Metabolic reaction pair prediction method based on deep learning and non-targeted metabonomics

Abstract

The invention discloses a metabolic reaction pair prediction method based on deep learning and non-targeted metabonomics, and belongs to the technical field of bioinformatics, and the method comprises the steps: (1) enabling reactants and products to form a metabolite reaction pair, and converting the metabolite reaction pair into a molecular fingerprint pair, thereby constructing a metabolic reaction pair data set; (2) establishing a deep learning model of which the input is molecular fingerprints and the output is a metabolic reaction pair prediction result, and training the model by utilizing the established metabolic reaction pair data set to realize accurate modeling; and (3) inputting a molecular fingerprint pair of a compound to be predicted or a metabolic characteristic into the trained model to generate a prediction result of whether a reaction pair is formed or not. The method can predict whether a metabolic reaction pair is formed between a compound pair or a metabolic characteristic pair represented by non-targeted metabonomics secondary mass spectrometry data. According to the method, the information contained in the non-targeted metabonomics data can be fully mined, the utilization efficiency of the metabonomics data is remarkably improved, and the prediction precision is high.

Description

Technical Field

[0001] The present application relates to a metabolic reaction pair prediction method, which belongs to the technical field of bioinformatics, and particularly to a metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics data. Background Art

[0002] Metabolism is the general term for all chemical reactions required to maintain life in organisms. Metabolic reactions usually convert one chemical substance (substrate) into another chemical substance (product) through enzymes, where the substrate-product is often referred to as a metabolic reaction pair. Metabolic reactions are not only essential for the maintenance of host physiology, but also play a key role in shaping the host's defense system. Metabolic reactions are very complex. In-depth research on metabolic reactions and a better understanding of the metabolic processes in organisms will not only help reveal disease mechanisms, but also help in the development of new treatment strategies and drug design optimization.

[0003] Currently, there is limited knowledge about metabolic reaction pairs. On the one hand, the database information of metabolic reactions is limited, and the number of metabolites with reaction pair information is far less than the number of metabolites that have been discovered. Compared with traditional targeted metabolite analysis, non-targeted metabolomics technology based on high-resolution mass spectrometry has higher sensitivity and coverage, and can comprehensively and systematically detect metabolome information. This technology can capture rich metabolic features from complex biological samples, including about 1.8-20% of structurally annotated metabolic features (da Silva RR, Dorrestein PC, Quinn RA. Illuminating the dark matter in metabolomics [J]. Proceedings of the National Academy of Sciences, 2015, 112 (41): 12549-12550.S), as well as a large number of structurally unknown metabolic features that are not yet available. If we can effectively analyze and mine these unused non-targeted metabolomics data, identify and analyze the large metabolic reaction information, and then reveal the dynamic changes of metabolic networks, it will not only help deepen our understanding of metabolic processes in the body, but also have a positive role in promoting the development of personalized medicine and precision medicine. The usual manual analysis and biological experimental methods are very time-consuming and labor-intensive, and new technologies are urgently needed.

[0004] Related research on metabolic reaction prediction has been published in the literature. For example, the literature (Wishart DS, Tian S, Allen D, et al. BioTransformer 3.0—a web server for accurately predicting metabolic transformation products [J]. Nucleic acids research, 2022, 50 (W1): W115-W123.) combines machine learning methods with knowledge-based methods to predict a variety of metabolic or biotransformation processes such as enzymatic reactions, human intestinal microbial reactions, and environmental microbial transformations. Users only need to input the structural information of the molecule to be predicted to obtain its transformation product, but they cannot specify the reactants and products at the same time to determine whether the two constitute a reaction pair. Literature (Moriya Y, Shigemizu D, Hattori M, et al. PathPred: an enzyme-catalyzed metabolic pathway prediction server [J]. Nucleic acids research, 2010, 38 (suppl_2): W138-W143.) can determine whether two specified molecules can complete the conversion through a certain pathway, but is limited to exogenous compound biodegradation and secondary metabolite biosynthesis. In addition, both documents can only predict known structural compounds, and cannot determine unqualified metabolic features, which leads to a large number of metabolic features in metabolomics experiments cannot be used. And the reaction pair information of predicting metabolic features can be used for structural annotation and metabolic pathway information mining.

