Methods, training methods, apparatus, electronic devices, and computer programs for predicting reactant molecules.

Abstract

The present invention provides a method, a training method, an apparatus and an electronic device for predicting reactant molecules related to the reverse reaction technology field in the chemical field. The method includes the steps of: performing feature extraction on a product molecule to obtain the characteristics of the product molecule; predicting a conversion path from the product molecule to a plurality of reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule; predicting synthesizers and complementary synthesizers according to the conversion path to obtain a plurality of reactant molecules corresponding to the product molecule. The method provided by the present invention can not only reduce the prediction complexity of reactant molecules and improve the generalization performance of reactant molecule prediction, but also improve the prediction performance of reactant molecules.

Description

[Technical Field]

[0001] This application claims priority to the Chinese patent application filed on August 9, 2022, application number 202210952642.6, with the title of invention "Method for predicting, training, apparatus and electronic device for reactant molecules," and all contents of said Chinese patent application are incorporated into this application by reference.

[0002] Examples of the present invention relate to the field of chemical reverse reactions, and more specifically to methods, training methods, apparatus, and electronic devices for predicting reactant molecules. [Background technology]

[0003] Predicting retrosynthetic reactants in organic chemistry is a crucial step in new drug development and the production of novel materials. It aims to find a commercially viable set of reactant molecules used to synthesize product molecules. Traditional methods involve matching reactant molecules with reaction templates; however, these templates must be manually extracted by expert researchers, a time-consuming process, and the templates cannot cover all reaction types. In recent years, advancements in deep learning technology have made it possible to learn potential reaction types from vast databases of organic chemical reactions. Therefore, building robust reverse reaction prediction models using deep learning technology is particularly important.

[0004] Up until now, there have generally been two types of model structures that have been commonly adopted. One is the sequence translation model, which is based on the Simplified Molecular Input Line Entry System (SMILES) sequence, and SMILES is a standard for clearly describing molecular structures. The other is the graph generation model, which is based on graphs.

[0005] Graph generation models generally divide the reverse reaction prediction task in organic chemistry into two subtasks: synthesizer recognition and synthesizer completion. Typically, synthesizers are recognized by constructing one graph neural network, and the synthesizers of the product molecule are obtained, while the synthesizers can be completed atom by atom from the encoder by constructing one graph variation. However, constructing two independent networks, one for synthesizer recognition and the other for synthesizer completion, not only increases the prediction complexity but also prevents achieving good generalization performance because the two subtasks have different optimization goals. Furthermore, when synthesizer completion is performed atom by atom, the complexity is high and the prediction performance is limited. [Overview of the project] [Problems that the invention aims to solve]

[0006] Embodiments of the present invention provide a method, training method, apparatus, and electronic device for predicting reactant molecules that can reduce the predictive complexity of reactant molecules, improve the generalization performance of predicting reactant molecules, and also improve the predictive performance of reactant molecules. [Means for solving the problem]

[0007] In a first embodiment, the present invention provides a method for predicting reactant molecules, the prediction method is The steps include: performing feature extraction on the product molecule to obtain the characteristics of the product molecule; A step of predicting a conversion pathway from the product molecule to a plurality of reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule, wherein the conversion pathway includes an editing sequence and a synth completion sequence. A step of obtaining a plurality of synthesizers corresponding to the product molecule by editing the target to be edited according to the edited state indicated by each editing operation in the editing sequence, wherein the target to be edited is an atom or chemical bond in the product molecule. A step of obtaining a plurality of reactant molecules corresponding to the plurality of synthesizers by adding a basic graph indicated by each synthesizer, according to the interface atoms indicated by each synthesizer, based on at least one synthesizer completion operation corresponding to each synthesizer in the synthesizer completion sequence, wherein the basic graph includes a plurality of atoms or atomic edges for connecting the plurality of atoms.

[0008] In a second embodiment, the present invention provides a method for training an inverse response prediction model, the training method being: The steps include: performing feature extraction on the product molecule to obtain the characteristics of the product molecule; A step of predicting the conversion pathway from the product molecule to a plurality of reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule, Here, the transformation path includes an editing sequence and a synthesizer completion sequence, where each editing operation in the editing sequence is used to indicate the object to be edited and the edited state, where the object to be edited is an atom or chemical bond in the product molecule, and for a plurality of synthesizers of the product molecule obtained from the editing sequence, the synthesizer completion sequence includes at least one synthesizer completion operation corresponding to each of the plurality of synthesizers, where each synthesizer completion operation in the at least one synthesizer completion operation is used to indicate a base graph and interface atoms, where the base graph includes a plurality of atoms or atomic edges for connecting the plurality of atoms, and The process includes the step of training the inverse response prediction model based on the losses between the transformation path and the training path.

[0009] In a third aspect, the present invention provides a predictor of reactant molecules, the predictor of which An extraction unit configured to perform feature extraction on product molecules and obtain the characteristics of the product molecules, A prediction unit configured to predict a conversion pathway from the product molecule to a plurality of reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule, wherein the conversion pathway includes an editing sequence and a synth-complement sequence. An editing unit configured to edit the target to be edited according to the edited state indicated by each editing operation in the editing sequence, in order to obtain a plurality of synthesizers corresponding to the product molecule, wherein the target to be edited is an atom or chemical bond in the product molecule; An additional unit configured to obtain a plurality of reactant molecules corresponding to the plurality of synthesizers by adding a basic graph indicated by each synthesizer complementation operation, according to the interface atoms indicated by each synthesizer complementation operation in the at least one synthesizer complementation operation, based on at least one synthesizer complementation operation corresponding to each synthesizer in the synthesizer complementation sequence, wherein the basic graph includes a plurality of atoms or atomic edges for connecting the plurality of atoms.

[0010] In a fourth embodiment, the present invention provides a training device for an inverse response prediction model, the training device is An extraction unit configured to perform feature extraction on product molecules and obtain the characteristics of the product molecules, A prediction unit configured to predict the conversion pathway from the product molecule to a plurality of reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule, Here, the transformation path includes an editing sequence and a synthesizer completion sequence, each editing operation in the editing sequence is used to indicate the object to be edited and the edited state, the object to be edited is an atom or chemical bond in the product molecule, and for a plurality of synthesizers of the product molecule obtained from the editing sequence, the synthesizer completion sequence includes at least one synthesizer completion operation corresponding to each of the plurality of synthesizers, each synthesizer completion operation in the at least one synthesizer completion operation is used to indicate a base graph and interface atoms, the base graph includes a prediction unit which includes a plurality of atoms or atomic edges which connect the plurality of atoms, The system includes a training unit configured to train the inverse response prediction model based on the losses between the conversion path and the training path.

[0011] In a fifth embodiment, the present invention provides an electronic device, the electronic device being A processor that executes computer instructions, The system comprises a computer-readable storage medium in which computer instructions are stored, and the processor loads and executes computer instructions to perform the reactant molecule prediction method according to the first embodiment or the inverse reaction prediction model training method according to the second embodiment.

[0012] In a sixth embodiment, an embodiment of the present invention provides a computer-readable storage medium in which computer instructions are stored, and when the computer instructions are read and executed by the processor of the computer device, the computer device is made to execute the reactant molecule prediction method according to the first embodiment or the inverse reaction prediction model training method according to the second embodiment.

[0013] In a seventh embodiment, an embodiment of the present invention provides a computer program product or computer program, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, thereby causing the computer device to perform a reactant molecule prediction method according to the first embodiment or a reverse reaction prediction model training method according to the second embodiment. [Effects of the Invention]

[0014] Based on the above technical solutions, by introducing a reverse reaction prediction model for predicting the conversion pathway from the product molecule to multiple reactant molecules, it becomes possible to integrate the synthetic atom prediction task and the synthetic atom completion prediction task. That is, the reverse reaction prediction model introduced in the embodiment of the present invention can learn the potential relationship between the two subtasks of synthetic atom prediction and complementary synthetic atom, thereby significantly improving the generalization performance of the model, reducing the prediction complexity of reactant molecules, and improving the generalization performance of reactant molecule prediction. Furthermore, by introducing a basic graph and designing the basic graph to include a structure that includes multiple atoms or atomic edges for connecting the multiple atoms, it is possible to rationally construct a short and accurate conversion pathway, avoiding excessively long synthetic atom completion sequences, reducing the difficulty of predicting reactant molecules, improving the prediction accuracy of reactant molecules, and improving the prediction performance of reactant molecules.

[0015] Furthermore, by improving the predictive performance of reactant molecules, the following technical benefits can also be obtained.

[0016] 1. Synthetic routes can be planned for designed drug or novel material molecules, thereby increasing the efficiency of research on drug or novel material molecules.

[0017] 2. By presenting several potential scientific disciplines, it is possible to provide new scientific knowledge.

[0018] 3. We can provide more precise synthesis route plans than expert researchers, predict reliable reactant molecules even when reaction templates are lacking, and predict reaction types that expert researchers have not yet clarified, thereby significantly improving the efficiency of new drug and material development. [Brief explanation of the drawing]

[0019] [Figure 1] This is an example of a system framework provided by an embodiment of the present invention. [Figure 2] This is a schematic flowchart of the method for predicting reactant molecules provided by the embodiments of the present invention. [Figure 3] This is an example of a conversion path provided by an embodiment of the present invention. [Figure 4] This is another schematic flowchart of the reactant molecule prediction method provided by the embodiments of the present invention. [Figure 5] This is a schematic flowchart of the training method for the inverse response prediction model provided by the embodiment of the present invention. [Figure 6] Figure 6 is a schematic block diagram of a reactant molecule prediction device provided by an embodiment of the present invention. [Figure 7] Figure 7 is a schematic block diagram of a training apparatus for an inverse response prediction model provided by an embodiment of the present invention. [Figure 8] Figure 8 is a schematic block diagram of an electronic device provided by an embodiment of the present invention. [Modes for carrying out the invention]

[0020] Hereinafter, the technical solutions in embodiments of the present invention will be clearly and completely described with reference to the drawings of the embodiments.

[0021] The technology provided in this invention relates to the field of artificial intelligence (AI).

