analysis. Background Technology

[0002] Chromatography-mass spectrometry (GC-MS) is widely used for component analysis in proteomics and metabolomics. In the detection and quantification of compounds using GC-MS, mass spectrometry acquires the sample as a time-series mass spectrum. Mass spectrometry data has three dimensions: retention time (RT), mass-to-charge ratio (m / z), and intensity. Different analyte ions exhibit an approximately two-dimensional Gaussian distribution in the m / z and RT dimensions, forming three-dimensional mass spectrometry peaks. These peaks vary in size and shape, are mixed with noise signals, and overlap. Due to the high dimensionality, high throughput, and high complexity of mass spectrometry data, accurate ion quantification is a major challenge in non-targeted mass spectrometry data analysis. Non-targeted data analysis methods first analyze all signal peaks and quantification results in the mass spectrometry data, and then correlate these results with the actual compound ions.

[0003] Existing non-targeted mass spectrometry data analysis methods are mainly divided into three categories: 2D analysis, 2.5D analysis, and 3D analysis. Most methods use 2D and 2.5D analysis, such as MarkerView, Compound Discoverer, MS-DIAL, MZmine2, XCMS Online, PeakOnly, MaxQuant, and Dinosaur. These methods reduce computation time and complexity through a dimensional analysis approach. 2D and 2.5D methods first analyze the m / z dimension, detecting ions and intensities at a specific RT time, and then discover ion peaks in the RT dimension. However, these methods only consider the distribution of ion intensity in the RT dimension when detecting ion peaks, neglecting the signal distribution in the m / z dimension. This incomplete data analysis results in lower ion peak extraction accuracy.

[0004] In recent years, an increasing number of methods have begun to utilize 3D approaches to improve the accuracy of mass spectrometry data analysis by simultaneously analyzing the distribution of data signals in both the m / z and RT dimensions. Some methods use Gaussian fitting to match the acquired 3D mass spectrometry signals with theoretical peaks in a 2D Gaussian distribution, estimating the probability that each mass spectrometry signal belongs to a mass spectrometry peak by measuring the quality of the fit. Other methods employ image-based deep learning methods, such as DeepIso and SeA-M2Net, to convert the intensity dimension of the 3D mass spectrometry data into the color of an image, forming a 2D image in both m / z and RT coordinate systems, and then using computer vision-based object detection methods to select mass spectrometry peaks.

[0005] However, neither fitting methods nor image methods can obtain accurate mass spectrometry peak detection results. Summary of the Invention

[0006] To improve the accuracy of mass spectrometry peak detection and quantification, this invention proposes a mass spectrometry peak detection and analysis method, the specific scheme of which is as follows:

[0007] The mass spectrometry peak detection and analysis method includes the following steps:

[0008] Step S1: Treat the mass spectrometry data as a three-dimensional point cloud and divide the mass spectrometry data into windows;

[0009] Step S2: Extract point cloud data in each window area and preprocess the point cloud data;

[0010] Step S3: Analyze the preprocessed point cloud data using an end-to-end multi-task network based on deep learning.

[0011] Furthermore, the step S1 of dividing the mass spectrometry data into windows specifically includes the following steps:

[0012] Step S11: Divide the mass spectrometry data into grids at equal intervals starting from the minimum value in both the mass-to-charge ratio and retention time dimensions according to the user-defined grid width, and treat each grid as an initial segmentation window.

[0013] Step S12: Using the center of the initial segmentation window as the center, expand each initial window according to the predetermined expansion width in two dimensions: retention time and mass-to-charge ratio to obtain an expanded window.

[0014] Furthermore, the expansion width is greater than the maximum peak width of the signal peak in the corresponding dimension.

[0015] Furthermore, the preprocessing of the point cloud data in step S2 specifically includes:

[0016] The point cloud data is normalized based on the theoretical half-peak width in two dimensions: retention time and mass-to-charge ratio.

[0017] In terms of intensity, the logarithm of the intensity of each signal point is taken.

[0018] Furthermore, step S3 specifically includes the following steps:

[0019] Step S31: Extract point-by-point multi-scale spatial distribution features from the preprocessed point cloud data using the backbone network;

[0020] Step S32: Using the extracted results, perform mass spectrometry peak semantic prediction, mass spectrometry peak center prediction, and mass spectrometry peak boundary prediction respectively using the corresponding prediction branches;

[0021] Step S33: Aggregate the prediction results of mass spectrometry peak semantic prediction, mass spectrometry peak center prediction, and mass spectrometry peak boundary prediction through the result combination algorithm.

