A processing method and device for atomic reaction site prediction

Abstract

Embodiments of the present application relate to a kind of atomic reaction site prediction processing method and device, the method comprises: constructing reaction site prediction model;Through big data acquisition, first original data set is constructed;And based on first original data set, model training data set construction obtains first data set;And based on first data set, reaction site prediction model is trained;After training, the 3D molecular structure of any compound molecule input by client is input into reaction site prediction model to carry out atomic electrophilic / nucleophilic reaction site prediction processing and obtain corresponding current reaction site prediction tensor, and according to current 3D molecular structure and current reaction site prediction tensor, electrophilic / nucleophilic reaction site atomic information is summarized to obtain corresponding summary information to the current client feedback.Based on the reaction site prediction model provided in the application, the prediction efficiency can be improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a processing method and apparatus for predicting atomic-level reaction sites. Background Technology

[0002] The Fukui function is an important concept in Density Functional Theory (DFT). By calculating and analyzing the Fukui function, atomic-level electrophilic / nucleophilic reaction sites can be predicted in compound molecules. Currently, the traditional prediction method based on the Fukui function generally consists of three steps: 1) First, using the Fukui function calculation interface provided by quantum chemical computing tools (such as Gaussian, ORCA, CP2K, etc.), atomic-level Fukui function calculations are performed based on the 3D molecular structure of the compound molecule to obtain the nucleophilic and electrophilic reaction function values ​​of each atom. 2) The second step involves using quantum chemical calculation tools to estimate the three types of reaction sites (electrophilic, nucleophilic, and non-nucleophilic) of each atom on the compound molecule based on the three-dimensional spatial configuration of the 3D molecular structure and the calculation results of the Fukui function; 3) The third step involves verifying the correctness of the reaction site estimation results output by the quantum chemical calculation tools through chemical experiments, and outputting the tool's estimation results as the final prediction results when the verification is successful. Practice has shown that the second and third steps are relatively time-consuming and have high computational and experimental costs. In other words, if traditional prediction methods are used, problems such as long prediction cycles, high prediction costs, and low prediction efficiency will inevitably arise. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by providing a processing method, apparatus, electronic device, and computer-readable storage medium for atomic-level reaction site prediction. This invention first constructs a processing model for atomic-level electrophilic / nucleophilic reaction site prediction of 3D molecular structures, denoted as the corresponding reaction site prediction model. Then, a first raw dataset containing a large number of 3D molecular structures is constructed through big data acquisition. Atomic-level electrophilic / nucleophilic reaction sites are predicted for each 3D molecular structure in the first raw dataset using traditional prediction methods. Based on the prediction results, label data is constructed to obtain a model training dataset, i.e., the first dataset. The reaction site prediction model is then trained based on the first dataset. After model training, end-to-end atomic-level electrophilic / nucleophilic reaction site prediction is performed based on the reaction site prediction model for any compound's 3D molecular structure. The reaction site prediction model provided by this invention integrates the first step of the traditional prediction method but omits the second and third steps. Based on the reaction site prediction model provided by this invention, the prediction cycle can be shortened, the prediction cost reduced, and the prediction efficiency improved.

[0004] To achieve the above objectives, a first aspect of the present invention provides a method for predicting atomic-level reaction sites, the method comprising:

[0005] A processing model for predicting atomic-level electrophilic / nucleophilic reaction sites in 3D molecular structures is constructed, denoted as the corresponding reaction site prediction model.

[0006] A first original dataset is constructed by collecting large amounts of data on the 3D molecular structure; and a corresponding first dataset is constructed by using the first original dataset and preset quantum chemical calculation tools and experimental methods to train the model; and the reaction site prediction model is trained based on the first dataset.

[0007] After the model training is completed, the 3D molecular structure of any compound molecule input from the client is received as the corresponding current 3D molecular structure; the current 3D molecular structure is input into the reaction site prediction model to perform atomic-level electrophilic / nucleophilic reaction site prediction processing to obtain the corresponding current reaction site prediction tensor; and the atomic information of electrophilic / nucleophilic reaction sites is summarized based on the current 3D molecular structure and the current reaction site prediction tensor to obtain the corresponding summary information and feed it back to the current client.

[0008] Preferably, the quantum chemical calculation tools include at least Gaussian software, ORCA software, and CP2K software.

[0009] Preferably, the reaction site prediction model is used to perform atomic-level electrophilic / nucleophilic reaction site prediction processing based on the 3D molecular structure input to the model and output the corresponding reaction site prediction tensor;

[0010] The model input terminal of the reaction site prediction model is used to receive the 3D molecular structure input by the model, and the model output terminal is used to output the corresponding reaction site prediction tensor.

[0011] The 3D molecular structure is composed of a first set of atoms and a first set of chemical bonds;

[0012] The first set of atoms includes multiple first atoms; the atomic attributes of the first atom include at least a first atom identifier, a first atom type, a first atom coordinates, a first atom charge, and a first atom chiral attribute set; the first atom identifier is the unique identifier of the current first atom; the first atom type is the chemical element type of the current first atom; the first atom charge is the atomic charge number of the current first atom; the first atom chiral attribute set is composed of multiple first chiral attributes; each first chiral attribute is composed of a corresponding attribute type and attribute value; if the first atom is a chiral center atom, then all attribute values ​​in the corresponding first atom chiral attribute set are not empty; if the first atom is not a chiral center atom, then all attribute values ​​in the corresponding first atom chiral attribute set are empty;

[0013] The first set of chemical bonds includes multiple first chemical bonds; the chemical bond attributes of the first chemical bond include at least a first chemical bond identifier, a first chemical bond type, and a first atom identifier group; the first chemical bond identifier is a unique identifier for the current first chemical bond; the first chemical bond type includes multiple chemical bond types; the first atom identifier group is composed of the first atom identifiers of the two first atoms connected by the current first chemical bond;

[0014] The reaction site prediction tensor comprises multiple reaction site prediction vectors; each reaction site prediction vector corresponds one-to-one with the first atom; each reaction site prediction vector consists of three prediction probabilities that sum to 1, namely: electrophilic reaction site prediction probability, nucleophilic reaction site prediction probability, and non-reaction site prediction probability.

[0015] The reaction site prediction model includes a Fukui function calculation module, a molecular graph construction module, a molecular graph feature extraction network, a fully connected network, and a Softmax activation layer;

[0016] The input terminal of the Fukui function calculation module is connected to the input terminal of the model, and the output terminal is connected to the second input terminal of the molecular graph construction module; the first input terminal of the molecular graph construction module is connected to the input terminal of the model, and the output terminal is connected to the input terminal of the molecular graph feature extraction network; the output terminal of the molecular graph feature extraction network is connected to the input terminal of the fully connected network; the output terminal of the fully connected network is connected to the Softmax activation layer; the Softmax activation layer is connected to the output terminal of the model.

