Method for establishing prediction model of volatile organic compounds atmospheric oxidation reaction rate constant

Abstract

A method for establishing a volatile organic matter atmospheric oxidation reaction rate constant prediction model, comprising: obtaining reaction rate constant data of a plurality of volatile organic matters reacting with different oxidants in the atmosphere; converting SMILES codes of the volatile organic matters and SMILES codes of the oxidants into molecular graphs respectively; inputting the molecular graphs of the volatile organic matters and the oxidants into a graph neural network model for each reaction of each volatile organic matter and each oxidant in the atmosphere, and outputting reaction rate constant prediction data; and optimizing network parameters of the graph neural network model according to the reaction rate constant data and the reaction rate constant prediction data, to obtain the volatile organic matter atmospheric oxidation reaction rate constant prediction model. The present application establishes a prediction model of the molecular structure of two substances to the properties of the reaction process, saves the calculation time and resources of the molecular descriptor during model training, reduces the requirement for the basic calculation chemistry, improves the prediction performance, and expands the prediction range.

Description

Technical Field

[0001] This invention relates to the field of chemical environmental health risk assessment technology, and more specifically to a method for establishing and predicting a predictive model for the atmospheric oxidation reaction rate constant of volatile organic compounds. Background Technology

[0002] Volatile organic compounds (VOCs) are important precursors to near-surface ozone formation and PM2.5. Their emissions cause ozone and particulate pollution. In addition, VOCs also lead to environmental problems such as photochemical air pollution and acid deposition. Many VOCs are also toxic and carcinogenic, posing significant health risks. The main pathways for VOC removal from the atmosphere are oxidation reactions with OH radicals, NO3 radicals, and ozone, as well as photolysis reactions under sunlight. Numerous studies have indicated that the reaction rate of VOCs with their corresponding oxidizing agents is crucial to the persistence of these pollutants. The thermodynamic properties of compounds determine their behavior in the environment, and this environmental behavior is closely related to their toxic effects. However, there is a significant gap between the number of compounds whose properties have been determined and the number of compounds for which this information needs to be identified. In recent years, many experimental studies have been conducted to determine the environmental process properties of pollutants. However, experimental methods for determining the environmental process properties of substances are time-consuming and labor-intensive, and are only applicable to a limited number of compounds. With the increasing variety of pollutants, determining their environmental process properties experimentally is becoming increasingly challenging. To improve efficiency, new big data computing methods are urgently needed to conduct prediction work for a large number of pollutants whose properties have not been determined by experimental means.

[0003] With the development of machine learning, it has demonstrated its powerful potential to surpass traditional computational methods. Therefore, in recent decades, the field of chemical information science has begun to use machine learning methods in modeling chemical problems. Many studies have rapidly predicted compounds in the atmosphere whose oxidation reaction rate constants are yet to be determined by establishing quantitative structure-activity relationship (QSAR) models. However, these models are all built using traditional machine learning methods, such as multiple linear regression and support vector machines. In the last decade, the emergence of deep learning methods has shown that there is room for improvement in the predictive performance of traditional machine learning methods; moreover, these models predict the reaction rate constant for single oxides, using relatively small datasets, and lack sufficient generalization guarantees. Summary of the Invention

[0004] In view of this, the main objective of the present invention is to provide a method for establishing and predicting a prediction model for the atmospheric oxidation reaction rate constant of volatile organic compounds, in order to at least partially solve at least one of the aforementioned technical problems.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] As one aspect of the present invention, a method for establishing a prediction model of atmospheric oxidation reaction rate constants of volatile organic compounds is provided, comprising: acquiring reaction rate constant data of various volatile organic compounds reacting with different oxides in the atmosphere; converting the SMILES codes of the volatile organic compounds and the SMILES codes of the oxides into molecular graphs respectively; for each reaction of volatile organic compounds and each oxide in the atmosphere, inputting the molecular graphs of the volatile organic compounds and the oxides into a graph neural network model, and outputting reaction rate constant prediction data; optimizing the network parameters of the graph neural network model based on the reaction rate constant data and the reaction rate constant prediction data to obtain the prediction model of atmospheric oxidation reaction rate constants of volatile organic compounds.

[0007] In another aspect of the present invention, a method for predicting the atmospheric oxidation reaction rate constant of volatile organic compounds is provided, comprising: converting the SMILES codes of the volatile organic compounds to be predicted and the SMILES codes of the oxides to be predicted into molecular diagrams respectively; inputting the molecular diagrams of the volatile organic compounds to be predicted and the oxides to be predicted into a volatile organic compound atmospheric oxidation reaction rate constant prediction model to obtain the reaction rate constant of the volatile organic compounds to be predicted reacting with the oxides to be predicted in the atmosphere; wherein the volatile organic compound atmospheric oxidation reaction rate constant prediction model is obtained by the establishment method described above.