[0022] Furthermore, step S31 specifically includes the following steps:

[0023] Step S311: Apply local spatial coding blocks to extract local spatial features for each point;

[0024] Step S312: Apply multiple encoding modules to perform hierarchical spatial feature encoding on the point feature data. In each encoding module, the point features are aggregated through intensity probability sampling and point-by-point interpolation convolution.

[0025] Step S313: Apply multiple decoding blocks to propagate the extracted hierarchical spatial features to the original mass spectrum signal points. The point feature data extracted from different levels are combined through nearest neighbor interpolation and concatenation. Then, a shared MLP is applied to convert the concatenated features into the final decoded features.

[0026] Furthermore, step S33 specifically includes the following steps:

[0027] Step S331: Analyze the prediction results of mass spectrum peak semantic prediction and mass spectrum peak center prediction to select potential mass spectrum peak center points;

[0028] Step S332: By analyzing the relationship between the 3D coordinates of the potential mass spectrum peak center point and the polar coordinate mask of the predicted mass spectrum peak boundary, duplicate mass spectrum peaks are screened out.

[0029] Step S333: Predict the boundary of the intersection of the mass spectrometry peaks after separation and screening.

[0030] Another embodiment of the present invention discloses a mass spectrometry peak detection and analysis device, comprising,

[0031] The transformation partitioning unit is used to treat mass spectrometry data as a three-dimensional point cloud and to divide the mass spectrometry data into windows;

[0032] The data processing unit is used to extract point cloud data in each window area and preprocess the point cloud data.

[0033] The data analysis unit is used to analyze preprocessed point cloud data using an end-to-end multi-task network based on deep learning.

[0034] Furthermore, the data analysis unit includes

[0035] A spatial feature extraction backbone network is used to perform point-by-point multi-scale feature extraction on preprocessed point cloud data.

[0036] A prediction branch network is used to perform semantic prediction of mass spectrometry peaks, prediction of mass spectrometry peak centers, and prediction of mass spectrometry peak boundaries based on the extracted results.

[0037] The computational unit is used to aggregate the prediction results of the prediction branch network through a result combination algorithm.

[0038] Another embodiment of the present invention discloses a computer device, including a processor and a storage medium, wherein a computer program is stored on the storage medium, and the processor reads from the storage medium and runs the computer program to execute a mass spectrometry peak detection and analysis method as described in any of the above embodiments. Attached Figure Description

[0039] Figure 1 This is a flowchart of a mass spectrometry peak detection and analysis method proposed in this invention;

[0040] Figure 2 The flowchart shows the steps of analyzing preprocessed point cloud data based on a deep learning end-to-end multi-task network in the mass spectrometry peak detection and analysis method proposed in this invention.

[0041] Figure 3 This invention presents a flowchart of the steps in the mass spectrometry peak detection and analysis method to aggregate prediction results using a combination algorithm.

[0042] Figure 4 In the mass spectrometry peak detection and analysis method proposed in this invention, a backbone network is used to extract multi-scale features from the preprocessed point cloud data point by point.

[0043] Figures 5a-5c This is a schematic diagram illustrating the cutting and dilation of mass spectrometry data in the mass spectrometry peak detection and analysis method proposed in this invention;

[0044] Figure 6 This is a schematic diagram of the structure of the analysis unit in the mass spectrometry peak detection and analysis device proposed in this invention;

[0045] Figures 7a-7c To compare our method with popular non-targeted analysis software in tests on publicly available metabolomics evaluation datasets;

[0046] Figures 8a-8d To compare this method with popular non-targeted analysis software in tests on proteomics datasets and evaluation datasets;

[0047] Figure 9 This is a comparison chart showing the runtime of this invention with other deep learning software and conventional analysis software. Detailed Implementation

[0048] like Figure 1 As shown, this invention discloses a method for mass spectrometry peak detection and analysis, comprising the following steps:

[0049] Step S1: Treat the mass spectrometry data as a three-dimensional point cloud and divide the mass spectrometry data into windows;

[0050] Step S2: Extract point cloud data in each window area and preprocess the point cloud data;

[0051] Step S3: Analyze the preprocessed point cloud data using an end-to-end multi-task network based on deep learning.

