A method for predicting degradation of dissolved organic molecules based on time series graph neural networks

Abstract

The present application relates to the technical field of environmental geochemistry, and particularly relates to a dissolved organic molecule degradation prediction method based on a time sequence graph neural network. The dissolved organic molecule degradation prediction method provided by the present application is based on an existing water body degradation whole-process molecular composition dataset, and data standardization processing is performed according to a specified standard. Then, the standard data matrix is converted into a molecular graph network and input into a time sequence graph neural network to form a graph sequence in time sequence organization. Through the combination of the graph network features of the molecular structure and the time sequence constraints of the environmental conditions, a dynamic graph neural network model is constructed to realize high-precision prediction of the degradation path, rate and potential product of the organic molecules in the environment under different conditions. The present application introduces a rule rationality constraint and an explainability mechanism, which can effectively guarantee the chemical effectiveness of the prediction results and provide explainable output results, thereby further enhancing the reliability and explainability of the model.

Description

Technical Field

[0001] This invention relates to the field of environmental geochemistry, and in particular to a method for predicting the degradation of soluble organic molecules based on time-series graph neural networks. Background Technology

[0002] With the rapid development of industry and agriculture in my country, a large number of toxic and harmful organic molecules, represented by chemical fertilizers, organic pesticides, chemical drug residues, and other emerging organic pollutants (persistent organic pollutants, endocrine disruptors, antibiotics, microplastics, etc.), enter the water and soil environment through various methods such as agricultural runoff, urban drainage, industrial sewage discharge, landfill, and atmospheric deposition. They also widely enter critical ecological interfaces such as surface rivers and lakes, groundwater bodies, and oceans through hydrological cycles, leading to widespread environmental pollution, ecological degradation, and potential risks to biological health. As a planet with approximately 71% of its surface covered by water, the aquatic ecological environment is crucial for maintaining Earth's ecosystems, regulating regional climates, and supporting socio-economic activities. Therefore, predicting the degradation pathways of dissolved organic molecules in water and analyzing the environmental persistence and potential degradation pathways of organic molecules are of great significance.

[0003] With the development of modern analytical and detection technologies, various physicochemical testing methods have been widely applied to predict dissolved organic molecules in the environment. Current traditional techniques for analyzing and predicting environmental organic molecules mainly include migration process simulation experiments, organic matter half-life determination, chemical reaction degradation models, and ecosystem degradation simulation models. These methods are often based on laboratory physicochemical experiments, using basic analytical methods such as chromatography, spectroscopy, mass spectrometry, nuclear magnetic resonance spectroscopy, fluorescence spectroscopy, and electrochemical analysis to obtain the characteristics of organic molecules in water bodies over a certain period.

[0004] Traditional prediction methods, which rely on simplification, assumptions, and segmentation to study problems, struggle to reconstruct a global dynamic picture, exhibiting significant limitations, especially when dealing with ultra-complex, nonlinear environmental systems. In recent years, the widespread application of data-driven artificial intelligence (AI) technologies has provided a novel perspective for analyzing the properties of dissolved organic molecules in water and their correlation with molecular degradation. Graph Neural Networks (GNNs), with their unique three-layer architecture of vertex-edge-global, possess significant advantages in processing data structures with complex relationships, particularly network-like relationships. Summary of the Invention

[0005] To address the shortcomings of traditional methods for predicting the degradation of dissolved organic molecules in water, the present invention aims to provide a method for predicting the degradation of dissolved organic molecules based on time series graph neural networks, thereby achieving high-precision prediction of the degradation of organic molecules in the aquatic environment.

[0006] This invention provides a method for predicting the degradation of soluble organic molecules based on time-series graphical neural networks, comprising:

[0007] Data acquisition and preprocessing: High-resolution Fourier transform ion cyclotron resonance mass spectrometry was used to test water samples at different times and degradation stages to obtain the molecular formulas and relative abundances of existing organic molecules; the obtained data were then cleaned, standardized, and converted into a standard data matrix.

[0008] A time series graph neural network is constructed by inputting a standard data matrix into the time series graph neural network in the form of a molecular graph, and then organizing it in chronological order to form a graph sequence.

[0009] By training a time-series graph neural network model, we can learn and construct a network of transformation relationships between organic molecules, thereby predicting the degradation pathways of molecules under different time and environmental conditions.

[0010] Based on fundamental principles of physics and / or chemical rules, the predicted degradation pathways are validated for reasonableness; key reaction sites are visually identified to generate interpretable prediction results.

