Training method and apparatus for reaction product prediction model, application method, device, and computer program

Abstract

Embodiments of this application disclose a method for training a reaction product prediction model, an application method, an apparatus, and a device. Related embodiments are applied to various scenarios such as artificial intelligence and are used to improve prediction accuracy while reducing model training costs. Such a method includes constructing a positive sample reactant set and a negative sample reactant set through a first auxiliary network, calculating a reaction prediction loss value based on the positive sample reactant set and the negative sample reactant set, constructing a positive sample reaction group set and a negative sample reaction group set through a second auxiliary network, calculating a reaction relationship prediction loss value based on the positive sample reaction group set and the negative sample reaction group set, obtaining a predicted probability value and an atomic label of atoms in the sample reactant existing in the main product through a third auxiliary network, calculating an atomic prediction loss value, and training a reaction product prediction model based on the reaction prediction loss value, the reaction relationship prediction loss value, and the atomic prediction loss value to obtain a target reaction product prediction model.

Description

[Technical Field] 【0001】 [Cross-reference of related applications] This application claims priority to Chinese Patent Application No. 2022108264623, filed with the China National Intellectual Property Administration on 14 July 2022, titled "Method for parameter adjustment, application method, apparatus and device for reaction product prediction model," the entirety of which is incorporated herein by reference. This application relates to the field of artificial intelligence technology, and more particularly to reaction product prediction technology. [Background technology] 【0002】 The task of predicting organic chemical reaction products is of extremely important significance in fields such as computational chemistry and pharmaceuticals. 【0003】 Typical organic chemical reaction prediction methods use standard reaction templates to predict the possible product structures. However, organic chemical reactions are diverse, and as chemical research and development continues to advance, new reactions are constantly emerging. As a result, reaction templates cannot cover all reaction types and have become inapplicable to recently developed reaction types. With the advancement of deep learning technology, it is considered particularly important to utilize deep learning to learn latent reaction rules from organic chemical reaction data. 【0004】 However, currently used deep learning techniques based on organic chemical reaction data typically rely on a large amount of artificial labeling to train reaction product prediction models. Artificially labeled data is generally very expensive, and as the amount of chemical reaction data increases, more artificial labels are often needed to support it. This results in significant labor and time costs, leading to the problem of high costs for organic chemical reaction product prediction tasks. [Overview of the Initiative] [Means for solving the problem] 【0005】 Embodiments of this application provide a method for training a reaction product prediction model, an application method, an apparatus, and a device. This enables automatic labeling of data in a sample reaction data set without relying on artificial labeling, thereby reducing the cost required for the reaction product prediction task while improving the accuracy of reaction product prediction by the reaction product prediction model. 【0006】 According to one aspect of the embodiments of this application, a method for training a reaction product prediction model, performed by a computer device, comprising the following steps: A step of obtaining a sample reactant vector and a sample reaction product vector by performing a vector transformation on each reaction sequence in a sample reaction data set through an encoder network of a reaction product prediction model, wherein the sample reaction data set includes multiple reaction sequences, and each reaction sequence includes a sample reactant and a sample reaction product. The first step involves constructing a set of positive and negative sample reactants according to the sample reactant vector through a first auxiliary network. The steps include identifying the predicted reaction loss value based on the set of positive and negative sample reactants, The steps include constructing a positive sample reaction group set and a negative sample reaction group set according to the sample reactant vector and the sample reaction product vector through a second auxiliary network, The steps include identifying the reaction relationship prediction loss value based on the positive sample reaction group set and the negative sample reaction group set, The steps include identifying the predicted probability value and atomic label of atoms in the sample reactant present in the main product, according to the sample reactant vector and the sample reaction product vector, through a third auxiliary network, A step of identifying the predicted atomic loss value based on the predicted probability value and atomic label, The present invention provides a method for training a reaction product prediction model, which includes the steps of training a reaction product prediction model based on reaction prediction loss values, reaction relationship prediction loss values, and atom prediction loss values ​​to obtain a target reaction product prediction model. 【0007】 According to another aspect of this application, a method for applying a reaction product prediction model performed by a computer device, comprising the following steps: The steps include inputting the reactants to be measured into the aforementioned target reaction product prediction model, and outputting the predicted change probability of the adjacency matrix from the target reaction product prediction model, The steps include: identifying the predicted change in the adjacency matrix based on the predicted change probability of the adjacency matrix, and The present invention provides a method for applying a reaction product prediction model, which includes the step of identifying a target reaction product based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured. 【0008】 According to another aspect of this application, an acquisition unit is configured to train a reaction product prediction model, wherein the acquisition unit performs vector transformation on each reaction sequence in a sample reaction data set through an encoder network of the reaction product prediction model to obtain a sample reactant vector and a sample reaction product vector, further constructs a positive sample reactant set and a negative sample reactant set according to the sample reactant vector through a first auxiliary network, constructs a positive sample reaction group set and a negative sample reaction group set according to the sample reactant vector and the sample reaction product vector through a second auxiliary network, and identifies the predicted probability value and atomic label of atoms in the sample reactant present in the main product according to the sample reactant vector and the sample reaction product vector, wherein the sample reaction data set includes a plurality of reaction sequences, and each reaction sequence includes a sample reactant and a sample reaction product, A processing unit is configured to identify a predicted reaction loss value based on a set of positive and negative sample reactants, determine a predicted reaction relationship loss value based on a set of positive and negative sample reaction groups, and further identify a predicted atomic loss value based on a predicted probability value and an atomic label. The present invention provides a reaction product prediction model training apparatus, which includes a specific unit configured to train a reaction product prediction model based on reaction prediction loss values, reaction relationship prediction loss values, and atomic prediction loss values, and to obtain a target reaction product prediction model. 【0009】 According to another aspect of this application, an apparatus for applying a reaction product prediction model, An acquisition unit configured to input the reactants to be measured into a target reaction product prediction model and output the predicted change probability of the adjacency matrix from the target reaction product prediction model, A processing unit configured to identify the predicted change amount of the adjacency matrix based on the predicted change probability of the adjacency matrix, The present invention provides an application apparatus for a reaction product prediction model, which includes a specific unit configured to identify a target reaction product based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured. 【0010】 According to another aspect of this application, The present invention provides a computer device that includes memory for storing a program, a processor for implementing the methods described above when the program in memory is executed, and a bus system for connecting the memory and the processor to enable communication between the memory and the processor. 【0011】 According to another aspect of this application, a computer-readable storage medium is provided which, when operated on a computer, stores instructions causing the computer to perform the methods relating to each of the aforementioned aspects. 【0012】 According to another aspect of this application, a computer program product is provided which, when executed by a processor, includes a computer program for realizing the methods relating to each of the aforementioned aspects. [Effects of the Invention] 【0013】 As is clear from the above technical methods, the embodiments of this application have the following beneficial effects. 【0014】 After obtaining the sample reactant vector and the sample reaction product vector, a set of positive and negative sample reactants is constructed through a first auxiliary network, and the reaction prediction loss value is identified based on the positive and negative sample reactant sets. Then, a set of positive and negative sample reaction groups is constructed through a second auxiliary network, and the reaction relationship prediction loss value is identified based on the positive and negative sample reaction group sets. Then, the predicted probability value and atomic label of atoms in the sample reactants being present in the main product are obtained through a third auxiliary network, and the atomic prediction loss value is identified based on the predicted probability value and atomic label. Subsequently, a reaction product prediction model is trained based on the reaction prediction loss value, reaction relationship prediction loss value, and atomic prediction loss value, and a target reaction product prediction model is obtained. As described above, by discovering and constructing positive and negative sample sets and atomic labels from the characteristics of the sample reaction data set itself through the first, second, and third auxiliary networks, which are capable of self-supervised learning, it is possible to achieve automatic labeling of data in the sample reaction data set without relying on artificial labeling, thereby reducing the cost required for training the reaction product prediction model. Furthermore, the first auxiliary network helps the reaction product prediction model to better learn the distance relationships between reactants, thereby enabling the model to predict whether reactants can react; the second auxiliary network helps the reaction product prediction model to better learn the distance relationships between reactants and products, thereby enabling the model to predict the correspondence between reactants and products; and the third auxiliary network helps the reaction product prediction model to better learn the changes that can occur during the reaction process in the relationship between the bond positions of atoms in the reactants, thereby enabling the model to predict whether atoms remain in the main product after the reaction. Based on the auxiliary learning by these auxiliary networks, the accuracy of the reaction product prediction by the reaction product prediction model can be improved. [Brief explanation of the drawing] 【0015】 [Figure 1] This is a schematic diagram of the reaction data control system according to an embodiment of the present application. [Figure 2] This is a flowchart of one embodiment of the training method for a reaction product prediction model according to the embodiments of this application. [Figure 3] This is a flowchart of another embodiment of the training method for the reaction product prediction model according to the embodiment of this application. [Figure 4] This is a flowchart of another embodiment of the training method for the reaction product prediction model according to the embodiment of this application. [Figure 5] This is a flowchart of another embodiment of the training method for the reaction product prediction model according to the embodiment of this application. [Figure 6] This is a flowchart of another embodiment of the training method for the reaction product prediction model according to the embodiment of this application. [Figure 7] This is a flowchart of another embodiment of the training method for the reaction product prediction model according to the embodiment of this application. [Figure 8] This is a flowchart of another embodiment of the training method for the reaction product prediction model according to the embodiment of this application. [Figure 9] This is a flowchart of another embodiment of the training method for the reaction product prediction model according to the embodiment of this application. [Figure 10] This is a flowchart of another embodiment of the training method for the reaction product prediction model according to the embodiment of this application. [Figure 11] This is a flowchart of another embodiment of the training method for the reaction product prediction model according to the embodiment of this application. [Figure 12] This is a flowchart of another embodiment of the training method for the reaction product prediction model according to the embodiment of this application. [Figure 13] This is a flowchart illustrating one embodiment of the method for applying the reaction product prediction model according to the embodiments of this application. [Figure 14] This is a flowchart of another embodiment of the method for applying the reaction product prediction model according to the embodiment of this application. [Figure 15]This is a schematic diagram showing one auxiliary network for a training method of a reaction product prediction model according to an embodiment of this application. [Figure 16] This is a schematic diagram illustrating the conceptual framework of a reaction product prediction model, which is one of the methods for training a reaction product prediction model according to an embodiment of this application. [Figure 17] This is a schematic diagram of one embodiment of a training apparatus for a reaction product prediction model according to an embodiment of this application. [Figure 18] This is a schematic diagram of one embodiment of an application apparatus for the reaction product prediction model according to the embodiment of this application. [Figure 19] This is a schematic diagram of one embodiment of a computer device according to the embodiments of this application. [Modes for carrying out the invention] 【0016】 For the sake of understanding, we will first explain some of the terms or concepts referred to in the embodiments of this application. 