[0005] Therefore, current methods can only predict compounds with known structures and cannot be used for a large number of metabolic features with unknown structures in non-targeted metabolomics data. Summary of the invention

[0006] In view of the shortcomings of the prior art and to solve the above bottleneck problems, the present invention proposes a metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics data. This method can predict whether a metabolic reaction pair is formed between a compound pair or a metabolic feature pair represented by non-targeted metabolomics secondary mass spectrometry data. It can effectively overcome the current problem of insufficient metabolic reaction data information and difficulty in obtaining reaction pair information of metabolic features, and has the advantages of high prediction accuracy, high recovery rate, and convenience for fully exploring the information contained in metabolic features.

[0007] A method for predicting metabolic reaction pairs based on molecular fingerprints based on deep learning, comprising the following steps:

[0008] S1: Combining reactants and products into metabolite reaction pairs; converting the metabolite reaction pairs into molecular fingerprint pairs, thereby constructing a metabolic reaction pair data set;

[0009] S2: using deep learning technology to build a metabolic reaction pair prediction model, wherein the input of the metabolic reaction pair prediction model is a molecular fingerprint pair, and the output is a label of whether it is a reaction pair; using the metabolic reaction pair data set to train the metabolic reaction pair model;

[0010] S3: Input the molecular fingerprint pair of the test compound or non-targeted metabolomics data into the trained metabolic reaction pair prediction model, and output the prediction result of whether it constitutes a metabolic reaction pair.

[0011] Optionally, in step S1, the reactants and products are extracted from metabolic reaction data, and the metabolic reaction data is obtained through an open source metabolic reaction database.

[0012] Optionally, the open source metabolic reaction database includes but is not limited to one of the following: KEGG Reaction database, HumanCyc Reaction database.

[0013] Optionally, constructing the metabolic reaction pair data set further includes: label encoding the metabolic reaction pairs, wherein the label encoding uses a binary classification label, where "1" represents a reaction pair, i.e. a positive sample; and "0" represents not a reaction pair, i.e. a negative sample.

[0014] Optionally, for each metabolite, the number of constituent reaction pairs in the positive sample and the negative sample is ensured to be the same, so that the generated metabolic reaction pair model can judge the correctness of each reaction pair in a relatively balanced manner.

[0015] Optionally, obtaining the positive sample includes: predicting a two-step enzyme-catalyzed metabolic reaction of the compound to obtain a simulated reaction pair, and combining the existing reaction pair and the simulated reaction pair to form a positive sample.

[0016] Optionally, the negative samples are generated using random combinations of metabolites that are not reaction pairs.

[0017] Optionally, if the molecular fingerprint is derived from a compound of known structure, it is directly calculated; if the molecular fingerprint to be tested is derived from secondary mass spectrometry data of non-targeted metabolomics data, the corresponding molecular fingerprint is predicted by an open source software tool.

[0018] Alternatively, if the molecular fingerprint to be tested is derived from secondary mass spectrometry data of non-targeted metabolomics data, an open source software tool such as SIRIUS can be used to predict the corresponding molecular fingerprint.

[0019] Optionally, if the molecular fingerprint is derived from a compound of known structure, a molecular fingerprint of the metabolite is calculated based on the compound structure, and the molecular fingerprint calculation method may include at least one of SMILES, InChI, and InChIKey.

[0020] Optionally, in step S2, the metabolic reaction pair prediction model is based on a Siamese neural network architecture, which includes two sub-networks with the same parameters and weights, and the inputs of the two sub-networks correspond to the two molecular fingerprints of the molecular fingerprint pair respectively. After the calculation of the two sub-networks, two vectors are given, and the two vectors are calculated. The Cartesian product is subtracted from the Euclidean distance of the two input molecular fingerprints, and then pass through a layer of dimensionality reduction fully connected layer and a layer of Sigmoid function layer in sequence, and finally output a one-dimensional vector. When the value of the one-dimensional vector is less than 0.5, it is recorded as 0, indicating that the prediction result is a negative sample of the reaction pair; when the value of the one-dimensional vector is greater than or equal to 0.5, it is recorded as 1, indicating that the prediction result is a positive sample of the reaction pair.

[0021] Optionally, the structure of the sub-network includes, in sequence: a dimensionality reduction fully connected layer, four transformer encoder layers, a random dropout layer, a layer of activation function ReLU, and a dimensionality reduction fully connected layer.

[0022] Optionally, the non-targeted metabolomics data includes primary mass spectrometry MS1 and corresponding secondary mass spectrometry MS2 data.

