Method for predicting metabolic pathway by using tandem mass spectrum fragment tree based on deep learning

Abstract

The invention discloses a method for predicting a metabolic pathway through a secondary mass spectrum fragment tree based on deep learning. According to the method, a training data set is constructed based on initial data matching, an initial prediction model is trained based on deep learning, model parameters with the highest accuracy are selected through cross validation, and data to be measured are input to obtain a path prediction result. The method effectively solves the problem that at present, metabolic pathway information in spectroscopic data must be obtained through compound identification, so that the information utilization rate is insufficient. Compared with an existing method, the method has the advantages that experimental mass spectrum data and metabolic pathways are effectively associated, and biological function explanation based on metabonomics data is facilitated. The method has the advantages that the path prediction accuracy is high; the method can be used for efficient utilization of metabolome data, and is convenient for enrichment of metabolic pathways.

Description

Technical Field

[0001] The present invention relates to a method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning, and belongs to the field of metabolic pathway prediction. Background Art

[0002] Metabolic pathways are a series of interconnected enzymatic biochemical reactions that occur within cells, converting substrate molecules through a series of metabolic intermediates to produce products. Analyzing metabolic pathways helps to gain a deeper understanding of chemical changes in organisms, thereby revealing the basic principles of life activities and their intrinsic connections with health, disease, and other aspects. Metabolomics (Nicholson, JK et al., Metabonomics: a platform for studying drug toxicity and gene function. Nature Reviews Drug Discovery 2002, 1, 153-161.) is the science of studying the metabolite composition and changes of organisms under specific physiological states. The research object of metabolomics is small molecule metabolites (usually with a molecular weight <1500Da) in biological systems. Changes in metabolites in biological systems can reflect changes in related metabolic pathways, which helps us understand biological mechanisms and is of great significance for research such as drug development and disease occurrence. However, our current understanding of metabolic pathways is limited. Traditional metabolomics pathway enrichment analysis requires first characterizing ions in mass spectrometry data as known compounds. The process involves steps such as peak extraction and peak matching, which may result in a large amount of information loss, thereby affecting the results of pathway enrichment analysis. Moreover, due to the limited number of spectral databases and the difficulty in obtaining standards, the ability to identify metabolites is limited, which may lead to bias in the results of pathway enrichment analysis. In addition, traditional methods are limited by the known compound database, making it difficult to fully explore the diversity of metabolic pathways, thereby limiting the comprehensive understanding of the metabolic network of organisms. However, studying metabolic pathways through biological experiments is very costly. Therefore, it is crucial to develop new methods and technologies to enhance the understanding of various metabolic pathways and their potential applications.

[0003] Currently published literature mainly focuses on predicting 11 metabolic pathway categories in the Kyoto Encyclopedia of Genes and Genomes database and is unable to predict specific pathways. More importantly, these prediction models are limited to predicting compounds with known structures. In metabolomics research, identifying compound structures from metabolomics mass spectrometry data is very challenging. As mentioned above, the qualitative process loses a lot of information, and much of the detected ion information is discarded, making subsequent predictions impossible. This makes it difficult to associate metabolomics mass spectrometry data with metabolic pathways, limiting the subsequent interpretation of biological functions in metabolomics. Summary of the Invention

[0004] The present invention provides a method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning, which can effectively solve the problem that current prediction methods can only predict the metabolic pathway category of compounds with known structures but cannot predict the specific metabolic pathway.

[0005] It can effectively solve the problem that metabolomics mass spectrometry data collected in experiments are difficult to associate with metabolic pathways, which is not conducive to the subsequent interpretation of biological functions;

[0006] It can effectively solve the current problem that obtaining metabolic pathway information from spectral data requires compound identification, resulting in insufficient information utilization.

[0007] The present invention has the advantages of high prediction accuracy, no need for prior compound structure identification, and convenience for metabolic pathway enrichment.