[0022] Specifically, the technology provided in this invention relates to a technology for predicting reverse reactions based on AI in the field of chemistry.

[0023] Here, AI refers to theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, artificial intelligence is a comprehensive technology of computer science that seeks to understand the nature of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. Artificial intelligence involves studying the design principles and implementation methods of various intelligent machines so that they can have the functions of perception, reasoning, and decision-making.

[0024] Artificial intelligence is a comprehensive field of study encompassing a wide range of areas, including both hardware-level and software-level technologies. Foundational technologies in artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technologies, manipulative / interactive systems, and mechatronics. Artificial intelligence software technologies primarily include computer vision technologies, speech processing technologies, natural language processing technologies, and machine learning / deep learning.

[0025] With the ongoing research and advancement of artificial intelligence technology, it is being studied and applied in a wide range of fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robotics, smart healthcare, and smart calls. We believe that as the technology develops, artificial intelligence technology will be applied in even more fields and will become increasingly important.

[0026] Embodiments of the present invention may relate to computer vision (CV) technology in artificial intelligence technology.

[0027] Specifically, the embodiments of the present invention relate to a technical field for predicting reverse reactions based on the results of CV recognition in the chemical field.

[0028] Here, computer vision is the science of studying how machines "see," and furthermore, machine vision that uses cameras and computers in place of human eyes to perform recognition, prediction, and measurement of targets, etc., and furthermore, by performing graphics processing, it makes computer processing more suitable for transmitting images to instruments for observation or detection by the human eye. As a scientific field, computer vision studies related theories and technologies and attempts to build artificial intelligence systems that can extract information from images or multidimensional data. Computer vision technologies typically include technologies such as image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / action recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, synchronous positioning and map building, and furthermore, conventional biometric recognition technologies such as facial recognition and fingerprint recognition.

[0029] Embodiments of the present invention can further relate to machine learning (ML) in artificial intelligence.

[0030] Specifically, embodiments of the present invention may relate to the technology of performing reverse reaction prediction using machine learning predictive models.

[0031] Here, ML is an interdisciplinary field encompassing various disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specializes in the study of how computers can simulate or realize human learning behavior to acquire new knowledge or skills, reconstruct existing knowledge structures, and continuously improve their own performance. Machine learning is at the core of artificial intelligence, the fundamental path to giving intelligence to computers, and its applications extend to all fields of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, trust networks, reinforcement learning, transfer learning, inductive learning, and educational learning.

[0032] Figure 1 shows an example of a system framework 100 provided by an embodiment of the present invention.

[0033] The system framework 100 may be an application system, and the embodiments of the present invention do not limit the specific type of application. The system framework 100 comprises terminal 131, terminal 132, and server cluster 110. Both terminal 131 and terminal 132 are connected to the server cluster 110 via a wireless network or a wired network 120.

[0034] Terminals 131 and 132 may be at least one of a smartphone, a game console, a desktop computer, a tablet computer, an e-book reader, an MP3 player, an MP4 player, and a laptop computer. An application program is installed and executed on terminals 131 and 132. The application program may be one of the following: an online video program, a short video program, a photo sharing program, a sound social program, an animation program, a wallpaper program, a news push program, a supply and demand information push program, an academic exchange program, a technology exchange program, a policy exchange program, a program including a comment mechanism, a program including a perspective presentation mechanism, or a knowledge sharing program. Terminals 131 and 132 may be terminals used by user 141 and user 142, respectively, and user accounts are registered in the applications running on terminals 131 and 132.

[0035] The server cluster 110 includes at least one of a single server, multiple servers, a cloud computing platform, and a virtualization center. The server cluster 110 is used to provide background services to applications (e.g., applications on terminals 131 and 132). Exemplarily, the server cluster 110 is responsible for primary computing tasks and terminals 131 and 132 are responsible for secondary computing tasks, or the server cluster 110 is responsible for secondary computing tasks and terminals 131 and 132 are responsible for primary computing tasks, or terminals 131 and 132 and the server cluster 110 cooperate to perform calculations using a distributed computing architecture. For example, the computing tasks according to the present invention may be computing tasks related to predicting retrosynthesis reactants in organic chemistry or related auxiliary tasks.

[0036] For example, if the system framework 100 is a web browsing system, the server cluster 110 includes an access server 112, a web page server 111, and a data server 113. The access server 112 may be one or more, located in different nearby cities, and is used to receive service requests from terminals 131 and 132 and forward the service requests to the corresponding servers for processing. The web page server 111 is a server for providing web pages to terminals 131 and 132, which have tracking codes integrated into them, and the data server 113 is used to receive data reported from terminals 131 and 132 (e.g., service data).

[0037] Predicting retrosynthetic reactants in organic chemistry is a crucial step in new drug development and the production of novel materials. It aims to find a commercially viable set of reactant molecules used to synthesize product molecules. Traditional methods involve matching reactant molecules to reaction templates; however, these templates must be manually extracted by expert researchers, a time-consuming process, and the templates cannot cover all reaction types. In recent years, advancements in deep learning technology have made it possible to learn potential reaction types from vast databases of organic chemical reactions. Therefore, building robust reverse reaction prediction models using deep learning technology is extremely important.

[0038] Up until now, there have generally been two types of model structures that have been commonly adopted: one is the sequence translation model, which is based on the Simplified Molecular Input Line Entry System (SMILES) sequence, and SMILES is a standard for clearly describing molecular structures; the other is the graph generation model, which is based on graphs.

[0039] Graph generation models typically divide the reverse reaction prediction task in organic chemistry into two subtasks: synthesizer recognition and synthesizer completion. Typically, one graph neural network is constructed to recognize synthesizers and obtain the synthesizers of the product molecule, and one graph variational autoencoder is constructed to complete the synthesizers atom by atom. However, constructing two independent networks, one for synthesizer recognition and the other for synthesizer completion, not only increases prediction complexity but also prevents achieving good generalization performance because the two subtasks have different optimization goals. Furthermore, completing synthesizers atom by atom is complex and limits predictive performance.

[0040] In light of the above challenges, it is conceivable to complement the synthesizers using pre-extracted leaving groups in order to reduce the difficulty of complementary synthesizers and improve the performance of reverse reaction prediction. Here, a leaving group is an atom or functional group that is removed from a larger molecule in a chemical reaction, and a functional group is an atom or group of atoms that determines the chemical properties of an organic compound. Common functional groups include carbon-carbon double bonds, carbon-carbon triple bonds, hydroxyl groups, carboxyl groups, ether bonds, aldehyde groups, etc.

[0041] However, using pre-extracted leaving groups to complement the synths fails to improve the generalization performance of the model. Furthermore, it is limited by the properties of the leaving groups, resulting in an extremely unbalanced sample distribution, which in turn limits the predictive effect of the reactant molecules.

[0042] Furthermore, a single end-to-end model can jointly learn two subtasks; that is, the two subtasks of synthesizer recognition and complementary synthesizers are integrated into a single end-to-end model, complementing synthesizers either atom by atom or by employing benzene rings. Here, an end-to-end model refers to a model that optimizes training based on a single optimization goal.

[0043] However, while the end-to-end model architecture allows the model to achieve better generalization performance, the use of single atoms and smaller units such as benzene rings to complement the synthesizers increases the difficulty of prediction because it requires predicting longer graph editing sequences, preventing the top-ranked category accuracy from reaching the state of the art (sota). Here, top-ranked category accuracy refers to the accuracy with which the top-ranked category matches the actual results.

[0044] In view of this, embodiments of the present invention provide a method, training method, apparatus, and electronic device for predicting reactant molecules that can reduce the predictive complexity of reactant molecules and improve the generalization performance of predicting reactant molecules, as well as improve the predictive performance of reactant molecules. Specifically, embodiments of the present invention design a single end-to-end reverse reaction prediction model to jointly optimize two subtasks, synthetizer prediction and complementary synthetizer, wherein the synthetizer prediction task can be constructed to predict an editing sequence that can show the conversion process from product molecules to synthesizers, and the complementary synthetizer prediction task can be constructed to predict a synthetizer complementary sequence that can show the conversion process from synthesizers to reactant molecules, and the editing sequence and synthetizer complementary sequence construct a conversion path that can show the conversion process from product molecules to reactant molecules.

[0045] In particular, as the amount of data increases, the techniques provided by the embodiments of the present invention can be easily extended to more complex and diverse chemical reaction models. Exemplarily, the reverse reaction prediction model provided by the embodiments of the present invention can be extended to multi-step reverse reaction prediction tasks, for example, by employing a Monte Carlo tree search algorithm or the like.

[0046] Furthermore, by introducing a basic graph, the embodiments of the present invention can minimize the problem of limited predictive effectiveness of reactant molecules due to extreme imbalances in the leaving group sample, solve the problem of excessively high prediction complexity when using atomic complementary synthesizers, and improve the prediction accuracy of the reverse reaction prediction model.

[0047] Figure 2 is a schematic flowchart of a reactant molecule prediction method 200 provided by an embodiment of the present invention, the prediction method 200 may be performed by any electronic device having data processing capabilities. For example, the electronic device may be implemented as a server. The server may be an independent physical server, or a server cluster or distributed system consisting of multiple physical servers, and may also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, networking services, cloud communications, middleware services, domain services, security services, and big data and artificial intelligence platforms. The servers may be connected directly or indirectly by wired or wireless communication methods, and the present invention is not limited thereto. For convenience of explanation, the prediction method provided in the present invention will be described below using a reactant molecule prediction device as an example.

[0048] As shown in Figure 2, the prediction method 200 may include the following steps.

[0049] In step S210, feature extraction is performed on the product molecule to obtain the characteristics of the product molecule.

[0050] Illustratively, feature extraction is performed on the product molecule using a Simplified Molecular Input Line Entry System (SMILES) or a Graph Neural Network (GNN) to obtain the characteristics of the product molecule. Of course, feature extraction can also be performed on product molecules with other types of functions, and the embodiments of the present invention are not particularly limited to these.

[0051] For example, the characteristics of the product molecule can be determined based on the characteristics of each atom in the product molecule. For instance, a single global attention pool function can be used to calculate the characteristics of all atoms and obtain the characteristics of the product molecule.