[0017] The Fukui function calculation module is implemented based on the Fukui function calculation interface provided by the quantum chemical calculation tool. The Fukui function calculation module performs atomic-level Fukui function calculations based on the 3D molecular structure input from the model input terminal to obtain a corresponding first Fukui function value tensor, which is then sent to the molecular graph construction module. The first Fukui function value tensor consists of multiple first Fukui function value vectors. Each first Fukui function value vector corresponds one-to-one with the first atom of the 3D molecular structure. Each first Fukui function value vector includes nucleophilic reaction function values. and electrophilic reaction function value

[0018] The molecular graph construction module is used to construct a corresponding molecular graph node set based on the first atom set and the first Fukui function value tensor of the 3D molecular structure input from the model input terminal, and to construct a corresponding molecular graph node edge set based on the first chemical bond set of the 3D molecular structure input from the model input terminal, and to send the first molecular graph composed of the first node set and the first node edge set to the molecular graph feature extraction network. The first molecular graph includes the first node set and the first node edge set. The first node set includes multiple first nodes. Each first node corresponds to a first atom. The node attributes of the first node are composed of the corresponding first atom identifier, first atom type, first atom coordinates, first atom charge, first atom chiral attribute set, and first Fukui function value vector. The first node edge set includes multiple first node edges. The edge attributes of the first node edges are composed of the corresponding first chemical bond identifier, first chemical bond type, and first atom identifier group.

[0019] The molecular graph feature extraction network is implemented by sequentially connecting multiple GINEConv layers; the molecular graph feature extraction network is used to perform node feature extraction processing with edge feature fusion mechanism on the first molecular graph to obtain the corresponding first node feature tensor and send it to the fully connected network; the first node feature tensor includes multiple first node feature vectors; the first node feature vector corresponds one-to-one with the first node;

[0020] The fully connected network is implemented by one or more fully connected layers connected sequentially; the fully connected network is used to perform a linear transformation on the feature tensor of the first node from the feature tensor to the prediction tensor to obtain the corresponding prediction tensor of the first node, which is then sent to the Softmax activation layer; the prediction tensor of the first node includes multiple prediction vectors of the first node; the prediction vector of the first node corresponds one-to-one with the feature vector of the first node; the prediction vector of the first node consists of three types of prediction values, namely, prediction values ​​of electrophilic reaction sites, prediction values ​​of nucleophilic reaction sites, and prediction values ​​of non-reactive sites.

[0021] The Softmax activation layer consists of multiple parallel Softmax activation functions; each Softmax activation function corresponds one-to-one with the first node prediction vector; the Softmax activation layer is used to obtain the corresponding reaction site prediction vector by performing prediction probability normalization processing on each Softmax activation function according to the corresponding first node prediction vector; and to form the corresponding reaction site prediction tensor by all the obtained reaction site prediction vectors; and to send the reaction site prediction tensor to the model output terminal;

[0022] The Softmax activation layer is specifically used to obtain the corresponding reaction site prediction vector by normalizing the prediction probability of each reaction site prediction vector based on the corresponding first node prediction vector using each Softmax activation function. Specifically, it is set that the predicted values ​​of the electrophilic reaction site, nucleophilic reaction site, and non-reaction site for each first node prediction vector are z1, z2, and z3, respectively. Furthermore, it is set that the predicted probabilities of the electrophilic reaction site, nucleophilic reaction site, and non-reaction site for each reaction site prediction vector corresponding to each first node prediction vector are p1, p2, and p3, respectively. The calculation formulas for p1, p2, p3 and z1, z2, z3 are as follows:

[0023] Preferably, the first original dataset includes multiple first original records; each first original record includes a first original molecular formula and a first original molecular structure; each first original record corresponds to a compound molecule; the first original molecular formula is the chemical formula of the current compound molecule; the first original molecular structure is the 3D molecular structure of the current compound molecule;

[0024] The first dataset includes multiple first data records; each first data record includes a first training structure and a first label tensor; the first training structure is a 3D molecular structure; the first label tensor includes multiple first label vectors; each first label vector consists of three label probabilities that sum to 1, namely: electrophilic reaction site label probability, nucleophilic reaction site label probability, and non-reactive site label probability; only one of the three label probabilities in the first label vector is 1, and the other two are 0; each first label vector corresponds one-to-one with the first atom in the first atom set of the first training structure.

[0025] Preferably, the step of constructing the first original dataset through big data collection of the 3D molecular structure specifically includes:

[0026] Data on the 3D molecular structures of various compounds are collected through multiple publicly available data channels. The chemical formula and 3D molecular structure of each compound molecule are used as the corresponding first original molecular formula and first original molecular structure to form the corresponding first original record. All the obtained first original records form the corresponding first original dataset. The multiple publicly available data channels include at least various publicly available datasets that publish molecular 3D structures, journals / magazines / scientific and technological literature / papers.

[0027] Preferably, the step of constructing the corresponding first dataset based on the first original dataset and preset quantum chemical calculation tools and experimental methods for model training specifically includes:

[0028] Each of the first original records in the first original dataset is taken as the corresponding current original record; the first original molecular formula and the first original molecular structure of the current original record are taken as the corresponding current molecular formula and current molecular structure; and based on the quantum chemical calculation tool, a corresponding three-dimensional molecular model is constructed according to the current molecular formula and current molecular structure to obtain the corresponding current molecular model; and based on the Fukui function calculation interface provided by the quantum chemical calculation tool, atomic-level Fukui function calculations are performed according to the current molecular model, and the calculation results are taken as the corresponding current Fukui function value tensor; and based on the current Fukui function value tensor and the three-dimensional spatial configuration of the current molecular model, the three types of reaction site types of each atom on the current molecular structure are estimated to obtain the corresponding set of atomic reaction site types; and through chemical... The experimental method verifies the correctness of the estimated set of atomic reaction site types. Upon successful verification, the current molecular structure is used as a corresponding first training structure. A label tensor transformation is performed based on the set of atomic reaction site types corresponding to the current molecular structure to obtain the corresponding first label tensor. The first training structure corresponding to the current molecular structure and the first label vector together form a corresponding first data record. All obtained first data records constitute the corresponding first dataset. The set of atomic reaction site types includes multiple atomic reaction site types. Each atomic reaction site type corresponds one-to-one with the first atom in the first atom set of the corresponding 3D molecular structure. The atomic reaction site types include electrophilic reaction sites, nucleophilic reaction sites, and non-reactive sites.

[0029] Specifically, the step of performing a label tensor transformation based on the set of atomic reaction site types corresponding to the current molecular structure to obtain the corresponding first label tensor involves: taking each atomic reaction site type in the current set of atomic reaction site types as the corresponding current reaction site type; generating a corresponding first label vector for the current reaction site type as the corresponding current label vector; taking the electrophilic reaction site label probability, the nucleophilic reaction site label probability, or the non-reaction site label probability corresponding to the current reaction site type in the current label vector as the corresponding current label probability, setting the current label probability to 1, and setting the other two label probabilities to 0; and forming a corresponding first label tensor by combining all the first label vectors corresponding to the current set of atomic reaction site types.

[0030] Preferably, training the reaction site prediction model based on the first dataset specifically includes:

[0031] Step 71: Based on a preset first segmentation ratio, the first dataset is divided into two sub-datasets, denoted as the first training set and the first evaluation set.