【0017】 1. Transformers The transformer consists of an encoder and a decoder, can utilize a self-attention mechanism, and does not employ the hierarchical structure of a recurrent neural network (RNN) and a long short-term memory (LSTM). Therefore, the model can be trained in parallel and can possess global information. 【0018】 2. Contrastive Learning Contrastive learning is a type of self-supervised learning method that learns the general characteristics of a dataset by having a model learn which data points are similar or different, even when labeling is not performed. 【0019】 3. Graph Neural Networks Graph neural networks include graph convolutional networks, graph attention networks, graph autoencoders, graph generative networks, and graph spatial-temporal networks. Compared to the fully connected layer (MLP), the most basic layer of neural networks, graph neural networks have an additional adjacency matrix in addition to multiplying the feature matrix by the weight matrix. 【0020】 4. Auxiliary Task Auxiliary tasks are an important method for supporting the learning of the primary task in reinforcement learning. In learning, when the primary task offers relatively few rewards or the task is difficult, auxiliary tasks can be used to support the learning of feature representations. 【0021】 While specific embodiments of the present invention refer to data related to sample reaction data sets and the like, it should be understood that when the above embodiments of this application are applied to specific products or technologies, user permission or consent must be obtained, and the collection, use, and handling of related data must comply with the relevant laws and regulations and standards of the applicable country and region. 【0022】 It should be understood that the training method and application method for reaction product prediction models described in this application are applicable to a variety of scenarios, including but not limited to artificial intelligence, cloud technology, computational chemistry, and pharmaceuticals. By training a powerful reaction product prediction model and optimizing reaction product prediction tasks, it can be applied to scenarios such as verifying drug retrosynthesis, elucidating scientific laws, and developing pharmaceuticals. 【0023】 The method for training a reaction product prediction model according to this application can be applied to the reaction data control system shown in Figure 1. Referring to Figure 1, Figure 1 is a schematic configuration diagram of a reaction data control system according to an embodiment of this application. As shown in Figure 1, the server acquires sample reactant vectors and sample reaction product vectors based on a sample reaction data set provided from a terminal device, constructs a set of positive sample reactants and a set of negative sample reactants through a first auxiliary network, identifies a reaction prediction loss value based on the set of positive and negative sample reactants, constructs a set of positive and negative sample reaction groups through a second auxiliary network, identifies a reaction relationship prediction loss value based on the set of positive and negative sample reaction groups, acquires a predicted probability value and atomic label for atoms in the sample reactants present in the main product through a third auxiliary network, identifies an atomic prediction loss value based on the predicted probability value and atomic label, and then trains a reaction product prediction model based on the reaction prediction loss value, reaction relationship prediction loss value and atomic prediction loss value to obtain a target reaction product prediction model. As described above, by discovering and constructing positive and negative sample sets and atomic labels from the characteristics of the sample reaction data set itself through the first, second, and third auxiliary networks, which are capable of self-supervised learning, it is possible to achieve automatic labeling of data in the sample reaction data set without relying on artificial labeling, thereby reducing the cost required for training the reaction product prediction model.In this process, the first auxiliary network helps the reaction product prediction model to better learn the distance relationships between reactants, thereby enabling the model to predict whether reactants can react or not; the second auxiliary network helps the reaction product prediction model to better learn the distance relationships between reactants and products, thereby enabling the model to predict the correspondence between reactants and products; and the third auxiliary network helps the reaction product prediction model to better learn the changes that can occur during the reaction process in the relationship between the bond positions of atoms in the reactants, thereby enabling the model to predict whether atoms remain in the main product after the reaction. Based on the auxiliary learning provided by these auxiliary networks, the accuracy of the reaction product prediction by the reaction product prediction model can be improved. 【0024】 Although Figure 1 shows only one type of terminal device, in a real-world scenario, many more types of terminal devices may participate in the data handling process. Terminal devices include, but are not limited to, mobile phones, computers, smart voice interaction devices, smart home appliances, and in-car terminals. The specific number and types will be determined by the actual scenario, and it should be understood that there are no particular limitations here. Also, although Figure 1 shows only one server, in a real-world scenario, multiple servers may participate, and in particular in a multi-model training interaction scenario, the number of servers will be determined by the actual scenario, and there are no particular limitations here. 【0025】 In this embodiment, the server may be an independent physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Terminal devices and servers may be connected directly or indirectly by wired or wireless communication, and terminal devices and servers may be connected to form a blockchain network, but are not limited herein. 【0026】 Referring to the description above, the method for training the reaction product prediction model according to this application will now be described. This method can be performed by a computer device. This computer device may be a server or a terminal device. Referring to Figure 2, one embodiment of the method for training the reaction product prediction model according to the embodiment of this application includes the following steps. 【0027】 In step S101, a vector transformation is performed on each reaction sequence in the sample reaction data set through the encoder network of the reaction product prediction model to obtain the sample reactant vector and the sample reaction product vector. The sample reaction data set contains multiple reaction sequences, and each reaction sequence contains a sample reactant and a sample reaction product. 【0028】 Please understand that the sample reaction data set contains multiple reaction sequences. Each reaction sequence is (G r ,G p ) can be expressed as, and each reaction sequence is used to represent one organic chemical reaction data R1+R2+…→P1+P2+…, G r G represents a series of sample reactants. p This represents a series of products, i.e., the sample reaction products. 【0029】 Here, in the molecular graph G=(V,E), V represents an atomic set, and the size of the set is the number of atoms |V|=N, where each atom υ∈V is associated with one atomic feature, including atomic type, chargeability, and aromaticity. E represents a set of edges, where each edge is associated with one bond type, including single bonds, double bonds, triple bonds, and aromatic bonds. The goal of predicting organic chemical reactions is to give reactants and predict the products. 【0030】 Furthermore, following the atom-mapping principle in organic chemical reactions, atoms in the reactants and atoms in the products are mapped one-to-one. Therefore, the changes before and after the reaction are mainly due to changes in the linkage between atoms, i.e., changes in the adjacency matrix A, and no changes occur in the atoms involved. Consequently, when predicting the product (i.e., the reaction product), the overall structure of the product can be reproduced simply by predicting its corresponding adjacency matrix. The adjacency matrix used in this embodiment is different from a general adjacency matrix. Also, the change in the adjacency matrix A, i.e. 【0031】 【number】 Since predicting is easier than directly predicting product A, in this embodiment the problem of predicting organic chemical reaction products is addressed by probability distribution P 【0032】 【number】 It can be converted into a model. 【0033】 Subsequent probability distribution 【0034】 【number】 To make the modeling more effective, first, a sample reaction data set is obtained, and then this sample reaction data set is input into a reaction product prediction model. Vector transformation is then performed on each reaction sequence in the sample reaction data set through the encoder network of the reaction product prediction model, and sample reactant vectors and sample reaction product vectors corresponding to each reaction sequence are obtained, i.e., sample reactant vectors corresponding to the sample reactants in each reaction sequence and sample reaction product vectors corresponding to the sample reaction products. 【0035】 Here, the reaction product prediction model can be specifically represented as a VAE architecture that satisfies the law of conservation of electron transfer, as shown in Figure 16. Furthermore, it can be represented as other models such as a flow-based model, maximum likelihood training, or deep VAE, but is not particularly limited here. 【0036】 Furthermore, if the reaction product prediction model uses a VAE architecture that satisfies the law of conservation of electron transfer energy as shown in Figure 16, the sample reactant vector and sample reaction product vector can be obtained by performing a vector transformation on each reaction sequence in the sample reaction data set through the graph neural network (GNN) and transformer shown in Figure 16. 【0037】 In step S102, a set of positive and negative sample reactants is constructed through the first auxiliary network according to the sample reactant vector. 【0038】 In this embodiment, it is assumed that a predetermined reactant used for predicting reaction products will definitely cause a reaction during the modeling process. However, in an actual scenario, not all reactants will necessarily cause a reaction. Therefore, an accurate reaction product prediction model must have the ability to determine whether a reactant can react. Also, in the embedding space, the distance between molecules that do not react should be far, and the distance between molecules that can react should be close. Thus, this embodiment constructs an auxiliary task for predicting whether a reactant can react. That is, by constructing a positive sample reactant set and a negative sample reactant set through a first auxiliary network, it helps the reaction product prediction model learn the distance between molecules, thereby giving it the ability to determine whether a reactant can react and enhancing the generalization performance of the molecular representation. Thereby, after obtaining the sample reactant vector, the sample reactant vector is input into the first auxiliary network, and a positive sample reactant set and a negative sample reactant set are constructed through the first auxiliary network. 【0039】 Specifically, for Task1 shown in FIG. 15, that is, input the sample reactant vector into the first auxiliary network, and sample the sample reactant vectors corresponding to any two sample reactants from the same reaction sequence, for example, R 11 +R 12 +R 13 →P 11 to obtain positive sample reactant combinations, such as R 11 +R 12 、R 11 +R 13 and R 12 +R 13 etc., and then the positive sample reactant combinations can be added to the positive sample reactant set positive. 【0040】 Furthermore, for different reaction sequences, such as R 11 +R 12 +R 13 →P 11 and R 21 +R 22 +R 23 →P 21Sample reactant vectors corresponding to any two sample reactants are sampled from the negative sample reactant combination, for example, R 11 +R 22 , R 12 +R 23 and R 21 +R 13 By doing so, a negative sample reactant combination can be added to the negative sample reactant set (negative). 