[0052] For example, the characteristics of each atom can be obtained by encoding the original characteristics of each atom and the original characteristics of the chemical bonds between each atom and its neighboring nodes. For instance, an MPNN can be used to encode the original characteristics of each atom and the original characteristics of the chemical bonds between each atom and its neighboring nodes to obtain the characteristics of each atom. Here, the original characteristics of each atom are used to indicate at least one of the following: the type of atom (C, N, O, S, etc.), the degree of chemical bonding, chirality, and the number of hydrogen atoms. The original characteristics of the chemical bonds are used to indicate at least one of the following: the type of chemical bond (single bond, double bond, triple bond, and aromatic bond, etc.), configuration, and aromaticity.

[0053] Exemplary, the characteristics of the product molecule may further include characteristics that represent the type of chemical bonding in the product molecule. For example, the type of chemical bonding in the product molecule may include, but is not limited to, single bonds, double bonds, triple bonds, and non-chemical bonds.

[0054] Exemplary, the characteristics of the product molecule can be represented in the form of a vector or a matrix, and of course, they can also be represented by an array or other form of information, and the present invention is not particularly limited thereto.

[0055] The following illustrates the characteristics of the product molecule using feature vectors as an example.

[0056] For example, a molecule containing n atoms and m chemical bonds can be represented by the graph structure G=(V,E), where V is the set of atoms of size n and E is the set of chemical bonds of size m. Each atom v∈V has one feature vector x v Each chemical bond e∈E has a single feature vector x that represents information such as the type of atom (C, N, O, S, etc.), the degree of chemical bonding, chirality, and the number of hydrogen atoms. v,u This includes information such as chemical bond type (single, double, triple, and aromatic bonds), configuration, and aromaticity. Furthermore, a 4-dimensional one-hot vector can be defined to represent the type of chemical bond in the molecular graph, i.e., single, double, triple, and non-chemical bonds. A one-hot vector is a vector in which one element is 1, and only one element is 1, with all other elements being 0. Every atom and chemical bond has a single tag s∈{0,1} to indicate whether or not it is an editable target relevant when the product molecule is converted into a synthesizer.

[0057] For example, first, a Message Passing Neural Network (MPNN) with L layers can be used to encode each atom in the product molecule G, and a feature vector for each atom in the product molecule can be obtained. MPNNs are a graph-applicable monitoring learning framework. Furthermore, a Multilayer Perceptron (MLP) can be used to encode each chemical bond in the product molecule G, and a feature vector for each chemical bond can be obtained. MLPs are feedforward artificial neural network models that can map multiple input datasets into a single output dataset.

[0058] For example, the feature vector h of an atom.v and the characteristic vector h of the chemical bond v,u is obtained by calculating with the following formula.

[0059] [Number] Here, MPNN(·) represents the neural network of message passing, G represents the graph structure of the product molecule, L represents the number of layers of MPNN(·), and h L v represents the characteristic vector of atom v output by the L-th layer of MPNN(·), x v represents the characteristic vector of atom v before encoding, x v,u represents the characteristic vector of the chemical bond between atom v and atom u before encoding, N(v) represents the set of adjacent nodes of atom v, and h v,u represents the characteristic vector of the chemical bond between atom v and atom u after encoding, MLP bond (·) represents the multi-layer perceptron, || represents the concatenation operation, and h L u represents the characteristic vector of atom u output by the L-th layer of MPNN(·).

[0060] Furthermore, for the convenience of subsequent tasks, h v,u is represented in the form of a self-loop, that is,

[0061] [Number] Furthermore, to simplify the calculation, atoms and chemical bonds are represented in the same form, that is,

[0062] [Number] Here, i is the tag of the atom or the tag of the chemical bond.

[0063] Exemplarily, the tag of the atom can be the index of the atom, and the tag of the chemical bond can be the index of the chemical bond.

[0064] After obtaining the atomic feature vectors, the global attention pool function is used to obtain the feature vector h of the product molecule from all the atomic feature vectors. G It is possible to calculate this. Note that the feature vector of the compositer may be required, h G The acquisition method is the same as above, and using the global attention pool function, the feature vector h of the composite particle is obtained from the feature vector containing all atoms of the composite particle. syn It can also be calculated.

[0065] Of course, in other alternative embodiments, comparative learning strategies may be introduced, such as occluding the graph structure or features of the molecular graph during the encoding process of the product molecules or synthesizers, thereby improving the predictive performance of reactant molecules by expanding the feature dimension.

[0066] In step S220, based on the characteristics of the product molecule, a reverse reaction prediction model is used to predict the conversion pathway from the product molecule to a plurality of reactant molecules, the conversion pathway including an editing sequence and a synth completion sequence.

[0067] For example, the editing sequence is a sequence formed by editing operations, and the compositer completion sequence is a sequence formed by compositer completion operations.

[0068] In other words, the prediction task of the reverse reaction prediction model is defined as a reactant molecule generation task. That is, the prediction task of the reverse reaction prediction model predicts a transformation path to describe the transformation path from the product molecule graph to the reactant molecule graph, and this transformation path is defined by editing operations on the product molecules and synthesizer completion operations on the synthesizers. In other words, it is constructed by an Edit sequence formed by editing operations and a synthesizer completion sequence formed by synthesizer completion operations. In this way, a transformation path describing the transformation path from the product molecule graph to the reactant molecule graph can be predefined for each product molecule. Synthesizer completion operations are also called basic graph addition (AddingMotif) operations, and synthesizer completion sequences are also called basic graph addition sequences.

[0069] Alternatively, the conversion pathway includes an editing sequence and a synthetizer completion sequence. Here, the editing sequence is used to describe the chemical bonding and atomic changes from the product molecule to the synthetizer. The synthetizer completion sequence is used to describe the process of completing the synthetizer in the basic graph (motif). For the editing sequence, each change from the product molecule to the synthetizer is shown by introducing the editing operation, and for the synthetizer completion sequence, each completion operation in the process of completing the synthetizer in the basic graph (motif) is described by introducing the synthetizer completion operation.

[0070] Exemplary, the editing operation is an operation to edit atoms or chemical bonds in the product molecule during the process in which the product molecule is converted into multiple synthetic elements of the product molecule.

[0071] Exemplary, the synth-complementary operation is an operation that adds to the basic graph in the process from the multiple synths to the multiple reactant molecules.

[0072] Exemplary, the editing sequence may further include an editing completion operation, for example, the conversion path may sequentially include at least one editing operation, an editing completion operation, and at least one synthesizer completion operation. The editing completion operation is used to link or separate the at least one editing operation from the at least one synthesizer completion operation. In other words, the editing completion operation is used to trigger the start of the synthesizer completion task of the inverse reaction prediction model. In embodiments of the present invention, the conversion path is constructed by innovatively introducing an editing completion operation to link at least one editing operation to at least one synthesizer completion operation.

[0073] Exemplary, the editing sequence may further include a Start operation, for example, the transformation path including the Start operation, at least one editing operation, an editing completion operation, and at least one synthesizer completion operation in sequence. The Start operation is used to trigger the start of a synthesizer prediction task in the reverse reaction prediction model, or to trigger the start of a reactant molecule prediction task in the reverse reaction prediction model.

[0074] Of course, in other alternative embodiments, the editing completion operation may also be an operation in the compositer completion sequence, and the present invention is not particularly limited thereto.

[0075] Furthermore, the inverse response prediction model according to the present invention may be any deep learning or machine learning model for recognition, and the embodiments of the present invention are not limited to any specific type thereof.

[0076] For example, the inverse response prediction model includes, but is not limited to, conventional learning models, integrated learning models, or deep learning models. Exemplary examples include, but are not limited to, tree models (regression trees) or logistic regression (LR) models; integrated learning models include, but are not limited to, gradient boosting algorithms (XGBoost) or random forest models; and deep learning models include, but are not limited to, neural networks, dynamic Bayesian networks (DBNs), and stacked auto-encoder network (SAE) models. Of course, in other embodiments of the present invention, models from other classes of machine learning may be used instead.

[0077] For example, a model can be trained using a batch size.

[0078] Here, the batch size is the number of samples selected in a single training run. The batch size value determines the time required to complete each epoch during deep learning training and the smoothness of the gradient between each iteration. For a training set of size N, if we adopt the most common sampling method for each epoch, sampling once per ampoule with N samples, and set the batch size to b, then the number of iterations required for each epoch is N / b, and therefore the time required for each epoch also increases approximately with increasing iterations. If the batch size value is too small, it takes a long time and the gradient fluctuations are severe, which is detrimental to convergence. If the batch size value is too large, the gradient direction of different batch sizes does not change, and it is prone to falling into local minima. The embodiments of the present invention are not specifically limited to the value of the batch size. For example, an appropriate batch size can be determined according to actual needs or scenarios. Of course, in other alternative embodiments, for databases with a small number of samples, it is not necessary to adopt a batch size, and the effect is also good. However, for large databases, inputting all the data into the network at once can cause a memory explosion, so in this case, a batch size can be adopted to train the network.

[0079] To make the embodiments of the present invention easier to understand, the contents of the present invention will be explained.

[0080] Neural Network (NN): A neural network is a computational model composed of multiple neuron nodes interconnected. Here, the connections between nodes represent weights, which are the weights applied to the output signals from the input signals. Each node performs weighted summation (SUM) on different input signals and outputs them via a specific activation function (f).

[0081] Examples of neural networks include, but are not limited to, deep neural networks (DNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs).

[0082] Convolutional Neural Networks (CNN): It is a feedforward neural network that includes convolutional computations and has a depth structure, and is one of the representative algorithms of deep learning. Because convolutional neural networks can perform shift-invariant classification, they are also called Shift-Invariant Artificial Neural Networks (SIANN).

[0083] Recurrent Neural Network (RNN): RNNs are neural networks that model sequence data and have achieved remarkable results in the field of natural language processing, such as in machine translation and speech recognition. In concrete terms, the network stores information from past time points and uses it to calculate the current output. That is, the nodes between hidden layers are no longer connectionless but connected, and furthermore, the input to a hidden layer includes not only the input layer but also the output of the hidden layer at the previous time point. Commonly used RNNs include structures such as Long Short-Term Memory (LSTM) and gated recurrent unit (GRU).