[0011] Furthermore, the data preprocessing includes:

[0012] For the loss of molecular features due to incomplete detection signals or uncertain peak identification, a differentiated repair strategy is adopted according to the type of loss.

[0013] For systematic missing data, the adjacent sample averaging method or the interpolation method based on molecular structure similarity is used to fill in the missing data.

[0014] For random missing values, reasonable values ​​are restored by combining distribution fitting and regression imputation methods.

[0015] Outlier handling utilizes the Interquartile Range (IQR) method to identify outliers. Data exceeding 1.5 times the IQR range is corrected or removed.

[0016] Missing value imputation (structural similarity interpolation / regression imputation), outlier detection (IQR method), and numerical normalization are employed to ensure data consistency.

[0017] Furthermore, the step of inputting the standard data matrix into the time series graph neural network in the form of a molecular graph, and organizing it in chronological order to form a graph sequence, includes:

[0018] Individual organic molecules in the water sample are considered as nodes in a molecular graph, while the intermolecular transformation relationships existing at the current time are considered as edges;

[0019] Molecular maps from different time points are organized in chronological order to form a map sequence.

[0020] In time-series molecular prediction tasks, the fusion of graph neural networks (GNNs) and long short-term memory networks (LSTMs) provides a powerful framework for modeling the dynamic evolution of molecules. Unlike traditional molecular graphs that use atoms as nodes and chemical bonds as edges, this invention employs a high-level reaction graph representation method with molecules as nodes and intermolecular transformation relationships as edges. This graph structure can characterize the reactions, transformations, and interactions between molecules, thus better reflecting the macroscopic evolutionary features of complex chemical processes.

[0021] At each time step, the GNN is responsible for encoding the reaction graph at the current moment. It captures the structural similarity and reaction reachability between molecules by aggregating information from neighboring molecular nodes, generating a temporal feature representation for each molecule. These graph embeddings are then fed as input to the LSTM. The LSTM leverages its ability to model long-term dependencies in time series to learn the evolution of molecular states over time. In other words, the GNN learns the relational structure of the molecular network in the spatial dimension, while the LSTM captures the changing trends of molecular states over time.

[0022] This spatiotemporal coupling model of GNN–LSTM can integrate "intermolecular relationships" (modeled through graph structure) and "time-dependent relationships" (modeled through sequence) within the same framework, thereby enabling spatiotemporal dynamic prediction of molecular systems. The time-series graph neural network model can not only characterize short-term transformation pathways (such as the generation and disappearance of reaction intermediates), but also predict long-term system evolution trends (such as pollutant degradation processes or steady-state changes in metabolic networks).

[0023] This invention constructs a graph sequence of molecular topological changes during degradation (edge ​​breaking / generation, node / functional group appearance and disappearance), and directly learns the topological evolution pattern in the temporal dimension, thereby predicting the topological structure of new intermediates and final products.

[0024] Furthermore, the learning and construction of the organic intermolecular transformation network also includes:

[0025] The focus of node and edge feature learning is to extract the topological relationships and chemical properties between molecules, and to quantify the importance of nodes in global and local networks as well as the intrinsic properties of molecules.

[0026] Edge learning is based on local and global connectivity features.

[0027] Furthermore, the quantification of the importance of nodes in global and local networks and the intrinsic properties of molecules includes: introducing a Graphlet degree vector, which vectorizes the structure of nodes in local networks and counts the frequency of small connected subgraphs around them to accurately capture the topological features of the node's neighborhood.

[0028] The formula for the Graphlet degree vector is shown in equation (1):

[0029] (1);

[0030] in, The target node for calculating the feature vector; It represents a specific node role, that is, a node's unique structural position in a certain Graphlet; Represents a node Within its neighborhood The number of times the root node appears; This represents the total number of different nodes considered, and its value depends on the maximum size of the Graphlet.

[0031] Node features are integrated with molecular topological features, physicochemical properties, and local structural features of Graphlet Degree Vector (GDV).

[0032] Furthermore, the edge learning is based on local and global connectivity features, including:

[0033] Local connectivity demonstrates the similarity of molecular physicochemical properties and degradation stages by analyzing common neighbors between molecules, and is then transferred to molecular network evaluation using the Adamic Adar algorithm to quantitatively calculate the proximity of graph network nodes.

[0034] The Adamic Adar algorithm formula is shown in equation (2):

[0035] (2);

[0036] In the above formula, and Representing nodes respectively and The set of adjacent nodes; This represents each node when traversing the set of adjacent nodes; The degree of a node, i.e., its relation to the node. The number of adjacent nodes.

