Dissolved organic matter isomer vector prediction model construction method based on graph learning
Through a graph-based learning method, combined with liquid chromatography and electrostatic field orbital ion trap mass spectrometry, a molecular graph network of DOM was constructed, the properties and environmental characteristics of molecules were extracted, and the deep neural network prediction model was constructed, which solved the DOM isomers. The accuracy of identification achieves efficient prediction results.
Patent Information
- Application Number
- CN202410501288.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-04-23
- Filing Date
- 2024-04-25
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-04-25
AI Technical Summary
The prior art has difficulty accurately identifying and predicting the characteristics of soluble organic matter (DOM) isomers, especially differences in different sample environments, limiting the research and application of DOM.
Using graph-based learning methods, the molecular tags of DOM are obtained through liquid chromatography and electrostatic field orbital ion trap mass spectrometry technology, and a molecular graph network is constructed. Unsupervised learning and random walk algorithms are used to extract the properties and environmental characteristics of molecules, combined with deep nerves. Network builds prediction models.
Accurate prediction of dissolved organic matter isomer vectors is achieved, the recognition accuracy and recall rate are improved, and the application space of DI-FT-ICRMS is expanded.
Smart Images

Figure CN119028481A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of organic matter isomer identification, and specifically relates to a method for constructing a dissolved organic matter isomer vector prediction model based on graph learning. Background Art
[0002] As a highly complex mixture of organic molecules, dissolved organic matter (DOM) plays an important role in the cycle of carbon and various biogenic elements in the natural environment. The accurate characterization of DOM is a prerequisite for studying its source, transformation and degradation process, and fate. Ultra-high resolution Fourier transform cyclotron resonance mass spectrometry (DI-FT-ICR MS) in direct injection mode can provide accurate molecular formulas of thousands of molecules in DOM, which has greatly promoted the study of the composition characteristics and environmental significance of DOM in the past decade. However, a molecular formula usually represents multiple isomers, and the molecular formula provided by DI-FT-ICR MS may represent the mixed information of multiple isomers. Isomers have different structural characteristics, reactivity, and degradation history, which is of great value for exploring molecular evolution. In recent years, liquid chromatography and electrostatic field orbitrap mass spectrometry (LC-Orbitrap MS) can effectively separate DOM molecules according to polarity before mass spectrometry analysis, and provide characteristic vectors of different molecules according to retention time, so as to characterize DOM molecules in a fine-grained manner in terms of structure, polarity, and number of isomers, which is more helpful to explore the characteristics of DOM. However, compared with DI-FT-ICR MS technology, LC-Orbitrap MS has so far only been established in a few laboratories at home and abroad, which greatly limits its application space.
[0003] In recent years, the development and successful application of artificial intelligence technology has provided a new perspective for the identification of isomers of molecular formulas. However, building an accurate prediction model is extremely complex and requires comprehensive consideration of the following factors. First, each molecule of DOM has its own unique attribute characteristics, including explicit attribute characteristics (such as: the number of C, H, O, etc., AImod, etc.) and implicit attribute characteristics that have not yet been observed, and there are complex chemical transfer processes between molecules, that is, spatial structural characteristics; secondly, the same molecule has different effects and corresponding isomers in different sample environments, and is affected by neighboring molecules. It is necessary to distinguish the local environmental characteristics of the molecule in a specific sample. The same molecule has different effects and labels in different samples. How to extract the characteristics of each molecule in different samples and predict the ism of the molecule is a technical problem that needs to be solved urgently. Summary of the invention
[0004] In order to solve the above-mentioned technical problems existing in the prior art, the attribute characteristics, spatial structure characteristics and local environmental characteristics of each molecule in DOM are comprehensively considered, and a prediction model for predicting the dissolved organic matter isomer vector is accurately constructed. To this end, the present invention provides a method for constructing a prediction model for dissolved organic matter isomer vector based on graph learning.