[0008] Based on the above technical solution, the method for establishing and predicting the rate constant of volatile organic compound oxidation reaction of the present invention has at least one or a part of the following beneficial effects:

[0009] (1) This invention uses the molecular graph converted by SMILES encoding as the input of the graph neural network model. It does not require the molecularization parameters calculated by quantum chemical calculation methods in traditional machine learning as molecular descriptors, saving time and computational resources for molecular description calculation and selection, and reducing the requirements for computational chemistry foundation when applying it.

[0010] (2) The graph neural network used in this invention takes molecular graphs of volatile organic compounds and different oxides as input and reaction rate constant data as output. It learns the molecular characteristics of volatile organic compounds and oxides reacting in the atmosphere, thereby establishing a prediction model that can target a variety of oxides, with wider and more convenient applications and prediction performance.

[0011] (3) The atmospheric oxidation reaction rate constant prediction model of volatile organic compounds established in this invention is applicable to the prediction of atmospheric oxidation reaction rate constant of volatile organic compounds on a large scale. The method is simple, fast and efficient. This method has broad application prospects in the fields of chemical environmental health risk assessment and environmental assessment. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the method for establishing a prediction model for the atmospheric oxidation reaction rate constant of volatile organic compounds according to an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of the method for predicting the atmospheric oxidation reaction rate constant of volatile organic compounds based on SMILES encoding in Embodiment 1 of the present invention.

[0014] Figure 3 This is a schematic diagram of the neural network structure in Embodiment 1 of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0016] A literature review of existing models for predicting the atmospheric oxidation rate constant of volatile organic compounds reveals shortcomings in the methods and techniques employed. Predictive models built using traditional machine learning algorithms require manual calculation of molecular parameters as molecular descriptors using quantum chemical methods. This data preparation consumes significant time and computational resources. Furthermore, these models are limited to single oxides, resulting in narrow application scope, limited predictive capabilities, and a need for performance improvement.

[0017] Deep learning is a type of machine learning algorithm that uses hierarchical recombination of features to extract relevant information and learn the pattern properties of data representations. In recent years, deep learning has seen a surge in applications in computer-aided drug design, computational structural biology, quantum chemical computation, and molecular property prediction. In 2013, Lilienfeld et al. were the first to apply multi-task deep neural network (DNN) models to predict properties such as atomization energy, static polarizability, and electron affinity. Graph neural networks (GNNs) are a deep learning-based method that learns information based on graph representations, particularly suitable for protein-protein interaction networks and research areas where molecular properties are expressed using graphs and networks. Deep learning methods require encoding molecular structures mathematically, including graph-based representations and molecular descriptor-based representations, corresponding to graph neural network models and traditional models, respectively. Benchmark tests on multiple datasets show that graph-based models outperform other methods on most datasets, demonstrating a significant advantage. Applying graph neural network-based deep learning methods to predict molecular physicochemical properties and intermolecular interactions holds great potential.

[0018] Therefore, this invention provides a graph neural network model based on the SMILES encoding of different oxide and volatile organic compound molecular structures, which establishes a predictive model from the molecular structure of the two substances to the properties of the reaction process. This saves the computation time and resources for molecular descriptors during model training, reduces the requirements for computational chemistry, improves prediction performance, and expands the prediction range.

[0019] Specifically, according to some embodiments of the present invention, a method for establishing a predictive model for the atmospheric oxidation reaction rate constant of volatile organic compounds is provided. Figure 1 This is a flowchart illustrating the method for establishing a prediction model for the atmospheric oxidation reaction rate constant of volatile organic compounds according to an embodiment of the present invention. Figure 1 As shown, the establishment method includes the following steps S101 to S104.

[0020] In step S101, the reaction rate constant data of various volatile organic compounds reacting with different oxides in the atmosphere are obtained;

[0021] In step S102, the SMILES codes of volatile organic compounds and oxides are converted into molecular diagrams, respectively.

[0022] In step S103, for each volatile organic compound and each oxide that reacts in the atmosphere, the molecular diagrams of the volatile organic compounds and oxides are input into the graph neural network model, which can output predicted data of reaction rate constants.

[0023] In step S104, based on the reaction rate constant data and the reaction rate constant prediction data, the network parameters of the graph neural network model are optimized to obtain the prediction model of the atmospheric oxidation reaction rate constant of volatile organic compounds.

[0024] According to embodiments of the present invention, SMILES encoding is an encoding method used to describe molecular structures. Molecules are represented by simple symbols of chemical structures, such as element symbols, plus and minus signs, parentheses, etc., which can quickly and directly represent the topological structure and connectivity of molecules and can be easily recognized and processed by computer systems. However, features learned directly from SMILES encoding lose a significant amount of information. This invention converts SMILES encoding into molecular graphs, which can better characterize molecular structures, thereby extracting key information more efficiently and reducing unnecessary repetitive steps in computation. The conversion between SMILES encoding and molecular graphs makes it easier to use graph neural network models to learn the molecular features of volatile organic compounds and oxides reacting in the atmosphere, thus saving computation time.

