High resolution mass spectrometry - deep learning driven automatic analysis method for metabolomics data

Abstract

The present application relates to a kind of high-resolution mass spectrometry-depth learning driven metabolomics data automatic analysis method, belong to data science technical field.The method includes: obtaining original mass spectrum data file by liquid chromatography-high resolution mass spectrometry technique;Through automatic pre-processing module, peak extraction, peak alignment, denoising processing are carried out to original mass spectrum data file, and standardization metabolite feature quantitative matrix is generated based on feature extraction module;Based on the standardization metabolite feature quantitative matrix and the secondary mass spectrum spectrum information in original mass spectrum data file, metabolite feature is automatically identified by adaptive identification model;The metabolite and corresponding metabolite feature quantitative matrix identified are matched with the best result in database automatically, and the corresponding molecular formula composition and compound name are obtained.The speed and accuracy of metabolite identification are significantly improved, and the error caused by manual intervention is reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data science, and particularly relates to a high-resolution mass spectrometry-deep learning driven metabolomics data automatic analysis method. BACKGROUND

[0002] Current metabolomics research widely adopts liquid chromatography-high resolution mass spectrometry (LC-HRMS) technology, and its data analysis process includes:

[0003] Raw data preprocessing: peak extraction and alignment are performed through commercial software (such as XCMS, MS-DIAL), but the algorithm depends on fixed parameters, and the retention time drift needs to be manually adjusted;

[0004] Metabolite identification: relies on manual comparison of experimental MS / MS spectrum and public standard database (such as HMDB, MassBank), and the matching result needs to be reviewed by experts for isomer and isomer;

[0005] Downstream analysis: use independent tools (such as MetaboAnalyst) for pathway enrichment, which is disconnected with the data in the identification link.

[0006] The technical problems existing at present are:

[0007] 1. Insufficient efficiency and accuracy of preprocessing: the traditional peak alignment algorithm (such as LOESS) is sensitive to non-linear drift of retention time, and needs to optimize parameters repeatedly; the denoising process lacks standardized quality control (such as not forced RSD filtering), which leads to low reproducibility of features mixed into the quantitative matrix.

[0008] 2. Low degree of metabolite identification automation: the existing method cannot jointly use the primary quantitative characteristics (parent ion intensity / isotope pattern) and secondary structure information, which leads to the dependence of manual inspection of ion abundance ratio for distinguishing similar compounds (such as structural isomers), and the low-abundance metabolites are missed due to poor MS / MS spectrum quality.

[0009] 3. Database coverage and matching reliability defects: the public spectrum library has limited coverage (especially new metabolites), and has not integrated virtual fragment spectrum; the matching result lacks multi-dimensional verification (such as isotope distribution, parent ion intensity), and the false matching rate is high; there is no adaptive processing mechanism (such as expanding the mass-to-charge ratio tolerance range) for low confidence results.

[0010] 4. Analysis process fragmentation: preprocessing, identification and pathway analysis need to be operated across platforms, and manual data transmission introduces errors; there is lack of end-to-end automatic output (from raw data to compound name / pathway map).

[0011] 5. Model generalization and computational efficiency bottleneck: traditional identification models cannot incrementally learn, and new data types need to be retrained; CPU processing massive mass spectrometry data is slow and difficult to meet the high-throughput demand of samples. SUMMARY

[0012] To solve the above problems existing in the prior art, the present application provides a high-resolution mass spectrometry-deep learning driven metabolomics data automatic analysis method; the purpose of the present application can be realized by the following technical scheme, which comprises:

[0013] S1: obtaining an original mass spectrometry data file through liquid chromatography-high resolution mass spectrometry technology; performing peak extraction, peak alignment, and denoising processing on the original mass spectrometry data file through an automatic preprocessing module, and generating a standardized metabolite feature quantitative matrix based on a feature extraction module;

[0014] S2: based on the standardized metabolite feature quantitative matrix and the secondary mass spectrometry spectrum information in the original mass spectrometry data file, automatically identifying metabolite features through an adaptive recognition model;

[0015] S3: automatically matching the best result in the database for the identified metabolites and the corresponding metabolite feature quantitative matrix, and obtaining the corresponding molecular formula composition and compound name.