[0037] Furthermore, when there are too few directly adjacent nodes in the graph network, the Katz exponent is used to calculate the number of all paths of all lengths between two nodes, and the shorter path is given greater weight by using a decay coefficient to capture indirect transformation relationships.

[0038] To prevent differences in the numerical dimensions of features from affecting model convergence and performance, the node features and edge weights were normalized using Min-Max.

[0039] Edge features are modeled using a combination of Adamic-Adar and Katz indices to express direct and indirect degradation associations between molecules.

[0040] By capturing cross-temporal dependencies through sequence training, high-precision prediction of molecular degradation rates and pathways can be achieved.

[0041] Furthermore, the fundamental physical principles include the thermodynamic standard molar Gibbs free energy change and the kinetic Arrhenius equation;

[0042] The thermodynamic standard molar Gibbs free energy change formula is shown in equation (3):

[0043] (3);

[0044] in, This is the standard molar Gibbs free energy change, in kJ·mol. -1 , which is the change in free energy when the reaction proceeds by 1 mol of reaction under standard conditions; is the molar gas constant, with a value of 8.314 J·mol⁻¹. -1 ·K -1 ; Temperature is the thermodynamic temperature, and its unit is Kelvin (K). The standard equilibrium constant is a dimensionless index that measures the degree to which a reaction tends toward equilibrium.

[0045] The dynamic Arrhenius equation is shown in equation (4):

[0046] (4);

[0047] In the above formula, It is the reaction rate constant, which can be used to quantitatively characterize the speed of a chemical reaction; The pre-collision factor is related to the frequency of collisions between reactant molecules and their orientation, representing the tendency for molecules to undergo effective collisions at higher temperatures. Activation energy, measured in kJ / mol, is the minimum energy required for reactant molecules to transition from their normal state to their activated state. is the molar gas constant, with a value of 8.314 J·mol⁻¹.-1 ·K -1 ; Temperature is the thermodynamic temperature, and its unit is Kelvin (K).

[0048] The chemical rules include valence bond conservation, energy-driven constraints, and priority of known reaction templates.

[0049] Furthermore, the visualization and identification of key response sites to generate interpretable prediction results includes:

[0050] The correlation of degradation reactions is learned through the reaction site attention model, and key reaction nodes, more active molecules or reaction pathways are given higher weights.

[0051] The interpretability of a model can be improved by using the SHAP analysis method. The SHAP analysis method calculates the contribution of features to the model's predictions and assigns an independent contribution value to each input feature, which can clearly show which features play a key role in the model's decision-making.

[0052] Furthermore, a reaction site attention module is embedded in the spatial feature aggregation layer of the time series graph neural network model. Through the reaction site attention module, the time series graph neural network model can automatically allocate weights according to the importance of nodes and edges, increase the attention to key functional groups and chemical bonds in molecules, and explicitly focus on reactive sites, thereby providing more accurate details for the prediction of molecular degradation pathways.

[0053] To avoid "black box" predictions and ensure the scientific rationality of model outputs, this invention introduces interpretability and verification mechanisms into time series graphical neural networks. By combining the response site attention module with interpretability subgraph generation, the prediction results maintain high accuracy while being intuitively interpretable.

[0054] In summary, the advantages and beneficial effects of the present invention are as follows:

[0055] This invention provides a method for predicting the degradation of soluble organic molecules based on a time-series graph neural network. It utilizes an existing dataset of molecular composition from the entire degradation process in water bodies, which is standardized and transformed into a standard data matrix. This standard data matrix is ​​then input into the time-series graph neural network in the form of a molecular graph, organized chronologically to form a graph sequence. By combining the graph network features of molecular structure with the time-series constraints of environmental conditions, a dynamic graph neural network model is constructed to achieve high-precision prediction of the degradation pathways, rates, and products of organic molecules in the environment under different conditions. By introducing the rational constraints and interpretability mechanisms of basic physical principles and / or chemical rules, the chemical validity of the prediction results can be effectively guaranteed, and interpretable output results can be provided, thereby further enhancing the reliability and interpretability of the model. Attached Figure Description

[0056] Figure 1 The flowchart of the degradation prediction method for soluble organic molecules based on time series graph neural networks disclosed in this invention;

[0057] Figure 2 This invention discloses a schematic diagram of the degradation prediction process of soluble organic molecules based on a time-series graph neural network. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0059] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0060] This invention provides a method for predicting the degradation of soluble organic molecules based on time-series graph neural networks, such as... Figure 1 As shown, the specific steps include: S10 to S40.