[0025] Specifically, the molecular graphs derived from the SMILES encoding of volatile organic compounds and oxides are used as input to the graph neural network model. The model is trained using known reaction rate constants of volatile organic compounds and different oxides in the atmosphere as labels, and the molecular features of volatile organic compounds and oxides reacting in the atmosphere are extracted. The resulting prediction model is suitable for predicting the reaction rate constants of volatile organic compounds and different oxides in the atmosphere on a large scale, reducing the computational chemistry requirements for the application of the prediction model.

[0026] According to an embodiment of the present invention, the molecular graph is a set of nodes and edges, where nodes correspond to atoms in the molecule and edges correspond to bonds in the molecule. Each node has node features for characterizing atomic features and each edge has edge features for characterizing bond features.

[0027] According to an embodiment of the present invention, step S102 specifically includes sub-steps S1021 to S1022:

[0028] In sub-step S1021, based on the SMILES encoding of volatile organic compounds, node features for characterizing atomic feature information and edge features for characterizing bond feature information are added to the undirected graph to obtain the molecular graph of volatile organic compounds.

[0029] In sub-step S1022, based on the SMILES encoding of the oxide, node features for characterizing atomic feature information and edge features for characterizing bond feature information are added to the undirected graph to obtain the molecular graph of the oxide.

[0030] According to embodiments of the present invention, the atomic feature information includes atomic symbol type, atomic implicit valence, number of free electrons, number of atomic connections, atomic formal charge, atomic hybridization mode, number of hydrogen atoms connected to the atom, atomic chirality and FCFP fingerprint characteristics; the bond feature information includes bond type, whether the bond is conjugated, whether the bond is on a ring, and the bond stereoconfiguration information.

[0031] According to an embodiment of the present invention, in sub-step S1021, the undirected graph G(V, E) is created using DGL (Deep Graph Library), where V represents the set of atoms and E represents the set of bonds.

[0032] Further, sub-step S1021 specifically includes: First, the number N of non-hydrogen atoms in the SMILES encoding of volatile organic compounds can be obtained using the RDKit toolkit, and N nodes are added to the undirected graph G. For each non-hydrogen atom corresponding to a node, atomic feature information is obtained using the RDKit toolkit, and a matrix of size 1×L is generated using one-hot encoding. Second, the node feature matrices corresponding to all nodes are connected to generate an N×L matrix, which is stored as the node features of the molecular graph G, where L is set as needed. Third, the combinations between all non-hydrogen atoms in the SMILES encoding of volatile organic compounds are traversed, and edges are added to the non-hydrogen atoms corresponding to all nodes in the molecular graph G. For the chemical bond represented by each edge, bond feature information is obtained using the RDKit toolkit, and a matrix of size 1×K is generated using one-hot encoding. Finally, the edge feature matrices corresponding to all edges are connected to generate a 2M×K matrix, which is stored as the edge features of the molecular graph G, where M is the number of chemical bonds, and K can be set as needed.

[0033] It is understandable that the operation of converting the SMILES encoding of oxides into a molecular diagram in sub-step S1022 is similar to that in sub-step S1021, and will not be described in detail here.

[0034] To better illustrate the process of converting SMILES codes into molecular diagrams, we will take the volatile organic compound isoprene as an example. The SMILES code for isoprene is C=CC(=C)C.

[0035] For example, the atomic feature information obtained for isoprene includes: atomic symbol type (10 bits), atomic implicit valence (2 bits), number of free electrons (2 bits), number of atomic connections (7 bits), atomic formal charge (3 bits), atomic hybridization type (4 bits), functional fingerprint (FCFP) feature (6 bits), number of hydrogen atoms connected to the atom (5 bits), atomic chirality (2 bits), and whether the atom is chiral (1 bit). Then, the first carbon atom feature in the SMILES code of isoprene is [True, False, False, False, False, False, False, False, False, False, False, False, False, True, True, False, False, False, False, False, False, False, False, False, False, False, False, False, 0, 0, 0, 0, 0, 0, False, False, False, False, True, 0]. One-hot encoding is used to generate a matrix of size 1×L = 1×42 [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0]. Similarly, the node feature matrices corresponding to all nodes are concatenated to generate an N×L = 5×42 matrix, which is stored as the node features of the molecular graph G.

[0036] For example, the bond feature information obtained for isoprene includes: bond type (4 bits), whether the bond is conjugated (1 bit), whether the bond is on a ring (1 bit), and bond stereoconfiguration information (4 bits). Therefore, the feature of the first edge added to the first carbon atom in the SMILES encoding of isoprene (i.e., the C=C bond) is [False, True, False, False, True, False, True, False, False, False]. A matrix of size 1×K = 1×10 is generated using one-hot encoding [0, 1, 0, 0, 1, 0, 1, 0, 0, 0]. Connecting all the edge feature matrices corresponding to the edges generates a 2M×K = 8×10 matrix, which is stored as the edge features of the molecular graph G.