【0041】 In step S103, the reaction prediction loss value is determined based on the set of positive and negative sample reactants. 【0042】 In this embodiment, the reaction prediction loss value represents the ability of the reaction product prediction model to determine whether the reactants can react or not. In other words, the reaction prediction loss value is data representing the accuracy of the reaction product prediction model's prediction of the likelihood of a reaction occurring, where the likelihood of a reaction occurring refers to whether the reactants input into the reaction product prediction model can react or not. 【0043】 Specifically, after obtaining the set of positive and negative sample reactants, the following loss function equation (1) (Also called the predetermined first loss function) The reaction prediction loss value can be calculated using this method. 【0044】 【number】 Here, 【0045】 【number】 This represents the set of positive sample reactants, 【0046】 【number】 represents the set of negative sample reactants, (i,j) represents the combination of positive or negative sample reactants sampled from the corresponding set of sample reactants, ε and γ are two margin hyperparameters, and h i and h j These are the sample reactant vector embeddings of sample reactants i and j, respectively. The first term of the loss function equation (1) described above should be understood as being used to reduce the embedding distance between positive sample molecules, and the second term as being used to increase the embedding distance between negative sample molecules. 【0047】 In step S104, a positive sample reaction group set and a negative sample reaction group set are constructed through a second auxiliary network according to the sample reactant vector and the sample reaction product vector. 【0048】 In this embodiment, the ranking of candidate products is important information, and since the reaction product prediction model can receive probability scores for different candidate products, this embodiment constructs an auxiliary task to predict whether reactants and products are correctly associated, that is, it helps the reaction product prediction model learn the relationship between reactants and reaction products through a second auxiliary network, thereby enabling it to determine whether reactants and reaction products are correctly associated and enhancing the generalization performance of the molecular representation. Thus, after obtaining the sample reactant vector and sample reaction product vector, the sample reactant vector and sample reaction product vector are input into the second auxiliary network, and the positive sample reaction group set and the negative sample reaction group set are constructed through the second auxiliary network. 【0049】 Specifically, in Task 2 shown in Figure 15, the sample reactant vector and sample reaction product vector are input to the second auxiliary network, and the corresponding sample reactant vector and sample reaction product vector for all reaction sequences are, for example, R in one reaction sequence. 11 +R 12 +R 13 →P 11 Regarding reactant R 11 +R 12 +R 13 and reaction product P 11 If there is a correct correspondence between the two, then it can be considered a set of positive sample reaction groups. 【0050】 Furthermore, by constructing a negative sample reaction group set by incorrectly combining reactants and reaction products, the matching score between reactants and reaction products with incorrect correspondences can be further reduced. For example, R in one reaction sequence 11 +R 12 +R 13 →P 11 Regarding reactant R 11 +R 12 +R 13 and reaction product P 11 This is a correct correspondence between the two, and R in another reaction sequence. 21 +R 22 +R 23 →P 21 Regarding reactant R 21 +R 22 +R 23 and reaction product P 21 If there is a correct correspondence between the reactants and reaction products, then by incorrectly mixing them, R 11 +R 12 +R 13 →P 21 Ya R 21 +R 22 +R 23 →P 11 For example, a set of negative sample reactions can be obtained in which the correspondence between reactants and reaction products is incorrect. 【0051】 In step S105, the reaction relationship prediction loss value is determined based on the positive sample reaction group set and the negative sample reaction group set. 【0052】 In this embodiment, the reaction relationship prediction loss value represents the ability of the reaction product prediction model to determine whether the reactants and reaction products have a correct correspondence. In other words, the reaction relationship prediction loss value is data representing the prediction accuracy of the reaction relationship by the reaction product prediction model, and this reaction relationship refers to the correspondence between the reactants and the reaction products. Note that the reaction relationship prediction loss value used here and the reaction prediction loss value mentioned in step S103 above are two parallel loss values, and both can be used to train the reaction product prediction model in subsequent steps. However, there is no inherent relationship between the two, and the information they represent is different. The reaction prediction loss value represents the ability of the reaction product prediction model to determine whether a reaction can occur between the reactants, while the reaction relationship prediction loss value represents the ability of the reaction product prediction model to determine whether a correct correspondence exists between the reactants and the reaction products. 【0053】 Specifically, after obtaining the positive sample reaction group set and the negative sample reaction group set, the following loss function equation (2) (Also called a predetermined second loss function) The reaction relationship prediction loss value can be calculated using this method. 【0054】 【number】 During the ceremony, 【0055】 【number】 This is a reactant network, 【0056】 【number】 This is a reaction product network, 【0057】 【number】 and 【0058】 【number】 The two networks have similar architectures but different parameters, and both satisfy the substitution invariance requirement. r This is a feature vector embedding corresponding to the reactant atoms, and h p ε is the feature vector embedding corresponding to the reaction product atoms, and ε and γ are two margin hyperparameters. The first term of this loss function is used to bring the sample reactant vector and sample reaction product vector closer together in identical reaction sequences, i.e., in a positive sample reaction group set, and the second term is used to move the sample reactant vector and sample reaction product vector further apart in miscombined reaction sequences, i.e., in a negative sample reaction group set. Here, it should be understood that the sample size of the target function can be set to B, since the task may be built online during the model training period. 【0059】 In step S106, through a third auxiliary network, the predicted probability values ​​and atomic labels for atoms in the sample reactants present in the main product are identified according to the sample reactant vector and the sample reaction product vector. 【0060】 In this embodiment, since reaction data in massive datasets such as the publicly available dataset USPTO-480K always ignores by-products, which tends to increase prediction errors for reaction products, this embodiment can add a third auxiliary network during the modeling process to predict whether atoms are present in the main product, thereby supplementing information on some by-products to some extent and improving the predictive ability of the reaction products. Thus, after obtaining the sample reactant vector and the sample reaction product vector, the sample reactant vector and the sample reaction product vector are input into the third auxiliary network, and the predicted probability value and atomic label of the atoms in the sample reactant being present in the main product are identified through the third auxiliary network. 【0061】 Specifically, in Task 3 shown in Figure 15, the sample reactant vector and the sample reaction product vector are input to the third auxiliary network, and the sample reactants are predicted through the third auxiliary network to obtain a predicted probability value of the presence of atoms in the sample reactants in the main product. Furthermore, based on the sample reactant vector and the sample reaction product vector, atoms in the sample reactants and atoms in the sample product are compared to obtain an atom comparison result, and atom labels are identified based on the atom comparison result. 【0062】 In step S107, the predicted atomic loss value is identified based on the predicted probability value and the atomic label. 【0063】 In this embodiment, the atomic prediction loss value represents the ability of the reaction product prediction model to determine whether an atom remains in the main product after the reaction. 【0064】 Specifically, after obtaining the predicted probability value and atomic label, the following loss function equation (3) is used. (Also called a predetermined third loss function) The atomic predicted loss value can be calculated using this method. 【0065】 【number】 During the ceremony, 【0066】 【number】 This represents the atomic label of the J atom and indicates whether that atom remains in the main product after the reaction. 【0067】 【number】 This represents the predicted probability value of the model based on atomic embedding. 【0068】 In step S108, a reaction product prediction model is trained based on the reaction prediction loss value, reaction relationship prediction loss value, and atom prediction loss value to obtain a target reaction product prediction model. 【0069】 Specifically, when reaction prediction loss values, reaction relationship prediction loss values, and atom prediction loss values ​​are obtained, a multi-task learning training method is adopted to weight the target functions such as reaction prediction loss values, reaction relationship prediction loss values, and atom prediction loss values ​​and add them directly to the overall target function. Subsequently, parameter adjustments are made to the reaction product prediction model based on the overall target function value. Specifically, parameter adjustments can be made using an iterative backpropagation method, or other methods can be adopted, but this is not limited to this method, and a target reaction product prediction model can be obtained. 【0070】 On the other hand, this embodiment can also employ a preliminary training strategy to pre-train the first, second, and third auxiliary networks based on the reaction prediction loss value, reaction relationship prediction loss value, and atom prediction loss value, respectively, and then transition the pre-trained networks to a reaction product prediction model, thereby obtaining a target reaction product prediction model. 【0071】 The embodiments of this application provide a method for training a reaction product prediction model. As described above, by discovering and constructing positive and negative sample sets and atomic labels from the characteristics of the data itself in the sample reaction data set through a first, second, and third auxiliary network capable of self-supervised learning, it is possible to achieve automatic labeling of data in the sample reaction data set without relying on artificial labeling, thereby reducing the cost required for training the reaction product prediction model. Within this framework, the first auxiliary network helps the reaction product prediction model better learn the distance relationships between reactants, thereby enabling the model to predict whether reactants can react. The second auxiliary network helps the reaction product prediction model better learn the distance relationships between reactants and products, thereby enabling the model to predict the correspondence between reactants and products. The third auxiliary network helps the reaction product prediction model better learn the changes that can occur during the reaction process in the relationship between the bond positions of atoms in the reactants, thereby enabling the model to predict whether atoms remain in the main product after the reaction. Based on the auxiliary learning provided by these auxiliary networks, the accuracy of the reaction product prediction model can be improved. The target reaction product prediction model thus trained can be used as an effective validation tool for drug retrosynthesis, thereby increasing the efficiency of research into new drug synthesis routes. Furthermore, the target reaction product prediction model can reliably predict candidate products even when none of the existing templates are usable, and can predict new reactions to significantly improve the efficiency of new drug development. 【0072】 As an option, based on the embodiment corresponding to Figure 2 described above, in another alternative embodiment of the method for training a reaction product prediction model according to the embodiment of this application, as shown in Figure 3, the method further includes steps S301 to S302 prior to step S101, in which step S101 includes step S303, where step S101 includes step S303. 【0073】 In step S301, data augmentation processing is performed on the sample reaction data set to obtain a sample composite reaction data set. 【0074】 In step S302, the sample reaction data set and the sample combined reaction data set are aggregated as an extended sample reaction data set. 【0075】 In step S303, a vector transformation is performed on each reaction sequence in the extended sample reaction data set through the encoder network of the reaction product prediction model to obtain the sample reactant vector and the sample reaction product vector. 