[0032] Wherein, both the first training set and the first evaluation set include multiple first data records; the ratio of the total number of records in the first training set to the total number of records in the first evaluation set satisfies the first segmentation ratio;

[0033] Step 72: Extract the first data record of the first training set as the corresponding current training record;

[0034] Step 73: Input the 3D molecular structure, which is the first training structure recorded in the current training record, into the reaction site prediction model to perform atomic-level electrophilic / nucleophilic reaction site prediction processing, and use the reaction site prediction tensor output by the model as the corresponding first prediction tensor; and substitute the first prediction tensor and the first label tensor recorded in the current training record into the preset first model loss function to calculate the corresponding first loss value.

[0035] Wherein, the loss function of the first model includes at least the cross-entropy loss function;

[0036] Step 74: Identify whether the first loss value meets the preset first loss value range; if the first loss value meets the first loss value range, identify whether the current training record is the last first data record of the first training set; if so, proceed to step 75; otherwise, extract the next first data record of the first training set as the new current training record and return to step 73; if the first loss value does not meet the first loss value range, perform a round of modulation on the model parameters of the molecular graph feature extraction network and the fully connected network of the reaction site prediction model based on the preset first model optimizer, and return to step 73 when the modulation ends.

[0037] The first model optimizer includes at least an SGD optimizer and an Adam optimizer;

[0038] Step 75: Perform a round of traversal on all the first data records in the first evaluation set; during this round of traversal, take the currently traversed first data record as the corresponding current evaluation record; and take the first training structure of the current evaluation record as the 3D molecular structure input to the reaction site prediction model for atomic-level electrophilic / nucleophilic reaction site prediction processing, and take the reaction site prediction tensor output by the model as the corresponding second prediction tensor; and form a corresponding prediction-label data pair by the second prediction tensor and the first label tensor of the current evaluation record; and at the end of this round of traversal, calculate the accuracy, precision, recall, and F1 score based on all the obtained prediction-label data pairs to obtain the corresponding first accuracy, first precision, first recall, and first F1 score;

[0039] Step 76: Identify the first accuracy, first precision, first recall, and first F1 score; if the first accuracy does not meet the preset first accuracy range, or the first precision does not meet the preset first precision range, or the first recall does not meet the preset first recall range, or the first F1 score does not meet the preset first F1 score range, then return to step 72 to continue training; if the first accuracy meets the first accuracy range, and the first precision meets the first precision range, and the first recall meets the first recall range, and the first F1 score meets the first F1 score range, then stop training and confirm that the model training has ended.

[0040] Preferably, the step of summarizing the electrophilic / nucleophilic reaction site atomic information based on the current 3D molecular structure and the current reaction site prediction tensor to obtain the corresponding summary information and feeding it back to the current client specifically includes:

[0041] The first set of atoms and the first set of chemical bonds corresponding to the current 3D molecular structure are taken as the corresponding current set of atoms and the current set of chemical bonds. A traversal is performed on all the first atoms in the current set of atoms. During this traversal, the first atom currently traversed is taken as the corresponding current atom. A reaction site type attribute is added to the atomic properties of the current atom as the corresponding first atom reaction site type. The reaction site prediction vector corresponding to the current atom in the current reaction site prediction tensor is taken as the corresponding current prediction vector. The electrophilic reaction site, nucleophilic reaction site, or non-reactive site with the highest probability in the current prediction vector is taken as the corresponding current reaction site type. The first atom reaction site type of the current atom is set based on the current reaction site type. At the end of this traversal, the new molecular structure composed of the latest set of atoms and the current set of chemical bonds is fed back to the current client as the corresponding summary information.

[0042] A second aspect of the present invention provides an apparatus for implementing the processing method for atomic-level reaction site prediction as described in the first aspect above, the apparatus comprising: a model building module, a model training module, and a model application module;

[0043] The model building module is used to build a processing model for predicting atomic-level electrophilic / nucleophilic reaction sites of 3D molecular structures, denoted as the corresponding reaction site prediction model.

[0044] The model training module is used to construct a first original dataset by collecting large amounts of data on the 3D molecular structure; and to construct a corresponding first dataset by using the first original dataset and preset quantum chemical calculation tools and experimental methods to build a model training dataset; and to train the reaction site prediction model based on the first dataset.

[0045] The model application module is used to receive the 3D molecular structure of any compound molecule input by the client as the corresponding current 3D molecular structure after the model training is completed; input the current 3D molecular structure into the reaction site prediction model to perform atomic-level electrophilic / nucleophilic reaction site prediction processing to obtain the corresponding current reaction site prediction tensor; and summarize the atomic information of electrophilic / nucleophilic reaction sites based on the current 3D molecular structure and the current reaction site prediction tensor to obtain the corresponding summary information and feed it back to the current client.

[0046] A third aspect of the present invention provides an electronic device, including: a memory, a processor, and a transceiver;

[0047] The processor is used to couple with the memory, read and execute instructions in the memory to implement the steps of the method described in the first aspect above;

[0048] The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

[0049] A fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions that, when executed by a computer, cause the computer to perform the instructions described in the first aspect.

[0050] This invention provides a method, apparatus, electronic device, and computer-readable storage medium for predicting atomic-level reaction sites. As described above, this invention first constructs a processing model for predicting atomic-level electrophilic / nucleophilic reaction sites in 3D molecular structures, denoted as the corresponding reaction site prediction model. Then, a first raw dataset containing a large number of 3D molecular structures is constructed through big data collection. Atomic-level electrophilic / nucleophilic reaction sites are predicted for each 3D molecular structure in the first raw dataset using traditional prediction methods. Label data is constructed based on the prediction results to obtain a model training dataset, i.e., the first dataset. The reaction site prediction model is then trained based on the first dataset. After model training, end-to-end atomic-level electrophilic / nucleophilic reaction site prediction is performed based on the reaction site prediction model according to the 3D molecular structure of any compound. The reaction site prediction model provided by this invention integrates the first step of the traditional prediction method but omits the second and third steps. Prediction task processing based on the reaction site prediction model provided by this invention shortens the prediction cycle, reduces prediction costs, and improves prediction efficiency. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of a method for predicting atomic-level reaction sites provided in Embodiment 1 of the present invention;

[0052] Figure 2 This is a block diagram of the reaction site prediction model provided in Embodiment 1 of the present invention;

[0053] Figure 3 This is a module structure diagram of a processing device for predicting atomic-level reaction sites provided in Embodiment 2 of the present invention;

[0054] Figure 4 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0056] Embodiment 1 of the present invention provides a processing method for predicting atomic-level reaction sites, such as... Figure 1 The schematic diagram shows a method for predicting atomic-level reaction sites provided in Embodiment 1 of the present invention. This method mainly includes the following steps:

[0057] Step 1: Construct a processing model for predicting atomic-level electrophilic / nucleophilic reaction sites in 3D molecular structures, denoted as the corresponding reaction site prediction model.