[0037] According to embodiments of the present invention, the molecular graph based on SMILES encoding conversion contains atomic feature information and bond feature information, which can better characterize the molecular structure of volatile organic compounds and oxides, thereby facilitating the subsequent learning of molecular features of volatile organic compounds and oxides reacting in the atmosphere by the graph neural network.

[0038] According to an embodiment of the present invention, the graph neural network model includes a message propagation neural network, an interaction layer, and a prediction layer connected in sequence. Step S103 specifically includes sub-steps S1031 to S1034:

[0039] In sub-step S1031, the molecular graph of volatile organic compounds is input into the first message propagation neural network to extract the first molecular feature tensor of volatile organic compounds.

[0040] In sub-step S1032, the molecular graph of the oxide is input into the second message propagation neural network to extract the second molecular feature tensor of the oxide;

[0041] In sub-step S1033, the first molecular feature tensor and the second molecular feature tensor are input into the interaction layer to extract the interaction features between volatile organic compounds and oxides.

[0042] In sub-step S1034, the interaction features are input into the prediction layer to obtain the reaction rate constant prediction data.

[0043] According to an embodiment of the present invention, sub-step S1031 specifically includes: updating the hidden state of each node feature for each node in the molecular graph of volatile organic compounds to obtain the updated atomic features corresponding to each node; and merging the updated atomic features corresponding to all nodes in the molecular graph of volatile organic compounds to obtain the first molecular feature tensor.

[0044] Specifically, the molecular diagram G(V,E) of volatile organic compounds has node features x v and edge features e vw The hidden state of the node is h v t At t=0, h v t equals x v According to the formula Obtain the updated state of the node at time t+1. According to the formula Obtain the hidden state of the node at time t+1. Where M t U t All are fully connected layers; after T time steps, according to the formula Add hidden states to nodes one by one to obtain F. v The first molecular characteristic tensor A represents the updated atomic features in volatile organic compound molecules, which are then combined to form the volatile organic compound's first molecular characteristic tensor.

[0045] Correspondingly, the operation in sub-step S1032 to extract the second molecular feature tensor of the oxide is similar to that in sub-step S1031. The final obtained second molecular feature tensor B of the oxide will not be described in detail here.

[0046] According to embodiments of the present invention, since the chemical reaction characteristics exhibited by each atom in volatile organic compounds and oxides in chemical reactions are related to its neighboring atoms and / or bonds, the message propagation neural network (MPNN) is a general framework for supervised learning of graph structure data. By extracting the molecular features of volatile organic compounds and oxides through the message propagation neural network, the correlation between each atom and its neighboring atoms and / or bonds can be learned better, thereby improving the accuracy of reaction rate constant prediction.

[0047] According to an embodiment of the present invention, sub-step S1033 specifically includes: dimensionally filling the first molecular feature tensor A and the second molecular feature tensor B to transform them into two feature tensors A1 and B1 with the same shape; splicing the two feature tensors A1 and B1 into tensor Z and inputting it into a fully connected layer to obtain the interaction feature tensor I; multiplying the interaction feature tensor I with the second molecular feature tensor B to obtain the first interaction molecular feature tensor A' used to characterize the action of oxides on volatile organic compounds; and multiplying the interaction feature tensor I with the first molecular feature tensor A to obtain the second interaction molecular feature tensor B' used to characterize the action of volatile organic compounds on oxides.

[0048] According to an embodiment of the present invention, by dimensionally augmenting and filling the shapes of the first molecular feature tensor A and the second molecular feature tensor B to make them consistent before splicing them, the molecular features of volatile organic compounds and oxides can be combined, and subsequent feature extraction can be performed based on the combined features. Further, for ease of understanding, the above-described dimensional augmentation and splicing operations are illustrated with examples below:

[0049] The first molecule characteristic tensor A is represented as ([[1.,1.,1.],

[0050] [0.,0.,0.]]).

[0051] The shape of the first molecule characteristic tensor is (2, 3).

[0052] The second molecular characteristic tensor B is represented as ([[3.,3.,3.,3.],

[0053] [4.,4.,4.,4.],

[0054] [5.,5.,5.,5.]]).

[0055] The shape of the second molecule characteristic tensor B is (3, 4).