[0008] To achieve the above objectives, the present invention provides a method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning, comprising the following steps:

[0009] S1: Construct an initial MS / MS fragment tree and metabolic pathway dataset, and match the two to form a training dataset containing the MS / MS fragment tree and metabolic pathway information;

[0010] S2: Build an initial prediction model based on deep learning and train it using the training dataset;

[0011] S3: Use cross-validation to select the deep learning model parameters with the highest accuracy as the final prediction model parameters;

[0012] S4: Inputting the secondary mass spectrum fragment tree of the compound to be predicted into the final prediction model to obtain a prediction result, i.e., metabolic pathway information of the compound to be predicted.

[0013] Optionally, the data used to construct the initial secondary mass spectrometry fragment tree in step S1 is obtained from an open source mass spectrometry database, such as GNPS, MASSBANK, etc., and at least includes information of the secondary mass spectrum and its corresponding compound structure information.

[0014] Optionally, the data used to construct the metabolic pathway dataset in step S1 is obtained from an open source metabolic pathway database, such as KEGG, PlantCyc and other databases, and at least includes metabolic pathway information and structural information of compounds contained in the pathway.

[0015] Optionally, the matching method in step S1 is specifically: matching identical compounds, wherein the structural representation method of the compound includes at least one of SMILES, InChI, and InChIKey.

[0016] Optionally, when performing matching, non-InChIKey type representations are uniformly converted to InChIKey using the RDKit toolkit.

[0017] Optionally, the initial secondary mass spectrum fragment tree is calculated by inputting the secondary mass spectrum into open source software (such as SIRIUS5), and after calculation, the initial secondary mass spectrum fragment tree is converted into a graph structure.

[0018] Optionally, in step S2, an initial prediction model is constructed based on deep learning, specifically based on a graph transformer neural network, with the input being a fragment tree of the secondary mass spectrum represented by a graph structure, and the output being a metabolic pathway label.

[0019] Optionally, the metabolic pathway label is encoded using a binary value, where "1" represents that the compound is a metabolite of the pathway, and "0" represents that the compound is not a metabolite of the pathway.

[0020] Optionally, in step S2, the neural network structure of the initial prediction model is specifically configured to include:

[0021] One layer of 8-head attention graph transformer network, one layer of batch normalization layer, one layer of activation function (Elu), one layer of 1-head attention graph transformer network, one layer of batch normalization layer, one layer of activation function (Elu), one layer of global pooling layer, one layer of fully connected layer, one layer of activation function (Relu), one layer of fully connected layer, one layer of activation function (Relu), and one layer of fully connected layer.

[0022] Optionally, in step S4, the secondary mass spectrum of the compound to be predicted is obtained by a mass spectrometer with a tandem mass spectrometry function.

[0023] The beneficial effects produced by the present invention include:

[0024] 1) The present invention provides a method for predicting metabolic pathways from a secondary mass spectrometry fragment tree based on deep learning. A deep learning model for predicting metabolic pathways from a secondary mass spectrometry fragment tree is constructed. This method effectively overcomes the problems of conventional metabolic pathway enrichment methods, such as the need to first characterize mass spectrometry ions as known compounds, resulting in a large amount of information loss during the qualitative process, as well as the lack of compound structure identification capabilities due to the difficulty in obtaining standards. This method provides assistance for metabolomics pathway enrichment analysis and biological function elucidation.

[0025] 2) The present invention provides a method for predicting metabolic pathways based on a deep learning-based MS / MS fragment tree. The deep learning model based on a graph transformer neural network uses a graph structure as input, effectively describing the fragmentation relationship of MS / MS ions, which is conducive to the training of the deep learning model.

[0026] 3) The present invention provides a method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning. The constructed metabolic pathway prediction model has the advantages of reliable prediction and high accuracy, and can realize the effective prediction of metabolic pathways. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the steps for predicting metabolic pathways from a secondary mass spectrometry fragment tree provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The present invention is described in detail below with reference to the embodiments, but the present invention is not limited to these embodiments.