【0076】 In this embodiment, while a typical reaction product prediction model can predict simple chemical reactions involving one or two or three reactants, it struggles to predict more reactants or more complex chemical reactions. In view of this, this embodiment performs data augmentation on the sample reaction data set during the modeling process to obtain a sample composite reaction data set. The sample reaction data set and the sample composite reaction data set are then aggregated as an augmented sample reaction data set. Subsequently, the augmented sample reaction data set is input into the reaction product prediction model, and vector transformation is performed on each reaction sequence in the augmented sample reaction data set through the encoder network of the reaction product prediction model to obtain sample reactant vectors and sample reaction product vectors. This not only augments the sample reaction data set but also enhances its complexity to some extent, thereby helping the reaction product prediction model master the ability to predict complex chemical reactions and thereby improving the accuracy of reaction product prediction to some degree. 【0077】 Specifically, as shown in Task 4 in Figure 15, when performing data augmentation processing on a sample reaction data set, two reaction sequences are randomly selected from the sample reaction data set, sample reactants in the two selected reaction sequences are combined to obtain a sample composite reactant, and sample reaction products in the two selected reaction sequences are combined to obtain a sample composite reaction product. Subsequently, a composite reaction sequence is obtained based on the sample composite reactant and sample composite reaction product, and a sample composite reaction data set is constructed based on the composite reaction sequence. Other augmentation methods may also be employed, but there are no particular limitations here. For example, R in one reaction sequence 11 +R 12 +R 13 →P 11 Reactant R in 11 +R 12 +R 13 and reaction product P 11 R in another reaction sequence 21 +R 22 +R 23 →P21 Reactant R in 21 +R 22 +R 23 and reaction product P 21 By combining them, a single complex reaction sequence is formed, for example, R 11 +R 12 +R 13 +R 21 +R 22 +R 23 →P 11 +P 21 You can obtain this. 【0078】 Furthermore, by aggregating the acquired sample reaction data set and sample composite reaction data set as an extended sample reaction data set, it becomes easier to later train a reaction product prediction model using the extended sample reaction data set, which helps the reaction product prediction model master the ability to predict complex chemical reactions. Specifically, by inputting the extended sample reaction data set into the reaction product prediction model, performing a vector transformation on each reaction sequence in the extended sample reaction data set through the reaction product prediction model's encoder network, and obtaining sample reactant vectors and sample reaction product vectors, the reaction product prediction model can then be trained by later connecting the aforementioned first auxiliary network, second auxiliary network, and third auxiliary network to the sample reactant vectors and sample reaction product vectors. 【0079】 As an option, in another alternative embodiment of the method for training a reaction product prediction model according to the embodiment of this application, based on the embodiment corresponding to Figure 2 or Figure 3 described above, step S102, which constructs a set of positive and negative sample reactants according to the sample reactant vector through a first auxiliary network as shown in Figure 4, includes the following steps: 【0080】 In step S401, sample reactant vectors corresponding to any two sample reactants are sampled from the same reaction sequence to form a positive sample reactant combination, and this positive sample reactant combination is added to the set of positive sample reactants. 【0081】 In step S402, sample reactant vectors corresponding to any two sample reactants are sampled from different reaction sequences to form a negative sample reactant combination, and this negative sample reactant combination is added to the negative sample reactant set. 【0082】 Specifically, in the embedding space, the distance between molecules where a reaction does not occur should be large, and the distance between molecules where a reaction can occur should be small. Therefore, after obtaining the sample reactant vector corresponding to each reaction sequence, chemical reaction data, for example, 【0083】 【number】 Therefore, here, 【0084】 【number】 and 【0085】 【number】 These represent sample reactants, 【0086】 【number】 and 【0087】 【number】 These represent sample reaction products, and reactant combination data is constructed by combining pairs of reactants. 【0088】 Furthermore, as shown in FIG. 15, for the convenience of display, in this embodiment, reaction data B of one batch size is given to construct two spaces. Among them, one is the positive sample space, that is, the positive sample reactant set, and the other is the negative sample space, that is, the negative sample reactant set. The task Task1 shown in FIG. 15, that is, input the sample reactant vector into the first auxiliary network, and the positive sample reactant set 【0089】 [Number] is constructed based on the predetermined reaction data in the chemical data B, that is, sample reactant vectors corresponding to any two sample reactants from the same reaction sequence are sampled and combined. Assuming that there are R reactants in each reaction, it is possible to obtain R(R - 1) combinations of positive sample reactant combinations. The negative sample reactant set 【0090】 [Number] is obtained by combining two sample reactants from different reaction sequences, that is, sample reactant vectors corresponding to any two sample reactants from different reaction sequences are sampled and combined. 【0091】 For example, as shown in FIG. 15, for the same reaction sequence, such as R 11 +R 12 +R 13 →P 11 sample the sample reactant vectors corresponding to any two sample reactants to form positive sample reactant combinations, such as R 11 +R 12 、R 11 +R 13 and R 12 +R 13 etc., so that they can be aggregated into the positive sample reactant set positive. 【0092】 Furthermore, for different reaction sequences, such as R 11+R 12 +R 13 →P 11 and R 21 +R 22 +R 23 →P 21 Sampling any two sample reactant vectors from a negative sample reactant combination, for example, R 11 +R 22 , R 12 +R 23 and R 21 +R 13 By doing so, the negative sample reactants can be aggregated into a negative set. 【0093】 Thus, by constructing a set of positive and negative sample reactants as described above, the first auxiliary network can better assist the reaction product prediction model in learning the distance relationships between reactants, allowing the reaction product prediction model to predict with greater accuracy whether reactants can react or not. 【0094】 As an option, in another alternative embodiment of the method for training a reaction product prediction model according to the embodiment of this application, based on the embodiment corresponding to Figure 2 or Figure 3 described above, step S104, which constructs a set of positive sample reaction groups and a set of negative sample reaction groups according to the sample reactant vector and the sample reaction product vector through a second auxiliary network as shown in Figure 5, includes the following steps: 【0095】 In step S501, the corresponding sample reactant vectors and sample reaction product vectors for all reaction sequences are set as the positive sample reaction group. 【0096】 In step S502, a negative sample reaction group is obtained by combining the corresponding sample reactant vectors and sample reaction product vectors of different reaction sequences. 【0097】 Specifically, by providing a batch-sized chemical reaction data B = {R1 → P1, R2 → P2, ...} and designating the combination of reactants, for example R1 and reaction product P1, given from that batch as a positive sample reaction group, i.e., R1 → P1, the chemical reaction data B can be made into a set of positive sample reaction groups. Specifically, the sample reactant vector and sample reaction product vector are input into Task 2, i.e., the second auxiliary network shown in Figure 15, and the corresponding sample reactant vector and sample reaction product vector for all reaction sequences are, for example, R in one reaction sequence. 11 +R 12 +R 13 →P 11 Regarding reactant R 11 +R 12 +R 13 and reaction product P 11 If there is a correct correspondence between the two, then it can be considered one of the positive sample reaction groups within the set of positive sample reaction groups. 【0098】 Furthermore, the negative sample reaction group set is constructed by incorrectly combining reactants and reaction products in the chemical reaction data B={R1→P1,R2→P2,…}, and by incorrectly combining the sample reactant vector and the sample reaction product vector to obtain the negative sample reaction group, the matching score between reactants and reaction products with incorrect correspondences can be made lower. For example, as shown in Figure 15, R in one reaction sequence 11 +R 12 +R 13 →P 11 Regarding reactant R 11 +R 12 +R 13 and reaction product P 11 This is a correct correspondence between the two, and R in another reaction sequence. 21 +R 22 +R 23 →P 21 Regarding reactant R 21 +R 22 +R 23 and reaction product P 21If there is a correct correspondence between the reactants and reaction products, then by incorrectly mixing them, R 11 +R 12 +R 13 →P 21 Ya R 21 +R 22 +R 23 →P 11 For example, a group of negative sample reactions can be obtained in which the correspondence between reactants and reaction products is incorrect. 【0099】 Thus, by constructing a positive sample reaction group set and a negative sample reaction group set as described above, the second auxiliary network can better assist the reaction product prediction model in learning the distance relationship between reactants and products, and the reaction product prediction model can predict the correspondence between reactants and products with greater accuracy. 【0100】 As an option, in another alternative embodiment of the method for training a reaction product prediction model according to the embodiment of this application, based on the embodiment described above in Figure 2 or Figure 3, step S106, as shown in Figure 6, identifies the predicted probability value and atomic label of atoms in the sample reactant present in the main product according to the sample reactant vector and the sample reaction product vector through a third auxiliary network, and includes the following steps: 【0101】 In step S601, the sample reactants are predicted through a third auxiliary network, and a predicted probability value is obtained for the presence of atoms in the sample reactants in the main product. 【0102】 In step S602, the atoms in the sample reactants and sample products are compared based on the sample reactant vector and the sample reaction product vector, and the atomic comparison results are obtained. 【0103】 In step S603, the atomic labels are identified based on the atomic comparison results. 【0104】 In this embodiment, predicting whether atoms are present in the main product and predicting the reaction center are two non-equivalent tasks. Predicting the reaction center only allows us to predict which bonds will break, and does not allow us to know which of the two atoms linked to those bonds is present in the byproduct. Therefore, a third auxiliary network for predicting whether atoms are present in the main product can teach the reaction product prediction model the relative importance of atoms in the reaction. Through this, the sample reactants can be predicted via the third auxiliary network, and a predicted probability value for the presence of atoms in the sample reactants in the main product can be obtained. Furthermore, based on the sample reactant vector and the sample reaction product vector, atoms in the sample reactants and atoms in the sample product can be compared to obtain corresponding atom comparison results, and then atom labels can be identified based on the atom comparison results. 【0105】 Specifically, when predicting sample reactants through a third auxiliary network and obtaining predicted probability values ​​for the presence of atoms in the sample reactants in the main product, the task of predicting whether or not atoms are present in the main product can be masked as existing information using a masking matrix. This allows the reaction product prediction model to focus only on the main product information. However, in reality, this masking matrix is ​​also the content that should be predicted by the reaction product prediction model. 【0106】 Furthermore, the method for obtaining atomic ground-truth labels can be specifically based on the sample reactant vector and the sample reaction product vector. This involves comparing the atoms of the sample reactant with those of the sample product, then marking atoms discarded from the reactant as 0 and atoms still present in the reaction product as 1. Other labeling methods can also be employed, but are not particularly limited here. 【0107】 Thus, as described above, the third auxiliary network can better assist the reaction product prediction model in learning the possible changes in the relationship between the bond positions of atoms in the reactants during the reaction process, allowing the reaction product prediction model to predict whether atoms will remain in the main product after the reaction. 【0108】 As an option, in another alternative embodiment of the training method for the reaction product prediction model according to the embodiment of this application, based on the embodiment corresponding to Figure 2 or Figure 3 described above, step S101, as shown in Figure 7, involves performing a vector transformation on each reaction sequence in the sample reaction data set through the reaction product prediction model's encoder network to obtain a sample reactant vector and a sample reaction product vector, and includes the following steps: 【0109】 In step S701, for each reaction sequence, information about the interactions between atoms of each molecule of the sample reactant is obtained by allowing them to interact with each other. 