[0084] In step S230, according to the edited state indicated by each editing operation in the editing sequence, the editing target indicated by each editing operation is edited to obtain a plurality of synthons corresponding to the product molecule, wherein the editing target is an atom or chemical bond in the product molecule.

[0085] For example, according to the order of editing operations in the editing sequence, and based on the edited state indicated by each editing operation in the editing sequence, the editing targets indicated by each editing operation are edited to obtain a plurality of synthons corresponding to the product molecule.

[0086] In some embodiments, when the object to be edited by the editing operation is an atom, the edited state indicated by the editing operation is a change in the number of charges on the atom or a change in the number of hydrogen atoms on the atom. When the object to be edited by the editing operation is a chemical bond, the edited state indicated by the editing operation is one of the following: an addition of a chemical bond, a deletion of a chemical bond, or a change in the type of chemical bond.

[0087] For example, the edited state specified by the editing operation is associated with the object to be edited specified by the editing operation.

[0088] For example, if the object to be edited by the editing operation is a chemical bond, the edited state indicated by the editing operation includes, but is not limited to, a state of chemical bond addition, chemical bond deletion, or a change in the type of chemical bond. Also, if the object to be edited by the editing operation is an atom, the edited state indicated by the editing operation is one of the following: a change in the number of charges on the atom or a change in the number of hydrogen atoms.

[0089] In some embodiments, the editing operation is indicated by an action tag for instructing editing, a tag for indicating the object to be edited, and a tag for indicating the state after editing.

[0090] For example, an editing operation in an editing sequence is defined as an editing triplet, i.e., (π1, o, τ), where π1 represents the operation predicted by the reverse reaction prediction model, o represents the tag to be edited corresponding to the editing operation predicted by the reverse reaction prediction model, and τ represents the tag of the edited state corresponding to the editing operation predicted by the reverse reaction prediction model. For example, let an editing triplet be (π2, b, none), where π1 represents the operation predicted by the reverse reaction prediction model, b represents the chemical bond of tag b that corresponds to the editing operation, and none represents the deletion of the chemical bond of tag b that corresponds to the editing operation.

[0091] Of course, in other alternative embodiments, the editing operation can be defined as a tuple or a numeric value in another format. For example, if the editing operation is defined as a tuple, it can specifically be defined as a tag to indicate the object to be edited and a tag to indicate the state after editing.

[0092] In step S240, for each of the plurality of synthesizers, based on at least one synthesizer completion operation corresponding to each synthesizer in the synthesizer completion sequence, a basic graph (Motif) indicated by each synthesizer completion operation is added according to the interface atom indicated by each synthesizer completion operation in the at least one synthesizer completion operation to obtain a plurality of reactant molecules corresponding to the plurality of synthesizers, the basic graph including a plurality of atoms or atomic edges for connecting the plurality of atoms.

[0093] For example, the atomic edge can represent the reciprocal force between two or more atoms connected to the atomic edge. This force is used to bond the atoms connected at the atomic edge.

[0094] Exemplary, the atomic edges are used to link chemical bonds between different atoms of the plurality of atoms. The atomic edges may be chemical bonds, including but not limited to ionic, covalent, and metallic bonds.

[0095] For example, a synthon is a molecular fragment obtained after a chemical bond has been cleaved in the product molecular graph.

[0096] For example, the motif is a subgraph of reactants, which may be, for instance, a subgraph on a reactant corresponding to the synthesizer. The reactant corresponding to the synthesizer may include a molecule or reactant containing the synthesizer.

[0097] The basic graph (motif) may include subgraphs obtained through the following methods.

[0098] 1. Cut the edges where there is a difference between the synthesizer and its corresponding reactant to obtain a series of subgraphs. These subgraphs hold interface atoms corresponding to the attached atoms on the synthesizer.

[0099] 2. If two linked atoms on a single subgraph each belong to two rings, the chemical bond between these two atoms is broken, thereby yielding two smaller subgraphs.

[0100] Furthermore, the rings in the embodiments of the present invention may be monorings, that is, those with only one ring in the molecule. Correspondingly, if two linked atoms in one subgraph each belong to two monorings, the chemical bond between these two atoms is broken, thereby yielding two smaller subgraphs. Moreover, the rings in the embodiments of the present invention may be cycloalkanes, which may be classified based on the number of carbon atoms on the ring. For example, a ring with 3 to 4 carbon atoms is called a small ring, a ring with 5 to 6 carbon atoms is called a regular ring, a ring with 7 to 12 carbon atoms is called a medium ring, and a ring with more than 12 carbon atoms is called a large ring.

[0101] 3. If, on one subgraph, one of two linked atoms belongs to a ring and the degree of the other atom is greater than 1, the linked chemical bond between these two atoms is broken, thereby yielding two smaller subgraphs.

[0102] Exemplary, the reverse reaction prediction model predicts the synthesizer completion sequence based on a first order of the plurality of synthesizers and adds a basic graph for each synthesizer in the plurality of synthesizers, if the traverse order of the plurality of synthesizers is a second order, then based on the first and second orders, at least one synthesizer completion operation corresponding to each synthesizer in the synthesizer completion sequence can be determined, and further, a basic graph indicated by each synthesizer completion operation can be added according to the interface atom indicated by each synthesizer completion operation in the at least one synthesizer completion operation. Here, the first and second orders may be the same or different. For example, based on the first order, the second order, and the number of at least one synthesizer completion operations corresponding to each synthesizer, at least one synthesizer completion operation corresponding to each synthesizer in the synthesizer completion sequence can be determined, and further, a basic graph indicated by each synthesizer completion operation can be added according to the interface atom indicated by each synthesizer completion operation in the at least one synthesizer completion operation.

[0103] Assuming that the first order is equal to the second order, and that the number of at least one synth completion operations corresponding to each synth is a predetermined number, then, in accordance with the traverse order of the plurality of synths used when predicting the synth completion sequence in the reverse reaction prediction model, a basic graph indicated by each synth completion operation can be added to each synth in the plurality of synths, according to the interface atoms indicated by each synth completion operation in the predetermined number of synth completion operations corresponding to each synth. For example, assuming that the plurality of synths include synth 1 and synth 2, and that the first synth completion operation in the synth completion sequence is a synth completion operation for synth 1, then the remaining synth completion operations in the synth completion sequence, other than the first synth completion operation, are synth completion operations for synth 2. In this case, a plurality of reactant molecules corresponding to synth 1 can be obtained by adding the basic graph indicated by the first synth completion operation according to the interface atoms indicated by the first synth completion operation. Furthermore, by sequentially adding the interface atoms indicated by the remaining synthesizer completion operations to the basic graph indicated by the remaining synthesizer completion operations, a plurality of reactant molecules corresponding to synthesizer 2 can be obtained.

[0104] For example, the interface atoms indicated by each of the aforementioned synthetic atom completion operations are atoms that become connection nodes on the basic graph indicated by each of the aforementioned synthetic atom completion operations when synthetic atom completion is performed using the basic graph indicated by each of the aforementioned synthetic atom completion operations.

[0105] For example, when adding a basic graph (Motif) indicated by the synthetizer complementation operation, multiple reactant molecules corresponding to the product molecule can be obtained by using the interface atom indicated by the synthetizer complementation operation and the atom attached to the synthetizer complementation operation as connecting nodes. Note that the interface atom indicated by the synthetizer complementation operation and the atom attached to the synthetizer complementation operation are the same atom in the reactant molecule.

[0106] Exemplary, the attached atoms in the embodiments of the present invention include the atom selected for editing and the atoms at both ends of the chemical bond selected for editing.

[0107] In some embodiments, the synthesizer completion operation is represented by an operation tag for instructing synthesizer completion, a tag for instructing the basic graph, and a tag for instructing the interface atom.

[0108] For example, in a synthesizer completion sequence, the synthesizer completion operation is defined as a synthesizer completion triplet, i.e., (π3, z, q). π3 represents that the operation predicted by the reverse reaction prediction model is the synthesizer completion operation. (outside 1) TIFF0007858996000004.tif8170 represents the tag of the base graph indicated by the synthetizer completion operation predicted by the reverse reaction prediction model. q represents the tag of the interface atom corresponding to the synthetizer completion operation predicted by the reverse reaction prediction model. For example, (π3, z1, q1). π3 indicates that the operation predicted by the reverse reaction prediction model is synthetizer completion. z1 represents the tag of the base graph indicated by the synthetizer completion operation predicted by the reverse reaction prediction model, i.e., the base graph with the tag z1. q1 represents the tag of the interface atom corresponding to the synthetizer completion operation predicted by the reverse reaction prediction model, i.e., the atom with the tag q1 in the base graph with the tag z1. Based on this, based on (π3, z1, q1), the atom with the tag q1 in the base graph with the tag z1 can be used as the interface atom, and synthetizer completion can be performed using the base graph of z1.

[0109] Of course, in other alternative embodiments, the synthesizer completion operation can be defined as a tuple or a numerical value in another format. For example, if the synthesizer completion operation is defined as a tuple, it can specifically be defined as a tag for indicating the basic graph and a tag for indicating the interface atom.

[0110] By introducing a reverse reaction prediction model for predicting the conversion pathway from the product molecule to multiple reactant molecules, it becomes possible to integrate the synthetic atom prediction task and the synthetic atom completion prediction task. That is, the reverse reaction prediction model introduced in the embodiment of the present invention can learn the latent relationship between the two subtasks of synthetic atom prediction and complementary synthetic atom, thereby significantly improving the generalization performance of the model, reducing the prediction complexity of reactant molecules, and improving the generalization performance of reactant molecule prediction. Furthermore, by introducing a basic graph and designing the basic graph to include a structure that includes multiple atoms or atomic edges for connecting the multiple atoms, it is possible to rationally construct a short and accurate conversion pathway, avoiding excessively long synthetic atom completion sequences, reducing the difficulty of predicting reactant molecules, improving the prediction accuracy of reactant molecules, and improving the prediction performance of reactant molecules.

