Methods and related equipment for predicting chemical reaction yields using molecular characterization learning
By combining two-dimensional and three-dimensional molecular information, feature vectors are generated and fused for prediction, solving the problems of high computational resources and lack of structural information in existing technologies, and achieving more accurate prediction of chemical reaction yields.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN INST OF ADVANCED TECH CHINESE ACAD OF SCI
- Filing Date
- 2023-11-01
- Publication Date
- 2026-05-26
AI Technical Summary
Existing models that use molecular characterization learning to predict chemical reaction yields require high computational resources, and two-dimensional molecular diagrams cannot fully cover the chemical structure of molecules, resulting in missing structural information.
A method combining two-dimensional and three-dimensional molecular information is adopted. The two-dimensional and three-dimensional structural information of molecules is processed by graph neural networks and spherical message passing networks respectively to generate feature vectors. Then, the chemical reaction yield is predicted by dimensionality reduction and fusion through a multilayer perceptron module.
It improves the accuracy and computational efficiency of chemical reaction yield prediction, enhances the understanding of molecular geometry by utilizing three-dimensional structural information, and improves the predictive ability of the model.
Smart Images
Figure CN117995303B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, system, terminal, and computer-readable storage medium for predicting chemical reaction yields using molecular characterization learning. Background Technology
[0002] The performance of chemical reactions, such as yield and selectivity, is influenced by dozens of control variables. These variables include not only the molecules directly involved in the reaction but also chemical molecules such as catalysts, solvents, and possible additives. Molecular characterization learning serves as a bridge between computer science and bioinformatics. It encodes molecular information into numerical vectors for use by downstream applications. By performing molecular characterization learning on molecules relevant to reaction conditions, researchers are able to conduct predictive studies on the performance of chemical reactions to better understand and optimize reaction yields and selectivity.
[0003] Currently, models using molecular characterization learning to predict chemical reaction yields mainly utilize molecular fingerprints and two-dimensional molecular graph structures. For example, molecular fingerprints have shown excellent performance in three different chemical reaction prediction tasks. Although combined molecular fingerprints contain more structural information, the feature dimension of a single molecule is as high as 71,374. For complex organic reactions involving multiple components (e.g., four components), the feature dimension can reach tens of thousands. Furthermore, state-of-the-art random forest models have 5,000-10,000 decision trees, both of which make this method computationally very demanding. For example, models using molecular graphs and graph neural networks can directly use molecular structure as input to predict reaction yields and stereoselectivity, but they only utilize two-dimensional molecular structures. However, two-dimensional molecular graphs cannot perfectly cover the chemical structure and coordinates of molecules and lack the ability to learn molecular geometry.
[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention
[0005] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for predicting chemical reaction yields using molecular characterization learning. This aims to address the problems in existing technologies where models using molecular characterization learning to predict chemical reaction yields have high computational resource requirements and where the use of two-dimensional information obtained from molecular diagrams leads to a lack of structural information.
[0006] To achieve the above objectives, the present invention provides a method for predicting chemical reaction yields using molecular characterization learning, the method comprising the following steps:
[0007] Obtain a dataset consisting of multiple reactions, where each reaction contains two-dimensional and three-dimensional structural information of four molecules and the actual yield of each reaction;
[0008] The dataset is input into the two-dimensional information module, and the dataset is processed by the two-dimensional information module to obtain the corresponding two-dimensional feature vector;
[0009] The dataset is input into the three-dimensional information module, which processes the dataset to obtain the resulting three-dimensional feature vector.
[0010] The two-dimensional feature vector and the three-dimensional feature vector are respectively input into the multilayer perceptron module. The multilayer perceptron module performs dimensionality reduction processing on the two-dimensional feature vector and the three-dimensional feature vector to obtain the two-dimensional yield prediction value and the three-dimensional yield prediction value. The two-dimensional feature vector and the three-dimensional feature vector of the four molecules are fused again to predict the yield of the entire reaction.
[0011] Optionally, in the method for predicting chemical reaction yields using molecular characterization learning, the two-dimensional structural information includes atomic attribute vectors and the connectivity relationships between atoms.