[0029] To demonstrate the effectiveness and feasibility of the present invention, we screened the secondary mass spectra from the Global Natural Products Social Molecular Networking (GNPS) (https: / / gnps.ucsd.edu) database, which originated from the GNPS Sigma's Mass Spectrometry Metabolite Library (hereinafter referred to as GNPS-MSMLS), GNPS Matches to NIST14 (hereinafter referred to as GNPS-NIST14), and ReSpect Spectral Library (hereinafter referred to as ReSpect). The calculated fragmentation trees were used as tests, and the fragmentation trees calculated from the remaining secondary mass spectra were used for training to predict metabolic pathways.

[0030] According to one embodiment of the present invention, a method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning is provided. Figure 1 As shown, the specific steps include:

[0031] S1: Construct an initial MS / MS fragment tree and metabolic pathway dataset, and match the two to form a training dataset containing the MS / MS fragment tree and metabolic pathway information;

[0032] The data used to construct the initial secondary mass spectrum fragment tree was obtained from the open source database GNPS, which includes secondary mass spectrum information and its corresponding compound structure information, and was downloaded in April 2023;

[0033] The initial secondary mass spectrum fragment tree is obtained by inputting the secondary mass spectrum into the open source software SIRIUS5 for calculation, and after calculation, the initial secondary mass spectrum fragment tree is converted into a graph structure representation;

[0034] The data used to construct the metabolic pathway dataset were obtained from the open source metabolic pathway database KEGG, including metabolic pathway information and structural information of compounds contained in the pathway, and the download time was December 2022;

[0035] The method for matching the two is specifically as follows: matching the same compound, wherein the representation method of the compound structure may include SMILES, InChI, and InChIKey, and converting the non-InChIKey type representation method into InChIKey through the RDKit toolkit during matching;

[0036] In this example, after matching the initial secondary mass spectrometry fragment tree with the metabolic pathway dataset, a training dataset was constructed including 33,221 graph data, which were distributed on 150 metabolic pathways. The labels of each pathway are shown in Table 1.

[0037] Table 1

[0038]

[0039]

[0040]

[0041]

[0042] S2: Build an initial prediction model based on deep learning and train it using the training dataset.

[0043] The method for constructing the initial prediction model is specifically as follows: it is constructed based on a graph transformer neural network, the input is a fragment tree of the secondary mass spectrum represented by a graph structure, and the output is a metabolic pathway label.

[0044] The metabolic pathway label is encoded using a binary value, where "1" represents that the compound is a metabolite of the pathway, and "0" represents that the compound is not a metabolite of the pathway.

[0045] In this embodiment, the specific neural network structure is that the input graph passes through a layer of 8-head attention graph transformer network, a layer of batch normalization layer, a layer of activation function (Elu), a layer of 1-head attention graph transformer network, a layer of batch normalization layer, a layer of activation function (Elu), a layer of global pooling layer, a layer of fully connected layer, a layer of activation function (Relu), a layer of fully connected layer, a layer of activation function (Relu), a layer of fully connected layer, and finally outputs a 150-dimensional vector. These 150 numbers are recorded as 0 if they are less than 0.5, indicating that they are not on this pathway label; they are recorded as 1 if they are greater than 0.5, indicating that they are on this pathway label. The Adam optimizer is used for training, the initial learning rate is 0.001; the loss function is MultiLabelSoftMarginLoss; and the number of training times is 350.

[0046] S3: Use cross-validation to select the deep learning model parameters with the highest accuracy as the final prediction model parameters, specifically:

[0047] We used 5-fold cross-validation and saved the network parameters for each run. The results of the 5-fold cross-validation are shown in Table 2. The deep learning model with the highest accuracy and balanced accuracy was selected as the final prediction model.

[0048] Table 2

[0049]

[0050] S4: Inputting the secondary mass spectrum fragment tree of the compound to be predicted into the final prediction model to obtain a prediction result, i.e., metabolic pathway information of the compound to be predicted.