[0111] Furthermore, by improving the predictive performance of reactant molecules, the following technical benefits can also be obtained.

[0112] 1. Synthetic routes can be planned for designed drug or novel material molecules, thereby increasing the efficiency of research on drug or novel material molecules.

[0113] 2. By presenting several potential scientific disciplines, it is possible to provide new scientific knowledge.

[0114] 3. We can provide more precise synthesis route plans than expert researchers, predict reliable reactant molecules even when reaction templates are lacking, and predict reaction types that expert researchers have not yet clarified, thereby significantly improving the efficiency of new drug and material development.

[0115] In some embodiments, step S220 is, The process includes: a step of obtaining input features for the t-th operation based on the (t-1)-th operation predicted by the reverse reaction prediction model, where t is an integer greater than 1; and a step of predicting the t-th operation and obtaining the conversion path, based on the input features corresponding to the t-th operation and the hidden features corresponding to the t-th operation, until the operation predicted by the reverse reaction prediction model is a synthesizer complementation operation and the process has traversed all attached atoms on the plurality of synthesizers and all attached atoms on the basic graph added to the plurality of synthesizers, where the hidden features of the t-th operation are associated with operations predicted by the reverse reaction prediction model prior to the t-th operation.

[0116] Exemplary, the attached atoms include the atom selected for editing and the atoms at both ends of the chemical bond selected for editing.

[0117] For example, the inverse reaction prediction model may be a Recurrent Neural Network (RNN), and based on this, the (t-1)th action can correspond to time t-1, and the tth action can correspond to time t. That is, based on the action obtained by the prediction of the RNN at time t-1, the input features of the RNN at time t are obtained, and based on the input features of the RNN at time t and the hidden features of the RNN at time t, the action predicted by the RNN is a synthesizer completion action, and the action of the RNN at time t is predicted until all attached atoms on the plurality of synthesizers and all attached atoms on the basic graph added to the plurality of synthesizers have been traversed, thereby obtaining the transformation path.

[0118] A key feature of multilayer sensors and convolutional neural networks is the assumption that the input is an independent, context-less unit; for example, if the input is an image, the network recognizes whether it is a dog or a cat. However, in the case of sequential inputs with some clear contextual features, such as predicting the playback content of the next frame of a video, it is obvious that such an output must depend on the previous input. That is, the network needs to have some memory capacity, and RNNs can provide just such memory capacity to the network.

[0119] Of course, in other alternative embodiments, the reverse reaction prediction model may be a different model, and the present invention is not particularly limited thereto.

[0120] For example, if the inverse response prediction model predicts the t-th action, the output features u are based on the input features corresponding to the t-th action and the hidden features corresponding to the t-th action. t Next, we obtain the output feature u t and the characteristics of the product molecule h G The data is combined and processed to create a feature Ψ for recognizing the t-th action. t You can obtain this.

[0121] For example, according to the following method, the output feature u t and features Ψ for recognizing the t-th action t You can obtain it.

[0122]

number

[0123] For example, the inverse reaction prediction model can recognize the t-th action using the following equation.

[0124]

number

[0125] In the embodiment of the present invention, the output feature u t and the characteristics of the product molecule h G By combining and processing the features of the product molecule and the output features of the RNN, and using this to predict subsequent motion, global topology information can be integrated into the motion prediction process, thereby improving the accuracy of motion prediction.

[0126] In some embodiments, the conversion path can be determined using the following method.

[0127] According to the beam search method of hyperparameter k, the k first prediction results with the highest scores are obtained from the prediction results of the (t-1)th operation, and based on the k first prediction results, k first input features corresponding to the tth operation are determined, and the tth operation is predicted based on each of the k first input features and the hidden features corresponding to the tth operation. According to the beam search method of hyperparameter k, the k second prediction results with the highest scores are obtained from the prediction results obtained by the prediction, and based on the k second prediction results, k second input features corresponding to the (t+1)th operation are determined, and based on each of the k second input features and the hidden features corresponding to the (t+1)th operation, the operation obtained by the reverse reaction prediction model is determined to be a synthesizer complementation operation, and the (t+1)th operation is predicted until all attached atoms on the plurality of synthesizers and all attached atoms on the basic graph added to the plurality of synthesizers have been traversed, and the conversion path is obtained.

[0128] For example, the prediction device predicts the t-th action based on each of the k first input features and the hidden feature corresponding to the t-th action, and if 2k prediction results are obtained, it can then, according to a beam search scheme of hyperparameter k, obtain the k second prediction results with the highest scores from the 2k predicted results and determine them as the k second input features corresponding to the (t+1)-th action.

[0129] For example, when obtaining the k first prediction results with the highest scores from the prediction results of the (t-1)th operation according to a beam search method of hyperparameter k, the prediction results of the (t-1)th operation can first be sorted in order of score, and then the k prediction results with the highest scores from the (t-1)th operation can be selected as the k first prediction results. For example, when sorting the prediction results of the (t-1)th operation in order of score, the cumulative sum of the scores of all already predicted prediction results on the path in which each prediction result exists can first be calculated to obtain the cumulative score sum corresponding to each prediction result, and then the k prediction results with the highest cumulative sum scores from the (t-1)th operation can be selected as the k first prediction results.

[0130] Exemplary, a beam search with hyperparameter k is used to select multiple alternatives for an input sequence at each time step based on conditional probabilities. The number of alternatives depends on the hyperparameter k, called the beam width. At each time step, the beam search selects the k best alternatives with the highest probabilities as the most promising options at the current time step. That is, at each time step, the best result with the highest score k based on the log-likelihood score function is selected as the input for the next time step. In other words, this process can be described as the construction of a search tree, where the highest-scoring leaf node is expanded with its subnodes, while other leaf nodes are removed.

[0131] In some embodiments, if the (t-1)th operation is the editing operation, the input features of the tth operation are determined based on the features of the subgraph obtained by the editing performed by the (t-1)th operation. Based on the input features of the tth operation and the hidden features of the tth operation, the inverse response prediction model is used to predict the editing target and the edited state indicated by the tth operation, up to the point where the operation predicted by the inverse response prediction model is the last editing operation in the editing sequence, thereby obtaining the editing sequence.

[0132] For example, the last editing action mentioned above is the editing completion action.

[0133] In other words, if the (t-1)th operation is the editing operation, the input features of the tth operation are determined based on the features of the subgraph obtained by the editing performed by the (t-1)th operation. Based on the input features of the tth operation and the hidden features of the tth operation, the inverse response prediction model is used to predict the editing target and the edited state indicated by the tth operation, up to the point where the operation predicted by the inverse response prediction model is the editing completion operation, thereby obtaining the editing sequence.

[0134] For example, t is an integer greater than 1 or an integer greater than 2.

[0135] For example, if the (t-1)th operation is the first editing operation, then the (t-1)th operation is the initiation operation, and in this case, the features of the subgraph obtained by editing the (t-1)th operation are the features of the product molecule.

[0136] For example, if the reverse reaction prediction model predicts the t-th action, the intermediate molecular fragment processed based on the (t-1)-th action can be decoded, and the features of the intermediate molecular fragment edited based on the (t-1)-th action can be obtained.

[0137] For example, if the t-th operation is the editing operation, the reverse reaction prediction model scores s based on each chemical bond and each atom. i ^ By assigning a value, it is possible to predict the editing target corresponding to the t-th operation.

[0138] For example, if the t-th action is the editing action, the reverse reaction prediction model, in predicting the editing target corresponding to the t-th action, first assigns a score s to each chemical bond and each atom. i ^A score is assigned, which represents the probability that a chemical bond or atom is considered for editing in the (t-1)th operation, and then a score s is assigned based on each chemical bond and each atom. i ^ Assign a value and predict the editing target corresponding to the t-th operation.

[0139] For example, the reverse reaction prediction model assigns a score s to each chemical bond and each atom using the following formula. i ^ It is possible to find this.

[0140]

number

[0141] Next, the reverse reaction prediction model determines the edited state r for the edited object corresponding to the t-th operation. b ^ To predict.

[0142] For example, the inverse reaction prediction model predicts the edited state r of the target to be edited corresponding to the t-th operation, using the following equation. b ^ It can be predicted.

[0143]

number

[0144]

number

[0145] For example, the reverse reaction prediction model applies the edit target corresponding to the predicted t-th action and the edited state corresponding to the t-th action to the unedited intermediate molecular fragment corresponding to the t-th action to obtain the edited intermediate molecular fragment corresponding to the t-th action, and again calculates the characteristics h of the edited intermediate molecular fragment corresponding to the t-th action using MPNM(·). t syn Next, the characteristics h of the edited intermediate molecular fragment corresponding to the obtained t-th operation are obtained. t syn Based on the editing target corresponding to the t-th operation and the edited state corresponding to the t-th operation, input features corresponding to the (t+1)th operation t We seek.

[0146] For example, the input feature corresponding to the (t+1)th action is defined by the following formula. t You can obtain this.

[0147]

number

[0148]

number

[0149] In some embodiments, if the (t-1)th operation is the last editing operation in the editing sequence or the synthesizer completion operation, the input features of the tth operation are determined based on the features of the subgraph obtained by editing using the (t-1)th operation and the features of the attached atoms corresponding to the (t-1)th operation. Based on the input features of the tth operation and the hidden features of the tth operation, the fundamental graph and interface atoms indicated by the tth operation are predicted up to the point where the operation predicted by the reverse reaction prediction model is a synthesizer completion operation and all attached atoms on the plurality of synthesizers and all attached atoms on the fundamental graph added to the plurality of synthesizers have been traversed, thereby obtaining the synthesizer completion sequence.

[0150] For example, the last editing action mentioned above is the editing completion action.

[0151] In other words, if the (t-1)th operation is an editing completion operation or a synthesizer completion operation, the input features of the tth operation are determined based on the features of the subgraph obtained by editing using the (t-1)th operation and the features of the attached atoms corresponding to the (t-1)th operation. Based on the input features of the tth operation and the hidden features of the tth operation, the fundamental graph and interface atoms indicated by the tth operation are predicted up to the point where the operation predicted by the reverse reaction prediction model is a synthesizer completion operation and all attached atoms on the plurality of synthesizers and all attached atoms on the fundamental graph added to the plurality of synthesizers have been traversed, thereby obtaining the synthesizer completion sequence.