【0110】 In step S702, based on the sample reactant interaction information, different sample reactants are made to interact, and a sample reactant vector is obtained. 【0111】 In step S703, for each reaction sequence, information about the interactions between atoms of each molecule of the sample reaction product is obtained by allowing them to interact with each other. 【0112】 In step S704, based on the sample reaction product interaction information, different sample reactants are made to interact, and a sample reaction product vector is obtained. 【0113】 Specifically, if the reaction product prediction model uses a VAE architecture that satisfies the law of conservation of electron transfer energy as shown in Figure 16, in this embodiment, the encoder network that performs vector transformations on each reaction sequence in the sample reaction data set may be a combination of a graph neural network (GNN) and a transformer. 【0114】 Here, the Transformer consists of an encoder and a decoder, can utilize a self-awareness mechanism, and does not employ the hierarchical structure of recurrent neural networks (RNNs) and long-term short-term memory networks (LSTMs). Therefore, the model can be trained in parallel and can possess global information. 【0115】 Here, graph neural networks (GNNs) include graph convolutional networks, graph attention networks, graph autoencoders, graph generative networks, and graph spatiotemporal networks. Compared to the fully connected layer (MLP), which is the most basic layer of a neural network, graph neural networks have an additional adjacency matrix in addition to multiplying the feature matrix by the weight matrix. 【0116】 This allows the information within each molecule to interact with each other via the GNN using equation (4) below. In other words, for each reaction sequence, the information within each molecule of the sample reactants is interacted with each other to obtain sample reactant interaction information. Then, information is interacted between different reactants via a Transformer, that is, different sample reactants are interacted based on the sample reactant interaction information to obtain a sample reactant vector. 【0117】 【number】 In the formula, h R represents the sample reactant vector, GR This represents the sample reactant. 【0118】 Similarly, using equation (5) below, information within each molecule can be made to interact between atoms via the GNN; that is, for each reaction sequence, information within each molecule of the sample reaction product can be made to interact between atoms to obtain sample reaction product interaction information. Then, information can be made to interact between different reaction products via a Transformer; that is, different sample reaction products can interact based on the sample reaction product interaction information to obtain a sample reaction product vector. 【0119】 【number】 In the formula, h P represents the sample reaction product vector, G P This represents the sample reaction product. 【0120】 Thus, by obtaining the sample reactant vector and sample reaction product vector in the manner described above, it is possible to ensure that the obtained sample reactant vector and sample reaction product vector more accurately reflect the characteristics of the corresponding sample reactant and sample reaction product. Furthermore, this can help the first, second, and third auxiliary networks to better learn the relevant capabilities based on these sample reactant vectors and sample reaction product vectors, thereby improving the performance of the trained reaction product prediction model. 【0121】 As an option, in another alternative embodiment of the training method for the reaction product prediction model according to the embodiment of this application, based on the embodiment corresponding to Figure 3 described above, step S301, in which data augmentation processing is performed on the sample reaction data set to obtain a sample composite reaction data set, as shown in Figure 8, includes the following steps. 【0122】 In step S801, two reaction sequences are randomly selected from the sample reaction data set. 【0123】 In step S802, the sample reactants in the two selected reaction sequences are combined to obtain a sample composite reactant. 【0124】 In step S803, the sample reaction products from the two selected reaction sequences are combined to obtain a sample composite reaction product. 【0125】 In step S804, a composite reaction sequence is obtained based on the sample composite reactant and the sample composite reaction product, and a sample composite reaction data set is constructed based on the composite reaction sequence. 【0126】 In this embodiment, when performing data augmentation processing on the sample reaction data set, two reaction sequences may be randomly selected from the sample reaction data set, sample reactants in the two selected reaction sequences may be combined to obtain a sample composite reactant, and sample reaction products in the two selected reaction sequences may be combined to obtain a sample composite reaction product. Subsequently, a composite reaction sequence may be obtained based on the sample composite reactant and sample composite reaction product, and a sample composite reaction data set may be constructed based on this. Even with such a data augmentation method, it is not possible to obtain a completely accurate chemical reaction. For example, the reaction between new reactant combinations may not be A and B, E and F, but rather A and E, B and F, or other combinations may occur. Alternatively, even if A and B, E and F react, they may only produce an intermediate product, and the reaction may proceed further to obtain the final product. In other words, while new reactions obtained through a randomly combined data augmentation method are not necessarily accurate, they can still enhance the generalization performance of the reaction product prediction model. Furthermore, since erroneous reaction data always exist in the commonly used USPTO-480K dataset (for example, the ground-truth product given is an intermediate product, not the final product), the inaccurate new reaction combinations obtained through the aforementioned data augmentation method do not significantly affect the prediction accuracy of the reaction product prediction model. Therefore, the sample composite reaction data set obtained through data augmentation and the original sample reaction data set can be aggregated to obtain an augmented sample reaction data set, which can then be applied to subsequent parameter tuning of the reaction product prediction model. This enhances the robustness and generalization performance of the reaction product prediction model, provides strong scalability to larger datasets, and enables supervised learning tasks without relying on artificial labeling. 【0127】 Specifically, as shown in Task 4 in Figure 15, R in one reaction sequence 11 +R 12 +R 13 →P 11 Reactant R in 11 +R 12 +R13 and reaction product P 11 R in another reaction sequence 21 +R 22 +R 23 →P 21 Reactant R in 21 +R 22 +R 23 and reaction product P 21 By combining them, a single complex reaction sequence is formed, for example, R 11 +R 12 +R 13 +R 21 +R 22 +R 23 →P 11 +P 21 You can obtain this. 【0128】 As an option, in another alternative embodiment of the method for training a reaction product prediction model according to the embodiment of this application, based on the embodiment corresponding to Figure 2 described above, step S108, which trains a reaction product prediction model based on reaction prediction loss values, reaction relationship prediction loss values ​​and atomic prediction loss values ​​to obtain a target reaction product prediction model, includes the following steps, as shown in Figure 9. 【0129】 In step S901, the reconstruction loss value and the divergence loss value are obtained. 【0130】 In step S902, a reaction product prediction model is trained based on the reconstruction loss value, divergence loss value, reaction prediction loss value, reaction relationship prediction loss value, and atom prediction loss value to obtain a target reaction product prediction model. 【0131】 Specifically, if the reaction product prediction model uses a VAE architecture that satisfies the law of conservation of electron transfer energy as shown in Figure 16, this embodiment can employ a multi-task learning training method, and reconstruction loss values ​​and divergence loss values ​​can be obtained based on the VAE architecture. 【0132】 Furthermore, based on the overall target function equation (6) shown below, target functions such as reconstruction loss, divergence loss, reaction prediction loss, reaction relationship prediction loss, and atom prediction loss can be weighted and added directly to the overall target function, thereby obtaining the overall target function value. 【0133】 【number】 During the ceremony, here, 【0134】 【number】 This represents the total target function value, 【0135】 【number】 This represents the reconstruction loss value, 【0136】 【number】 represents the divergence loss value, L A L represents the reaction prediction loss value. B L represents the reaction relationship prediction loss value. C θ represents the predicted atomic loss value, and α, β, λ, and θ are weighting parameters set according to practical usage needs and are not particularly limited here. 【0137】 Furthermore, when adjusting the parameters of the reaction product prediction model based on the total target function value, the parameters can be adjusted using a specific backpropagation iterative method, or other methods can be employed without any particular restriction, thereby enabling the acquisition of a target reaction product prediction model. 【0138】 Thus, by training a reaction product prediction model based on the aforementioned reconstruction loss value, divergence loss value, reaction prediction loss value, reaction relationship prediction loss value, and atom prediction loss value, it is possible to improve the performance of the trained reaction product prediction model and also improve the training efficiency of the reaction product prediction model. 【0139】 As an option, in another alternative embodiment of the training method for the reaction product prediction model according to the embodiment of this application, based on the embodiment corresponding to Figure 9 described above, step S901, in which the reconstruction loss value and divergence loss value are obtained, as shown in Figure 10, includes the following steps. 【0140】 In step S1001, the sample reactant vector and the sample reaction product vector are processed using a cautionary mechanism to obtain a first hidden vector. 【0141】 In step S1002, the second hidden vector is determined based on the first hidden vector and the sample reactant vector. 【0142】 In step S1003, the sample prediction change probability of the adjacency matrix is ​​identified according to the second hidden vector through the decoder network of the reaction product prediction model. 【0143】 In step S1004, the sample predicted reaction product adjacency matrix is ​​identified based on the sample predicted change probability. 【0144】 In step S1005, the loss is calculated based on the reaction product adjacency matrix of the sample reaction product and the sample predicted reaction product adjacency matrix to obtain the reconstruction loss value and the divergence loss value. 【0145】 Specifically, as shown in Figure 16, an attention mechanism can be applied to the sample reactant vector and the sample reaction product vector, that is, reactant embedding and product embedding can be introduced into a single-layer cross-attention mechanism. Here, in order to obtain the parameter vectors μ and logσ of the Gaussian distribution, cross-attention can be directly implemented through a transformer decoder based on the following equations (7), (8), and (9), and a first hidden vector that fits the Gaussian distribution can be obtained using the reparameterization technique. 【0146】 【number】 【0147】 【number】 【0148】 【number】 In the formula, h z This represents the first hidden vector, 【0149】 【number】 These represent model parameters. 【0150】 Furthermore, when determining the second hidden vector based on the first hidden vector and the sample reactant vector, specifically, by adding the first hidden vector and the sample reactant vector and introducing them into the Transformer layer based on equation (10) below, a new hidden vector, i.e., the second hidden vector h L You can obtain this. 【0151】 【number】 Furthermore, as shown in Figure 16, the second hidden vector is input to the decoder network of the reaction product prediction model, and the second hidden vector is processed through the decoder network. 【0152】 【number】 The following is decoded. Here, the decoder network (decoder) mainly consists of two separate self-attention mechanisms. One self-attention mechanism outputs an attention matrix as the probability of an increase in shared electrons between two atoms, and the other self-attention mechanism outputs an attention matrix as the probability of a decrease in shared electrons between two atoms. Since energy is conserved during electron transfer, this property is modeled using the weighting matrix properties of the self-attention mechanisms, and the sample prediction change probability of the adjacency matrix is ​​obtained by performing a specific decoding operation using equations (11) and (12) below. 【0153】 【number】 【0154】 【number】 During the ceremony, 【0155】 【number】 This represents the predicted change probability for one sample, i.e., the probability matrix of electron increase between each pair of atoms. 【0156】 【number】 This represents the probability of another sample change, i.e., the probability matrix of electron loss between each pair of atoms. 