[0058] Here, the 3D molecular structures mentioned in the embodiments of the present invention are all composed of a first set of atoms and a first set of chemical bonds; wherein: 1) the first set of atoms includes multiple first atoms; the atomic properties of the first atom include at least a first atom identifier, a first atom type, a first atom coordinates, a first atom charge, and a first atom chiral property set; the first atom identifier is the unique identifier of the current first atom; the first atom type is the chemical element type of the current first atom; the first atom charge is the atomic charge number of the current first atom; the first atom chiral property set is composed of multiple first chiral properties; each first chiral property has a corresponding property type and property value. Composition; if the first atom is a chiral central atom, then all attribute values ​​in the corresponding first atom chiral attribute set are not empty; if the first atom is not a chiral central atom, then all attribute values ​​in the corresponding first atom chiral attribute set are empty; 2) The first chemical bond set includes multiple first chemical bonds; the chemical bond attributes of the first chemical bond include at least the first chemical bond identifier, the first chemical bond type, and the first atom identifier group; the first chemical bond identifier is the unique identifier of the current first chemical bond; the first chemical bond type includes multiple chemical bond types; the first atom identifier group is composed of the first atom identifiers of the two first atoms connected by the current first chemical bond.

[0059] The reaction site prediction model of this invention is used to perform atomic-level electrophilic / nucleophilic reaction site prediction processing based on the 3D molecular structure input to the model and output the corresponding reaction site prediction tensor; wherein, the reaction site prediction tensor is composed of multiple reaction site prediction vectors; the reaction site prediction vector corresponds one-to-one with the first atom; the reaction site prediction vector is composed of three prediction probabilities that sum to 1, namely: electrophilic reaction site prediction probability, nucleophilic reaction site prediction probability, and non-reaction site prediction probability.

[0060] The reaction site prediction model's input is a 3D molecular structure, and its output is the corresponding reaction site prediction tensor. The model components include a Fukui function calculation module, a molecular graph construction module, a molecular graph feature extraction network, a fully connected network, and a Softmax activation layer, such as... Figure 2 The block diagram of the reaction site prediction model provided in Embodiment 1 of the present invention is shown.

[0061] The connection relationships of each model component are as follows: the input end of the Fukui function calculation module is connected to the model input end, and the output end is connected to the second input end of the molecular graph construction module; the first input end of the molecular graph construction module is connected to the model input end, and the output end is connected to the input end of the molecular graph feature extraction network; the output end of the molecular graph feature extraction network is connected to the input end of the fully connected network; the output end of the fully connected network is connected to the Softmax activation layer; and the Softmax activation layer is connected to the model output end.

[0062] The functions of each model component are shown below.

[0063] 1) Fukui Function Calculation Module:

[0064] The Fukui function calculation module is implemented based on the Fukui function calculation interface provided by quantum chemical calculation tools; here, the quantum chemical calculation tools mentioned in this embodiment of the invention include at least Gaussian software, ORCA software, and CP2K software.

[0065] The Fukui function calculation module performs atomic-level Fukui function calculations based on the 3D molecular structure input from the model input terminal to obtain the corresponding first Fukui function value tensor, which is then sent to the molecular graph construction module. The first Fukui function value tensor consists of multiple first Fukui function value vectors; each first Fukui function value vector corresponds one-to-one with the first atom of the 3D molecular structure; and each first Fukui function value vector includes nucleophilic reaction function values. and electrophilic reaction function value

[0066] The Fukui function calculation module of the reaction site prediction model in this embodiment of the invention is actually the same as the first step of the traditional prediction method. However, the subsequent molecular graph construction module, molecular graph feature extraction network, fully connected network and Softmax activation layer are used to replace the second and third steps of the traditional prediction method.

[0067] 2) Molecular map construction module:

[0068] The molecular graph construction module is used to construct the corresponding molecular graph node set based on the first set of atoms and the first Fukui function value tensor of the 3D molecular structure input from the model input terminal, and to construct the corresponding molecular graph node edge set based on the first set of chemical bonds of the 3D molecular structure input from the model input terminal, and to send the corresponding first molecular graph composed of the first node set and the first node edge set to the molecular graph feature extraction network.

[0069] Here, the first molecular graph in this embodiment of the invention includes a first node set and a first node edge set; wherein, the first node set includes multiple first nodes; each first node corresponds one-to-one with a first atom; the node attributes of the first node consist of the corresponding first atom identifier, first atom type, first atom coordinates, first atom charge, first atom chiral attribute set, and first Fukui function value vector; the first node edge set includes multiple first node edges; the edge attributes of the first node edges consist of the corresponding first chemical bond identifier, first chemical bond type, and first atom identifier group.

[0070] 3) Molecular graph feature extraction network:

[0071] The molecular graph feature extraction network is implemented by sequentially connecting multiple graph isomorphism network convolution (GINEConv) layers with edge features.

[0072] The molecular graph feature extraction network is used to perform node feature extraction processing with edge feature fusion mechanism based on the first molecular graph to obtain the corresponding first node feature tensor and send it to the fully connected network; wherein, the first node feature tensor includes multiple first node feature vectors; the first node feature vectors correspond one-to-one with the first nodes.

[0073] 4) Fully connected network:

[0074] A fully connected network is achieved by connecting one or more fully connected layers sequentially.

[0075] The fully connected network is used to perform a linear transformation from the feature tensor to the prediction tensor of the first node to obtain the corresponding prediction tensor of the first node, which is then sent to the Softmax activation layer. The prediction tensor of the first node includes multiple prediction vectors of the first node. Each prediction vector of the first node corresponds one-to-one with the feature vector of the first node. The prediction vector of the first node consists of three types of prediction values: prediction values ​​of electrophilic reaction sites, prediction values ​​of nucleophilic reaction sites, and prediction values ​​of non-reaction sites.

[0076] 5) Softmax activation layer:

[0077] The Softmax activation layer consists of multiple parallel Softmax activation functions; each Softmax activation function corresponds one-to-one with the prediction vector of the first node.

[0078] The Softmax activation layer is used to obtain the corresponding reaction site prediction vector by normalizing the prediction probability of each Softmax activation function based on the prediction vector of the corresponding first node; and the corresponding reaction site prediction tensor is composed of all the obtained reaction site prediction vectors; and the reaction site prediction tensor is sent to the model output.

[0079] Here, the Softmax activation layer is specifically used to obtain the corresponding reaction site prediction vector by normalizing the prediction probability of each Softmax activation function based on the corresponding first node prediction vector: Let the predicted values ​​of electrophilic reaction sites, nucleophilic reaction sites, and non-reactive sites of each first node prediction vector be z1, z2, and z3, respectively; and let the predicted probabilities of electrophilic reaction sites, nucleophilic reaction sites, and non-reactive sites of each first node prediction vector be p1, p2, and p3, respectively; and set the calculation formulas for p1, p2, p3 and z1, z2, and z3 as follows:

[0080] Step 2: Construct the first original dataset by collecting large amounts of data on 3D molecular structures; and construct the corresponding first dataset by using the first original dataset and preset quantum chemical calculation tools and experimental methods to build a model training dataset; and train the reaction site prediction model based on the first dataset.