[0152] For example, if the operation obtained by the reverse reaction prediction model is the final editing operation, it indicates that the atom prediction stage is complete and the reactant prediction process has simultaneously entered the atom completion stage. In this case, all attached atoms are sorted according to the tag order of the product molecule for atom completion.

[0153] For example, if the action predicted by the inverse response prediction model is the last editing action, the input feature corresponding to action (t+1) is calculated according to the following formula. t It is possible to make a decision.

[0154]

number

[0155] For example, in the synthetizer completion stage, the reverse reaction prediction model traverses sequentially through all attached atoms on the synthetizer and all attached atoms on the added base graph, assigning a base graph (motif) to each attached atom. Predicting the base graph can be considered a multi-classification task in a pre-stored dictionary Z. Predicting the base graph based on the base graph results in the base graph Z. ^ After obtaining the basic graph z ^ Upper attached atom a t The corresponding interface atom q ^ It is possible to make a decision.

[0156] For example, the reverse reaction prediction model corresponds to the t-th operation, according to the following equation, and is represented by the basic graph z ^ It can be predicted.

[0157]

number

[0158] For example, the reverse reaction prediction model corresponds to the t-th operation, according to the following equation, and is represented by the basic graph z ^ Upper attached atom a t The corresponding interface atom q ^ It can be predicted.

[0159]

number

[0160] When the (t - 1)-th operation is the last editing operation or the synthon complementation operation, and the interface atom of the basic graph corresponding to the (t - 1)-th operation contains only one atom, based on the features of the subgraph obtained by editing using the (t - 1)-th operation and the features of the attachment atom corresponding to the (t - 1)-th operation, determine the input features of the t-th operation.

[0161] Exemplarily, when the predicted basic graph (outer 3) TIFF0007858996000016.tif8170 contains only one interface atom, according to the following formula, the input feature input t corresponding to the (t + 1)-th operation can be determined.

[0162]

Equation

[0163] If the (t-1)th operation is the last editing operation or the synthesizer completion operation, and the interface atoms of the basic graph corresponding to the (t-1)th operation contain only a few interface atoms, the input features of the tth operation are determined based on the features of the subgraph obtained by editing using the (t-1)th operation, the features of the basic graph corresponding to the (t-1)th operation, and the features of the attached atoms corresponding to the (t-1)th operation.

[0164] For example, a predicted base graph (outside 4) If TIFF0007858996000018.tif9170 contains only multiple interface atoms, then the input feature corresponding to the operation of (t+1) number is determined according to the following formula. t It is possible to make a decision.

[0165]

number

[0166] As can be seen from the above techniques, once the editing sequence and synthesizer completion sequence are obtained by the prediction of the reverse reaction prediction model, a transformation pathway can be obtained, and by applying this transformation pathway to the product molecular graph, a reactant molecular graph can be obtained.

[0167] Figure 3 shows an example of a conversion path provided by an embodiment of the present invention.

[0168] As shown in Figure 3, when predicting reactants using the reverse reaction prediction model for the product molecule shown on the left side of Figure 3(a), the transformation path shown in Figure 3(b) is obtained according to the flow shown in Figure 3(c), and further, the two reactant molecules shown on the right side of Figure 3(a) can be obtained. Here, π1 represents that the prediction operation by the reverse reaction prediction model is an editing operation, π2 represents that the prediction operation by the reverse reaction prediction model is an editing completion operation, π3 represents that the operation predicted by the reverse reaction prediction model is a synthesizer completion operation, a1~a3 represent attached atoms, q1~q4 represent interface atoms, z1~z3 represent basic graphs, b represents that the editing target corresponding to the editing operation is the chemical bond with tag b, and none represents that the edited state corresponding to the editing operation is the deletion of the chemical bond with tag b. Attached atoms are interfaces for adding basic graphs (motif).

[0169] Specifically, if the input to the reverse reaction prediction model is the starting operation, the tag of the first predicted editing operation is triplet (π1,b, none), then triplet (π1,b, none) is input, and the tag of the second predicted operation is π2, then all attached atoms are sorted in tag order in the product molecule, and tuples (π3,a1), (π3,a2), (π3,a2) are input in order, and the tags of the predicted operations in order are triplet (π3,z1,q1), triplet (π3,z2,q2), triplet (π3,z3,q4), and based on this, the resulting transformation path is defined as the path shown in Figure 3(b): triplet (π1,b, none), π2, triplet (π3,z1,q1), triplet (π3,z2,q2), triplet (π3,z3,q4). This allows the conversion pathway to act on the product molecule, yielding the reactant molecule shown on the right side of Figure 3(a).

[0170] In other words, in the reactant prediction process shown in Figure 3, the editing sequence contains only one editing operation, which is defined as a triplet (π2, b, none). The synthesizer completion sequence contains three synthesizer completion operations, which are defined as triplets (π3, z1, q1), triplets (π3, z2, q2), and triplets (π3, z3, q4), respectively.

[0171] Furthermore, when adding a basic graph (Motif) indicated by the aforementioned synthetizer completion operation, the interface atom indicated by the synthetizer completion operation and the atom attached to the synthetizer completion operation can be used as connection nodes. Note that the interface atom indicated by the synthetizer completion operation and the atom attached to the synthetizer completion operation are the same atom in the reactant molecule. For example, when adding z1 based on a triplet (π3, z1, q1), a1 and q1 are the same atom in the reactant molecule (i.e., N atom). Similarly, when adding z2 based on a triplet (π3, z2, q2), a2 and q2 are the same atom in the reactant molecule (i.e., O atom). After adding z2, the interface atom q3 is changed to the attached atom a3. In this case, when adding z3 based on a triplet (π3, z3, q4), a3 and q4 are the same atom in the reactant molecule (i.e., C atom).

[0172] Figure 3 is merely an example of the present invention and should not be understood as a limitation of the present invention.

[0173] For example, in other alternative embodiments, the transformation pathway may further include a number of editing or synthesizer completion operations, and consequently, a number of synthesizers or reactant molecules, and the embodiments of the present invention are not particularly limited to these.

[0174] Figure 4 is another schematic flowchart of the reactant molecule prediction method provided by the embodiment of the present invention. Here, π1 represents the operation of the prediction by the reverse reaction prediction model as an editing operation, π2 represents the operation of the prediction by the reverse reaction prediction model as an editing completion operation, π3 represents the operation predicted by the reverse reaction prediction model as a synthesizer completion operation, a1~a3 represent attached atoms, q1~q3 represent interface atoms, z1~z3 represent basic graphs, g represents the editing target corresponding to the editing operation as the chemical bond of tag g, and none represents the edited state corresponding to the editing operation as the chemical bond of tag g being deleted. Attached atoms are interfaces to which the basic graph (motif) has been added.

[0175] Specifically, if the input to the reverse reaction prediction model is the starting operation, the tag of the first predicted editing operation is triplet (π1,g, none), then triplet (π1,g, none) is input, and the tag of the second predicted operation is π2, then all attached atoms are sorted in tag order in the product molecule, and tuples (π3,a1), (π3,a2), (π3,a3) are input in order, and the tags of the predicted operations in order are triplet (π3,z1,q1), triplet (π3,z2,q2), triplet (π3,z3,q4), and based on this, the obtained transformation path is defined as the path shown in Figure 4: triplet (π1,g, none), π2, triplet (π3,z1,q1), triplet (π3,z2,q2), triplet (π3,z3,q4). Here, a1-a3, q1-q3, and z1-z3 are as shown in the figure, and based on this, the conversion pathway can be applied to the product molecule to obtain the final reactant molecule.

[0176] In the reactant prediction process shown in Figure 4, the editing sequence includes only one editing operation, which is defined as a triplet (π1, g, none). The synthetic component completion sequence includes three synthetic component completion operations, defined as triplets (π3, z1, q1), (π3, z2, q2), and (π3, z3, q4), respectively. When adding a basic graph (Motif) indicated by the synthetic component completion operation, the interface atom indicated by the synthetic component completion operation and the atom attached to the synthetic component completion operation can be used as connection nodes. The interface atom indicated by the synthetic component completion operation and the atom attached to the synthetic component completion operation are the same atom in the reactant molecule.

[0177] In other words, the reactant prediction process includes an editing stage and a base graph addition stage, where the editing stage describes the bonding and atomic changes from the product to the synthesizer, i.e., the synthesizer prediction process, while the base graph addition stage completes the reactant formation by adding an appropriate base graph to the synthesizer.

[0178] In the editing phase, the input molecular graph is first encoded by a graph neural network (GNN) to obtain the GNN output. Then, if the t-th operation is an editing operation, the circulating neural network (RNN) predicts the operation based on the GNN output corresponding to the t-th operation and the hidden state output by the previous node. If the t-th operation is an editing completion operation or a synthetic atom completion operation, the RNN predicts the operation based on the GNN output corresponding to the t-th operation, the attached atoms corresponding to the t-th operation, and the hidden state output by the previous node. In other words, in the editing phase, the RNN predicts the editing sequence step by step until it predicts and obtains an editing completion operation, then ends the editing phase and starts the basic graph addition phase. In the basic graph addition phase, the RNN sequentially adds the basic graph until all attached atoms have been traversed.

[0179] In the example in Figure 4, the first editing operation is applied to the chemical bond S=O, the new chemical bond type is none, and represents the deletion of the aforementioned chemical bond. In the synthesizer completion operation, the interface atoms (q1, q2, and q3) in the basic graph and the attached atoms (a1, a2, and a3) in the synthesizer / intermediate represent the same atom, and when the basic graph is attached to the synthesizer / intermediate, they are bonded to a single atom. For example, when z1 is added based on the triplet (π3, z1, q1), a1 and q1 are the same atom in the reactant molecule (i.e., S atom). Similarly, when z2 is added based on the triplet (π3, z2, q2), a2 and q2 are the same atom in the reactant molecule (i.e., O atom). After adding z2, the interface atom q3 is changed to the attached atom a3, and in this case, when z3 is added based on the triplet (π3, z3, q4), a3 and q4 are the same atom in the reactant molecule (i.e., C atom).