[0052] Specifically, a 3D point cloud is a data format composed of 3D coordinate points, and mass spectrometry data itself has a similar data structure to a 3D point cloud. In this embodiment, mass spectrometry data is approximated as a 3D point cloud, and then further analyzed. Since a holistic analysis of the mass spectrometry data is required, but considering the limitations of computer memory in practical applications, the mass spectrometry data needs to be divided into several windows. Point cloud data is extracted within each window region, where point cloud data refers to the mass spectrometry data treated as a 3D point cloud. After preprocessing the point cloud data, it can be analyzed using methods for analyzing 3D point clouds. In this embodiment, an end-to-end task network based on deep learning analyzes the preprocessed point cloud data. This embodiment uses a novel mass spectrometry data analysis method, abandoning destructive analysis methods such as previous dimension-by-dimensional analysis. By treating mass spectrometry data as a 3D point cloud, a deep learning model is used to perform direct spatial analysis of the mass spectrometry point cloud, achieving for the first time the analysis of non-destructive chromatography-mass spectrometry data, resulting in higher accuracy in mass spectrometry peak detection and quantification.

[0053] Furthermore, the step S1 of dividing the mass spectrometry data into windows specifically includes the following steps:

[0054] Step S11: Divide the mass spectrometry data into grids at equal intervals starting from the minimum value in both the mass-to-charge ratio and retention time dimensions according to the user-defined grid width, and treat each grid as an initial segmentation window.

[0055] Step S12: Using the center of the initial segmentation window as the center, expand each initial window according to the predetermined expansion width in two dimensions: retention time and mass-to-charge ratio to obtain an expanded window.

[0056] Although this embodiment does not limit the number of input points for mass spectrometry data, due to computer memory limitations, the entire mass spectrometry point cloud cannot be analyzed at once. Therefore, it is necessary to divide the mass spectrometry point cloud into multiple parts for analysis. Specifically, firstly, the segmentation width of the mass spectrometry data in the two dimensions of retention time (RT) and mass-to-charge ratio (m / z) is obtained. Starting from the minimum value of the above two dimensions, the data is divided into equally spaced grid windows according to the segmentation width. For example, as shown... Figures 5a-5c As shown, by Figure 5aAs can be seen, the horizontal direction represents the retention time dimension, and the vertical direction represents the mass-to-charge ratio dimension; Figure 5b To illustrate the division of the original mass spectrometry data distribution into four sub-windows along the m / z and RT dimensions, a long dashed line is used. For ease of explanation, in practical applications, this might involve tens of thousands of windows. Each grid is considered an initial window. The center point of each initial window is calculated, and then each initial window is expanded around that center point. Figure 5c The short dashed lines represent four overlapping expansion windows.

[0057] Furthermore, the expansion width is greater than the maximum peak width of the signal peak in the corresponding dimension.

[0058] Using the segmentation method of this embodiment, all signal points in the signal peak centered within the initial window will be completely contained within the expanded window. This embodiment uses the expanded window to segment the point cloud, and then performs three-dimensional mass spectrometry peak extraction on each segmented point cloud. In the search for mass spectrometry peaks, only the detection results centered within the initial window are retained. Using this point cloud segmentation and result merging method ensures the comprehensiveness and non-reproducibility of the detection.

[0059] The preprocessing of the point cloud data specifically includes:

[0060] The point cloud data is normalized based on the theoretical half-peak width in two dimensions: retention time and mass-to-charge ratio.

[0061] In terms of intensity, the logarithm of the intensity of each signal point is taken.

[0062] Specifically, regardless of whether the point cloud data is normalized or logarithmized, all preprocessing operations are reversible. Therefore, in this embodiment, the preprocessing of the point cloud data will not reduce the accuracy of the mass spectrometry data, and accurate mass spectrometry peak detection results can be obtained.

[0063] like Figure 2 As shown, step S3 specifically includes the following steps:

[0064] Step S31: Extract point-by-point multi-scale spatial distribution features from the preprocessed point cloud data using the backbone network;

[0065] Step S32: Using the extracted results, perform mass spectrometry peak semantic prediction, mass spectrometry peak center prediction, and mass spectrometry peak boundary prediction respectively using the corresponding prediction branches; all three prediction branches are implemented using fully connected networks.