[0005] The specific scheme adopted is as follows:
[0006] A method for constructing a dissolved organic matter isomer vector prediction model based on graph learning, the method comprising the following steps:
[0007] Liquid chromatography coupled to electrostatic field orbital ion trap mass spectrometry was used to characterize the molecular composition of dissolved organic matter and obtain isomeric labels;
[0008] The molecules detected in each dissolved organic matter sample are converted into a molecular graph network through chemical formula conversion;
[0009] Unsupervised learning methods are used to learn the global molecular transfer process and the hierarchical topological structure in the molecular graph network to obtain the molecular representation matrix;
[0010] Traverse the molecules in each dissolved organic matter sample and find its neighbor nodes from the molecular graph network;
[0011] According to the obtained molecular characterization matrix, the global characterization vector of all neighboring nodes is obtained, and the environmental feature vector of the molecule is obtained by taking the average value of the characterization of all neighboring nodes;
[0012] The characteristic representation vector of the molecule is concatenated with the environmental characteristic vector to obtain a constructed characteristic vector;
[0013] Based on the constructed feature vectors and isomer labels, a deep neural network prediction model for predicting isomers is built.
[0014] Preferably, when generating a molecular graph network by converting chemical formulas, the molecular networks in each sample are aggregated and deduplicated, and a global molecular graph network is constructed based on whether a chemical transfer process exists between molecules.
[0015] Preferably, an unsupervised learning method is used to learn the global molecular transfer process and the hierarchical topological structure in the molecular graph network, and the specific method is:
[0016] The random walk algorithm is used to generate node sequences;
[0017] For each molecular node in the node sequence, learning can express the inherent attribute characteristics of the node and the topological structure characteristics in the molecular graph network.
[0018] Preferably, the specific method of using the random walk algorithm to generate a node sequence is: select a molecular node and randomly select its neighbor node; after the molecular node moves one step, select the neighbor node of the molecular node after the move again; repeat this process and record the passed nodes to obtain a random walk node sequence.
[0019] Furthermore, the skip-gram model is used to learn the molecular node representation on the node sequence. The specific method is:
[0020] Construct a d-dimensional vector for each molecular formula;
[0021] Using each molecular node in the node sequence as the central node molecular formula to predict the context node molecular formula;
[0022] Traverse each node in the node sequence to form a <central molecule node, background molecule node> data pair;
[0023] A conditional probability model is constructed based on the representation vector of the central node and the feature vector of the context background node, and the probability of each central node inferring the background node is calculated.
[0024] Preferably, the specific calculation formula of the constructed conditional probability model is as follows:
[0025]
[0026] f in the formula c represents the central molecule, f o Represents the context node, u o and u c The molecular representation vector representing the central molecular node and the context background node; (f o |f c ) represents a given central molecular node f c , whose context node is f o probability.
[0027] Given a sequence of molecular nodes, its joint probability distribution is as follows:
[0028]
[0029] Where: T represents the position of the center node of the window, and m represents the size of the window;
[0030] The maximum likelihood estimation method is used to logarithmically transform the above formula, turning it into a monotonically increasing function, and transforming the maximization into the minimization, that is:
[0031]
[0032]
[0033] The gradient descent method is used to optimize and solve the representation vector u of each molecular node of the molecule.
[0034] Preferably, a negative sampling method is used to randomly extract central molecular nodes and background molecular nodes to construct negative samples, and then a gradient descent method is used to solve the representation vector u of each molecular node.
[0035] Compared with the prior art, the technical solution of the present invention has the following advantages:
[0036] A. The prediction model established by the method of the present invention includes molecular graph network construction, graph node representation learning based on random walk, and local environmental feature extraction and prediction. It comprehensively considers the attribute characteristics of each molecule while also considering the spatial structural characteristics of molecular movement and local environmental characteristics. The constructed prediction model can accurately predict the dissolved organic matter isomer vector.