[0012] Optionally, in the method for predicting chemical reaction yields using molecular characterization learning, the three-dimensional structural information includes the three-dimensional coordinates and atomic numbers of atoms.
[0013] Optionally, in the method for predicting chemical reaction yields using molecular characterization learning, the framework of the two-dimensional information module is a graph neural network, the hidden layer size of the two-dimensional information module is set to 64, the number of GNN network layers is set to 2, and the activation function is LeakyRelu.
[0014] Optionally, the method for predicting chemical reaction yields using molecular characterization learning, wherein inputting the dataset into a two-dimensional information module and processing the dataset through the two-dimensional information module to obtain a two-dimensional feature vector of the reaction specifically includes:
[0015] The molecular data input to the two-dimensional information module has a size of N*10, where N is the number of atoms in each molecule;
[0016] After the molecular data is input and encoded, the data size becomes N*64. After decoding, the size becomes N*16, and the feature vector of each atom is obtained.
[0017] The feature vector of the entire molecule is obtained by averaging the feature vectors of all atoms.
[0018] The four types of molecules involved in a reaction are sequentially input into the two-dimensional information module to obtain the feature vectors of the four molecules. The feature vectors of the four molecules are concatenated to obtain the two-dimensional feature vector of the reaction, with a dimension of 1*64.
[0019] Optionally, in the method for predicting chemical reaction yields using molecular characterization learning, the framework of the three-dimensional information module is a spherical message passing network.
[0020] Optionally, the method for predicting chemical reaction yields using molecular characterization learning, wherein inputting the dataset into a three-dimensional information module and processing the dataset through the three-dimensional information module to obtain a three-dimensional feature vector of the reaction specifically includes:
[0021] Based on the input three-dimensional data, the atomic coordinates are converted into interatomic spacing, bond angles, and torque.
[0022] Based on the spherical Bessel function, the distance embedding, bond angle embedding and torque embedding are calculated according to the interatomic spacing, bond angle and torque respectively. After linear transformation and nonlinear transformation, the feature vector of each side in the molecule is obtained.
[0023] Aggregate the feature vectors of all edges in the molecule into the feature vectors of all atomic nodes, and aggregate the feature vectors of all atomic nodes into a 1*16 molecular feature vector.
[0024] The reaction conditions and molecular three-dimensional structure information of each reaction are sequentially input into the three-dimensional information module to obtain multiple three-dimensional feature vectors. Each three-dimensional feature vector is then concatenated to obtain the three-dimensional feature vector of the reaction, with a dimension of 1*64.
[0025] Furthermore, to achieve the above objectives, the present invention also provides a system for predicting chemical reaction yields using molecular characterization learning, wherein the system for predicting chemical reaction yields using molecular characterization learning comprises:
[0026] The dataset acquisition module is used to acquire a dataset consisting of multiple reactions, where each reaction contains two-dimensional structure information, three-dimensional structure information, and actual yield of four molecules.
[0027] A two-dimensional information processing module is used to input the dataset into the two-dimensional information module, and after the two-dimensional information module processes the dataset, a corresponding two-dimensional feature vector is obtained;
[0028] A three-dimensional information processing module is used to input the dataset into the three-dimensional information module, and after the three-dimensional information module processes the dataset, a corresponding three-dimensional feature vector is obtained;
[0029] The reaction yield prediction module is used to input the two-dimensional feature vector and the three-dimensional feature vector into the multilayer perceptron module, respectively. The multilayer perceptron module performs dimensionality reduction processing on the two-dimensional feature vector and the three-dimensional feature vector to obtain the two-dimensional yield prediction value and the three-dimensional yield prediction value. The two-dimensional feature vector and the three-dimensional feature vector of the four molecules are fused again to predict the yield of the entire reaction.
[0030] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a program stored in the memory and executable on the processor for predicting chemical reaction yields using molecular characterization learning, wherein when the program for predicting chemical reaction yields using molecular characterization learning is executed by the processor, it implements the steps of the method for predicting chemical reaction yields using molecular characterization learning as described above.