【0157】 Furthermore, when calculating the sample predicted reaction product adjacency matrix based on the sample predicted change probability, the sample predicted change amount in the adjacency matrix is ​​calculated based on the sample predicted change probability, and the sample predicted reaction product adjacency matrix is ​​calculated based on the sample predicted change amount in the adjacency matrix and the adjacency matrix of the sample reactants. Subsequently, the cross-entropy loss is calculated based on the reaction product adjacency matrix of the sample reaction products and the sample predicted reaction product adjacency matrix, thereby obtaining the reconstruction loss value and the divergence loss value. 【0158】 In this embodiment, the numerical values ​​in the adjacency matrix are not only 0 or 1 (0 indicates no linkage between corresponding atom pairs, and 1 indicates a linkage between corresponding atom pairs), but also four numerical values: 0, 1, 2, and 3, which represent unlinked, linked by single bonds, linked by double bonds, and linked by triple bonds, respectively. Linkage by aromatic bonds is represented by 1, and the linked atom is denoted as an aromatic atom to distinguish it from single-bonded links. 【0159】 As an option, in another alternative embodiment of the training method for the reaction product prediction model according to the embodiment of this application, based on the embodiment described above in Figure 10, step S1004, which identifies the sample predicted reaction product adjacency matrix based on the sample predicted change probability, includes the following steps, as shown in Figure 11. 【0160】 In step S1101, the sample predicted change amount of the adjacency matrix is ​​calculated based on the sample predicted change probability. 【0161】 In step S1102, the sample predicted reaction product adjacency matrix is ​​calculated based on the sample predicted change amount and the sample reactant adjacency matrix of the adjacency matrix. 【0162】 Specifically, after obtaining the sample predicted change probability, the sample predicted change amount of the adjacency matrix is ​​calculated based on the sample predicted change probability using equation (13) below, i.e., the probability matrix of electron increase between each pair of atoms. 【0163】 【Mathematics】 The probability matrix of electron reduction between each pair of atoms 【0164】 【Mathematics】 Multiply the result of subtracting by 4 to obtain the predicted sample change amount between each pair of atoms 【0165】 【Mathematics】 Can be obtained 【0166】 【Mathematics】 Here, 4 represents the maximum change amount between each pair of atoms 【0167】 Furthermore, using the following formula (14), based on the predicted sample change amount of the adjacency matrix and the adjacency matrix of the sample reactants, calculate the predicted sample reaction product adjacency matrix 【0168】 【Mathematics】 Here 【0169】 【Mathematics】 Represents the predicted sample reaction product adjacency matrix 【0170】 【Mathematics】 Represents the adjacency matrix of the sample reactants 【0171】 Furthermore, since the adjacency matrix must be symmetric, the predicted sample reaction product adjacency matrix can be symmetrized using the following formula (15) 【0172】 【Number】 As an option, based on the embodiment corresponding to FIG. 2 described above, in another alternative embodiment of the training method of the reaction product prediction model according to the embodiment of the present application, as shown in FIG. 12, based on the reaction prediction loss value, the reaction relationship prediction loss value, and the atom prediction loss value, the step S108 of training the reaction product prediction model to obtain the target reaction product prediction model includes the following steps. 【0173】 In step S1201, parameter adjustment is performed on the first auxiliary network based on the reaction prediction loss value to obtain a first sub-model. 【0174】 In step S1202, parameter adjustment is performed on the second auxiliary network based on the reaction relationship prediction loss value to obtain a second sub-model. 【0175】 In step S1203, parameter adjustment is performed on the third auxiliary network based on the atom prediction loss value to obtain a third sub-model. 【0176】 In step S1204, the first sub-model, the second sub-model, and the third sub-model are transitioned to the reaction product prediction model to obtain the target reaction product prediction model. 【0177】 Specifically, this embodiment further employs a preliminary training strategy to pre-train the first, second, and third auxiliary networks based on the reaction prediction loss value, reaction relationship prediction loss value, and atom prediction loss value, respectively. That is, the first submodel is obtained by adjusting the parameters of the first auxiliary network based on the reaction prediction loss value, the second submodel is obtained by adjusting the parameters of the second auxiliary network based on the reaction relationship prediction loss value, and the third submodel is obtained by adjusting the parameters of the third auxiliary network based on the atom prediction loss value. Subsequently, the first, second, and third submodels are transitioned into a reaction product prediction model. Specifically, the skeletal networks of the first, second, and third submodels are transitioned into a reaction product prediction model. After that, a decoder network layer that predicts the reaction products is placed after the layer that outputs embeddings from the skeletal networks, thereby obtaining a target reaction product prediction model. 【0178】 By pre-training the first, second, and third auxiliary networks based on this preliminary training strategy, and then transitioning the trained first, second, and third auxiliary networks to the reaction product prediction model, the training efficiency of the reaction product prediction model can be effectively improved, and this also contributes to improving the performance of the reaction product prediction model. 【0179】 Next, a method for applying the reaction product prediction model according to this application will be described. This method may be performed by a computer device, which may be a server or a terminal device. Referring to Figure 13, one embodiment of the method for applying the reaction product prediction model according to the embodiment of this application includes the following steps. 【0180】 In step S1301, the reactants to be measured are input into the target reaction product prediction model, and the predicted change probability of the adjacency matrix is ​​output from the target reaction product prediction model. 【0181】 In step S1302, the predicted change in the adjacency matrix is ​​determined based on the predicted change probability of the adjacency matrix. 【0182】 In step S1303, the target reaction product is identified based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured. 【0183】 In this embodiment, after obtaining the reactants to be measured, the reactants to be measured are input into a target reaction product prediction model. The target reaction product prediction model outputs the predicted change probability of the adjacency matrix, and the predicted change amount of the adjacency matrix can be calculated based on the predicted change probability of the adjacency matrix. Subsequently, the target reaction product can be identified based on the predicted change amount of the adjacency matrix and the adjacency matrix of the reactants to be measured. The target reaction product obtained in this way can be applied to scenarios such as chemical pharmaceuticals and drug verification. 【0184】 It should be understood that target reaction product prediction models can be an effective validation tool for drug retrosynthesis, increasing the efficiency of research into new drug synthesis pathways; they can reveal several underlying scientific laws and provide new scientific knowledge; they can provide more accurate predictions than experts; they can reliably predict candidate products even when none of the existing templates are usable, and they can also predict new reactions, thereby significantly improving the efficiency of new drug development. 【0185】 Specifically, the reactants to be measured are input into a target reaction product prediction model, and the predicted change probability of the adjacency matrix is ​​output from the target reaction product prediction model, i.e., the probability matrix of electron increase between each pair of atoms. 【0186】 【number】 and the probability matrix of electron loss between each pair of atoms 【0187】 【number】 Based on this, the predicted change in the adjacency matrix is ​​then calculated using the aforementioned equation (13). 【0188】 【number】 It is possible to calculate this. 【0189】 As an option, in another alternative embodiment of the method for applying the reaction product prediction model according to the embodiment of this application, based on the embodiment described above in Figure 13, step S1303, which identifies the target reaction product based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured, includes the following steps, as shown in Figure 14. 【0190】 In step S1401, the predicted reaction product adjacency matrix is ​​calculated based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured. 【0191】 In step S1402, the predicted reaction product adjacency matrix is ​​symmetrized to obtain the target reaction product adjacency matrix. 【0192】 In step S1403, the target reaction product is identified based on the target reaction product adjacency matrix. 【0193】 Specifically, using equation (14) described above, the predicted reaction product adjacency matrix can be calculated based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured. In the equation, 【0194】 【number】 This represents the predicted reaction product adjacency matrix, 【0195】 【number】 This represents the adjacency matrix of the reactants to be measured. 【0196】 Furthermore, since the adjacency matrix must be symmetric, we can use equation (15) above to perform a symmetrization process on the predicted reaction product adjacency matrix to obtain the target reaction product adjacency matrix, and then the target reaction product adjacency matrix 【0197】 【number】 Based on this, the target reaction product can be derived. 【0198】 Next, the training apparatus for the reaction product prediction model according to this application will be described in detail. Referring to Figure 17, Figure 17 is a schematic diagram of one embodiment of the training apparatus for the reaction product prediction model according to an embodiment of this application. The training apparatus for the reaction product prediction model 20 is an acquisition unit 201 configured to perform vector transformation on each reaction sequence in the sample reaction data set through a reaction product prediction model encoder network to obtain a sample reactant vector and a sample reaction product vector, further construct a set of positive sample reactants and a set of negative sample reactants according to the sample reactant vector through a first auxiliary network, construct a set of positive sample reaction groups and a set of negative sample reaction groups according to the sample reactant vector and the sample reaction product vector through a second auxiliary network, and determine the predicted probability value and atomic label of atoms in the sample reactants present in the main product according to the sample reactant vector and the sample reaction product vector, wherein the set of sample reaction data includes a plurality of reaction sequences, and each reaction sequence includes a sample reactant and a sample reaction product, A processing unit 202 is configured to determine a reaction prediction loss value based on a set of positive sample reactants and a set of negative sample reactants, to determine a reaction relationship prediction loss value based on a set of positive sample reaction groups and a set of negative sample reaction groups, and further to determine an atomic prediction loss value based on a prediction probability value and an atomic label. The system includes a specific unit 203 configured to train a reaction product prediction model based on reaction prediction loss values, reaction relationship prediction loss values, and atom prediction loss values, in order to obtain a target reaction product prediction model. 【0199】 As an option, based on the embodiment corresponding to Figure 17 described above, in another embodiment of the training apparatus for the reaction product prediction model according to the embodiment of this application, The processing unit 202 is further configured to perform data augmentation processing on the sample reaction data set to obtain a sample composite reaction data set. The processing unit 202 is further configured to aggregate the sample reaction data set and the sample composite reaction data set as an extended sample reaction data set. Specifically, the acquisition unit 201 is configured to perform vector transformations on each reaction sequence in the extended sample reaction data set through an encoder network of reaction product prediction models to obtain sample reactant vectors and sample reaction product vectors. 【0200】 As an option, based on the embodiment corresponding to Figure 17 described above, in another embodiment of the training apparatus for the reaction product prediction model according to the embodiment of this application, the acquisition unit 201 specifically is: The system is configured to sample sample reactant vectors corresponding to any two sample reactants from the same reaction sequence to form a positive sample reactant combination, and then add this positive sample reactant combination to the set of positive sample reactants. The system is configured to sample sample reactant vectors corresponding to any two sample reactants from different reaction sequences to form a negative sample reactant combination, and then add this negative sample reactant combination to the negative sample reactant set. 【0201】 As an option, based on the embodiment corresponding to Figure 17 described above, in another embodiment of the training apparatus for the reaction product prediction model according to the embodiment of this application, the acquisition unit 201 specifically is: The system is configured such that the corresponding sample reactant vectors and sample reaction product vectors for all reaction sequences form a positive sample reaction group set. The system is configured to obtain a negative sample reaction group set by combining corresponding sample reactant vectors and sample reaction product vectors for different reaction sequences. 