[0056] First, the first molecular feature tensor A is dimension-enhanced to become a tensor X1 with shape (1, 2, 3), and the second molecular feature tensor B is dimension-enhanced to become a tensor Y1 with shape (3, 1, 4). Second, tensor X1 copies the contents of the first molecular feature tensor A and changes its first dimension to be consistent with the first dimension of the second molecular feature tensor B, generating a feature tensor A1 with shape (3, 2, 3). Tensor Y1 copies the contents of the second molecular feature tensor B and changes its second dimension to be consistent with the first dimension of the first molecular feature tensor A, generating a feature tensor B1 with shape (3, 2, 4).

[0057] At this point, the two feature tensors A1 and B1 have the same first two dimensions and can be concatenated to obtain tensor Z, where:

[0058] The feature tensor A1 is represented as ([[[1.,1.,1.],

[0059] [0.,0.,0.]],

[0060] [[1.,1.,1.],

[0061] [0.,0.,0.]],

[0062] [[1.,1.,1.],

[0063] [0.,0.,0.]]])

[0064] The feature tensor B1 is represented as ([[[3.,3.,3.,3.],

[0065] [3.,3.,3.,3.]],

[0066] [[4.,4.,4.,4.],

[0067] [4.,4.,4.,4.]],

[0068] [[5.,5.,5.,5.],

[0069] [5.,5.,5.,5.]]])

[0070] Tensor Z is represented as ([[[1.,1.,1.,3.,3.,3.,3.],

[0071] [0.,0.,0.,3.,3.,3.,3.]],

[0072] [[1.,1.,1.,4.,4.,4.,4.],

[0073] [0.,0.,0.,4.,4.,4.,4.]],

[0074] [[1.,1.,1.,5.,5.,5.,5.],

[0075] [0.,0.,0.,5.,5.,5.,5.]]])

[0076] According to an embodiment of the present invention, the prediction layer comprises two pooling layers and multiple fully connected layers, wherein all fully connected layers include a dropout layer with coefficient P to prevent model overfitting. Further, to accurately predict the reaction rate constant, sub-step S1034 specifically includes:

[0077] Connect the first interacting molecule feature tensor A' to the first molecule feature tensor A and connect the second interacting molecule feature tensor B' to the second molecule feature tensor B to obtain two connected molecule feature tensors;

[0078] The two connected molecular feature tensors are input into the pooling layer one-to-one to obtain two pooled molecular feature tensors A” and B”, where the pooling layer can be, for example, a Set2Set pooling layer;

[0079] The two pooled molecular feature tensors A” and B” are connected and then sequentially input into multiple fully connected layers to obtain the reaction rate constant prediction data.

[0080] According to an embodiment of the present invention, further, in sub-step S1034, each node in the current layer of the fully connected layer is connected to all nodes in the previous layer, and the number of nodes in each layer is f. A f B 1. Except for the last layer, the output of each layer uses the ReLU activation function to convert the linear features in the neural network into nonlinear features; among them, the last layer of the fully connected layer is the output layer, which outputs the predicted value ki of the atmospheric chemical reaction rate constant of the corresponding chemical substance and oxide.

[0081] According to an embodiment of the present invention, in step S104, the network parameters of the graph neural network model include model parameters and model hyperparameters. In order to determine the optimal prediction model, the reaction rate constant data are divided into a training set and a validation set. Step S104 specifically includes sub-steps S1041 to S1044:

[0082] In sub-step S1041, based on preset values ​​of a set of model hyperparameter combinations, the graph neural network model is iteratively trained using the reaction rate constant data and the reaction rate constant prediction data in the training set, and the model parameters are optimized to obtain the trained graph neural network model.

[0083] In sub-step S1042, the trained graph neural network model is used to predict the validation set to obtain statistical parameters;

[0084] In sub-step S1043, based on the automatic hyperparameter tuning algorithm, the preset values ​​of the model hyperparameter combination are adjusted, and the training and prediction operations are repeated iteratively until the optimal statistical parameters are determined. The set of model hyperparameter combinations corresponding to the optimal statistical parameters is the optimal model hyperparameter.

[0085] In sub-step S1044, the trained graph neural network model obtained based on the optimal model hyperparameters is determined as the prediction model for the atmospheric oxidation reaction rate constant of volatile organic compounds.

[0086] According to an embodiment of the present invention, the model hyperparameter combination is {α, λ, P, S, T}, where α is the learning rate, used to control the progress of convergence to a local minimum; λ is the weight decay parameter, used to reduce model complexity; P is the random inactivation rate of neurons in the fully connected layer, used to prevent model overfitting; S is the batch data volume; and T is the iteration step size in the message propagation network.

[0087] According to an embodiment of the present invention, in order to improve the model prediction performance, the automatic parameter tuning algorithm is preferably a Bayesian optimization algorithm; the statistical parameters include at least one of Mean Squared Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Edterification (R²). The model parameters corresponding to the lowest RMSE in the validation set results during the E-generation (epaoch) iterations are selected and saved.

[0088] According to an embodiment of the present invention, the reaction rate constant data are further divided into a test set, and the method of establishing the present invention further includes: using the test set to evaluate the performance of the prediction model for the atmospheric oxidation reaction rate constant of volatile organic compounds.