[0037] B. The method of the present invention uses machine learning technology to build a prediction model based on the isomeric label data obtained by LC-Orbitrap MS, simulates the relationship between the molecular formula and the isomeric label vector, and enables it to be extended to the isomeric identification on DI-FT-ICRMS.
[0038] C. The precision and recall of the prediction model provided by the present invention in predicting isomer label vectors are 69.8% and 71.2% respectively. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] In order to more clearly illustrate the specific embodiments of the present invention, the drawings required for use in the specific embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0040] Figure 1 It is a block diagram of the prediction model construction method provided by the present invention. DETAILED DESCRIPTION
[0041] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0042] like Figure 1 As shown, the present invention provides a method for constructing a dissolved organic matter isomer vector prediction model based on graph learning, comprising the following steps:
[0043]
S01
[0044] For example: Based on 8 DOM samples from the Yangtze River Estuary to the East China Sea, high performance liquid chromatography coupled to electrostatic field orbital ion trap mass spectrometry (LC-Orbitrap MS) was used to characterize the molecular composition of DOM and obtain isomeric labels. The specific acquisition method is as follows (of course not limited to): a C18 reverse phase column (Thermo Scientific, Hypersil Gold Vanquish, 1.9μm, 100×2.1mm) was used to separate the samples before mass spectrometry analysis. There were two types of mobile eluents: eluent A: water containing 0.1% formic acid (v:v); eluent B: methanol containing 0.1% formic acid (v:v), with a flow rate of 0.3mL / min.
[0045] The proportion of eluent B was initially 10%, continued for 3 minutes, and gradually increased to 100% within 25 minutes; then maintained at 100% for 2 minutes, and decreased to 90% within 0.5 minutes, and maintained for 6 minutes. In specific operations, the total ion flow graph within the effective retention time of 9 minutes was divided into 9 time periods with an interval of 1 minute, and each molecule was observed to see whether it appeared in these 9 time periods. If it appeared once, it was recorded as an "isomer cluster" of the molecule.
[0046]
S02
[0047] For example, C 10 H 10 N2O3 can be converted to C by removing one O 10 H 10 In this example, molecular data of 8 samples were used to generate 8 molecular networks G l (V l ,E l ),V l represents the molecular set in the sample, E l Represents the molecular edge set in the sample. Then, the molecular network in each sample is aggregated and deduplicated, and a global molecular graph network is constructed based on whether there is a chemical transfer process between molecules. The generation process only focuses on whether there is a chemical transfer process between two molecules. If so, there is an edge, otherwise there is no edge. The generated global network has a total of 25,349 molecules and 5,933,073 edges.
[0048] Use G g (V g ,E g) represents the constructed graph, V g Represents the set of all molecular nodes, E g represents the set of chemical transfer edges in all samples.
[0049]
S03
[0050]
S031
[0051] Specifically, a node is selected, and then its neighboring nodes are randomly selected. After moving one step, the neighboring nodes of the node after the move are selected again to move, and this process is repeated, and the nodes passed are recorded, forming a random walk sequence.
[0052] The present invention adopts a biased random walk algorithm, returning to the current node from the current node with a probability of α=1, returning to the previous node with a probability of α=1 / p, and selecting the next node with a probability of α=1 / q. p and q are hyperparameters, p in the present invention is set to 4, and q is set to 1. The number of random walks is expressed as the length of the sequence, which is set to 80 in the present invention, that is, 80 molecular nodes are randomly walked. The above process is repeated n times, n is set to 30, and 30 node sequences can be generated for each molecular node.
[0053]
S032
[0054] The purpose of this invention is to learn a representation for each molecular node, which can express the inherent attribute characteristics of the node and its topological structure characteristics in the network. Specifically, in a given global graph G g , construct the mapping function f:V g →R d .