[0031] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a program for predicting chemical reaction yields using molecular characterization learning, and when the program for predicting chemical reaction yields using molecular characterization learning is executed by a processor, it implements the steps of the method for predicting chemical reaction yields using molecular characterization learning as described above.
[0032] In this invention, a dataset consisting of multiple reactions is acquired, where each reaction includes two-dimensional and three-dimensional structural information of four molecules, as well as the actual yield of each reaction. The dataset is input into a two-dimensional information module, which processes the dataset to obtain a two-dimensional feature vector for the reaction. The dataset is then input into a three-dimensional information module, which processes the dataset to obtain a three-dimensional feature vector for the reaction. The two-dimensional and three-dimensional feature vectors are respectively input into a multilayer perceptron module, which performs dimensionality reduction processing on the two-dimensional and three-dimensional feature vectors to obtain two-dimensional and three-dimensional yield predictions. The fused two-dimensional and three-dimensional feature vectors of the four molecules are then fused again to predict the yield of the entire reaction. This invention aims to simultaneously utilize the features of both two-dimensional and three-dimensional molecular structural information to enhance molecular structural information, predict the yield of chemical reactions, and improve molecular characterization learning using three-dimensional molecular information to more accurately predict the yield of chemical reactions. Attached Figure Description
[0033] Figure 1 This is a flowchart of a preferred embodiment of the method for predicting chemical reaction yields using molecular characterization learning according to the present invention;
[0034] Figure 2This is a flowchart illustrating the overall process of predicting chemical reaction yields using both two-dimensional and three-dimensional molecular structural information features in a preferred embodiment of the method for predicting chemical reaction yields using molecular characterization learning according to the present invention.
[0035] Figure 3 This is a schematic diagram illustrating the principle of a preferred embodiment of the system for predicting chemical reaction yields using molecular characterization learning according to the present invention.
[0036] Figure 4 This is a schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. Detailed Implementation
[0037] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0038] Existing research has shown that previous methods for predicting chemical reaction yields have not utilized the geometrical information of molecules related to reaction conditions. However, molecular geometrical information, including parameters such as the spatial arrangement of atoms, bond lengths, bond angles, and torque, has a potentially significant impact on chemical reaction yields. Three-dimensional molecular diagrams provide missing geometrical information by explicitly encoding spatial structure. Utilizing three-dimensional structural information to enhance molecular characterization learning is valuable. Therefore, this invention proposes a method for predicting chemical reaction yields using the geometrical information of molecules related to reaction conditions. Specifically, it involves first cascading reaction condition-related molecules in a chemical reaction, then simultaneously inputting their two-dimensional and three-dimensional information into a model. Different methods are used to learn the two-dimensional and three-dimensional features of the same molecule. Finally, all molecular features in the reaction are aggregated, and cross-features are learned through a multilayer perceptron to predict the yield.
[0039] The method for predicting chemical reaction yields using molecular characterization learning, as described in the preferred embodiment of the present invention, is as follows: Figure 1 and Figure 2 As shown, the method for predicting chemical reaction yields using molecular characterization learning includes the following steps:
[0040] Step S10: Obtain a dataset consisting of multiple reactions, wherein each reaction contains two-dimensional structural information, three-dimensional structural information and actual yield of four molecules.
[0041] Specifically, the dataset used in this invention consists of 3955 reactions, wherein, as Figure 2 As shown, each reaction contains two-dimensional structural information of four molecules (including atomic property vectors, the connection relationships between atoms, i.e.) Figure 2The 2D structural information and the 3D structural information (including the 3D coordinates of atoms, atomic number, etc.) are both included in the structure. Figure 2 The dataset contains 3D structural information and the actual yield of the reaction. Since the data format of this dataset can be directly read by PyTorch and the PyTorch_geometric library, no other processing is required. Simply read the data for each reaction (i.e., the 3955 reactions in the dataset) into the yield prediction model, which includes a two-dimensional information module and a three-dimensional information module.
[0042] Step S20: Input the dataset into the two-dimensional information module, and process the dataset through the two-dimensional information module to obtain the corresponding two-dimensional feature vector.