[0081] Specifically, this includes: Step 21, constructing the first original dataset by collecting big data on 3D molecular structures;

[0082] Specifically, this includes: collecting 3D molecular structures of various compound molecules through multiple publicly available data channels, and using the chemical formula and 3D molecular structure of each compound molecule collected as the corresponding first original molecular formula and first original molecular structure to form the corresponding first original record; and using all the obtained first original records to form the corresponding first original dataset;

[0083] Among them, the various types of publicly available data channels include at least various publicly available datasets, journals / magazines / scientific and technological literature / papers that publish molecular three-dimensional structures; the first original dataset includes multiple first original records; each first original record includes a first original molecular formula and a first original molecular structure; each first original record corresponds to a compound molecule; the first original molecular formula is the chemical molecular formula of the current compound molecule; the first original molecular structure is the 3D molecular structure of the current compound molecule;

[0084] Step 22, and construct the corresponding first dataset by using the first original dataset and the preset quantum chemical calculation tools and experimental methods to train the model;

[0085] The first dataset includes multiple first data records; each first data record includes a first training structure and a first label tensor; the first training structure is a 3D molecular structure; the first label tensor includes multiple first label vectors; each first label vector consists of three label probabilities that sum to 1, namely: electrophilic reaction site label probability, nucleophilic reaction site label probability, and non-reactive site label probability; only one of the three label probabilities in the first label vector is 1, and the other two are 0; each first label vector corresponds one-to-one with the first atom of the first atom set of the first training structure.

[0086] Specifically, this includes: using each of the first original records in the first original dataset as the corresponding current original record; using the first original molecular formula and first original molecular structure of the current original record as the corresponding current molecular formula and current molecular structure; constructing a corresponding three-dimensional molecular model based on the current molecular formula and current molecular structure using quantum chemical calculation tools; performing atomic-level Fukui function calculations based on the current molecular model using the Fukui function calculation interface provided by the current quantum chemical calculation tools, and using the calculation results as the corresponding current Fukui function value tensor; estimating the three types of reaction sites for each atom on the current molecular structure based on the current Fukui function value tensor and the three-dimensional spatial configuration of the current molecular model to obtain the corresponding set of atomic reaction site types; verifying the correctness of the estimated set of atomic reaction site types through chemical experiments; and when the correctness verification is successful, using the current molecular structure as a corresponding first training structure, and performing label tensor transformation based on the set of atomic reaction site types corresponding to the current molecular structure to obtain the corresponding first label tensor, and forming a corresponding first data record by combining the first training structure corresponding to the current molecular structure and the first label vector; and forming the corresponding first dataset by combining all the obtained first data records.

[0087] The atomic reaction site type set includes multiple atomic reaction site types; each atomic reaction site type corresponds one-to-one with the first atom of the first atom set of the corresponding 3D molecular structure; the atomic reaction site types include electrophilic reaction sites, nucleophilic reaction sites, and non-reactive sites.

[0088] In step 22, the first label tensor is obtained by performing a label tensor transformation based on the set of atomic reaction site types corresponding to the current molecular structure. Specifically, each atomic reaction site type in the current set of atomic reaction site types is taken as the corresponding current reaction site type; a corresponding first label vector is generated for the current reaction site type as the corresponding current label vector; the electrophilic reaction site label probability, nucleophilic reaction site label probability, or non-reaction site label probability corresponding to the current reaction site type in the current label vector is taken as the corresponding current label probability, and the current label probability is set to 1, while the other two label probabilities are set to 0; and a corresponding first label tensor is formed by all the first label vectors corresponding to the current set of atomic reaction site types.

[0089] Step 23, and train the reaction site prediction model based on the first dataset;

[0090] Specifically, it includes: Step 231, dividing the first dataset into two sub-datasets based on a preset first segmentation ratio, denoted as the corresponding first training set and first evaluation set;

[0091] Wherein, the first segmentation ratio is a pre-set ratio parameter, such as 8:2; both the first training set and the first evaluation set include multiple first data records; the ratio of the total number of records in the first training set to the total number of records in the first evaluation set satisfies the first segmentation ratio;

[0092] Step 232: Extract the first data record of the first training set as the corresponding current training record;

[0093] Step 233: Input the 3D molecular structure of the first training structure of the current training record into the reaction site prediction model for atomic-level electrophilic / nucleophilic reaction site prediction, and use the reaction site prediction tensor output by the model as the corresponding first prediction tensor; and input the first prediction tensor and the first label tensor of the current training record into the preset first model loss function to calculate the corresponding first loss value.

[0094] The first model loss function includes at least the cross-entropy loss function;

[0095] Step 234: Identify whether the first loss value meets the preset first loss value range; if the first loss value meets the first loss value range, identify whether the current training record is the last first data record of the first training set; if so, proceed to step 235; otherwise, extract the next first data record of the first training set as the new current training record and return to step 233; if the first loss value does not meet the first loss value range, perform a round of modulation on the model parameters of the molecular graph feature extraction network and the fully connected network of the reaction site prediction model based on the preset first model optimizer, and return to step 233 when the modulation ends.

[0096] The first loss value range is a pre-set numerical range; the first model optimizer includes at least the SGD optimizer and the Adam optimizer;

[0097] Step 235: Perform a round of traversal on all first data records in the first evaluation set; during this round of traversal, take the currently traversed first data record as the corresponding current evaluation record; input the first training structure of the current evaluation record as the 3D molecular structure input to the reaction site prediction model for atomic-level electrophilic / nucleophilic reaction site prediction processing, and take the reaction site prediction tensor output by the model as the corresponding second prediction tensor; and form a corresponding prediction-label data pair by the second prediction tensor and the first label tensor of the current evaluation record; and at the end of this round of traversal, calculate the accuracy, precision, recall, and F1 score based on all the obtained prediction-label data pairs to obtain the corresponding first accuracy, first precision, first recall, and first F1 score;

[0098] Step 236: Identify the first accuracy, first precision, first recall, and first F1 score; if the first accuracy does not meet the preset first accuracy range, or the first precision does not meet the preset first precision range, or the first recall does not meet the preset first recall range, or the first F1 score does not meet the preset first F1 score range, then return to step 232 to continue training; if the first accuracy meets the first accuracy range, and the first precision meets the first precision range, and the first recall meets the first recall range, and the first F1 score meets the first F1 score range, then stop training and confirm that the model training has ended.

[0099] Here, the first accuracy range, the first precision range, the first recall range, and the first F1 score range are four pre-set numerical ranges.

[0100] Step 3: After the model training is completed, the 3D molecular structure of any compound molecule input from the client is received as the corresponding current 3D molecular structure; the current 3D molecular structure is input into the reaction site prediction model for atomic-level electrophilic / nucleophilic reaction site prediction processing to obtain the corresponding current reaction site prediction tensor; and the atomic information of electrophilic / nucleophilic reaction sites is summarized based on the current 3D molecular structure and the current reaction site prediction tensor to obtain the corresponding summary information and feed it back to the current client.