[0066] Step S33: Aggregate the prediction results of mass spectrometry peak semantic prediction, mass spectrometry peak center prediction, and mass spectrometry peak boundary prediction through the result combination algorithm.

[0067] Specifically, the extraction results can be effectively improved by performing semantic prediction of mass spectrometry peaks, prediction of mass spectrometry peak centers, and prediction of mass spectrometry peak boundaries, thereby reducing the interference of various factors. Finally, the multi-dimensional prediction results are aggregated through the result combination algorithm.

[0068] Furthermore, step S31 specifically includes the following steps:

[0069] Step S311: Apply local spatial coding blocks to extract local spatial features for each point; specifically, local spatial coding blocks use a shared multilayer perceptron (MLP) to encode the 3D coordinates of the nearest neighbors into local spatial features;

[0070] Step S312: Apply multiple encoding modules to perform hierarchical spatial feature encoding on the point feature data. In each encoding module, the point features are aggregated by intensity probability sampling and point-by-point interpolation convolution. Although the point cloud is sampled into fewer points, each sampled point obtains a larger scale of features by aggregating features from a larger receptive field.

[0071] Step S313: Apply multiple decoding blocks to propagate the extracted hierarchical spatial features to the original mass spectrum signal points. The point features extracted from different levels are combined through nearest neighbor interpolation and concatenation. Then, a shared MLP is applied to convert the concatenated features into the final decoded features.

[0072] The following provides a detailed description of the local space coding module, the coding module, and the decoding module.

[0073] like Figure 4 As shown, the local spatial coding module performs a K-nearest neighbor search for each point in the mass spectral point cloud of N input points, and calculates the density features of the K-nearest neighbors. The K-nearest neighbor points are then input into a shared MLP for spatial feature encoding. After that, the encoded features are obtained through average pooling. The local spatial coding result is obtained by combining the current point coordinates, the encoded features, and the nearest neighbor density.

[0074] The encoding module first downsamples the point cloud with local spatial features. For each sampled point, it performs two-dimensional interpolation convolution and shared MLP encoding in the dimensions of retention time and mass-to-charge ratio to obtain a larger range of spatial features. The coordinates of the current point are combined with the encoded larger range of features to obtain the encoding result.

[0075] The decoding module needs to pass the high-dimensional features obtained by the encoding module to the point cloud before downsampling. For each point in the point cloud before downsampling, a K-nearest neighbor search is performed in the point cloud after downsampling. The high-dimensional features obtained by the search are weighted and summed according to the distance of the nearest neighbors. The obtained high-dimensional features are combined with the current point coordinates and the original features to obtain the decoding result.

[0076] In step S32, the mass spectrum peak semantic prediction branch is used to predict the probability that each mass spectrum signal point belongs to a mass spectrum peak, in order to distinguish mass spectrum peaks from noise. The mass spectrum peak center prediction branch is used to predict the probability that each mass spectrum signal point is the center point of a mass spectrum peak, in order to distinguish different mass spectrum peaks. The mass spectrum peak boundary prediction branch is used to predict the boundary range of the mass spectrum peak to which each mass spectrum signal point belongs, in order to obtain the quantitative range of each mass spectrum peak.

[0077] Furthermore, such as Figure 3 As shown, step S33 specifically includes the following steps:

[0078] Step S331: Analyze the prediction results of mass spectrum peak semantic prediction and mass spectrum peak center prediction to select potential mass spectrum peak center points;

[0079] Specifically, the steps for selecting the center point of a mass spectrometry peak are as follows: The results of semantic prediction and center prediction are filtered according to a threshold. Data points whose semantic prediction and center prediction results are both greater than the user-defined threshold are selected as potential mass spectrometry peak centers. This joint filtering method can obtain more accurate peak center prediction results compared to single-branch prediction.

[0080] Step S332: By analyzing the relationship between the 3D coordinates of the potential mass spectrum peak center point and the polar coordinate mask of the predicted mass spectrum peak boundary, duplicate mass spectrum peaks are screened out.

[0081] Preferably, for several potential mass spectrum peak centers falling in the same polar coordinate mask, only the center point with the highest intensity is retained.