[0061] S10: Data acquisition and preprocessing. High-resolution Fourier transform ion cyclotron resonance mass spectrometry was used to test water samples at different times and degradation stages to obtain the molecular formulas and relative abundances of existing organic molecules. The obtained data was then cleaned, standardized, and converted into a standard data matrix.

[0062] S20: Construct a time series graph neural network. Input the standard data matrix into the time series graph neural network in the form of a molecular graph, and organize it in chronological order to form a graph sequence.

[0063] S30: By training a time-series graph neural network model, we learn and construct a network of transformation relationships between organic molecules, thereby predicting the degradation pathways of molecules under different time and environmental conditions;

[0064] S40: Based on fundamental physical principles and / or chemical rules, validate the rationality of predicted degradation pathways; visually identify key reaction sites and generate interpretable prediction results.

[0065] In this embodiment, data is first acquired, then processed into a data matrix, and input into a time-series graph neural network in the form of a molecular graph. After being organized chronologically, a graph sequence is formed. Through a multi-layer graph convolutional network (GCN), the time-series graph neural network model can learn the temporal dependencies between these molecules, thereby predicting the degradation pathways of molecules under different environmental conditions. The feasibility of the degradation pathways is verified, key reaction sites are visualized and identified, and interpretable prediction results are generated.

[0066] Compared to cross-sectional data, time series data provides continuous data over multiple periods while also reflecting past trends, making it a good source of data for predicting the future, classifying samples, and detecting anomalies.

[0067] For example, based on samples collected over many years from rivers (the Yangtze River, the Yellow River, and their major first- and second-level tributaries), estuaries (the estuaries of major outflowing rivers in my country), and surrounding oceans (the Bohai Sea, the Yellow Sea, the East China Sea, and the South China Sea), the natural degradation process of environmental organic molecules is artificially simulated in the laboratory by setting conditions such as temperature, light, and microbial activity. Samples are taken at different time points (e.g., 0, 7, 14, 49, 180, 360 days) and Fourier transform mass spectrometry (FT-ICR MS) is used to determine the molecular formula and relative abundance of dissolved organic matter in the water. Simultaneously, multiple open molecular degradation databases can be integrated, and after data cleaning and standardization, they can be combined with laboratory data to construct a comprehensive DOM (dissolved organic matter) molecular degradation database. With the continuous growth of data volume, the accuracy and generalization ability of the time-series graph neural network model will be further improved, enabling high-precision and customized predictions for different water body types (oceans, rivers, estuaries, etc.) in the future.

[0068] Based on the large-scale FT-ICR MS dataset, data suitable for temporal graph neural networks is obtained through a series of methods such as data standardization and feature extraction.

[0069] When performing time series forecasting based on GNN, the target water sample's molecular domain (DOM) must first be monitored using FT-ICR MS at a specified time (ti). After data cleaning and processing, a graph network (GNNi) for each monitoring section is generated. In each network, nodes represent different molecules of the DOM, and edges represent the correlations between molecules. This network can reveal the molecular change patterns at the current time point and provide strong support for subsequent time series forecasting.

[0070] In one embodiment, step S10, data preprocessing, includes:

[0071] For the loss of molecular features due to incomplete detection signals or uncertain peak identification, a differentiated repair strategy is adopted according to the type of loss.

[0072] For systematic missing data, the adjacent sample averaging method or the interpolation method based on molecular structure similarity is used to fill in the missing data.

[0073] For random missing values, reasonable values ​​are restored by combining distribution fitting and regression imputation methods.

[0074] Outlier handling utilizes the Interquartile Range (IQR) method to identify outliers. When data exceeds 1.5 times the IQR range, it is corrected or removed as appropriate.

[0075] In one embodiment, step S20 involves inputting a standard data matrix into a time-series graph neural network in the form of a molecular graph, organizing it chronologically to form a graph sequence, including:

[0076] Individual organic molecules in the water sample are considered as nodes in a molecular graph, while the intermolecular transformation relationships existing at the current time are considered as edges;

[0077] Molecular maps from different time points are organized in chronological order to form a map sequence.

[0078] For example, each molecule in a graph neural network not only acts as a static node but also updates its feature representation over time. Within each time step, the network updates the edge weights based on the transformation relationships between molecules (such as chemical reactions, degradation processes, etc.), thus dynamically adjusting the node features and fully utilizing the evolutionary information of molecules over time. Through multilayer graph convolutional networks (GCNs), time-series graph neural network models can learn the temporal dependencies between these molecules, thereby predicting the degradation paths of molecules under different environmental conditions.