[0043] Specifically, the framework of the two-dimensional information module of the present invention is a graph neural network, the hidden layer size of the two-dimensional information module is set to 64, the number of GNN network layers is set to 2, and the activation function is LeakyRelu.
[0044] like Figure 2 As shown, the initial molecular data input to the two-dimensional information module has a size of N*10, where N is the number of atoms in each molecule. After input encoding, the molecular data size becomes N*64, and then after decoding, the size becomes N*16, thus obtaining the feature vector of each atom. Then, the feature vectors of all atoms are averaged to obtain the feature vector of the entire molecule. Four types of molecules involved in a reaction are sequentially input into the two-dimensional information module, resulting in feature vectors for the four molecules. These feature vectors are then concatenated (Concat) to obtain the two-dimensional feature vector of the reaction (i.e.,...). Figure 2 Output1 in the dataset has a dimension of 1*64.
[0045] The two-dimensional information module of this invention uses a graph neural network as a framework, which encodes and decodes the two-dimensional information of each molecule, and finally generates the feature vector of the entire reaction.
[0046] Step S30: Input the dataset into the three-dimensional information module, and process the dataset through the three-dimensional information module to obtain the resulting three-dimensional feature vector.
[0047] Specifically, the framework of the three-dimensional information module of the present invention is a spherical message passing network.
[0048] like Figure 2As shown, the atomic coordinates are first converted into interatomic spacing, bond angles, and torque based on the input 3D data. Using a spherical Bessel function, the distance embedding Ψ(d), bond angle embedding Ψ(d, θ), and torque embedding Ψ(d, θ, φ) are calculated based on the interatomic spacing, bond angles, and torque, respectively. After linear and nonlinear transformations, the feature vectors of each edge in the molecule are obtained. Then, the feature vectors of all edges in the molecule are aggregated into feature vectors of all atomic nodes. Finally, the feature vectors of all atomic nodes are aggregated into a 1*16 molecular feature vector. Next, the reaction conditions and 3D molecular structure information of each reaction are sequentially input into the 3D information module to obtain multiple 3D feature vectors. These 3D feature vectors are then concatenated to obtain the 3D feature vector of the reaction (i.e.,...). Figure 2 Output2 in the dataset has a dimension of 1*64.
[0049] The three-dimensional information module of this invention uses a spherical message passing network as a framework. This module transforms atomic coordinates into atomic spacing, bond angles, and torques. After calculation, distance embedding, bond angle embedding, and torque embedding are obtained, generating feature vectors for each edge, and finally obtaining the feature vectors of the entire reaction.
[0050] Step S40: Input the two-dimensional feature vector and the three-dimensional feature vector into the multilayer perceptron module respectively. The multilayer perceptron module performs dimensionality reduction processing on the two-dimensional feature vector and the three-dimensional feature vector to obtain the two-dimensional yield prediction value and the three-dimensional yield prediction value. Then, the two-dimensional feature vector and the three-dimensional feature vector of the four molecules are fused again to predict the yield of the entire reaction.
[0051] Specifically, such as Figure 2 As shown, response feature vectors of different dimensions (i.e., the two-dimensional feature vector and the three-dimensional feature vector) are respectively input into the multilayer perceptron module (i.e., Figure 2 The MLP (Multilayer Perceptron) is used to obtain richer feature representations, resulting in two-dimensional and three-dimensional yield predictions (e.g., reaction yield predictions are obtained when the features are reduced to 1 in the input MLP layer). Finally, the fused three-dimensional and two-dimensional features of the four molecules are fused again to predict the overall reaction yield.
[0052] This invention inputs molecular information of different dimensions into different constructed modules to obtain feature vectors of their respective dimensions. The feature vectors of each reaction are then concatenated to form a single reaction feature. Finally, the reaction features of the two dimensions are fused and used for regression prediction. This invention relates to the fusion of two-dimensional and three-dimensional information. A multilayer perceptron module is used to learn multidimensional feature representations and predict yields. Different-dimensional reaction feature vectors are input into the multilayer perceptron module to obtain richer feature representations. This invention combines two-dimensional and three-dimensional molecular information, enabling more comprehensive prediction of chemical reaction yields. Traditional methods typically use only two-dimensional information, while the method of this invention is better able to capture the influence of molecular geometry on reaction performance. This invention utilizes three-dimensional molecular structure information, increasing the understanding of reaction mechanisms and the accuracy of yield prediction.