[0023] The beneficial effects of this application include:

[0024] 1) The present application provides a metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics, which constructs a deep learning model for predicting metabolic reaction pairs from the molecular fingerprints of metabolites or metabolic features, effectively overcoming the problems of insufficient information in existing metabolic reaction-related databases, small number of qualitative data in metabolomics experiments that are difficult to associate with metabolic reaction databases, and difficulty in obtaining metabolic reaction pair information due to unqualified metabolic features, and can provide assistance for the structural characterization of metabolic features and the interpretation of biological functions related to metabolic pathways.

[0025] 2) The present application provides a metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics. The deep learning model based on the Siamese neural network structure uses molecular fingerprints as input, effectively describes the structural information contained in metabolites and metabolic characteristics, and is effectively encoded after being processed by the transformer encoder, which is conducive to the training of the deep learning model.

[0026] 3) The present application provides a metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics. The constructed reaction pair prediction model has the advantages of reliable prediction and high accuracy, and can effectively predict whether a reaction pair is formed between metabolites and metabolic characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of a process for predicting metabolic reaction pairs based on deep learning and non-targeted metabolomics in one embodiment of the present application. DETAILED DESCRIPTION

[0028] The present application is described in detail below in conjunction with embodiments and drawings, but the present application is not limited to these embodiments.

[0029] The present application provides a metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics, comprising the following steps:

[0030] S1: Combining reactants and products into metabolite reaction pairs; converting the metabolite reaction pairs into molecular fingerprint pairs, thereby constructing a metabolic reaction pair data set;

[0031] S2: using deep learning technology to build a metabolic reaction pair prediction model, wherein the input of the metabolic reaction pair prediction model is a molecular fingerprint pair, and the output is a label of whether it is a reaction pair; using the metabolic reaction pair data set to train the metabolic reaction pair model;

[0032] S3: Input the molecular fingerprint pair of the test compound or non-targeted metabolomics data into the trained metabolic reaction pair prediction model, and output the prediction result of whether it constitutes a metabolic reaction pair.

[0033] The present application can predict whether a metabolite constitutes a reaction pair based on a molecular fingerprint.

[0034] Example 1

[0035] In order to verify the effectiveness and feasibility of the present invention, reaction pairs consisting of compounds in the KEGG Reaction database were used for training, and reaction pairs consisting of compounds in the HumanCyc database were used for prediction.

[0036] See also Figure 1 , which shows a flowchart of a metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics provided by the present application in one embodiment, the method comprising the following steps:

[0037] S1: A metabolic reaction pair dataset for training, namely the training set, is constructed based on compounds in the KEGG Reaction database.

[0038] S11: Download the information of each reaction in the KEGG Reaction database, extract the reaction formula and the compounds represented by SMILES, each reactant and product can form a reaction pair, and obtain the combination of all reaction pairs.

[0039] S12: In order to train the model more fully, data enhancement was performed. The tool Biotransformer was used to predict the two-step enzyme-catalyzed metabolic reactions of the compounds downloaded above to obtain simulated reaction pairs.

[0040] S13: Merge the downloaded and simulated reaction pairs. Compounds that constitute more than 10 reaction pairs were deleted to prevent bias in training, and 37,989 reaction pairs were obtained as positive samples. Next, negative samples were generated by randomly combining metabolites that were not reaction pairs. Here, it was ensured that the number of reaction pairs composed of each metabolite in the positive and negative samples was the same, so that the model could more balancedly judge the correctness of each reaction pair, and 43,665 reaction pairs were obtained as negative samples. The molecular fingerprints of the metabolites in the positive and negative samples were obtained, and the metabolite-metabolite pairs were converted into molecular fingerprint-molecular fingerprint pairs. The molecular fingerprint sites here select the intersection of the molecular fingerprints of the two open source software MS2Quant and SIRIUS 5.6.3. The positive ion mode and negative ion mode are two sets of molecular fingerprint sites, the same below.

[0041] S14: Merge positive samples and negative samples to obtain a complete training set.

[0042] Step S2: Based on deep learning technology, an initial prediction model is constructed, and the metabolic reaction pair data set is used to train the model.