[0055] The present invention preferably adopts the skip-gram model to learn the molecular node representation on the node sequence. Specifically, a d-dimensional vector is constructed for each molecular formula (d is set to 128), and the central node molecular formula in the node sequence is used to predict the context node molecular formula, that is: take each molecular node as the central node, traverse the node sequence to form a <central molecular node, background molecular node> data pair. In this embodiment, 10 context background molecular nodes are constructed for each molecular node, that is, the first 5 and last 5 molecular nodes constitute the data pair. A conditional probability model is constructed based on the representation vector of the central node and the context background node feature vector to calculate whether the data pair is established. The specific calculation process is as follows:
[0056]
[0057] f in the formula c represents the central molecule, f o Represents the context node, u o and u c The molecular representation vector representing the central molecular node and the context background node; (f o |f c ) represents a given central molecular node f c , whose context node is f o The probability of
[0058] Subsequently, given a sequence of molecular nodes, its joint probability distribution can be written as follows:
[0059]
[0060] Where T represents the position of the center node of the window, and m represents the size of the window. In this way, the probability of each center node inferring the background node can be calculated. The input gives the vector of the background node, and at this time, it is only necessary to maximize the output probability of the background node. For this reason, the present invention preferably adopts the method of maximum likelihood estimation. Specifically, the above formula is logarithmically transformed to become a monotonically increasing function, and the maximization is converted into minimization, that is:
[0061]
[0062]
[0063] Since V represents the number of molecular nodes in the network, it is extremely time-consuming to traverse all molecular nodes. Therefore, the present invention preferably adopts a negative sampling method to reduce the amount of calculation, that is, randomly extracting central molecular nodes-background molecular nodes to construct negative samples.
[0064] Finally, the gradient descent method is used to solve the representation vector u of each molecular node.
[0065]
S04
[0066]
S05
[0067] Since the properties and number of isomers of each molecule in its environment vary greatly, the role of the molecule is determined by environmental factors. To this end, the present invention traverses the molecules in each sample and finds its neighbor nodes from the sample molecule network. The molecular characterization matrix learned in [S03] is used to obtain the global characterization vector of the neighbor node. Then the average of all neighbor node characterizations is taken to obtain the environmental feature vector of the molecule.
[0068]
S06
[0069] The characteristic representation vector of the molecule itself is concatenated with the environmental characteristic vector to obtain the constructed characteristic vector x, that is, The constructed feature vector x covers both its own attribute characteristics and its environmental factors, which is helpful for predicting the isomer vector of the molecular formula.
[0070]
S07
[0071] The constructed feature vector x is used as input, combined with the isomer labels, to construct a deep neural network prediction model. The present invention preferably constructs a 6-layer fully connected network, the input layer is a 256-dimensional feature vector, and the number of nodes in the hidden layer is 512, 512, 256, 256, and 128 respectively. The number of nodes in the output layer is 9. Finally, the stochastic gradient descent method SGD is used to train the network parameters to obtain a deep neural network prediction model.
[0072] Combined with the constructed deep neural network prediction model for predicting isomers, the dissolved organic matter isomer vector can be predicted based on the constructed feature vector x.
[0073] The present invention establishes an isomer vector prediction model through the acquisition of LC-Orbitrap MS isomer label data, data processing, graph network construction, random walk-based graph node representation learning, and local environment feature extraction and prediction. While comprehensively considering the attribute characteristics of each molecule (including explicit characteristics and implicit characteristics), the spatial structure characteristics of molecular movement and local environment characteristics are also considered. The constructed prediction model can accurately predict the dissolved organic matter isomer vector. After testing, the precision and recall rate in the prediction of isomer label vector are 69.8% and 71.2%, respectively.
[0074] The matters not described in the present invention are applicable to the prior art.
[0075] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. However, the obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.