【0202】 As an option, based on the embodiment corresponding to Figure 17 described above, in another embodiment of the training apparatus for the reaction product prediction model according to the embodiment of this application, the acquisition unit 201 specifically is: It is configured to predict the sample reactants through a third auxiliary network and obtain a predicted probability value for the presence of atoms in the sample reactants in the main product. Based on the sample reactant vector and the sample reaction product vector, it is configured to compare atoms in the sample reactant and the sample reaction product and obtain an atomic comparison result. It is configured to identify atomic labels based on the results of atomic comparison. 【0203】 As an option, based on the embodiment corresponding to Figure 17 described above, in another embodiment of the training apparatus for the reaction product prediction model according to the embodiment of this application, the acquisition unit 201 specifically is: For each reaction sequence, it is configured to obtain information about the interaction between atoms of each molecule of the sample reactants within that sequence, thereby obtaining information about the interaction between the sample reactants. Based on sample reactant interaction information, it is configured to interact with different sample reactants and obtain a sample reactant vector. For each reaction sequence, the system is configured to obtain interaction information between atoms by allowing information about the internal structure of each molecule of the sample reaction product within that sequence to be generated. Based on sample reaction product interaction information, the system is configured to interact with different sample reactants to obtain a sample reaction product vector. 【0204】 As an option, based on the embodiment corresponding to Figure 17 described above, in another embodiment of the training apparatus for the reaction product prediction model according to the embodiment of this application, the processing unit 202 specifically, It is configured to randomly select two reaction sequences from a sample reaction data set. It is configured to combine sample reactants in two selected reaction sequences to obtain a sample composite reactant. It is configured to combine the sample reaction products in two selected reaction sequences to obtain a sample composite reaction product. The system is configured to obtain a complex reaction sequence based on the sample complex reactant and the sample complex reaction product, and to construct a sample complex reaction data set based on the complex reaction sequence. 【0205】 As an option, based on the embodiment corresponding to Figure 17 described above, in another embodiment of the training apparatus for the reaction product prediction model according to the embodiment of this application, the specific unit 203 is specifically: The system is configured to acquire reconstruction loss values ​​and divergence loss values. The system is configured to train a reaction product prediction model based on reconstruction loss values, divergence loss values, reaction prediction loss values, reaction relationship prediction loss values, and atom prediction loss values, in order to obtain a target reaction product prediction model. 【0206】 As an option, based on the embodiment corresponding to Figure 17 described above, in another embodiment of the training apparatus for the reaction product prediction model according to the embodiment of this application, the specific unit 203 is specifically: The system is configured to process the sample reactant vector and the sample reaction product vector using a cautionary mechanism to obtain a first hidden vector. It is configured to determine a second hidden vector based on the first hidden vector and the sample reactant vector. The reaction product prediction model is configured to identify the sample prediction change probability of the adjacency matrix according to a second hidden vector through a decoder network. It is configured to identify the sample predicted reaction product adjacency matrix based on the sample predicted change probability, The system is configured to calculate the loss based on the reaction product adjacency matrix of the sample reaction product and the sample predicted reaction product adjacency matrix, and to obtain the reconstructed loss value and the divergence loss value. 【0207】 As an option, based on the embodiment corresponding to Figure 17 described above, in another embodiment of the training apparatus for the reaction product prediction model according to the embodiment of this application, the specific unit 203 is specifically: It is configured to calculate the sample-predicted change in the adjacency matrix based on the sample-predicted change probability, The system is configured to calculate the sample prediction reaction product adjacency matrix based on the sample prediction change in the adjacency matrix and the sample reactant adjacency matrix. 【0208】 As an option, based on the embodiment corresponding to Figure 17 described above, in another embodiment of the training apparatus for the reaction product prediction model according to the embodiment of this application, the specific unit 203 is specifically: The system is configured to perform parameter adjustments on the first auxiliary network based on the reaction prediction loss value to obtain a first submodel. The system is configured to perform parameter adjustments on a second auxiliary network based on the reaction relationship prediction loss value to obtain a second submodel. The system is configured to perform parameter adjustments on a third auxiliary network based on predicted atomic loss values ​​to obtain a third submodel. The system is configured to transition the first submodel, the second submodel, and the third submodel into a reaction product prediction model to obtain a target reaction product prediction model. 【0209】 Next, the application apparatus for the reaction product prediction model in this application will be described in detail. Referring to Figure 18, Figure 18 is a schematic diagram of one embodiment of the application apparatus for the reaction product prediction model according to the embodiment of this application. The application apparatus 30 for the reaction product prediction model consists of the following units: The acquisition unit 301 is configured to input the reactants to be measured into a target reaction product prediction model and output the predicted change probability of the adjacency matrix from the target reaction product prediction model, wherein the target reaction product prediction model is trained by the reaction product prediction model training device shown in Figure 17. A processing unit 302 is configured to identify the predicted change amount of the adjacency matrix based on the predicted change probability of the adjacency matrix, The system includes a identifying unit 303 configured to identify a target reaction product based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured. 【0210】 As an option, based on the embodiment corresponding to Figure 18 described above, in another embodiment of the application apparatus for the reaction product prediction model according to the embodiment of this application, the specific unit 303 is specifically: It is configured to calculate the predicted reaction product adjacency matrix based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured. The system is configured to perform a symmetrization process on the predicted reaction product adjacency matrix to obtain the target reaction product adjacency matrix. It is configured to identify the target reaction product based on the target reaction product adjacency matrix. 【0211】 Another aspect of this application provides a schematic diagram of another computer device. As shown in Figure 19, Figure 19 is a schematic configuration diagram of a computer device according to an embodiment of this application. This computer device 300 can vary considerably in configuration or performance and may include one or more central processing units (CPUs) 310 (e.g., one or more processors), memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 331 or data 332. Here, the memory 320 and the storage media 330 may be temporary or permanent storage. The programs stored in the storage media 330 may include one or more modules (not shown) that include a set of instruction operations in the computer device 300. Furthermore, the central processing unit 310 communicates with the storage media 330 and is configured to execute the set of instruction operations in the storage media 330 on the computer device 300. 【0212】 The computer device 300 includes one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or Windows Server. TM Mac OS X TM Unix TM Linux (registered trademark) TM FreeBSD TM It may also include one or more operating systems such as 333. 【0213】 The aforementioned computer device 300 may also be used to perform the steps in the corresponding embodiments shown in Figures 2 to 13, and the steps in the corresponding embodiment shown in Figure 14. 【0214】 In another aspect of this application, a computer-readable storage medium in which a computer program is stored is provided. When the computer program is executed by a processor, the steps of the method in the embodiments shown in Figures 2 to 13 and the steps of the method in the embodiment shown in Figure 14 are realized. 【0215】 According to another aspect of this application, a computer program product including a computer program is provided. When the computer program is executed by a processor, the steps of the method in the embodiments shown in Figures 2 to 13 and the steps of the method in the embodiment shown in Figure 14 are realized. 【0216】 For the convenience and brevity of this explanation, and so that those skilled in the art can clearly understand it, the specific operating processes of the systems, apparatus, and units described above should be referenced to the corresponding processes in the embodiments of the methods described above. Further explanation is omitted here. 【0217】 In some embodiments of this application, it should be understood that the disclosed systems, apparatus, and methods may be implemented in other ways. For example, the embodiments of the apparatus described above are merely schematic. For example, the division of the units is merely a logical division of functions, and other division methods may be used in actual implementation. For example, multiple units or components may be combined or integrated into another system, and some features may be ignored or not implemented. Furthermore, the mutual coupling, direct coupling, or communication connection shown or considered may be via several interfaces, and the indirect coupling or communication connection of apparatus or units may be electrical, mechanical, or of other forms. 【0218】 The units described above as individual components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Depending on the actual needs, some or all of these units can be selected to achieve the objectives of the technical method of this embodiment. 【0219】 Furthermore, each functional unit according to each embodiment of this application may be integrated into a single processing unit, each unit may exist physically independently, or two or more units may be integrated into a single unit. The aforementioned integrated unit may be implemented in hardware form or in the form of a software functional unit. 【0220】 The integrated unit may be implemented in the form of a software function unit and, if sold or used as an independent product, may be stored on a computer-readable storage medium. Based on this understanding, the essence of the technical method of the present application, or any part that can contribute to the prior art, or all or part of the technical method, may be expressed in the form of a software product, which is stored on a storage medium and contains several instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage mediums include various media capable of storing program code, such as U disks, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