[0089] According to some embodiments of the present invention, a method for predicting the atmospheric oxidation reaction rate constant of volatile organic compounds is also provided, including steps S201 to S202:

[0090] In step S201, the SMILES codes of the volatile organic compounds to be predicted and the SMILES codes of the oxides to be predicted are converted into molecular diagrams, respectively.

[0091] In step S202, the molecular diagrams of the volatile organic compounds to be predicted and the oxides to be predicted are input into the atmospheric oxidation reaction rate constant prediction model of volatile organic compounds to obtain the reaction rate constants of the volatile organic compounds to be predicted and the oxides to be predicted in the atmosphere.

[0092] The prediction model for the atmospheric oxidation reaction rate constant of volatile organic compounds was obtained using the method described above.

[0093] The above prediction method can reliably predict the reaction rate constants of volatile organic compounds and various oxides in atmospheric reactions.

[0094] The following detailed description provides several specific embodiments to illustrate the technical solution of the present invention. It should be noted that the specific embodiments described below are merely examples and are not intended to limit the scope of the invention.

[0095] Example 1

[0096] Please see Figure 2 The method for predicting the atmospheric reaction rate constant of volatile organic compounds based on graph neural networks in this embodiment includes the following steps:

[0097] (1) Acquisition and processing of atmospheric reaction rate constant data for volatile organic compounds

[0098] Download atmospheric reaction rate constant data for organic compounds from an open-source database, including SMILES codes and reaction rate constant data for volatile organic compounds. Data at 298K was selected, resulting in a dataset containing 1586 organic compounds and 4 oxides, with 2802 corresponding organic-oxide reaction rate constants. The logarithms of all data were taken and normalized.

[0099] (2) Chemical structure is converted into molecular diagram

[0100] Use DGL to create an undirected graph G(V, E), where the set of atoms corresponds to V and the set of bonds corresponds to E;

[0101] The number of non-hydrogen atoms N in the molecule is obtained using the RDKit toolkit. N nodes are added to the molecular graph G. For each non-hydrogen atom, the atomic symbol, implicit valence, number of free electrons, number of connections, formal charge, hybridization, number of hydrogen atoms connected to it, and chirality information are obtained using the RDKit toolkit. These are encoded using one-hot encoding to generate a 1×42 matrix. The feature matrix of all non-hydrogen atoms is then connected to generate an N×42 matrix, which is stored as the node feature of the molecular graph G, where N is the number of atoms and the feature dimension is 42.

[0102] Iterate through all combinations between non-hydrogen atoms and add edges between all non-hydrogen atoms in the molecular graph G accordingly. For each chemical bond represented by an edge, use the RDKit toolkit to obtain information on the bond type, whether the bond is conjugated, whether the bond is on a ring, and the chirality of the bond. Encode this information using one-hot encoding to generate a 1×10 matrix. Connect the feature matrices of all bonds to generate a 2M×10 matrix, where M is the number of chemical bonds and the feature dimension is 10. Store this matrix as the edge features of the molecular graph G.

[0103] (3) Training and hyperparameter search of graph neural network models

[0104] The existing data were randomly divided into training, validation, and test sets according to oxide categories in an 8:1:1 ratio. The training set was used to train the graph neural network model; the validation set was used for hyperparameter search; and the test set was used to evaluate the model's predictive performance.

[0105] Please see Figure 3 The graph neural network model constructed in this invention includes a message propagation neural network, an interaction layer, and a prediction layer;

[0106] In the training set, molecular diagrams of volatile organic compounds and oxides are used as model inputs.

[0107] Message Propagation Neural Network (MPNN) is a general framework for supervised learning of graph-structured data, where the molecular graph G(V,E) has node features x v and edge features e vw The node status is h v t At t=0, h v t equals x v According to the formula Obtain the updated state of the node at time t+1. According to the formula Obtain the hidden state of the node at time t+1. With L nodes, all nodes use the Rectified Linear Unit (ReLU) activation function to convert linear features in the neural network into nonlinear features; after T time steps, according to the formula... By adding hidden states to each node, we can eventually obtain F. v The features of each atom in the molecule are combined into a feature tensor that represents the molecular features, and the output feature tensor is of size N×42.

[0108] The interaction layer transforms the first molecular characteristic tensor A of volatile organic compounds and the second molecular characteristic tensor B of oxides into tensors A1 and B1 with the same shape through dimensionality enhancement and filling. Then, A1 and B1 are concatenated to form tensor Z. The interaction characteristic tensor I is obtained through a fully connected layer, and the hyperbolic tangent activation function (Tanh) is used as the activation function. Multiplying the interaction characteristic tensor I by A yields the molecular characteristic tensor B' after the interaction of volatile organic compounds, and multiplying it by B yields the molecular characteristic tensor A' after the interaction of oxides.