[0079] As the target time point GNN network is constructed, degradation path maps at different time points (e.g., 0, 7, 14, 49 days) are organized chronologically to form a graph sequence, which serves as input to the time-series graph neural network to learn dynamic patterns of degradation behavior. In this setup, the molecular graph at each time point not only includes intermolecular transformation relationships but also state features such as the relative abundance of molecules within that time step. The graph sequence is trained using a time-series graph neural network model, which learns not only the molecular features at each time point but also captures dynamic change patterns across time steps. Subsequently, in the graph neural network, each layer's graph convolution operation is based on the current graph state and adjacency relationships in the time series. As the number of network layers increases, the time-series graph neural network model can progressively learn deeper temporal features extracted from the time series and ultimately generate an output that can accurately predict future molecular states. During this process, the time-series graph neural network dynamically adjusts the feature representations of nodes and edges in the graph to better adapt to the degradation trends of organic molecules in the water sample.

[0080] After constructing the graph neural network for each time step, the next step is to model the entire time series and explore the evolutionary relationships of molecules at different time points. To this end, a dynamic graph modeling approach is used to describe the time-varying molecular relationships. In this process, the graph neural network learns the dynamic changes at each time step by continuously updating the features of nodes and edges, and captures the patterns of molecular evolution through a temporal modeling mechanism. Combining this with a Long Short-Term Memory (LSTM) network model can further enhance the performance of the graph neural network in time series prediction, enabling it to accurately predict the molecular state and its degradation path at future time steps.

[0081] In one embodiment, step S30, the learning and construction of the organic intermolecular transformation network, further includes:

[0082] The focus of node and edge feature learning is to extract the topological relationships and chemical properties between molecules, and to quantify the importance of nodes in global and local networks as well as the intrinsic properties of molecules.

[0083] Edge learning is based on local and global connectivity features.

[0084] For example, common quantitative metrics include degree centrality, eigenvector centrality, and clustering coefficient, which are mainly used to measure the importance and connectivity of nodes in the overall network.

[0085] More specifically, the importance of nodes in global and local networks and the intrinsic properties of molecules are quantified, including:

[0086] By introducing Graphlet degree vectors, the structure of nodes in the local network is vectorized, and the frequency of small connected subgraphs around them is statistically analyzed to accurately capture the topological features of the node's neighborhood.

[0087] The formula for the Graphlet degree vector is shown in equation (1):

[0088] (1);

[0089] in, The target node for calculating the feature vector; It represents a specific node role, that is, a node's unique structural position in a certain Graphlet; Represents a node Within its neighborhood The number of times the root node appears; This represents the total number of different nodes considered, and its value depends on the maximum size of the Graphlet.

[0090] For example, u is the current specific target node; d is the total number of all possible structural positions; z refers to a specific role; and c(u,z) is the core statistic, recording the number of times node u actually plays the role z in its neighborhood (e.g., the node appears 5 times as a "triangle vertex" and 3 times as a "chain center"). This series of statistics can ultimately be pieced together into a feature vector, accurately describing the node's "identity" and importance in the local network topology.

[0091] By combining these topological features, a comprehensive feature vector for each node is formed, which includes both the importance information of the global network and the local topological structure, further enhancing the graph neural network model's ability to learn node relationships.

[0092] More specifically, edge learning is based on local and global connectivity features, including:

[0093] Local connectivity demonstrates the similarity of molecular physicochemical properties and degradation stages by analyzing common neighbors between molecules, and is then transferred to molecular network evaluation using the Adamic Adar algorithm to quantitatively calculate the proximity of graph network nodes.

[0094] The formula for the Adamic Adar algorithm is shown in equation (2):

[0095] (2);

[0096] In the above formula, and Representing nodes respectively and The set of adjacent nodes; This represents each node when traversing the set of adjacent nodes; The degree of a node, i.e., its relation to the node. The number of adjacent nodes.

[0097] The Adamic Adar algorithm evaluates the similarity between different nodes in a graph network by measuring the common neighbors between two nodes and assigning greater weight to common neighbors with lower degrees. Generally speaking, A higher value indicates a greater similarity between two nodes, and consequently, more shared neighbors. The Adamic Adar algorithm measures the closeness between nodes by calculating the common neighbors of two nodes and assigning higher weights to common neighbors with lower degrees.

[0098] In one embodiment, when there are too few directly common adjacent nodes in the graph network, the Katz exponent is used to calculate the number of all length paths between two nodes, and shorter paths are given greater weight by using a decay coefficient to capture indirect transformation relationships.

[0099] To prevent differences in the numerical dimensions of features from affecting model convergence and performance, the node features and edge weights were normalized using Min-Max.