[0180] Figures 3 and 4 are merely examples of the present invention and should not be understood as limitations of the present invention.

[0181] For example, in other alternative embodiments, the conversion path may not include a start operation or an edit completion operation.

[0182] Figure 5 is a schematic flowchart of a training method 300 for a reverse reaction prediction model provided by an embodiment of the present invention, the training method 300 of which can be performed by any electronic device having data processing capabilities. For example, the electronic device can be implemented as a server. The server may be an independent physical server, or a server cluster or distributed system consisting of multiple physical servers, or it may be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, networking services, cloud communications, middleware services, domain services, security services, and big data and artificial intelligence platforms, and the server may be connected directly or indirectly by wired or wireless communication, and the present invention is not limited thereto. For convenience of explanation, the prediction method provided in the present invention will be described in the following paragraphs using a training apparatus for a predictive model of reactant molecules as an example.

[0183] As shown in Figure 5, the training method 300 is In step S310, feature extraction is performed on the product molecule to obtain the characteristics of the product molecule.

[0184] In step S320, based on the characteristics of the product molecule, a reverse reaction prediction model is used to predict the conversion pathway from the product molecule to multiple reactant molecules.

[0185] Here, the transformation path includes an editing sequence and a synthesizer completion sequence, where each editing operation in the editing sequence is used to indicate the object to be edited and the edited state, where the object to be edited is an atom or chemical bond in the product molecule, and for a plurality of synthesizers of the product molecule obtained from the editing sequence, the synthesizer completion sequence includes at least one synthesizer completion operation corresponding to each of the plurality of synthesizers, where each synthesizer completion operation in the at least one synthesizer completion operation is used to indicate a basic graph and an interface atom, where the basic graph includes a plurality of atoms or atomic edges for linking the plurality of atoms.

[0186] In step S330, the inverse response prediction model is trained based on the loss between the conversion path and the training path.

[0187] Based on the above technical solutions, by introducing a conversion pathway from the product molecule to multiple reactant molecules, the reverse reaction prediction model can learn the potential relationship between the two subtasks of synthesizer prediction and complementary synthesizer prediction. This significantly improves the generalization performance of the model, reduces the prediction complexity of reactant molecules, and enhances the generalization performance of reactant molecule prediction.

[0188] Furthermore, by introducing a basic graph and designing it to include a structure with multiple atoms or atomic edges for connecting those multiple atoms, it is possible to rationally construct a short and accurate transformation pathway, avoid the length of the synthesizer completion sequence being too long, reduce the difficulty of predicting reactant molecules, improve the accuracy of predicting reactant molecules, and enhance the predictive performance of reactant molecules.

[0189] In some embodiments, before step S320, the method 300 is performed. The method may further include the steps of obtaining candidate reactant molecules corresponding to the product molecule, obtaining a basic graph dictionary by comparing the molecular structures of the product molecule and the candidate reactant molecule, and obtaining the training path based on the basic graph dictionary.

[0190] For example, the candidate reactant molecules may be all of the reactant molecules of the product molecule. That is, the candidate reactant molecules can be used to produce all of the reactant molecules of the product molecule.

[0191] In some embodiments, a linked tree is constructed based on the basic graph dictionary, and the linked tree includes a tree structure in which the plurality of compositers are root nodes and the basic graphs in the basic graph dictionary are subnodes, and the shortest path is determined as the training path by traversing the linked tree.

[0192] The reactant molecule is decomposed into a synthetizer and a motif, where the synthetizer is a molecular fragment whose chemical bonds have been broken in the product molecular graph, and the motif is a subgraph of the reactant. Therefore, the connectivity between the synthetizer and the motif is maintained by constructing a linked tree. The linked tree represents the synthetizer and the motif as a layered tree structure, where the synthetizer is the root node and the motif is a subnode. The edges between two nodes in the linked tree represent the direct connection between two subgraphs in the reactant molecular graph, where triplets of attached atoms, motifs, and interface atoms can be used to represent each edge.

[0193] In this embodiment, by constructing a tree structure (i.e., a linked tree), the connection relationships between the composites and the basic graph are represented, and the linked tree provides an effective strategy for constructing training paths and can reduce the complexity of training.

[0194] For example, a depth-first search can be used to traverse the entire connected tree, and the shortest path after the traverse can be determined as the training path.

[0195] Here, the depth-first traverse method can refer to traversing starting from a vertex v in the tree structure according to the following method.

[0196] 1. Access vertex v.

[0197] 2. Starting from the unaccessed adjacent points of v, perform a depth-first traverse of the tree structure until all vertices that have a path to v in the tree structure have been accessed.

[0198] 3. If there are still vertices in the tree structure that have not been accessed at this point, perform a depth-first traverse again, starting from the unaccessed vertices, until all vertices in the tree structure have been accessed.

[0199] Depth-first search follows a search strategy that searches the tree as "deeply" as possible. The basic idea is that to find the solution to a problem, you first select a possible situation and proceed forward (by subnode) exploration. If, during the exploration, it turns out that the original selection does not meet the requirements, you go back to the parent node, select a different node, and continue the search forward, repeating this process until the optimal solution is found. In other words, depth-first search is the concept of going as deep as possible, starting from vertex V0, going all the way to the end along a certain path, and if you find that you cannot reach the target solution, you go back to the previous node, then go to the end along a different path, and so on.

[0200] In some embodiments, molecular fragments other than multiple synths in the candidate reactant molecule are determined as multiple candidate subgraphs. If the first candidate subgraph in the multiple candidate subgraphs contains a first atom and a second atom, and the first atom and the second atom belong to different rings, the chemical bond between the first atom and the second atom is broken to obtain multiple first subgraphs. If the second candidate subgraph in the multiple candidate subgraphs contains linked third and fourth atoms, and one atom in the third atom and the fourth atom belongs to a ring, and the degree of another atom in the third atom and the fourth atom is greater than or equal to a predetermined value, the chemical bond between the third atom and the fourth atom is broken to obtain multiple second subgraphs. The candidate subgraphs excluding the first and second candidate subgraphs in the multiple candidate subgraphs, the multiple first subgraphs, and the multiple second subgraphs are determined as basic graphs in the basic graph dictionary.

[0201] Exemplary, a product molecule can be broken down into a set of incomplete subgraphs called synthesizers, and after a suitable motif is attached to each atom, the synthesizers can be reconstructed into a reactant molecular graph. In other words, a motif can be considered as a subgraph on a reactant molecular graph. Therefore, in the embodiments of the present invention, the process of extracting motifs is divided into the following steps.

[0202] 1. Cut the edges where there is a difference between the synthesizer and its corresponding reactant to obtain a series of subgraphs. These subgraphs hold interface atoms corresponding to the attached atoms on the synthesizer.

[0203] Furthermore, the reactants corresponding to the synthesizer may include molecules or reactants containing the synthesizer.

[0204] 2. If two linked atoms on one subgraph each belong to two rings, the linked chemical bond between these two atoms is broken, thereby obtaining two smaller subgraphs.

[0205] Furthermore, the rings in the embodiments of the present invention may be monorings, that is, those with only one ring in the molecule. Correspondingly, if two linked atoms in one subgraph each belong to two monorings, the chemical bond between these two atoms is broken, thereby yielding two smaller subgraphs. Moreover, the rings in the embodiments of the present invention may be cycloalkanes, which may be classified based on the number of carbon atoms on the ring. For example, a ring with 3 to 4 carbon atoms is called a small ring, a ring with 5 to 6 carbon atoms is called a regular ring, a ring with 7 to 12 carbon atoms is called a medium ring, and a ring with more than 12 carbon atoms is called a large ring.

[0206] 3. If, on one subgraph, one of two linked atoms belongs to a ring and the degree of the other atom is greater than 1, the linked chemical bond between these two atoms is broken, thereby yielding two smaller subgraphs.

[0207] Ultimately, the number of basic graphs (motifs) is determined by a predetermined number of dictionaries. (outside 5) TIFF0007858996000020.tif8170 can be extracted. For example, a dictionary with 210 basic graph (motif) counts. (outside 6) The file TIFF0007858996000021.tif8170 can be extracted.

[0208] In some embodiments, before step S330, the method 300 is performed. The loss between the transformation path and the training path is, The process may further include steps determined based on information such as the difference between the tags of the predicted actions in the conversion path and the tags of the training actions in the training path; the difference between the score of the edited target corresponding to the predicted action and the score of the edited target corresponding to the training action; the difference between the edited state corresponding to the predicted action and the edited state corresponding to the training action; the difference between the basic graph indicated by the predicted action and the basic graph indicated by the training action; and the difference between the interface atoms indicated by the predicted action and the interface atoms indicated by the training action.

[0209] In the embodiment of the present invention, the training objective of the reverse reaction prediction model is to provide a training pathway and predict the transformation pathway; therefore, reverse reaction prediction is modeled as a molecule generation problem based on autoregression. That is, one product molecule G P Given, for each step t, the autoregressive model creates a new graph structure G based on the graph structure of the history. t Obtain the reactant molecule G. R When this is predicted, it explains that the generation process is complete. Therefore, the generation process of the reactant molecular graph can be defined as the following binding probability likelihood function.

[0210]

number

[0211] Note that intermediate molecular fragment G tIt is not directly generated by the reverse reaction prediction model. Instead, based on historical operations, the reverse reaction prediction model generates a new graph editing operation, an editing target (i.e., a chemical bond, an atom, or a basic graph), and its state after editing (i.e., a new chemical bond type or an interface atom), and applies it to the intermediate molecular fragment in the previous step to obtain a new intermediate molecular fragment. Based on this, when providing the editing target, the state after editing, and the intermediate molecular fragment of the history, the likelihood function can be changed as follows.

[0212]

Number

[0213] For example, the loss between the conversion path and the training path can be determined using the following formula.

[0214]

Number

[0215] Note that the implementation method of step S320 in the training method 300 can refer to the implementation method of step S220 in the prediction method 200. To avoid duplication, the description is omitted here.