[0016] Specifically, the database matching in S3 comprises the following steps:

[0017] Based on the mass-to-charge ratio and retention time of the metabolite feature quantitative matrix, preliminary screening is performed in the local and cloud metabolite standard library;

[0018] Real-time comparison is performed between the experimental spectrum and the standard spectrum of the candidate compound screened, and the spectrum similarity score is calculated through a deep learning model;

[0019] A comprehensive confidence score is generated according to the similarity score, the consistency of the parent ion and the daughter ion intensity ratio, and the isotopic distribution coincidence degree;

[0020] The matching result with a comprehensive confidence score higher than a preset threshold is automatically selected, and the molecular formula and the compound name are output.

[0021] Specifically, the peak alignment in S1 adopts a dynamic time warping algorithm based on retention time correction, and the denoising processing includes blank sample background deduction and mixed sample relative standard deviation filtering.

[0022] Specifically, the input data of the adaptive recognition model in S2 comprises:

[0023] The parent ion intensity ratio and the isotopic distribution pattern extracted from the metabolite feature quantitative matrix;

[0024] The secondary mass spectrum spectrum information is vectorized into a sub-ion intensity sequence.

[0025] Further, the deep learning model fuses the primary mass spectrum feature vector and the secondary spectrum feature vector through a dot product vector algorithm.

[0026] Further, the metabolite standard library comprises a self-built standard sample library, a public database, and an algorithm-predicted virtual fragment spectrum library, wherein the virtual fragment spectrum is generated by a spectrum prediction tool.

[0027] Specifically, the adaptive recognition model in S2 is deployed with an online updating mechanism, and the incremental training is automatically triggered based on the user feedback of the identification result error samples.

[0028] Specifically, based on the identification result, a metabolic pathway visualization map is automatically generated, and the enrichment significance of the differential metabolites in the pathway is labeled.

[0029] Specifically, when the similarity score of the best result is lower than the preset threshold, the mass-to-charge ratio error tolerance range is expanded, and the isomer differentiation module is called to analyze the characteristic ion abundance ratio.

[0030] The method of the present application has significant advantages. In the pretreatment link, based on the dynamic time warping algorithm and standardization denoising processing of retention time correction, the efficiency and accuracy of pretreatment are effectively improved, avoiding the sensitive problem of traditional algorithm to the nonlinear drift of retention time and the defect of lack of standardization control in denoising. In the metabolite identification aspect, the adaptive recognition model can jointly use the primary quantitative features and secondary structure information, improve the discrimination ability of similar compounds, reduce the need for manual inspection, and the online updating mechanism ensures that the model can continuously learn and optimize. In the database matching process, the comprehensive confidence score is generated by combining multiple dimensions of information, which increases the reliability of the matching result, and has an adaptive processing mechanism for low confidence results. In addition, the entire analysis process realizes end-to-end automation, directly outputs the compound name and metabolic pathway visualization map from the raw data, avoiding the error introduced by data fragmentation and manual data transmission. In summary, the high-resolution mass spectrometry - deep learning driven metabolomics data automatic analysis method of the present application provides an efficient, accurate and highly automated solution for metabolomics research. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to facilitate the understanding of those skilled in the art, the present application will be further described below with reference to the accompanying drawings.

[0032] Fig. 1 is a flowchart of a high-resolution mass spectrometry - deep learning driven metabolomics data automatic analysis method of the present application.

[0033] Fig. 2A schematic diagram of data processing operation of the present application. DETAILED DESCRIPTION

[0034] To further illustrate the technical means and effects taken by the present application to achieve the predetermined inventive purposes, the specific embodiments, structures, features and effects thereof according to the present application are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0035] Please refer to Figs. 1-2 A high-resolution mass spectrometry-deep learning driven metabolomics data automatic analysis method, comprising:

[0036] S1: Obtain an original mass spectrometry data file through liquid chromatography-high resolution mass spectrometry technology; perform peak extraction, peak alignment, and noise removal processing on the original mass spectrometry data file through an automatic preprocessing module, and generate a standardized metabolite feature quantitative matrix based on a feature extraction module;

[0037] S2: Based on the standardized metabolite feature quantitative matrix and the secondary mass spectrometry spectrum information in the original mass spectrometry data file, automatically identify metabolite features through an adaptive identification model;

[0038] S3: Automatically match the best results of the identified metabolites and the corresponding metabolite feature quantitative matrix in a database, and obtain the corresponding molecular formula composition and compound name.