[0100] The Katz index captures potential connections between even distant nodes by calculating the number of all possible paths and assigning greater weight to shorter paths.

[0101] To prevent differences in the numerical dimensions of features from affecting model convergence and performance, this embodiment strictly performs Min-Max normalization on node features, edge weights, etc.

[0102] In one embodiment, in step S40, the basic principles of physics include the thermodynamic standard molar Gibbs free energy change and the kinetic Arrhenius equation;

[0103] The thermodynamic standard molar Gibbs free energy change formula is shown in equation (3):

[0104] (3);

[0105] in, This is the standard molar Gibbs free energy change, in kJ·mol. -1 , which is the change in free energy when the reaction proceeds by 1 mol of reaction under standard conditions; is the molar gas constant, with a value of 8.314 J·mol⁻¹. -1 ·K -1 ; Temperature is the thermodynamic temperature, and its unit is Kelvin (K). The standard equilibrium constant is a dimensionless index that measures the degree to which a reaction tends toward equilibrium.

[0106] The dynamic Arrhenius equation is shown in equation (4):

[0107] (4);

[0108] In the above formula, It is the reaction rate constant, which can be used to quantitatively characterize the speed of a chemical reaction; The pre-collision factor is related to the frequency of collisions between reactant molecules and their orientation, representing the tendency for molecules to undergo effective collisions at higher temperatures. Activation energy, measured in kJ / mol, is the minimum energy required for reactant molecules to transition from their normal state to their activated state. is the molar gas constant, with a value of 8.314 J·mol⁻¹. -1 ·K -1 ; Temperature is the thermodynamic temperature, and its unit is Kelvin (K).

[0109] Chemical rules include valence bond conservation, energy-driven constraints, and priority of known reaction templates.

[0110] The degradation pathway needs to be analyzed for feasibility using thermodynamic standard molar Gibbs free energy change calculations (Equation 3) and the kinetic Arrhenius equation (Equation 4). In practical analysis, if... The reaction can be considered to be spontaneous. It is believed that its reverse reaction is more likely to be spontaneous; This indicates that the situation is in equilibrium.

[0111] First use Determine whether a reaction is possible under given conditions. If thermodynamically feasible, the reaction rate can then be evaluated using the Arrhenius equation. (Whether the value is large enough). This embodiment predicts the chemical reaction process from the topological level, and with the dual constraints of thermodynamics and kinetics, it realizes the scientific rationality verification of the prediction, which can significantly improve the reliability of the prediction.

[0112] In one embodiment, step S40 involves visually identifying key response sites and generating interpretable prediction results, including:

[0113] The correlation of degradation reactions is learned through the reaction site attention model, and key reaction nodes, more active molecules, or reaction pathways are given higher weights.

[0114] More specifically, a reaction site attention module is embedded in the spatial feature aggregation layer of the time series graph neural network model. Through the reaction site attention module, the time series graph neural network model can automatically allocate weights according to the importance of nodes and edges, increase the attention to key functional groups and chemical bonds in molecules, and give explicit attention to reactive sites, thereby providing more accurate details for the prediction of molecular degradation pathways.

[0115] When learning the temporal representation of a molecule, the model not only aggregates the features of its neighborhood nodes, but also needs to determine which molecules or bonds in the neighborhood contribute the most to the degradation changes of the target molecule.

[0116] This embodiment introduces learnable attention parameters, enabling the model to automatically learn these "reaction correlations" and assign higher weights to key reaction nodes, more active molecules, or reaction pathways. This mechanism ensures that the model can effectively identify the most influential molecules and chemical bonds in complex chemical reaction networks, thereby improving the accuracy and interpretability of predictions.

[0117] To achieve interpretable output for degradation pathway prediction and to visualize key reaction sites, this embodiment introduces a Reaction-Site Attention Mechanism (RSAM) into the time-series graph neural network structure. This mechanism embeds a learnable attention module (Graph Attention Layer, GAL) into the graph convolutional coding layer, enabling the model to automatically assign weights based on the importance of nodes and edges, significantly improving the attention given to key functional groups and chemical bonds in the molecule. In this way, the GNN can not only effectively aggregate neighborhood information but also explicitly focus on reactive sites through the attention mechanism, thus providing more accurate details for predicting molecular degradation pathways.