[0109] The prediction layer consists of a Set2Set pooling layer and three fully connected layers, all of which contain dropout layers with a coefficient of P.

[0110] Among them, the volatile organic compound molecule feature tensor A' and the oxide molecule feature tensor B' are respectively connected to the original first molecule feature tensor and the second molecule feature tensor A and B, respectively, and the pooled molecular feature tensors A” and B” are obtained through a Set2Set pooling layer, and then the pooled molecular feature tensors A” and B” are connected together through three fully connected layers;

[0111] In a fully connected layer, each node in the current layer is connected to all nodes in the previous layer. The number of nodes in each layer is 256, 128, and 1, respectively. Except for the last layer, the output of each layer uses the ReLU activation function to convert the linear features in the neural network into non-linear features.

[0112] The last layer of the fully connected layer is the output layer, which outputs the predicted value ki of the atmospheric chemical reaction rate constant between the corresponding chemical substance and the oxide.

[0113] During training, the adaptive moment estimation optimizer (Adam) method is used to update the neural network parameters based on gradients, with a learning rate of α and a weight decay parameter of λ. A learning rate scheduler (ReduceLROnPlateau) method is used to modify the learning rate over time, with a patience parameter of 5, meaning the learning rate will not be modified after a cumulative 5 adjustments. Each training iteration consists of a maximum of 200 iterations. The model from each iteration is used for prediction on the validation set, and the MSE, RMSE, MAE, and R² parameters are calculated. Training is stopped early when the MSE on the validation set no longer decreases for a cumulative 15 iterations, and the model with the lowest MSE value on the validation set is saved.

[0114] To improve the model's generalization ability, Bayesian optimization was performed on the hyperparameters used. A preset set of model hyperparameters was set as {α, λ, P, S, T}, where α is the learning rate, used to control the progress of convergence to a local minimum, with a value range of 0.0005–0.005; λ is the weight decay parameter, used to reduce model complexity, with a value range of 0–0.01; P is the random inactivation rate of neurons in the fully connected layer, used to prevent overfitting, with a value range of 0–0.5; S is the batch size of data, set as an integer ranging from 8 to 256; and T is the iteration step size in the message propagation network, set with a value range of {4, 5, 6, 7, 8, 9, 10}. Using Bayesian search, the optimal set of hyperparameters was selected as {α: 0.0043, λ: 6 × 10⁻⁶}. -7 , P: 0.0016926, S: 67, T: 4}

[0115] (5) Comparison of prediction performance with other machine learning methods

[0116] To better demonstrate the superiority of the graph neural network-based method for predicting the atmospheric oxidation reaction rate constant of volatile organic compounds described in this invention, this method is compared with the commonly used existing multiple linear regression (MLR) model. The results are shown in Table 1 below. Compared with traditional methods, the method described in this invention can predict the reaction rate of various oxides. Moreover, when targeting a single oxide, the root mean square error and coefficient of determination evaluation index on the test set are superior to those of traditional methods.

[0117] Table 1

[0118]

[0119] In summary, this invention establishes a predictive model for the atmospheric oxidation reaction rate constant of volatile organic compounds, enabling the prediction of the atmospheric reaction rate constants of volatile organic compounds and various oxides based on SMILES encoding of volatile organic compounds and oxides.

[0120] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for establishing a predictive model for the atmospheric oxidation reaction rate constant of volatile organic compounds, comprising: To obtain the reaction rate constants of various volatile organic compounds reacting with different oxides in the atmosphere; The SMILES codes of the volatile organic compounds and the SMILES codes of the oxides were converted into molecular diagrams, respectively. For each volatile organic compound and each oxide that reacts in the atmosphere, the molecular diagrams of the volatile organic compounds and oxides are input into a graph neural network model, which outputs predicted reaction rate constants. Based on the reaction rate constant data and the reaction rate constant prediction data, the network parameters of the graph neural network model are optimized to obtain the atmospheric oxidation reaction rate constant prediction model for volatile organic compounds. The graph neural network model comprises a message propagation neural network, an interaction layer, and a prediction layer connected in sequence. The molecular diagrams of the volatile organic compounds and the oxides are input into the graph neural network model, and the output reaction rate constant prediction data includes: The molecular graph of the volatile organic compound is input into a first message propagation neural network to extract the first molecular feature tensor of the volatile organic compound; The molecular graph of the oxide is input into a second message propagation neural network to extract the second molecular feature tensor of the oxide; The interaction features between the volatile organic compound (VOC) and the oxide are extracted by inputting the first molecular feature tensor and the second molecular feature tensor into the interaction layer. This includes: dimensionally augmenting and filling the first and second molecular feature tensors to create two feature tensors with the same shape; concatenating the two feature tensors and inputting them into a fully connected layer to obtain an interaction feature tensor; multiplying the interaction feature tensor by the second molecular feature tensor to obtain a first interaction molecular feature tensor characterizing the interaction between the oxide and the VOC; and multiplying the interaction feature tensor by the first molecular feature tensor to obtain a second interaction molecular feature tensor characterizing the interaction between the VOC and the oxide. The interaction features are input into the prediction layer to obtain the predicted reaction rate constant data.