[0043] A deep learning model is built based on the Siamese neural network architecture. The input is a pair of molecular fingerprints, and the output is the label of whether it is a reaction pair. "1" represents a positive sample of the reaction pair, and "0" represents a negative positive sample of the reaction pair. The model includes two sub-networks with the same parameters and weights. The sub-networks pass through a layer of dimension reduction fully connected layer, four layers of transformer encoder layer, one layer of random inactivation layer, one layer of activation function (ReLU), and one layer of dimension reduction fully connected layer. After the two sub-networks are calculated, two vectors are given. The two vectors are calculated Cartesian product and the Euclidean distance of the two input molecular fingerprints is subtracted. Then, they pass through a layer of dimension reduction fully connected layer and a layer of Sigmoid function layer, and finally a vector with a dimension of 1 is output. This vector with a dimension of 1 is recorded as 0 if it is less than 0.5, indicating that the prediction result is a negative sample of the reaction pair; it is recorded as 1 if it is greater than or equal to 0.5, indicating that the prediction result is a positive sample of the reaction pair. The training uses the Adam optimizer, with an initial learning rate of 0.0001; the loss function is FocalLoss; and the number of training times is 100. Use 5-fold cross validation to save the network parameters each time. The same training conditions were used in both positive and negative ion modes, with only the input molecular fingerprints being different.

[0044] The 5-fold cross-validation results in the positive ion and negative ion modes are shown in Tables 1 and 2, respectively. It can be seen that the indicators are slightly different in each cross-validation. The first deep learning model is used as the test model for the next step.

[0045] Table 1 Cross-validation model performance in positive ion mode

[0046]

[0047] Table 2 Cross-validation model performance in negative ion mode

[0048]

[0049] S3: Apply the trained metabolic reaction pair prediction model to the molecular fingerprint pair input of the compound-compound composition to be tested, and output a prediction result of whether a metabolic reaction pair is constituted.

[0050] We downloaded 490 reaction pairs consisting of metabolites from the HumanCyc database, obtained the molecular fingerprint-molecular fingerprint pairs formed by them, and input them into the trained positive and negative ion prediction models. In the positive ion mode, the prediction accuracy was 0.9081; in the negative ion mode, the prediction accuracy was 0.9040.

[0051] The above results show that the deep learning model constructed by the present invention effectively extracts molecular fingerprint features, realizes the discrimination and prediction tasks from molecular fingerprints to metabolic reaction pairs, and the results are accurate and reliable.

[0052] Example 2

[0053] In order to demonstrate the effectiveness and feasibility of the present invention, reaction pairs consisting of compounds in the KEGG Reaction database were used for training, and metabolic features in the non-targeted metabolomics data collected from NIST SRM 1950 plasma were used to predict reaction pairs with known metabolites in the KEGG database.

[0054] Step S1 and step S2: Step S1 and step S2 of Example 2 are exactly the same as those of Example 1. After S1 and S2, a trained metabolic reaction pair prediction model is obtained.

[0055] Step S3: Apply the trained metabolic reaction pair prediction model to the molecular fingerprint pair input consisting of the compound to be tested and the non-targeted metabolomics data, the non-targeted metabolomics data including the primary mass spectrometry MS1 and the corresponding secondary mass spectrometry MS2 data, and output a prediction result of whether a metabolic reaction pair is constituted.

[0056] Here, liquid chromatography-mass spectrometry was used to collect non-targeted metabolomics data of NIST SRM1950 plasma in positive and negative ion modes. The liquid chromatography separation system used was ACQUITY ultra-high performance liquid chromatography (Waters Corp, Milford, USA). The mass spectrometer used was Q Exactive HF (Thermo Fisher Scientific, Rockford, IL, USA), which was analyzed in ddMS2 mode. The specific pretreatment process and detailed instrument parameters are consistent with the literature (Wang X, Li C, Li Z, et al. A Structure-Guided Molecular Network Strategy for Global Untargeted Metabolomics Data Annotation [J]. Analytical Chemistry, 2023, 95 (31): 11603-11612.).

[0057] After collecting non-targeted metabolomics data, the primary mass spectrometry MS1 and the corresponding secondary mass spectrometry MS2 data are obtained, input into the software SIRIUS 5.6.3, and the molecular fingerprint is predicted. Then, the molecular fingerprint predicted by SIRIUS 5.6.3 and the molecular fingerprint of MS2Quant are taken as the intersection of the molecular fingerprint sites, that is, the same as the molecular fingerprint sites in Example 1, to meet the requirements of model input. Afterwards, the molecular fingerprints of these non-targeted metabolomics data and the molecular fingerprints of the compounds in the KEGG COMPOUND database are combined into molecular fingerprint-molecular fingerprint pairs and input into the trained prediction model. In positive ion mode, 3026 of the 4305 metabolic features with MS1 and MS2 are predicted to constitute reaction pairs with the compounds in the KEGG COMPOUND database. In negative ion mode, 2281 of the 3503 metabolic features with MS1 and MS2 are predicted to constitute reaction pairs with the compounds in the KEGG COMPOUND database.