[Claim 1] A method for training a reaction product prediction model performed by a computer device, A step of performing a vector transformation on each reaction sequence in a sample reaction data set through an encoder network of a reaction product prediction model to obtain a sample reactant vector and a sample reaction product vector, wherein the sample reaction data set includes a plurality of such reaction sequences, and each such reaction sequence includes a sample reactant and a sample reaction product. The steps include constructing a set of positive and negative sample reactants according to the sample reactant vector through a first auxiliary network, The steps include: determining the reaction prediction loss value using a predetermined first loss function based on the set of positive sample reactants and the set of negative sample reactants; The steps include constructing a positive sample reaction group set and a negative sample reaction group set through a second auxiliary network according to the sample reactant vector and the sample reaction product vector, The steps include identifying a reaction relationship prediction loss value using a predetermined second loss function based on the set of positive sample reaction groups and the set of negative sample reaction groups, The steps include identifying, through a third auxiliary network, the predicted probability value and atomic label of the atoms in the sample reactant present in the main product, according to the sample reactant vector and the sample reaction product vector, The steps include: determining the predicted atomic loss value using a predetermined third loss function based on the predicted probability value and the atomic label; The step of training the reaction product prediction model based on the reaction prediction loss value, the reaction relationship prediction loss value, and the atom prediction loss value to obtain a target reaction product prediction model, is included. The step of constructing a set of positive and negative sample reactants according to the sample reactant vector through a first auxiliary network is: The steps include sampling sample reactant vectors corresponding to any two sample reactants from the same reaction sequence to form a positive sample reactant combination, and adding the positive sample reactant combination to the set of positive sample reactants, The process includes the steps of sampling sample reactant vectors corresponding to any two sample reactants from different reaction sequences to form a negative sample reactant combination, and adding the negative sample reactant combination to the negative sample reactant set, The step of constructing a positive sample reaction group set and a negative sample reaction group set through a second auxiliary network according to the sample reactant vector and the sample reaction product vector is: The steps include: setting the corresponding sample reactant vector and sample reaction product vector for all reaction sequences as the positive sample reaction group set; A training method comprising the step of obtaining the negative sample reaction group set by combining the corresponding sample reactant vectors and sample reaction product vectors of different reaction sequences. [Claim 2] Prior to the step of performing a vector transformation on each reaction sequence in the sample reaction data set through an encoder network of reaction product prediction models to obtain sample reactant vectors and sample reaction product vectors, the training method further: The steps include: performing data augmentation processing on the aforementioned sample reaction data set to obtain a sample composite reaction data set; The step includes aggregating the aforementioned sample reaction data set and the aforementioned sample composite reaction data set as an extended sample reaction data set, The step of performing a vector transformation on each reaction sequence in the sample reaction data set through an encoder network of reaction product prediction models to obtain a sample reactant vector and a sample reaction product vector is: The step includes performing a vector transformation on each reaction sequence in the extended sample reaction data set through the encoder network of the reaction product prediction model to obtain the sample reactant vector and the sample reaction product vector, The step of performing data augmentation processing on the aforementioned sample reaction data set to obtain a sample composite reaction data set is, The steps include randomly selecting two reaction sequences from the aforementioned sample reaction data set, The steps include: combining sample reactants from the two selected reaction sequences to obtain a sample composite reactant; The steps include: combining the sample reaction products in the two selected reaction sequences to obtain a sample composite reaction product; The training method according to claim 1, comprising the steps of obtaining a composite reaction sequence based on the sample composite reactant and the sample composite reaction product, and constructing the sample composite reaction data set based on the composite reaction sequence. [Claim 3] The step of identifying the predicted probability value and atomic label of atoms in the sample reactant present in the main product, according to the sample reactant vector and the sample reaction product vector, through a third auxiliary network, The steps include predicting the sample reactants through the third auxiliary network and obtaining a predicted probability value for the presence of atoms in the sample reactants in the main product, The steps include comparing atoms in the sample reactant and the sample reaction product based on the sample reactant vector and the sample reaction product vector, and obtaining an atomic comparison result, A training method according to claim 1 or 2, comprising the step of identifying the atomic label based on the atomic comparison result. [Claim 4] The step of performing a vector transformation on each reaction sequence in the sample reaction data set through an encoder network of reaction product prediction models to obtain a sample reactant vector and a sample reaction product vector is: For each of the aforementioned reaction sequences, the step of obtaining sample reactant interaction information by allowing the internal information of each molecule of the sample reactant within it to interact between atoms, Based on the aforementioned sample reactant interaction information, the step of causing different sample reactants to interact and obtaining the sample reactant vector, For each of the aforementioned reaction sequences, the step of obtaining sample reaction product interaction information by allowing the internal information of each molecule of the sample reaction product within it to interact with each other. A training method according to claim 1 or 2, comprising the step of causing different sample reactants to interact based on the sample reaction product interaction information to obtain the sample reaction product vector. [Claim 5] The step of training the reaction product prediction model and obtaining a target reaction product prediction model based on the reaction prediction loss value, the reaction relationship prediction loss value, and the atom prediction loss value is as follows: Steps include obtaining reconstruction loss values ​​and divergence loss values, The step of training the reaction product prediction model based on the reconstruction loss value, the divergence loss value, the reaction prediction loss value, the reaction relationship prediction loss value, and the atom prediction loss value to obtain the target reaction product prediction model, is included. The step of obtaining the reconstruction loss value and the divergence loss value is: A step of obtaining a first hidden vector by processing the sample reactant vector and the sample reaction product vector using a cautionary mechanism, The steps include determining a second hidden vector based on the first hidden vector and the sample reactant vector, The steps include identifying the sample prediction change probability of the adjacency matrix according to the second hidden vector through the decoder network of the reaction product prediction model, The steps include identifying the sample prediction reaction product adjacency matrix based on the sample prediction change probability, The training method according to claim 1 or 2, comprising the steps of calculating a loss based on the reaction product adjacency matrix of the sample reaction product and the sample predicted reaction product adjacency matrix to obtain the reconstruction loss value and the divergence loss value. [Claim 6] The step of identifying the sample predicted reaction product adjacency matrix based on the sample predicted change probability is: A step of calculating the sample prediction change amount of the adjacency matrix based on the sample prediction change probability, The training method according to claim 5, comprising the step of calculating the sample predicted reaction product adjacency matrix based on the sample predicted change amount of the adjacency matrix and the sample reactant adjacency matrix. [Claim 7] The step of training the reaction product prediction model and obtaining a target reaction product prediction model based on the reaction prediction loss value, the reaction relationship prediction loss value, and the atom prediction loss value is as follows: The steps include: adjusting the parameters of the first auxiliary network based on the reaction prediction loss value to obtain a first submodel; The steps include: adjusting the parameters of the second auxiliary network based on the predicted loss value of the reaction relationship to obtain a second submodel; The steps include: adjusting the parameters of the third auxiliary network based on the predicted atomic loss values ​​to obtain a third submodel; The training method according to claim 1 or 2, comprising the step of transitioning the first submodel, the second submodel and the third submodel to the reaction product prediction model to obtain the target reaction product prediction model. [Claim 8] A method for applying a reaction product prediction model performed by a computer device, The steps include inputting the reactants to be measured into a target reaction product prediction model obtained by the training method described in claim 1 or 2, and outputting the predicted change probability of the adjacency matrix from the target reaction product prediction model, A step of identifying the predicted change amount of the adjacency matrix based on the predicted change probability of the adjacency matrix, A method for applying a reaction product prediction model, comprising the step of identifying a target reaction product based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured. [Claim 9] The step of identifying a target reaction product based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured is as follows: The steps include: calculating the predicted reaction product adjacency matrix based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured; The steps include performing a symmetrization process on the predicted reaction product adjacency matrix to obtain a target reaction product adjacency matrix, The application method according to claim 8, comprising the step of identifying the target reaction product based on the target reaction product adjacency matrix. [Claim 10] A training device for a reaction product prediction model, An acquisition unit configured to perform vector transformation on each reaction sequence in a sample reaction data set through an encoder network of a reaction product prediction model to obtain a sample reactant vector and a sample reaction product vector, further construct a set of positive sample reactants and a set of negative sample reactants according to the sample reactant vector through a first auxiliary network, construct a set of positive sample reaction groups and a set of negative sample reaction groups according to the sample reactant vector and the sample reaction product vector through a second auxiliary network, and identify the predicted probability value and atomic label of atoms in the sample reactant present in the main product according to the sample reactant vector and the sample reaction product vector, wherein the sample reaction data set includes a plurality of reaction sequences, and each reaction sequence includes a sample reactant and a sample reaction product, A processing unit is configured to determine a reaction prediction loss value using a predetermined first loss function based on the set of positive sample reactants and the set of negative sample reactants, to determine a reaction relationship prediction loss value using a predetermined second loss function based on the set of positive sample reaction groups and the set of negative sample reaction groups, and further to determine an atomic prediction loss value using a predetermined third loss function based on the prediction probability value and the atomic label. A specific unit configured to train the reaction product prediction model based on the reaction prediction loss value, the reaction relationship prediction loss value, and the atom prediction loss value, and to obtain a target reaction product prediction model, is included. Through the first auxiliary network, constructing a set of positive and negative sample reactants according to the sample reactant vector is possible. The process involves sampling sample reactant vectors corresponding to any two sample reactants from the same reaction sequence to form a positive sample reactant combination, and adding the positive sample reactant combination to the positive sample reactant set. This includes sampling sample reactant vectors corresponding to any two sample reactants from different reaction sequences to form a negative sample reactant combination, and adding the negative sample reactant combination to the negative sample reactant set, Through a second auxiliary network, constructing a positive sample reaction group set and a negative sample reaction group set according to the sample reactant vector and the sample reaction product vector is: The set of positive sample reaction groups consists of the corresponding sample reactant vector and sample reaction product vector for all reaction sequences, A training apparatus comprising: obtaining the negative sample reaction group set by combining the corresponding sample reactant vectors and sample reaction product vectors of different reaction sequences. [Claim 11] An acquisition unit configured to input the reactants to be measured into a target reaction product prediction model obtained by the training method described in claim 1 or 2, and to output the predicted change probability of the adjacency matrix from the target reaction product prediction model, A processing unit configured to identify the predicted change amount of the adjacency matrix based on the predicted change probability of the adjacency matrix, Applicator for a reaction product prediction model, comprising: a identifying unit configured to identify a target reaction product based on the predicted change in the adjacency matrix and the adjacency matrix of the reactants to be measured. [Claim 12] The memory in which computer programs are stored, When the computer program is executed, a processor that implements the training method described in claim 1 or 2 is provided. A computer device including a bus system for connecting the memory and the processor in order to enable communication between the memory and the processor. [Claim 13] A computer program, when executed by a processor, for realizing the training method described in claim 1 or 2.

Images