2. The method for establishing according to claim 1, wherein, The molecular graph includes node features for characterizing atomic features and edge features for characterizing bond features; Converting the SMILES codes of the volatile organic compounds and the SMILES codes of the oxides into molecular diagrams includes: Based on the SMILES encoding of the volatile organic compounds, node features for characterizing atomic features and edge features for characterizing bond features are added to the undirected graph to obtain the molecular graph of the volatile organic compounds; and Based on the SMILES encoding of the oxide, node features for characterizing atomic features and edge features for characterizing bond features are added to the undirected graph to obtain the molecular graph of the oxide.

3. The method for establishing according to claim 2, wherein: The atomic feature information includes atomic symbol type, atomic implicit valence, number of free electrons, number of atomic connections, atomic formal charge, atomic hybridization mode, number of hydrogen atoms connected to the atom, atomic chirality, and FCFP fingerprint characteristics. The key feature information includes the type of key, whether the key is conjugated, whether the key is on a ring, and the key's three-dimensional configuration information.

4. The method for establishing according to claim 1, wherein: The molecular graph of the volatile organic compound is input into a first message propagation neural network, and the extraction of the first molecular feature tensor of the volatile organic compound includes: For each node in the molecular graph of the volatile organic compound, the hidden state of each node's features is updated to obtain the updated atomic features corresponding to each node; and The updated atomic features corresponding to all nodes in the molecular graph of the volatile organic compound are merged to obtain the first molecular feature tensor; The molecular graph of the oxide is input into a second message propagation neural network, and the extraction of the second molecular feature tensor of the oxide includes: For each node in the molecular graph of the oxide, the hidden state of each node's features is updated to obtain the updated atomic features corresponding to each node; and The updated atomic features corresponding to all nodes in the molecular graph of the oxide are merged to obtain the second molecular feature tensor.

5. The method for establishing according to claim 1, wherein, The prediction layer comprises two pooling layers and multiple fully connected layers; The interaction features are input into the prediction layer to obtain the predicted reaction rate constant data, including: Connect the first interacting molecular feature tensor to the first molecular feature tensor and connect the second interacting molecular feature tensor to the second molecular feature tensor to obtain two connected molecular feature tensors; The two connected molecular feature tensors are input into the pooling layer in a one-to-one correspondence to obtain two pooled molecular feature tensors. The two pooled molecular feature tensors are concatenated and then sequentially input into the multiple fully connected layers to obtain the reaction rate constant prediction data.

6. The method for establishing according to claim 1, wherein, The network parameters of the graph neural network model include model parameters and model hyperparameters, and the reaction rate constant data are divided into training set and validation set; Based on the reaction rate constant data and the predicted reaction rate constant data in the validation set, the network parameters of the graph neural network model are optimized to obtain the predicted reaction rate constant model for volatile organic compounds at atmosphere, including: Based on a set of preset values ​​for a combination of model hyperparameters, the graph neural network model is iteratively trained using the reaction rate constant data and the reaction rate constant prediction data in the training set, and the model parameters are optimized to obtain the trained graph neural network model. The trained graph neural network model is used to predict the validation set to obtain statistical parameters; Based on the automatic hyperparameter tuning algorithm, the preset values ​​of the model hyperparameter combination are adjusted, and the iterative training and prediction operations are repeated until the optimal statistical parameters are determined. The set of model hyperparameter combinations corresponding to the optimal statistical parameters are the optimal model hyperparameters. The trained graphical neural network model obtained based on the optimal model hyperparameters is determined as the prediction model for the atmospheric oxidation reaction rate constant of volatile organic compounds.

7. The method for establishing according to claim 6, wherein: The automatic parameter tuning algorithm is a Bayesian optimization algorithm; The statistical parameters include at least one of the following: mean square error, root mean square error, mean absolute error, and coefficient of determination.

8. A method for predicting the rate constant of atmospheric oxidation of volatile organic compounds, comprising: The SMILES codes of the volatile organic compounds to be predicted and the SMILES codes of the oxides to be predicted are converted into molecular diagrams, respectively. The molecular diagrams of the volatile organic compounds to be predicted and the oxides to be predicted are input into the atmospheric oxidation reaction rate constant prediction model for volatile organic compounds to obtain the reaction rate constants of the volatile organic compounds to be predicted and the oxides to be predicted in the atmosphere. The predicted model for the atmospheric oxidation rate constant of volatile organic compounds is obtained by the method described in any one of claims 1 to 7.

Images