[0058] The above results show that the deep learning model constructed in this application can provide reaction pair information between metabolic features in non-targeted metabolomics data and metabolites with known structures or pathway information, which is of great significance for metabolomics data utilization and knowledge mining.

[0059] The above are only a few embodiments of the present application and do not constitute any form of limitation to the present application. Although the present application is disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the profession, without departing from the scope of the technical solution of the present application, 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 metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics, characterized in that: Includes steps: S1: Combining reactants and products into metabolite reaction pairs; converting the metabolite reaction pairs into molecular fingerprint pairs, thereby constructing a metabolic reaction pair data set; S2: using deep learning technology to build a metabolic reaction pair prediction model, wherein the input of the metabolic reaction pair prediction model is a molecular fingerprint pair, and the output is a label of whether it is a reaction pair; using the metabolic reaction pair data set to train the metabolic reaction pair model; S3: Input the molecular fingerprint pair of the test compound or non-targeted metabolomics data into the trained metabolic reaction pair prediction model, and output the prediction result of whether it constitutes a metabolic reaction pair.

2. The metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics according to claim 1, characterized in that: In step S1, the reactants and products are extracted from metabolic reaction data, and the metabolic reaction data is obtained through an open source metabolic reaction database; Preferably, the open source metabolic reaction database includes at least one of the following: KEGG Reaction database, HumanCyc Reaction database.

3. The metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics according to claim 2, characterized in that: The construction of the metabolic reaction pair data set further includes: label encoding the metabolic reaction pairs, wherein the label encoding uses a binary classification label, where "1" represents a reaction pair, i.e., a positive sample; and "0" represents not a reaction pair, i.e., a negative sample; Preferably, the number of reaction pairs of each metabolite in the positive sample and the negative sample is ensured to be the same; Preferably, obtaining the positive sample comprises: predicting a two-step enzyme-catalyzed metabolic reaction of the compound to obtain a simulated reaction pair, and combining the existing reaction pair and the simulated reaction pair to form a positive sample; Preferably, the negative samples are generated using random combinations of metabolites that are not reaction pairs.

4. The metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics according to claim 1, characterized in that: If the molecular fingerprint is derived from a compound of known structure, it is directly calculated; if the molecular fingerprint to be tested is derived from secondary mass spectrometry data of non-targeted metabolomics data, the corresponding molecular fingerprint is predicted by an open source software tool; Preferably, the open source software tool comprises SIRIUS.

5. The metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics according to claim 4, characterized in that: The direct calculation includes: performing molecular fingerprint calculation on the metabolite based on the compound structure, and the molecular fingerprint calculation method includes at least one of SMILES, InChI, and InChIKey.

6. The metabolic reaction pair prediction method based on deep learning and non-targeted metabolomics according to claim 1, characterized in that: In step S2, the metabolic reaction pair prediction model is based on a Siamese neural network architecture, which includes two sub-networks with the same parameters and weights. The inputs of the two sub-networks correspond to the two molecular fingerprints of the molecular fingerprint pair, respectively. After the calculation of the two sub-networks, two vectors are given. The two vectors calculate the Cartesian product and subtract the Euclidean distance of the two input molecular fingerprints, and then pass through a layer of dimensionality reduction fully connected layer and a layer of Sigmoid function layer in sequence, and finally output a one-dimensional vector. When the one-dimensional vector value is less than 0.5, it is recorded as 0, indicating that the prediction result is a negative sample of the reaction pair; when the one-dimensional vector value is greater than or equal to 0.5, it is recorded as 1, indicating that the prediction result is a positive sample of the reaction pair.

7. The method for predicting metabolic reaction pairs based on deep learning and non-targeted metabolomics according to claim 6, characterized in that: The structure of the sub-network includes: a dimension reduction fully connected layer, four transformer encoder layers, a random dropout layer, a ReLU activation function layer, and a dimension reduction fully connected layer.

8. The method for predicting metabolic reaction pairs based on deep learning and non-targeted metabolomics according to claim 1, characterized in that: The non-targeted metabolomics data include primary mass spectrometry MS1 and corresponding secondary mass spectrometry MS2 data.

Images