[0101] Specifically, this includes: Step 31, after the model training is completed, receiving the 3D molecular structure of any compound molecule input from the client as the corresponding current 3D molecular structure; and inputting the current 3D molecular structure into the reaction site prediction model to perform atomic-level electrophilic / nucleophilic reaction site prediction processing to obtain the corresponding current reaction site prediction tensor;

[0102] Step 32: Based on the current 3D molecular structure and the predicted tensor of the current reaction site, the atomic information of the electrophilic / nucleophilic reaction site is summarized and the corresponding summary information is fed back to the current client;

[0103] Specifically, this includes: using the first set of atoms and the first set of chemical bonds corresponding to the current 3D molecular structure as the corresponding current set of atoms and the current set of chemical bonds; performing a traversal of all first atoms in the current set of atoms; during this traversal, using the first atom currently traversed as the corresponding current atom; adding a reaction site type attribute to the atomic properties of the current atom as the corresponding first atom reaction site type; using the reaction site prediction vector corresponding to the current atom in the current reaction site prediction tensor as the corresponding current prediction vector; using the electrophilic reaction site, nucleophilic reaction site, or non-reactive site with the highest probability in the current prediction vector as the corresponding current reaction site type; setting the first atom reaction site type of the current atom based on the current reaction site type; and at the end of this traversal, feeding back the new molecular structure composed of the latest set of atoms and the current set of chemical bonds as the corresponding summary information to the current client.

[0104] Figure 3 This is a module structure diagram of a processing device for predicting atomic-level reaction sites provided in Embodiment 2 of the present invention. This device can be a terminal device or server implementing the aforementioned method embodiments, or it can be a device that enables the aforementioned terminal device or server to implement the aforementioned method embodiments. For example, the device can be a device or chip system of the aforementioned terminal device or server. Figure 3 As shown, the device includes: a model building module 201, a model training module 202, and a model application module 203.

[0105] The model building module 201 is used to build a processing model for predicting atomic-level electrophilic / nucleophilic reaction sites of 3D molecular structures, denoted as the corresponding reaction site prediction model.

[0106] The model training module 202 is used to construct a first original dataset by collecting big data on 3D molecular structures; and to construct a corresponding first dataset by constructing a model training dataset based on the first original dataset and preset quantum chemical calculation tools and experimental methods; and to train the reaction site prediction model based on the first dataset.

[0107] The model application module 203 is used to receive the 3D molecular structure of any compound molecule input by the client as the corresponding current 3D molecular structure after the model training is completed; input the current 3D molecular structure into the reaction site prediction model to perform atomic-level electrophilic / nucleophilic reaction site prediction processing to obtain the corresponding current reaction site prediction tensor; and summarize the atomic information of electrophilic / nucleophilic reaction sites based on the current 3D molecular structure and the current reaction site prediction tensor to obtain the corresponding summary information and feed it back to the current client.

[0108] The processing device for predicting atomic-level reaction sites provided in this embodiment of the invention can execute the method steps in the above method embodiments. Its implementation principle and technical effect are similar, and will not be repeated here.

[0109] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing elements; they can be fully implemented in hardware; or some modules can be implemented by processing elements calling software, while others are implemented in hardware. For example, the model building module can be a separate processing element, or it can be integrated into a chip in the above device. Alternatively, it can be stored as program code in the memory of the above device, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0110] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a System-on-a-Chip (SOC).

[0111] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the foregoing method embodiments are generated. The computer described above can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The aforementioned computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the aforementioned computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, Bluetooth, microwave, etc.) means. The aforementioned computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The aforementioned available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0112] Figure 4 This is a schematic diagram of an electronic device provided in Embodiment 3 of the present invention. This electronic device can be a terminal device or server implementing the methods of the aforementioned embodiments, or it can be a terminal device or server connected to the aforementioned terminal device or server implementing the methods of the aforementioned embodiments. Figure 4As shown, the electronic device may include: a processor 301 (e.g., CPU), a memory 302, and a transceiver 303; the transceiver 303 is coupled to the processor 301, and the processor 301 controls the transmission and reception operations of the transceiver 303. The memory 302 may store various instructions for performing various processing functions and implementing the processing steps described in the foregoing embodiments. Preferably, the electronic device involved in the embodiments of the present invention further includes: a power supply 304, a system bus 305, and a communication port 306. The system bus 305 is used to realize communication connections between components. The communication port 306 is used for communication between the electronic device and other peripherals.

[0113] exist Figure 4 The system bus 305 mentioned can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus. The communication interface is used to enable communication between the database access device and other devices (e.g., clients, read-write libraries, and read-only libraries). Memory may include Random Access Memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0114] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), graphics processing units (GPUs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0115] It should be noted that the embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when run on a computer, cause the computer to perform the methods and processes provided in the above embodiments.

[0116] This invention provides a method, apparatus, electronic device, and computer-readable storage medium for predicting atomic-level reaction sites. As described above, this invention first constructs a processing model for predicting atomic-level electrophilic / nucleophilic reaction sites in 3D molecular structures, denoted as the corresponding reaction site prediction model. Then, a first raw dataset containing a large number of 3D molecular structures is constructed through big data collection. Atomic-level electrophilic / nucleophilic reaction sites are predicted for each 3D molecular structure in the first raw dataset using traditional prediction methods. Label data is constructed based on the prediction results to obtain a model training dataset, i.e., the first dataset. The reaction site prediction model is then trained based on the first dataset. After model training, end-to-end atomic-level electrophilic / nucleophilic reaction site prediction is performed based on the reaction site prediction model according to the 3D molecular structure of any compound. The reaction site prediction model provided by this invention integrates the first step of the traditional prediction method but omits the second and third steps. Prediction task processing based on the reaction site prediction model provided by this invention shortens the prediction cycle, reduces prediction costs, and improves prediction efficiency.