[0051] The secondary mass spectrum of the compound to be predicted is obtained by a mass spectrometer with tandem mass spectrometry function.

[0052] Here, the secondary mass spectrometry fragmentation trees from GNPS-MSMLS, GNPS-NIST14, and ReSpect in the GNPS database were tested, and the results are shown in Table 3.

[0053] Table 3

[0054]

[0055] The above results show that the final prediction model constructed using the method of the present invention effectively extracted the secondary mass spectrometry fragment tree features, realized the multi-label prediction task from the secondary mass spectrometry fragment tree to metabolic pathways, and the results were accurate and reliable.

[0056] The above description is only a partial embodiment of the present invention and does not limit the present invention in any form. Although the present invention is disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any technician familiar with the profession, without departing from the scope of the technical solution of the present invention, using the technical content disclosed above to make slight changes or modifications are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning, characterized in that: The steps include: S1: construct an initial secondary mass spectrometry fragment tree and metabolic pathway data set, and match the two to form a training data set containing secondary mass spectrometry fragment tree and metabolic pathway information; S2: constructing an initial prediction model based on deep learning and using the training data set for training; S3: Use cross-validation to select the deep learning model parameters with the highest accuracy as the final prediction model parameters; S4: Inputting the secondary mass spectrum fragment tree of the compound to be predicted into the final prediction model to obtain a prediction result, namely, metabolic pathway information of the compound to be predicted.

2. The method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning according to claim 1, characterized in that: The data used to construct the initial secondary mass spectrum fragment tree in step S1 is obtained from an open source mass spectrum database, and at least includes secondary mass spectrum information and its corresponding compound structure information.

3. The method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning according to claim 1, characterized in that: The data used to construct the metabolic pathway data set in step S1 is obtained from an open source metabolic pathway database, and at least includes metabolic pathway information and structural information of compounds contained in the pathway.

4. The method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning according to claim 1, characterized in that: The matching method in step S1 is specifically as follows: matching identical compounds, wherein the method for representing the structure of the compound includes at least one of SMILES, InChI, and InChIKey.

5. The method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning according to claim 4, characterized in that: When matching, non-InChIKey type representations are uniformly converted to InChIKey through the RDKit toolkit.

6. The method for predicting metabolic pathways based on secondary mass spectrometry fragment tree based on deep learning according to claim 1, characterized in that: The initial secondary mass spectrum fragment tree is obtained by inputting the secondary mass spectrum into open source software for calculation, and after calculation, the initial secondary mass spectrum fragment tree is converted into a graph structure representation.

7. The method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning according to claim 6, characterized in that: In step S2, an initial prediction model is constructed based on deep learning, specifically: it is constructed based on a graph transformer neural network, the input is a fragment tree of the secondary mass spectrum represented by a graph structure, and the output is a metabolic pathway label.

8. The method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning according to claim 7, characterized in that: The metabolic pathway label is encoded using a binary value, where "1" represents that the compound is a metabolite of the pathway, and "0" represents that the compound is not a metabolite of the pathway.

9. The method for predicting metabolic pathways based on a secondary mass spectrometry fragment tree based on deep learning according to claim 1, characterized in that: In step S2, the neural network structure of the initial prediction model is specifically configured to include: One layer of 8-head attention graph transformer network, one layer of batch normalization layer, one layer of activation function, one layer of 1-head attention graph transformer network, one layer of batch normalization layer, one layer of activation function, one layer of global pooling layer, one layer of fully connected layer, one layer of activation function, one layer of fully connected layer, one layer of activation function, one layer of fully connected layer.

10. The method for predicting metabolic pathways based on secondary mass spectrometry fragment tree based on deep learning according to claim 1, characterized in that: In the step S4, the secondary mass spectrum of the compound to be predicted is obtained by a mass spectrometer with a tandem mass spectrometry function.

Images