[0082] Step S333: Predict the boundary of the intersection of the mass spectrometry peaks after separation and screening.

[0083] Specifically, for cases where polar coordinate masks of different mass spectral peaks intersect in the prediction results, the two intersection points of the intersecting mass spectral peak boundaries are calculated, and the intersecting mass spectral peak boundaries are separated by a straight line passing through the two intersection points, so that each mass spectral signal corresponds to at most one mass spectral peak.

[0084] The mass spectrum peak semantic prediction, mass spectrum peak center prediction, and mass spectrum peak boundary prediction branches predict the probability that each signal point in the mass spectrum data belongs to a mass spectrum peak, the probability that it is a mass spectrum peak center point, and the polar coordinate boundary of the mass spectrum peak centered at the current point, respectively. The prediction results of the three prediction branches are integrated through a result fusion mechanism.

[0085] To achieve the above method, this embodiment also discloses a mass spectrometry peak detection and analysis device, including...

[0086] The transformation partitioning unit is used to treat mass spectrometry data as a three-dimensional point cloud and to divide the mass spectrometry data into windows;

[0087] The data processing unit is used to extract point cloud data in each window area and preprocess the point cloud data.

[0088] The data analysis unit is used to analyze preprocessed point cloud data using an end-to-end multi-task network based on deep learning.

[0089] Furthermore, such as Figure 6 As shown, the data analysis unit includes,

[0090] A spatial feature extraction backbone network is used to perform point-by-point multi-scale feature extraction on preprocessed point cloud data.

[0091] A prediction branch network is used to perform semantic prediction of mass spectrometry peaks, prediction of mass spectrometry peak centers, and prediction of mass spectrometry peak boundaries based on the extracted results.

[0092] The computational unit is used to aggregate the prediction results of the above three branches through a result combination algorithm.

[0093] The beneficial effects of this invention include:

[0094] 1. For the first time, direct analysis of non-destructive mass spectrometry data was achieved.

[0095] This invention employs a novel mass spectrometry data analysis method, abandoning the original destructive analysis methods such as dimension-by-dimensional analysis. By treating mass spectrometry data as a three-dimensional point cloud, a deep learning model is used to perform direct spatial analysis of the mass spectrometry point cloud. This is the first time that non-destructive chromatography-mass spectrometry data analysis has been achieved, resulting in higher accuracy in mass spectrometry peak detection and quantification.

[0096] 2. A higher number of mass spectrometry peaks detected

[0097] In tests on publicly available metabolomics and proteomics evaluation datasets, this patented method achieved the highest number of mass spectrometry peak identifications on both metabolomics and proteomics datasets compared to popular non-targeted analysis software. In both metabolomics datasets, this invention achieved the highest mass spectrometry peak detection percentages, reaching 95.7% and 98.9% respectively, significantly higher than the second-ranked 91.1% and 96.6%. Figure 7a In the proteomics dataset, this invention also achieved the highest detection rate of 98.8%, higher than the second-place 98.12%. Figure 8a ).

[0098] 3. Higher accuracy and stability in mass spectrometry peak quantification

[0099] In tests on publicly available metabolomics and proteomics evaluation datasets, the method of this patent exhibits the highest accuracy in mass spectrometry peak quantification. Figure 7a(8b). In the quantification of results from repeated technical experiments, the method of this patent has the lowest intensity detection fluctuation ( Figure 7b In comparing measured intensity with theoretical concentration, the method of this patent exhibits the highest consistency and best mass spectrometry peak quantification performance. Figure 7c ).

[0100] 4. Higher detection quantity and accuracy of biomarkers

[0101] In tests on publicly available metabolomics evaluation datasets, this patent achieved the highest number of biomarker identifications on both metabolomics datasets, while having no misidentified biomarkers, achieving the highest number and accuracy of biomarker identifications (7a).

[0102] 5. Robustness and universality

[0103] All the test results for this patent were based on the same pre-trained model. The training dataset was derived from the first sample of the first metabolomics dataset, which was discarded in subsequent comparisons to avoid affecting the results. The three evaluation datasets were collected by different mass spectrometry systems and have different data distributions. This patent achieved the highest accuracy in mass spectrometry peak detection and quantification on all test datasets, demonstrating excellent universality and robustness.