[0117] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0118] 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 description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting atomic-level reaction sites, characterized in that, The method includes: A processing model for predicting atomic-level electrophilic / nucleophilic reaction sites in 3D molecular structures is constructed, denoted as the corresponding reaction site prediction model. A first original dataset is constructed by collecting large amounts of data on the 3D molecular structure; and a corresponding first dataset is constructed by using the first original dataset and preset quantum chemical calculation tools and experimental methods to train the model; and the reaction site prediction model is trained based on the first dataset. After the model training is completed, the 3D molecular structure of any compound molecule input from the client is received as the corresponding current 3D molecular structure; the current 3D molecular structure is input into the reaction site prediction model to perform atomic-level electrophilic / nucleophilic reaction site prediction processing to obtain the corresponding current reaction site prediction tensor; and the atomic information of electrophilic / nucleophilic reaction sites is summarized based on the current 3D molecular structure and the current reaction site prediction tensor to obtain the corresponding summary information and feed it back to the current client; The reaction site prediction model is used to perform atomic-level electrophilic / nucleophilic reaction site prediction based on the 3D molecular structure input to the model and output the corresponding reaction site prediction tensor. The 3D molecular structure consists of a first set of atoms and a first set of chemical bonds. The first set of atoms includes multiple first atoms. The atomic properties of the first atoms include at least a first atom identifier, a first atom type, a first atom coordinate, a first atom charge, and a first atom chiral property set. The first set of chemical bonds includes multiple first chemical bonds. The chemical bond properties of the first chemical bonds include at least a first chemical bond identifier, a first chemical bond type, and a first atom identifier group. The first atom identifier group consists of the first atom identifiers of the two first atoms currently connected by the first chemical bond. The reaction site prediction model includes a Fukui function calculation module, a molecular graph construction module, a molecular graph feature extraction network, a fully connected network, and a Softmax activation layer; The Fukui function calculation module is used to perform atomic-level Fukui function calculations based on the 3D molecular structure input from the model input terminal to obtain the corresponding first Fukui function value tensor, which is then sent to the molecular graph construction module. The molecular graph construction module is used to construct a corresponding molecular graph node set based on the first set of atoms and the first Fukui function value tensor of the 3D molecular structure input from the model input terminal, and to construct a corresponding molecular graph node edge set based on the first set of chemical bonds of the 3D molecular structure input from the model input terminal, and to send the first molecular graph composed of the first node set and the first node edge set to the molecular graph feature extraction network. The molecular graph feature extraction network is implemented by sequentially connecting multiple GINEConv layers; the molecular graph feature extraction network is used to perform node feature extraction processing with edge feature fusion mechanism on the first molecular graph to obtain the corresponding first node feature tensor and send it to the fully connected network; The fully connected network is implemented by one or more fully connected layers connected sequentially; the fully connected network is used to perform a linear transformation on the feature tensor of the first node from the feature tensor to the prediction tensor to obtain the corresponding prediction tensor of the first node and send it to the Softmax activation layer; the prediction tensor of the first node includes multiple prediction vectors of the first node. The Softmax activation layer consists of multiple parallel Softmax activation functions; each Softmax activation function corresponds one-to-one with the first node prediction vector; the Softmax activation layer is used to obtain the corresponding reaction site prediction vector by performing prediction probability normalization processing on each Softmax activation function according to the corresponding first node prediction vector; and to form the corresponding reaction site prediction tensor by all the obtained reaction site prediction vectors; and to send the reaction site prediction tensor to the model output.

2. The processing method for predicting atomic-level reaction sites according to claim 1, characterized in that, The quantum chemical calculation tools include at least Gaussian software, ORCA software, and CP2K software.

3. The processing method for predicting atomic-level reaction sites according to claim 1, characterized in that, The model input terminal of the reaction site prediction model is used to receive the 3D molecular structure input by the model, and the model output terminal is used to output the corresponding reaction site prediction tensor. The first atom identifier is the unique identifier of the current first atom; The first atom type is the chemical element type of the current first atom; The first atomic charge is the number of atomic charges of the first atom at present; The first set of chiral properties of the atom consists of multiple first chiral properties; each first chiral property consists of a corresponding property type and a property value; if the first atom is a chiral center atom, then all property values ​​in the corresponding set of chiral properties of the first atom are not empty; If the first atom is not a chiral center atom, then all attribute values ​​in the corresponding chiral attribute set of the first atom are empty; The first chemical bond identifier is a unique identifier for the current first chemical bond; the first chemical bond type includes multiple chemical bond types; The reaction site prediction tensor comprises multiple reaction site prediction vectors; each reaction site prediction vector corresponds one-to-one with the first atom; each reaction site prediction vector consists of three prediction probabilities that sum to 1, namely: electrophilic reaction site prediction probability, nucleophilic reaction site prediction probability, and non-reaction site prediction probability. The input terminal of the Fukui function calculation module is connected to the input terminal of the model, and the output terminal is connected to the second input terminal of the molecular graph construction module; the first input terminal of the molecular graph construction module is connected to the input terminal of the model, and the output terminal is connected to the input terminal of the molecular graph feature extraction network; the output terminal of the molecular graph feature extraction network is connected to the input terminal of the fully connected network; the output terminal of the fully connected network is connected to the Softmax activation layer; the Softmax activation layer is connected to the output terminal of the model. The Fukui function calculation module is implemented based on the Fukui function calculation interface provided by the quantum chemical calculation tool; The first Fukui function value tensor is composed of multiple first Fukui function value vectors; each first Fukui function value vector corresponds one-to-one with the first atom of the 3D molecular structure; the first Fukui function value vector includes nucleophilic reaction function values. and electrophilic reaction function value The first molecular graph includes a first node set and a first node edge set; the first node set includes multiple first nodes; each first node corresponds one-to-one with a first atom; the node attributes of the first node consist of the corresponding first atom identifier, first atom type, first atom coordinates, first atom charge, first atom chiral attribute set, and first Fukui function value vector; the first node edge set includes multiple first node edges; the edge attributes of the first node edges consist of the corresponding first chemical bond identifier, first chemical bond type, and first atom identifier group. The first node feature tensor includes multiple first node feature vectors; The feature vector of the first node corresponds one-to-one with the first node; The first node prediction vector corresponds one-to-one with the first node feature vector; the first node prediction vector consists of three types of prediction values, namely, electrophilic reaction site prediction value, nucleophilic reaction site prediction value, and non-reactive site prediction value; The Softmax activation layer is specifically used to obtain the corresponding reaction site prediction vector by normalizing the prediction probability of each reaction site prediction vector based on the corresponding first node prediction vector using each Softmax activation function. Specifically, it is set that the predicted values ​​of the electrophilic reaction site, nucleophilic reaction site, and non-reaction site for each first node prediction vector are z1, z2, and z3, respectively. Furthermore, it is set that the predicted probabilities of the electrophilic reaction site, nucleophilic reaction site, and non-reaction site for each reaction site prediction vector corresponding to each first node prediction vector are p1, p2, and p3, respectively. The calculation formulas for p1, p2, p3 and z1, z2, z3 are as follows:

4. The processing method for predicting atomic-level reaction sites according to claim 3, characterized in that, The first original dataset includes multiple first original records; each first original record includes a first original molecular formula and a first original molecular structure; each first original record corresponds to a compound molecule; the first original molecular formula is the chemical molecular formula of the current compound molecule; the first original molecular structure is the 3D molecular structure of the current compound molecule; The first dataset includes multiple first data records; each first data record includes a first training structure and a first label tensor; the first training structure is a 3D molecular structure; the first label tensor includes multiple first label vectors; each first label vector consists of three label probabilities that sum to 1, namely: electrophilic reaction site label probability, nucleophilic reaction site label probability, and non-reactive site label probability; only one of the three label probabilities in the first label vector is 1, and the other two are 0; each first label vector corresponds one-to-one with the first atom in the first atom set of the first training structure.

5. The processing method for predicting atomic-level reaction sites according to claim 4, characterized in that, The construction of the first original dataset through big data collection of the 3D molecular structure specifically includes: Data on the 3D molecular structures of various compounds are collected through multiple publicly available data channels. The chemical formula and 3D molecular structure of each compound molecule are used as the corresponding first original molecular formula and first original molecular structure to form the corresponding first original record. All the obtained first original records form the corresponding first original dataset. The multiple publicly available data channels include at least various publicly available datasets that publish molecular 3D structures, journals / magazines / scientific and technological literature / papers.