[0118] The attention mechanism is integrated into the spatial feature aggregation layer (Graph Encoder) of the time-series graph neural network, specifically the node update stage of the GNN model. In this stage, node features are updated not only by information from neighboring nodes but also by considering the "importance" of each neighboring node—that is, attention allocation based on reaction sites—making the representation of each node more accurate across different time steps. This adaptive weight allocation better simulates the dynamic evolution of different reaction sites during molecular degradation, especially nodes with high reactivity or influence. Through multi-level propagation of graph convolutional layers and the fusion of node attention, the model can progressively construct the evolutionary trajectory of molecules at different time steps, thereby accurately predicting future degradation paths. Because the model can focus on the most critical reaction nodes, this method significantly improves the interpretability of predictions, helping users understand which molecular features or reaction pathways have the most significant impact on the degradation process.

[0119] By introducing the Reaction Site Attention (RSAM) mechanism, this invention not only improves the accuracy of molecular degradation prediction, but also enables interpretable analysis of complex molecular reactions, greatly enhancing the application value of the model and user trust. It has broad application prospects, especially in environmental monitoring, drug degradation, materials science and other fields.

[0120] The interpretability of a model can be improved by using the SHAP analysis method. The SHAP analysis method calculates the contribution of features to the model's predictions and assigns an independent contribution value to each input feature, which can clearly show which features play a key role in the model's decision-making.

[0121] The SHAP (SHapley Additive exPlanations) analysis method is employed to further enhance model interpretability, especially when dealing with complex time-series data coupled with Graph Neural Networks (GNNs) and Long Short-Term Memory Networks (LSTMs). As a game theory-based interpretability method, SHAP assigns an independent contribution value to each input feature by calculating the contribution of features to the model's predictions. This clearly reveals which features play a key role in the model's decision-making. Compared to traditional model interpretation methods (such as feature importance ranking and LIME), SHAP provides a more accurate and systematic explanation. Traditional methods typically calculate feature importance based on local or global heuristics, while SHAP, by considering all possible feature combinations, provides a precise contribution value for each feature to the model's predictions. This method is not only more theoretically grounded but also provides users with a more reliable and detailed model interpretation.

[0122] like Figure 2 As shown, in this embodiment of the invention, nodes represent molecules at a certain time (t-2, t-1, or t), while edges represent the relationships between molecules, reflecting their interactions or similarities in physical and chemical properties. As time progresses, the graph neural network can capture the temporal changes of molecules, helping to reveal molecular degradation processes, reaction pathways, or other spatiotemporal evolution characteristics. In this way, time-series graph neural networks can more accurately predict future molecular states, providing important analytical basis.

[0123] A molecular degradation pathway tree is constructed based on the output of a time-series graph neural network, displaying the transformation chain from the parent molecule to various intermediates and the final product. Path tree nodes carry time step information and thermodynamic parameters, and can be updated in real time to output chemical evolution trends. It supports automatic screening of high-probability product branches and quantitative analysis of key reaction pathways.

[0124] Finally, it should be noted that any modification or equivalent substitution of some or all of the technical features made based on the methods and technical solutions disclosed in this invention and the embodiments thereof, without departing from the corresponding technical solutions of this invention, shall fall within the patent scope of the device structure and the embodiments thereof.

Claims

1. A method for predicting the degradation of soluble organic molecules based on time-series graphical neural networks, characterized in that, include: Data acquisition and preprocessing: High-resolution Fourier transform ion cyclotron resonance mass spectrometry was used to test water samples at different times and degradation stages to obtain the molecular formulas and relative abundances of existing organic molecules; the obtained data were then cleaned, standardized, and converted into a standard data matrix. A time series graph neural network is constructed by inputting a standard data matrix into the time series graph neural network in the form of a molecular graph, and then organizing it in chronological order to form a graph sequence. By training a time-series graph neural network model, we learn and construct a network of transformation relationships between organic molecules, thereby predicting the degradation pathways of molecules under different time and environmental conditions. The feature learning of nodes and edges focuses on extracting the topological relationships and chemical properties between molecules, quantifying the importance of nodes in the global and local networks, and the intrinsic properties of molecules. Edge learning is based on local and global connectivity features. Graphlet degree vectors are introduced to vectorize the structure of nodes in the local network and count the frequency of small connected subgraphs around them to accurately capture the topological features of the node's neighborhood. The formula for the Graphlet degree vector is shown in equation (1): （1）； in, The target node for calculating the feature vector; It represents a specific node role, that is, a node's unique structural position in a certain Graphlet; Represents a node Within its neighborhood The number of times the root node appears; This represents the total number of different nodes considered, and its value depends on the maximum size of the Graphlet; Based on fundamental principles of physics and / or chemical rules, the predicted degradation pathways are validated for reasonableness; key reaction sites are visually identified to generate interpretable prediction results.