[0104] 6. Reasonable calculation time:

[0105] This patent utilizes a deep learning-based analysis method. In a comparison of runtime, this patent's runtime is approximately 5 times faster than other deep learning-based software, and comparable to that of conventional analysis software. Figure 9 ).

[0106] In another embodiment of the present invention, a computer device is disclosed, including a processor and a storage medium, wherein a computer program is stored on the storage medium, and the processor reads from the storage medium and runs the computer program to execute a mass spectrometry peak detection and analysis method described in any of the above embodiments.

[0107] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for mass spectrometry peak detection and analysis, characterized in that, The mass spectrometry peak detection and analysis method includes the following steps: Step S1: Treat the mass spectrometry data as a three-dimensional point cloud and divide the mass spectrometry data into windows; Step S2: Extract point cloud data in each window area and preprocess the point cloud data; Step S3: Analyze the preprocessed point cloud data using an end-to-end multi-task network based on deep learning; Specifically, step S3 includes the following steps: Step S31: Extract point-by-point multi-scale spatial distribution features from the preprocessed point cloud data using the backbone network; Step S32: Using the extracted results, perform mass spectrometry peak semantic prediction, mass spectrometry peak center prediction, and mass spectrometry peak boundary prediction respectively using the corresponding prediction branches; Step S33: Aggregate the prediction results of mass spectrometry peak semantic prediction, mass spectrometry peak center prediction, and mass spectrometry peak boundary prediction through the result combination algorithm; Specifically, step S31 includes the following steps: Step S311: Apply local spatial coding blocks to extract local spatial features for each point; Step S312: Apply multiple encoding modules to perform hierarchical spatial feature encoding on the feature data of the points. In each encoding module, the point features are aggregated by intensity probability sampling and point-by-point interpolation convolution. Step S313: Apply multiple decoding blocks to propagate the extracted hierarchical spatial features to the original mass spectrum signal points. The point feature data extracted from different levels are combined through nearest neighbor interpolation and concatenation. Then, a shared MLP is applied to convert the concatenated features into the final decoded features.

2. The mass spectrometry peak detection and analysis method according to claim 1, characterized in that, Step S1, which involves dividing the mass spectrometry data into windows, specifically includes the following steps: Step S11: Divide the mass spectrometry data into grids at equal intervals starting from the minimum value in both the mass-to-charge ratio and retention time dimensions according to the user-defined grid width, and treat each grid as an initial segmentation window. Step S12: Using the center of the initial segmentation window as the center, expand each initial window according to the predetermined expansion width in two dimensions: retention time and mass-to-charge ratio to obtain an expanded window.

3. The mass spectrometry peak detection and analysis method according to claim 2, characterized in that, The expansion width is greater than the maximum peak width of the signal peak in the corresponding dimension.

4. The mass spectrometry peak detection and analysis method according to claim 1, characterized in that, The preprocessing of point cloud data in step S2 specifically includes: The point cloud data is normalized based on the theoretical half-peak width in two dimensions: retention time and mass-to-charge ratio. In terms of intensity, the logarithm of the intensity of each signal point is taken.

5. The mass spectrometry peak detection and analysis method according to claim 1, characterized in that, Step S33 specifically includes the following steps: Step S331: Analyze the prediction results of mass spectrum peak semantic prediction and mass spectrum peak center prediction to select potential mass spectrum peak center points; Step S332: By analyzing the relationship between the 3D coordinates of the potential mass spectrum peak center point and the polar coordinate mask of the predicted mass spectrum peak boundary, duplicate mass spectrum peaks are screened out. Step S333: Predict the boundary of the intersection of the mass spectrometry peaks after separation and screening.

6. A mass spectrometry peak detection and analysis device, characterized in that, The mass spectrometry peak detection and analysis method for performing any one of claims 1 to 5 includes: The transformation partitioning unit is used to treat mass spectrometry data as a three-dimensional point cloud and to divide the mass spectrometry data into windows; The data processing unit is used to extract point cloud data in each window area and preprocess the point cloud data. The data analysis unit is used to analyze preprocessed point cloud data using an end-to-end multi-task network based on deep learning.

7. A computer device, characterized in that, It includes a processor and a storage medium, on which a computer program is stored, and the processor reads from the storage medium and runs the computer program to execute a mass spectrometry peak detection and analysis method as described in any one of claims 1 to 5.

Images