Claims
1. A method for constructing a vector prediction model of dissolved organic matter isomers based on graph learning, characterized in that: The method comprises the following steps: Liquid chromatography coupled to electrostatic field orbital ion trap mass spectrometry was used to characterize the molecular composition of dissolved organic matter and obtain isomeric labels; The molecules detected in each dissolved organic matter sample are converted into a molecular graph network through chemical formula conversion; Unsupervised learning methods are used to learn the global molecular transfer process and the hierarchical topological structure in the molecular graph network to obtain the molecular representation matrix; Traverse the molecules in each dissolved organic matter sample and find its neighbor nodes from the molecular graph network; According to the obtained molecular characterization matrix, the global characterization vector of all neighboring nodes is obtained, and the environmental feature vector of the molecule is obtained by taking the average value of the characterization of all neighboring nodes; The characteristic representation vector of the molecule is concatenated with the environmental characteristic vector to obtain a constructed characteristic vector; Based on the constructed feature vectors and isomer labels, a deep neural network prediction model for predicting isomers is built.
2. The method for constructing a dissolved organic matter isomer vector prediction model based on graph learning according to claim 1, characterized in that: When generating a molecular graph network through chemical formula conversion, the molecular networks in each sample are aggregated and deduplicated, and a global molecular graph network is constructed based on whether there is a chemical transfer process between molecules.
3. The method for constructing a dissolved organic matter isomer vector prediction model based on graph learning according to claim 1, characterized in that: An unsupervised learning method is used to learn the global molecular transfer process and the hierarchical topological structure in the molecular graph network. The specific method is: The random walk algorithm is used to generate node sequences; For each molecular node in the node sequence, learning can express the inherent attribute characteristics of the node and the topological structure characteristics in the molecular graph network.
4. The method for constructing a dissolved organic matter isomer vector prediction model based on graph learning according to claim 3, characterized in that: The specific method of using the random walk algorithm to generate a node sequence is: select a molecular node and randomly select its neighbor node; after the molecular node moves one step, select the neighbor node of the molecular node after the move again; repeat this process and record the passed nodes to obtain a random walk node sequence.
5. The method for constructing a dissolved organic matter isomer vector prediction model based on graph learning according to claim 4, characterized in that: The skip-gram model is used to learn the molecular node representation on the node sequence. The specific method is: Construct a d-dimensional vector for each molecular formula; Using each molecular node in the node sequence as the central node molecular formula to predict the context node molecular formula; Traverse each node in the node sequence to form a <central molecule node, background molecule node> data pair; A conditional probability model is constructed based on the representation vector of the central node and the feature vector of the context background node, and the probability of each central node inferring the background node is calculated.
6. The method for constructing a dissolved organic matter isomer vector prediction model based on graph learning according to claim 5, characterized in that: The specific calculation formula of the constructed conditional probability model is as follows: f in the formula c represents the central molecule, f o Represents the context node, u o and u c The molecular representation vector representing the central molecular node and the context background node; (f o |f c ) represents a given central molecular node f c , whose context node is f o The probability of; V represents the number of molecular nodes in the network; Given a sequence of molecular nodes, its joint probability distribution is as follows: Where: T represents the position of the center node of the window, and m represents the size of the window; The maximum likelihood estimation method is used to logarithmically transform the above formula, turning it into a monotonically increasing function, and transforming the maximization into the minimization, that is: The gradient descent method is used to optimize and solve the representation vector u of each molecular node.
7. The method for constructing a dissolved organic matter isomer vector prediction model based on graph learning according to claim 6, characterized in that: The negative sampling method is used to randomly extract the central molecular node-background molecular node to construct negative samples, and then the gradient descent method is used to solve the representation vector u of each molecular node.
Citation Information
Patent Citations
Link prediction method based on heterogeneous network representation learning
CN111325326A
Molecular graph representation learning method based on contrastive learning
US20230052865A1
AU2020103335A4
Cited By
Prediction method for photochemical reactivity of soluble organic matter molecules and related medium
CN121506280A
Degradation prediction method for soluble organic molecules based on time sequence chart neural network
CN121583361A