2. The method for predicting the degradation of soluble organic molecules based on time-series graphical neural networks according to claim 1, characterized in that, The data preprocessing includes: For the loss of molecular features due to incomplete detection signals or uncertain peak identification, a differentiated repair strategy is adopted according to the type of loss. For systematic missing data, the adjacent sample averaging method or the interpolation method based on molecular structure similarity is used to fill in the missing data. For random missing values, reasonable values ​​are restored by combining distribution fitting and regression imputation methods. Outlier handling utilizes the interquartile range method to identify outliers. When data exceeds 1.5 times the range, it is corrected or removed.

3. The method for predicting the degradation of soluble organic molecules based on time-series graphical neural networks according to claim 1, characterized in that, The step of inputting the standard data matrix into the time series graph neural network in the form of a molecular graph, and organizing it in chronological order to form a graph sequence includes: Individual organic molecules in the water sample are considered as nodes in a molecular graph, while the intermolecular transformation relationships existing at the current time are considered as edges; Molecular maps from different time points are organized in chronological order to form a map sequence.

4. The method for predicting the degradation of soluble organic molecules based on time-series graph neural networks according to claim 2, characterized in that, The edge learning is based on local and global connectivity features, including: Local connectivity demonstrates the similarity of molecular physicochemical properties and degradation stages by analyzing common neighbors between molecules, and is then transferred to molecular network evaluation using the Adamic Adar algorithm to quantitatively calculate the proximity of graph network nodes. The Adamic Adar algorithm formula is shown in equation (2): （2）； In the above formula, and Representing nodes respectively and The set of adjacent nodes; This represents each node when traversing the set of adjacent nodes; The degree of a node, i.e., its relation to the node. The number of adjacent nodes.

5. The method for predicting the degradation of soluble organic molecules based on time-series graphical neural networks according to claim 4, characterized in that, When there are too few directly adjacent nodes in a graph network, the Katz exponent is used to calculate the number of all paths of all lengths between two nodes, and the shorter paths are given greater weight by using a decay coefficient to capture indirect transformation relationships. To prevent differences in the numerical dimensions of features from affecting model convergence and performance, the node features and edge weights were normalized using Min-Max.

6. The method for predicting the degradation of soluble organic molecules based on time-series graphical neural networks according to claim 1, characterized in that, The fundamental principles of physics include the thermodynamic standard molar Gibbs free energy change and the dynamical Arrhenius equation; The thermodynamic standard molar Gibbs free energy change formula is shown in equation (3): （3）； in, This is the standard molar Gibbs free energy change, in kJ·mol. -1 , which is the change in free energy when the reaction proceeds by 1 mol of reaction under standard conditions; is the molar gas constant, with a value of 8.314 J·mol⁻¹. -1 ·K -1 ; Temperature is the thermodynamic temperature, and its unit is Kelvin (K). The standard equilibrium constant is a dimensionless index that measures the degree to which a reaction tends toward equilibrium. The dynamic Arrhenius equation is shown in equation (4): （4）； In the above formula, It is the reaction rate constant, which can be used to quantitatively characterize the speed of a chemical reaction; The pre-collision factor is related to the frequency of collisions between reactant molecules and their orientation, and represents the tendency of molecules to undergo effective collisions at higher temperatures. Activation energy, measured in kJ / mol, is the minimum energy required for reactant molecules to transition from their normal state to their activated state. is the molar gas constant, with a value of 8.314 J·mol⁻¹. -1 ·K -1 ; Temperature is the thermodynamic temperature, and its unit is Kelvin (K). The chemical rules include valence bond conservation, energy-driven constraints, and priority of known reaction templates.

7. The method for predicting the degradation of soluble organic molecules based on time-series graphical neural networks according to claim 1, characterized in that, The visualization and identification of key response sites to generate interpretable prediction results includes: The correlation of degradation reactions is learned through the reaction site attention model, and key reaction nodes, more active molecules or reaction pathways are given higher weights. The interpretability of a model can be improved by using the SHAP analysis method. The SHAP analysis method calculates the contribution of features to the model's predictions and assigns an independent contribution value to each input feature, which can clearly show which features play a key role in the model's decision-making.

8. The method for predicting the degradation of soluble organic molecules based on time-series graphical neural networks according to claim 7, characterized in that, A reaction site attention module is embedded in the spatial feature aggregation layer of the time series graph neural network model. Through the reaction site attention module, the time series graph neural network model can automatically allocate weights according to the importance of nodes and edges, increase the attention to key functional groups and chemical bonds in molecules, and give explicit attention to reactive sites, thereby providing more accurate details for the prediction of molecular degradation pathways.

Images