[0053] Furthermore, multiple different prediction models, such as deep neural networks, support vector machines, and decision trees, can be used, and their predictions can be fused together to improve the accuracy and stability of yield predictions. Introducing more information into the model, such as reaction conditions like temperature, pressure, and catalyst type, can also help the model better understand the reaction mechanism and factors influencing yield.
[0054] Furthermore, such as Figure 3 As shown, based on the above-described method for predicting chemical reaction yields using molecular characterization learning, this invention also provides a system for predicting chemical reaction yields using molecular characterization learning, wherein the system for predicting chemical reaction yields using molecular characterization learning includes:
[0055] The dataset acquisition module 51 is used to acquire a dataset consisting of multiple reactions, wherein each reaction contains two-dimensional structure information, three-dimensional structure information and actual yield of four molecules.
[0056] The two-dimensional information processing module 52 is used to input the dataset into the two-dimensional information module, and process the dataset through the two-dimensional information module to obtain the corresponding two-dimensional feature vector;
[0057] The three-dimensional information processing module 53 is used to input the dataset into the three-dimensional information module, and process the dataset through the three-dimensional information module to obtain the corresponding three-dimensional feature vector;
[0058] The reaction yield prediction module 54 is used to input the two-dimensional feature vector and the three-dimensional feature vector into the multilayer perceptron module respectively. The multilayer perceptron module performs dimensionality reduction processing on the two-dimensional feature vector and the three-dimensional feature vector to obtain the two-dimensional yield prediction value and the three-dimensional yield prediction value. The two-dimensional feature vector and the three-dimensional feature vector of the four molecules are fused again to predict the yield of the entire reaction.
[0059] Furthermore, such as Figure 4 As shown, based on the above-described method and system for predicting chemical reaction yields using molecular characterization learning, the present invention also provides a terminal, which includes a processor 10, a memory 20, and a display 30. Figure 4 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.
[0060] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a program 40 for predicting chemical reaction yields using molecular characterization learning. This program 40 can be executed by the processor 10 to implement the method for predicting chemical reaction yields using molecular characterization learning as described in this application.
[0061] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the method of predicting chemical reaction yields using molecular characterization learning.
[0062] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.
[0063] In one embodiment, when processor 10 executes program 40 in memory 20 that uses molecular characterization learning to predict chemical reaction yields, the following steps are performed:
[0064] Obtain a dataset consisting of multiple reactions, where each reaction contains two-dimensional and three-dimensional structural information of four molecules and the actual yield of each reaction;
[0065] The dataset is input into the two-dimensional information module, and the dataset is processed by the two-dimensional information module to obtain the corresponding two-dimensional feature vector;
[0066] The dataset is input into the three-dimensional information module, which processes the dataset to obtain the resulting three-dimensional feature vector.
[0067] The two-dimensional feature vector and the three-dimensional feature vector are respectively input into the multilayer perceptron module. The multilayer perceptron module performs dimensionality reduction processing on the two-dimensional feature vector and the three-dimensional feature vector to obtain the two-dimensional yield prediction value and the three-dimensional yield prediction value. The two-dimensional feature vector and the three-dimensional feature vector of the four molecules are fused again to predict the yield of the entire reaction.
[0068] The two-dimensional structural information includes atomic attribute vectors and the connection relationships between atoms.
[0069] The three-dimensional structural information includes the three-dimensional coordinates and atomic numbers of the atoms.
[0070] The framework of the two-dimensional information module is a graph neural network. The hidden layer size of the two-dimensional information module is set to 64, the number of GNN network layers is set to 2, and the activation function is LeakyRelu.