6. The processing method for predicting atomic-level reaction sites according to claim 4, characterized in that, The construction of the corresponding first dataset based on the first original dataset and preset quantum chemical calculation tools and experimental methods to obtain the model training dataset specifically includes: Each of the first original records in the first original dataset is taken as the corresponding current original record; the first original molecular formula and the first original molecular structure of the current original record are taken as the corresponding current molecular formula and current molecular structure; and based on the quantum chemical calculation tool, a corresponding three-dimensional molecular model is constructed according to the current molecular formula and current molecular structure to obtain the corresponding current molecular model; and based on the Fukui function calculation interface provided by the quantum chemical calculation tool, atomic-level Fukui function calculations are performed according to the current molecular model, and the calculation results are taken as the corresponding current Fukui function value tensor; and based on the current Fukui function value tensor and the three-dimensional spatial configuration of the current molecular model, the three types of reaction site types of each atom on the current molecular structure are estimated to obtain the corresponding atomic reaction site type set; and through chemical... The correctness of the estimated set of atomic reaction site types is verified by experimental methods. Upon successful verification, the current molecular structure is used as a corresponding first training structure. A label tensor transformation is performed based on the set of atomic reaction site types corresponding to the current molecular structure to obtain the corresponding first label tensor. The first training structure corresponding to the current molecular structure and the first label vector form a corresponding first data record. All obtained first data records form the corresponding first dataset. The set of atomic reaction site types includes multiple atomic reaction site types. Each atomic reaction site type corresponds one-to-one with the first atom in the first atom set of the corresponding 3D molecular structure. The atomic reaction site types include electrophilic reaction sites, nucleophilic reaction sites, and non-reactive sites. Specifically, the step of performing a label tensor transformation based on the set of atomic reaction site types corresponding to the current molecular structure to obtain the corresponding first label tensor involves: taking each atomic reaction site type in the current set of atomic reaction site types as the corresponding current reaction site type; generating a corresponding first label vector for the current reaction site type as the corresponding current label vector; taking the electrophilic reaction site label probability, the nucleophilic reaction site label probability, or the non-reaction site label probability corresponding to the current reaction site type in the current label vector as the corresponding current label probability, setting the current label probability to 1, and setting the other two label probabilities to 0; and forming a corresponding first label tensor by combining all the first label vectors corresponding to the current set of atomic reaction site types.

7. The processing method for predicting atomic-level reaction sites according to claim 4, characterized in that, The training of the reaction site prediction model based on the first dataset specifically includes: Step 71: Based on a preset first segmentation ratio, the first dataset is divided into two sub-datasets, denoted as the first training set and the first evaluation set. Wherein, both the first training set and the first evaluation set include multiple first data records; the ratio of the total number of records in the first training set to the total number of records in the first evaluation set satisfies the first segmentation ratio; Step 72: Extract the first data record of the first training set as the corresponding current training record; Step 73: Input the 3D molecular structure, which is the first training structure recorded in the current training record, into the reaction site prediction model to perform atomic-level electrophilic / nucleophilic reaction site prediction processing, and use the reaction site prediction tensor output by the model as the corresponding first prediction tensor; and substitute the first prediction tensor and the first label tensor recorded in the current training record into the preset first model loss function to calculate the corresponding first loss value. Wherein, the loss function of the first model includes at least the cross-entropy loss function; Step 74: Identify whether the first loss value meets the preset first loss value range; if the first loss value meets the first loss value range, identify whether the current training record is the last first data record of the first training set; if so, proceed to step 75; otherwise, extract the next first data record of the first training set as the new current training record and return to step 73; if the first loss value does not meet the first loss value range, perform a round of modulation on the model parameters of the molecular graph feature extraction network and the fully connected network of the reaction site prediction model based on the preset first model optimizer, and return to step 73 when the modulation ends. The first model optimizer includes at least an SGD optimizer and an Adam optimizer; Step 75: Perform a round of traversal on all the first data records in the first evaluation set; during this round of traversal, take the currently traversed first data record as the corresponding current evaluation record; and take the first training structure of the current evaluation record as the 3D molecular structure input to the reaction site prediction model for atomic-level electrophilic / nucleophilic reaction site prediction processing, and take the reaction site prediction tensor output by the model as the corresponding second prediction tensor; and form a corresponding prediction-label data pair by the second prediction tensor and the first label tensor of the current evaluation record; and at the end of this round of traversal, calculate the accuracy, precision, recall, and F1 score based on all the obtained prediction-label data pairs to obtain the corresponding first accuracy, first precision, first recall, and first F1 score; Step 76: Identify the first accuracy, first precision, first recall, and first F1 score; if the first accuracy does not meet the preset first accuracy range, or the first precision does not meet the preset first precision range, or the first recall does not meet the preset first recall range, or the first F1 score does not meet the preset first F1 score range, then return to step 72 to continue training; if the first accuracy meets the first accuracy range, and the first precision meets the first precision range, and the first recall meets the first recall range, and the first F1 score meets the first F1 score range, then stop training and confirm that the model training has ended.

8. The processing method for predicting atomic-level reaction sites according to claim 3, characterized in that, The process of summarizing the electrophilic / nucleophilic reaction site atomic information based on the current 3D molecular structure and the current reaction site prediction tensor to obtain corresponding summary information, and then feeding it back to the current client, specifically includes: The first set of atoms and the first set of chemical bonds corresponding to the current 3D molecular structure are taken as the corresponding current set of atoms and the current set of chemical bonds. A traversal is performed on all the first atoms in the current set of atoms. During this traversal, the first atom currently traversed is taken as the corresponding current atom. A reaction site type attribute is added to the atomic properties of the current atom as the corresponding first atom reaction site type. The reaction site prediction vector corresponding to the current atom in the current reaction site prediction tensor is taken as the corresponding current prediction vector. The electrophilic reaction site, nucleophilic reaction site, or non-reactive site with the highest probability in the current prediction vector is taken as the corresponding current reaction site type. The first atom reaction site type of the current atom is set based on the current reaction site type. At the end of this traversal, the new molecular structure composed of the latest set of atoms and the current set of chemical bonds is fed back to the current client as the corresponding summary information.

9. An apparatus for performing the processing method for predicting atomic-level reaction sites according to any one of claims 1-8, characterized in that, The device includes: a model building module, a model training module, and a model application module; The model building module is used to build a processing model for predicting atomic-level electrophilic / nucleophilic reaction sites of 3D molecular structures, denoted as the corresponding reaction site prediction model. The model training module is used to construct a first original dataset by collecting large amounts of data on the 3D molecular structure; and to construct a corresponding first dataset by using the first original dataset and preset quantum chemical calculation tools and experimental methods to build a model training dataset; and to train the reaction site prediction model based on the first dataset. The model application module is used to receive the 3D molecular structure of any compound molecule input by the client as the corresponding current 3D molecular structure after the model training is completed; input the current 3D molecular structure into the reaction site prediction model to perform atomic-level electrophilic / nucleophilic reaction site prediction processing to obtain the corresponding current reaction site prediction tensor; and summarize the atomic information of electrophilic / nucleophilic reaction sites based on the current 3D molecular structure and the current reaction site prediction tensor to obtain the corresponding summary information and feed it back to the current client.

10. An electronic device, characterized in that, include: Memory, processor, and transceiver; The processor is configured to be coupled to the memory, read and execute instructions in the memory to implement the method according to any one of claims 1-8; The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a computer, cause the computer to perform the method described in any one of claims 1-8.

Images