[0039] In this embodiment, the automatic preprocessing module uses a dynamic time warping algorithm based on retention time correction for peak alignment, which can effectively handle the nonlinear drift of retention time and avoid the problems sensitive to traditional algorithms. In the noise removal process, blank sample deduction and relative standard deviation filtering are strictly implemented to ensure the high quality of the quantitative matrix. After receiving the parent ion intensity ratio and isotope distribution pattern extracted from the metabolite feature quantitative matrix, as well as the vectorized secondary mass spectrometry spectrum sub-ion intensity sequence, the adaptive identification model can accurately automatically identify metabolite features. Its online update mechanism will trigger incremental training in time according to the user feedback of identification error samples, so that the model always maintains high accuracy and adaptability.

[0040] In the database matching step, first, the mass-to-charge ratio and the retention time of the metabolite feature quantitative matrix are based on the initial screening in the self-built standard library, the public database and the algorithm prediction virtual fragment spectrum library. The self-built standard library combines the data of the standard in the actual research, the public database ensures the wide coverage, and the virtual fragment spectrum is calculated by the spectrum prediction tool to make up for the lack of coverage of the public spectrum library. After real-time comparison of the experimental spectrum with the reference spectrum of the candidate compound, the deep learning model fuses the primary mass spectrum feature vector and the secondary spectrum feature vector through the dot product vector algorithm to calculate the spectrum similarity score. Then, the comprehensive confidence score is generated according to the similarity score, the ion intensity ratio consistency and the isotope distribution coincidence, and the matching result with the comprehensive confidence score higher than the preset threshold is automatically selected to output the molecular formula and the compound name. If the similarity score of the best matching result is lower than the preset threshold, the system will expand the mass-to-charge ratio error tolerance range and call the isomer differentiation module to analyze the characteristic ion abundance ratio to ensure the effective processing of the low confidence result.

[0041] After the metabolite identification is completed, the system will automatically generate a metabolic pathway visualization map based on the identification result, and mark the enrichment significance of the differential metabolites in the pathway. Through these steps, the method of the application truly realizes the end-to-end automation of metabolomics data from raw data to final analysis results, and provides great convenience and efficient solutions for metabolomics research.

[0042] Specifically, the database matching in S3 includes the following steps:

[0043] Based on the mass-to-charge ratio and the retention time of the metabolite feature quantitative matrix, initial screening is performed in the local and cloud metabolite standard library;

[0044] Real-time comparison of the experimental spectrum with the standard spectrum of the candidate compound is performed, and the deep learning model is used to calculate the spectrum similarity score;

[0045] According to the similarity score, the ion intensity ratio consistency of the parent ion and the daughter ion, and the isotope distribution coincidence, a comprehensive confidence score is generated;

[0046] The matching result with the comprehensive confidence score higher than the preset threshold is automatically selected to output the molecular formula and the compound name.

[0047] Specifically, the peak alignment in S1 adopts the dynamic time warping algorithm based on the retention time correction, and the denoising processing includes blank sample background subtraction and mixed sample relative standard deviation filtering.

[0048] Specifically, the input data of the adaptive recognition model in S2 includes:

[0049] The parent ion pair intensity ratio and the isotope distribution pattern extracted from the metabolite feature quantitative matrix;

[0050] The secondary mass spectrum spectrum information is vectorized into a sub-ion intensity sequence.

[0051] Further, the deep learning model fuses the primary mass spectrum feature vector and the secondary spectrum feature vector through a dot product vector algorithm.

[0052] Further, the metabolite standard library comprises a self-built standard sample library, a public database, and a virtual fragment spectrum library predicted by an algorithm, wherein the virtual fragment spectrum is generated by a spectrum prediction tool.

[0053] Specifically, the adaptive recognition model in S2 is deployed with an online updating mechanism, and the incremental training is automatically triggered based on the user feedback of the identification result error samples.

[0054] Specifically, based on the identification result, a metabolic pathway visualization map is automatically generated, and the enrichment significance of the differential metabolites in the pathway is labeled.

[0055] Specifically, when the similarity score of the best result is lower than the preset threshold, the mass-to-charge ratio error tolerance range is expanded, and the isomer differentiation module is called to analyze the characteristic ion abundance ratio.