[0071] Specifically, the step of inputting the dataset into a two-dimensional information module and processing the dataset through the two-dimensional information module to obtain the resulting two-dimensional feature vector includes:
[0072] The molecular data input to the two-dimensional information module has a size of N*10, where N is the number of atoms in each molecule;
[0073] After the molecular data is input and encoded, the data size becomes N*64. After decoding, the size becomes N*16, and the feature vector of each atom is obtained.
[0074] The feature vector of the entire molecule is obtained by averaging the feature vectors of all atoms.
[0075] The four types of molecules involved in a reaction are sequentially input into the two-dimensional information module to obtain the feature vectors of the four molecules. The feature vectors of the four molecules are concatenated to obtain the two-dimensional feature vector of the reaction, with a dimension of 1*64.
[0076] The framework of the three-dimensional information module is a spherical message passing network.
[0077] Specifically, the step of inputting the dataset into the 3D information module and processing the dataset to obtain the resulting 3D feature vector includes:
[0078] Based on the input three-dimensional data, the atomic coordinates are converted into interatomic spacing, bond angles, and torque.
[0079] Based on the spherical Bessel function, the distance embedding, bond angle embedding and torque embedding are calculated according to the interatomic spacing, bond angle and torque respectively. After linear transformation and nonlinear transformation, the feature vector of each side in the molecule is obtained.
[0080] Aggregate the feature vectors of all edges in the molecule into the feature vectors of all atomic nodes, and aggregate the feature vectors of all atomic nodes into a 1*16 molecular feature vector.
[0081] The reaction conditions and molecular three-dimensional structure information of each reaction are sequentially input into the three-dimensional information module to obtain multiple three-dimensional feature vectors. Each three-dimensional feature vector is then concatenated to obtain the three-dimensional feature vector of the reaction, with a dimension of 1*64.
[0082] The present invention also provides a computer-readable storage medium storing a program for predicting chemical reaction yields using molecular characterization learning, wherein the program for predicting chemical reaction yields using molecular characterization learning, when executed by a processor, implements the steps of the method for predicting chemical reaction yields using molecular characterization learning as described above.
[0083] In summary, this invention provides a method and related equipment for predicting chemical reaction yields using molecular characterization learning. The method includes: acquiring a dataset consisting of multiple reactions, wherein each reaction contains two-dimensional structural information and three-dimensional structural information of four molecules, as well as the actual yield of each reaction; inputting the dataset into a two-dimensional information module, which processes the dataset to obtain a two-dimensional feature vector of the reaction; inputting the dataset into a three-dimensional information module, which processes the dataset to obtain a three-dimensional feature vector of the reaction; inputting the two-dimensional and three-dimensional feature vectors into a multilayer perceptron module, which performs dimensionality reduction on the two-dimensional and three-dimensional feature vectors to obtain two-dimensional and three-dimensional yield prediction values; and fusing the fused two-dimensional and three-dimensional feature vectors of the four molecules again to predict the yield of the entire reaction. This invention aims to simultaneously utilize the features of two-dimensional and three-dimensional molecular structural information to enhance the structural information of molecules, predict the yield of chemical reactions, and improve molecular characterization learning using three-dimensional molecular information to more accurately predict the yield of chemical reactions.
[0084] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.
[0085] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.
[0086] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.
Claims
1. A method of predicting chemical reaction yield using molecular characterization learning, characterized by, The method for predicting chemical reaction yields using molecular characterization learning includes: Obtain a dataset consisting of multiple reactions, where each reaction contains two-dimensional and three-dimensional structural information of four molecules and the actual yield of each reaction; The dataset is input into the two-dimensional information module, which processes the dataset. The two-dimensional information of the four types of molecules contained in a reaction is sequentially input into the two-dimensional information module to obtain the feature vectors of the four molecules. The feature vectors of the four molecules are concatenated to obtain the two-dimensional feature vector of the reaction, with a dimension of 1*64. The framework of the two-dimensional information module is a graph neural network. The dataset is input into the 3D information module, which processes the dataset and sequentially inputs the 3D molecular structure information of each reaction into the 3D information module to obtain multiple 3D feature vectors. Each 3D feature vector is then concatenated to obtain the 3D feature vector of the reaction, with a dimension of 1*64. The framework of the 3D information module is a spherical message passing network. The two-dimensional feature vector and the three-dimensional feature vector are respectively input into the multilayer perceptron module. The multilayer perceptron module performs dimensionality reduction processing on the two-dimensional feature vector and the three-dimensional feature vector to obtain the two-dimensional yield prediction value and the three-dimensional yield prediction value. The two-dimensional feature vector and the three-dimensional feature vector of the four molecules are fused again to predict the yield of the entire reaction.