[0056] In this embodiment, human serum metabolomics analysis is taken as an example, and 50 cases of healthy people / 50 cases of liver cancer patients are collected;

[0057] Liquid chromatography and high-resolution mass spectrometry are collected by instruments;

[0058] Real-time data preprocessing is performed, ion peaks with a signal-to-noise ratio greater than 3 are extracted, an initial feature list is generated, a dynamic time warping algorithm (DTW) is used to correct the retention time drift, and a maximum tolerance deviation is set; denoising processing is used to remove background contamination features, delete features with an RSD greater than 30% in QC samples, and retain high reproducibility metabolites; finally, output the standardized metabolite feature quantification matrix (containing multi-dimensional features).

[0059] Metabolite identification is performed, and model input is constructed:

[0060] Primary features: extract the intensity ratio of parent ions (such as [M+H]+ / [M+Na]+) and 13C isotope distribution pattern;

[0061] Secondary spectrum: vectorize the MS / MS sub-ion intensity sequence into a 256-dimensional tensor;

[0062] A variety of metabolite features are preliminarily identified;

[0063] Regarding database matching and result optimization, candidate compounds are initially screened using a self-built standard sample library, public databases such as HMDB, and theoretical prediction databases. A comprehensive score is then used for evaluation. The pseudocode for calculating the comprehensive score is as follows:

[0064] {

[0065] #Pseudocode: Formula for Overall Confidence Score

[0066] final_score = 0.5*cosine_similarity(exp_ms2, ref_ms2) # Spectral similarity

[0067] + 0.25*ion_ratio_consistency # Ion strength ratio consistency

[0068] + 0.25*isotope_pattern_match # Isotope distribution matching degree

[0069] }

[0070] Metabolites with high confidence results are output based on comprehensive scoring, and the molecular formula and compound name are associated with the database.

[0071] Downstream analysis involved screening for differential metabolites. KEGG pathway analysis showed significant enrichment of glycine / serine metabolism, and pathway maps were automatically generated and differential metabolites were labeled.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A high-resolution mass spectrometry-deep learning driven metabolomics data automatic analysis method, characterized in that, The method comprises the following steps: S1: obtaining an original mass spectrum data file through liquid chromatography-high resolution mass spectrometry; Peak extraction, peak alignment and noise reduction processing are performed on the original mass spectrum data file through an automatic preprocessing module, and a standardized metabolite feature quantitative matrix is generated based on a feature extraction module; The peak alignment adopts a dynamic time warping algorithm based on retention time correction, and the noise reduction processing includes blank sample background deduction and mixed sample relative standard deviation filtering; S2: based on the standardized metabolite feature quantitative matrix and the secondary mass spectrum information in the original mass spectrum data file, metabolite features are automatically identified through an adaptive identification model; The adaptive identification model is deployed in an online updating mechanism, and automatically triggers incremental training based on user feedback of identification result error samples; The input data of the adaptive identification model includes fragment peak area values and isotope distribution patterns extracted from the metabolite feature quantitative matrix, and fragment ion intensity sequences obtained by vectorizing the secondary mass spectrum information; S3: the identified metabolites and corresponding metabolite feature quantitative matrix are automatically matched with the best results in a database, and the corresponding molecular formula composition and compound name are obtained; based on the identification results, a metabolic pathway visualization map is automatically generated, and the enrichment significance of the differential metabolites in the pathway is labeled; The database automatic matching comprises: Based on the mass-to-charge ratio and retention time of the metabolite feature quantitative matrix, the initial screening is performed in the local and cloud metabolite standard library; The experimental spectrum is compared with the standard spectrum of the candidate compound in real time, and the spectrum similarity score is calculated through a deep learning model; According to the similarity score, ion intensity ratio consistency and isotope distribution coincidence, a comprehensive confidence score is generated; The matching results with a comprehensive confidence score higher than a preset threshold are automatically selected, and the molecular formula and compound name are outputted; The deep learning model fuses the primary mass spectrum feature vector and the secondary spectrum feature vector through a point vector product algorithm.

2. The method of claim 1, wherein, The metabolite standard library comprises a self-built standard sample experimental library, a public database and an algorithm predicted virtual fragment spectrum library, wherein the virtual fragment spectrum is generated by a spectrum prediction tool.

3. The method of claim 1, wherein, When the similarity score of the best result is lower than a preset threshold, the mass-to-charge ratio error tolerance range is expanded, and the isomer differentiation module is called to analyze the feature ion abundance ratio.

Images