2. The method of predicting a yield of a chemical reaction using molecular signature learning of claim 1, wherein, The two-dimensional structural information includes atomic attribute vectors and the connection relationships between atoms.
3. The method for predicting chemical reaction yields using molecular characterization learning according to claim 1, characterized in that, The three-dimensional structural information includes the three-dimensional coordinates and atomic numbers of the atoms.
4. The method for predicting chemical reaction yields using molecular characterization learning according to claim 1, characterized in that, The hidden layer size of the two-dimensional information module is set to 64, the number of GNN network layers is set to 2, and the activation function is LeakyRelu.
5. The method for predicting chemical reaction yields using molecular characterization learning according to claim 1, characterized in that, The process of inputting the dataset into the two-dimensional information module and processing the dataset by the two-dimensional information module specifically includes: The molecular data input to the two-dimensional information module has a size of N*10, where N is the number of atoms in each molecule; After the molecular data is input and encoded, the data size becomes N*64. After decoding, the size becomes N*16, and the feature vector of each atom is obtained. The eigenvectors of all atoms are averaged to obtain the eigenvectors of the entire molecule.
6. The method for predicting chemical reaction yields using molecular characterization learning according to claim 1, characterized in that, The step of inputting the dataset into the 3D information module and processing the dataset by the 3D information module specifically includes: Based on the input three-dimensional data, the atomic coordinates are converted into interatomic spacing, bond angles, and torque. Based on the spherical Bessel function, the distance embedding, bond angle embedding and torque embedding are calculated according to the interatomic spacing, bond angle and torque respectively. After linear transformation and nonlinear transformation, the feature vector of each side in the molecule is obtained. Aggregate the feature vectors of all edges in the molecule into feature vectors of all atomic nodes, and aggregate the feature vectors of all atomic nodes into a 1*16 molecular feature vector.
7. A system for predicting chemical reaction yields using molecular characterization learning, characterized in that, The system for predicting chemical reaction yields using molecular characterization learning is used to implement the method for predicting chemical reaction yields using molecular characterization learning according to any one of claims 1-6, wherein the system for predicting chemical reaction yields using molecular characterization learning comprises: The dataset acquisition module is used to acquire a dataset consisting of multiple reactions, where each reaction contains two-dimensional structure information, three-dimensional structure information, and actual yield of four molecules. A two-dimensional information processing module is used to input the dataset into the two-dimensional information module, and after the two-dimensional information module processes the dataset, a corresponding two-dimensional feature vector is obtained; A three-dimensional information processing module is used to input the dataset into the three-dimensional information module, and after the three-dimensional information module processes the dataset, a corresponding three-dimensional feature vector is obtained; The reaction yield prediction module is used to input the two-dimensional feature vector and the three-dimensional feature vector into the multilayer perceptron module, respectively. The multilayer perceptron module performs dimensionality reduction processing on the two-dimensional feature vector and the three-dimensional feature vector to obtain the two-dimensional yield prediction value and the three-dimensional yield prediction value. The two-dimensional feature vector and the three-dimensional feature vector of the four molecules are fused again to predict the yield of the entire reaction.
8. A terminal, characterized in that, The terminal includes: a memory, a processor, and a program stored in the memory and executable on the processor for predicting chemical reaction yields using molecular characterization learning. When the program for predicting chemical reaction yields using molecular characterization learning is executed by the processor, it implements the steps of the method for predicting chemical reaction yields using molecular characterization learning as described in any one of claims 1-6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for predicting chemical reaction yields using molecular characterization learning, which, when executed by a processor, implements the steps of the method for predicting chemical reaction yields using molecular characterization learning as described in any one of claims 1-6.