[0216] Hereinafter, preferred embodiments of the present invention will be described while referring to the accompanying drawings. However, the present invention is not limited to the specific details of the above embodiments, and within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solution of the present invention, and all of these simple modifications are included in the protection scope of the present invention.

[0217] For example, each specific technical feature described in the above specific embodiments can be combined in any appropriate manner as long as there is no contradiction. To avoid unnecessary duplication, the present invention omits further description of various possible combination methods. Also, for example, any combination is possible among various different embodiments of the present invention, and as long as it does not go against the idea of the present invention, it should be regarded as the content disclosed in the present invention as well.

[0218] For example, in order to reduce the convergence difficulty, the model can be trained by adopting a teacher-forcing strategy by the teacher.

[0219] Note that an RNN has two training modes: a free-running mode and a teacher-forcing mode. The free-running mode refers to using the output of the previous state as the input of the next state. The operating principle of the teacher-forcing mode is to use the desired output or actual output y(t) of the training dataset instead of the output h(t) generated by the model as the input x(t + 1) for the next time step during the training process.

[0220] Also, for example, after pre-training based on the basic graph dictionary, the benchmark dataset can also be fine-tuned.

[0221] Also, for example, hyperparameters in the reverse reaction prediction model, such as the number of layers of GNN and GRU, can be adjusted.

[0222] In the various embodiments of the present invention, the size of the sequence number of each process does not indicate the order of execution, and the execution order of each process is determined by its function and inherent logic, and does not limit the implementation process of the embodiments of the present invention in any way.

[0223] The above describes the method provided by the embodiments of the present invention, and the following describes the apparatus provided by the embodiments of the present invention.

[0224] Figure 6 is a schematic block diagram of the reactant molecule prediction device 400 provided by an embodiment of the present invention.

[0225] As shown in Figure 6, the reactant molecule prediction device 400 is An extraction unit 410 is configured to perform feature extraction on the product molecule and obtain the characteristics of the product molecule, A prediction unit 420 is configured to predict a conversion pathway from the product molecule to a plurality of reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule, wherein the conversion pathway includes an editing sequence and a synth unit completion sequence, the editing sequence being a sequence formed by an editing operation, and the synth unit completion sequence being a sequence formed by a synth unit completion operation. An editing unit 430 is configured to edit the editing target indicated by each editing operation in the editing sequence according to the edited state indicated by each editing operation, in order to obtain a plurality of synthesizers corresponding to the product molecule, wherein the editing target is an atom or chemical bond in the product molecule. An additional unit 440 is configured to obtain a plurality of reactant molecules corresponding to the plurality of synthesizers by adding a basic graph indicated by each synthesizer complementation operation, according to the interface atoms indicated by each synthesizer complementation operation in the at least one synthesizer complementation operation, based on at least one synthesizer complementation operation corresponding to each synthesizer in the synthesizer complementation sequence, wherein the basic graph includes a plurality of atoms or atomic edges for connecting the plurality of atoms.

[0226] In some embodiments, the prediction unit 420 specifically, Based on the (t-1)th action obtained by the prediction using the aforementioned inverse reaction prediction model, the input features for the tth action are obtained, where t is an integer greater than 1. Based on the input features corresponding to the t-th operation and the hidden features corresponding to the t-th operation, the system is configured to predict the t-th operation and obtain the transformation path, up to the point where the operation predicted by the inverse reaction prediction model is a synthesizer completion operation, and when traversing all attached atoms on the plurality of synthesizers and all attached atoms on the basic graph added to the plurality of synthesizers has been determined.

[0227] Here, the hidden features of the t-th action are associated with the action predicted by the inverse response prediction model prior to the t-th action.

[0228] In some embodiments, the prediction unit 420 specifically, According to the beam search method for hyperparameter k, the k first prediction results with the highest scores are obtained from the prediction results of the (t-1)th operation. Based on the k first prediction results, it is configured to determine k first input features corresponding to the t-th operation. Here, obtaining the aforementioned conversion path means Based on each of the k first input features and the hidden feature corresponding to the tth operation, the tth operation is predicted, and the k second prediction results with the highest scores are obtained from the 2k predicted prediction results according to the beam search method of hyperparameter k. Based on the k second prediction results, k second input features corresponding to the (t+1)th operation are determined. Based on the k second input features and the hidden feature corresponding to the operation of (t+1) in each of the k second input features, the operation obtained by the prediction by the reverse reaction prediction model is a synthesizer completion operation, and the operation of (t+1) is predicted up to the point where all attached atoms on the plurality of synthesizers and all attached atoms on the basic graph added to the plurality of synthesizers have been traversed, thereby obtaining the transformation path.

[0229] In some embodiments, the prediction unit 420 specifically, If the (t-1)th operation is the editing operation, the input features of the tth operation are determined based on the features of the subgraph obtained by the editing performed by the (t-1)th operation. Based on the input features of the tth operation and the hidden features of the tth operation, the inverse response prediction model is used to predict the editing target and the edited state indicated by the tth operation until the operation predicted by the inverse response prediction model is the last editing operation in the editing sequence, thereby obtaining the editing sequence. When the (t-1)-th operation is the last editing operation or the synthon complementation operation, based on the characteristics of the subgraph obtained by the editing in the (t-1)-th operation and the characteristics of the attachment atoms corresponding to the (t-1)-th operation, determine the input characteristics of the t-th operation. Based on the input characteristics of the t-th operation and the hidden characteristics of the t-th operation, if the operation obtained by the prediction by the reverse reaction prediction model is a synthon complementation operation, and until traversing all the attachment atoms on the plurality of synthons and all the attachment atoms on the basic graph added to the plurality of synthons, predict the basic graph and the interface atoms indicated by the t-th operation, and is configured to obtain the synthon complementation sequence.

[0230] In some embodiments, the editing operation is represented by an operation tag for instructing editing, a tag for instructing the editing target, and a tag for instructing the state after editing, and the synthon complementation operation is represented by an operation tag for instructing synthon complementation, a tag for instructing the basic graph, and a tag for instructing the interface atoms.

[0231] In some embodiments, when the editing target indicated by the editing operation is an atom, the state after editing indicated by the editing operation is to change the number of charges on the atom or the number of hydrogen atoms on the atom. When the editing target indicated by the editing operation is a chemical bond, the state after editing indicated by the editing operation is any one of adding a chemical bond, deleting a chemical bond, and changing the type of a chemical bond.

[0232] FIG. 7 is a schematic block diagram of a training device 500 for a reverse reaction prediction model provided by an embodiment of the present invention.

[0233] As shown in FIG. 7, the training device 500 for the reverse reaction prediction model includes an extraction unit 510 configured to perform feature extraction on a product molecule to obtain the features of the product molecule, A prediction unit 520 is configured to predict the conversion pathway from the product molecule to a plurality of reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule, Here, the transformation path includes an editing sequence and a synthesizer completion sequence, each editing operation in the editing sequence is used to indicate the object to be edited and the edited state, the object to be edited is an atom or chemical bond in the product molecule, and for a plurality of synthesizers of the product molecule obtained from the editing sequence, the synthesizer completion sequence includes at least one synthesizer completion operation corresponding to each of the plurality of synthesizers, each synthesizer completion operation in the at least one synthesizer completion operation is used to indicate a base graph and interface atoms, the base graph includes a plurality of atoms or atomic edges for connecting the plurality of atoms, the prediction unit 520, The system includes a training unit 530 configured to train the inverse response prediction model based on the losses between the conversion path and the training path.

[0234] In some embodiments, before the prediction unit 520 obtains the conversion path, the device A candidate reactant molecule corresponding to the aforementioned product molecule is obtained, By comparing the molecular structures of the product molecule and the candidate reactant molecule, a basic graph dictionary is obtained. The system is configured to acquire the training path based on the aforementioned basic graph dictionary.

[0235] In some embodiments, the prediction unit 520 specifically, A linked tree is constructed based on the aforementioned basic graph dictionary, and the linked tree includes a tree structure in which the plurality of compositers are the root node and the basic graphs in the aforementioned basic graph dictionary are the subnodes. The system is configured to determine the shortest path as the training path by traversing the aforementioned linked tree.

[0236] In some embodiments, the prediction unit 520 specifically, Multiple molecular fragments other than the multiple synths in the candidate reactant molecule are determined as multiple candidate subgraphs. If the first candidate subgraph among the plurality of candidate subgraphs contains a first atom and a second atom, and the first atom and the second atom belong to different rings, the chemical bond between the first atom and the second atom is broken to obtain a plurality of first subgraphs. If the second candidate subgraph in the plurality of candidate subgraphs includes a linked third atom and a fourth atom, and one of the third atom and the fourth atom belongs to a ring, and the degree of the other atom of the third atom and the fourth atom is greater than or equal to a predetermined value, then the chemical bond between the third atom and the fourth atom is broken to obtain a plurality of second subgraphs. The system is configured to determine the candidate subgraphs other than the first candidate subgraph and the second candidate subgraph among the plurality of candidate subgraphs, and the plurality of first subgraphs and the plurality of second subgraphs as the basic graphs in the basic graph dictionary.

[0237] In some embodiments, the training unit 530 further trains the inverse response prediction model based on the loss of the conversion path and the training path. The loss between the transformation path and the training path is, The system is configured to be determined based on the following information: the difference between the tags of the predicted operation in the conversion path and the tags of the training operation in the training path; the difference between the score of the edited target corresponding to the predicted operation and the score of the edited target corresponding to the training operation; the difference between the edited state corresponding to the predicted operation and the edited state corresponding to the training operation; the difference between the basic graph indicated by the predicted operation and the basic graph indicated by the training operation; and the difference between the interface atom indicated by the predicted operation and the interface atom indicated by the training operation.

[0238] Note that the apparatus examples and method examples can correspond to each other, and the same explanation can be found in the method examples. To avoid duplication, the explanation is omitted here. Specifically, the reactant molecule prediction apparatus 400 corresponds to a corresponding entity that performs method 200 according to an embodiment of the present invention, and each unit in the prediction apparatus 400 is configured to perform the corresponding flow in method 200. Similarly, the inverse reaction prediction model training apparatus 500 corresponds to a corresponding entity that performs method 300 of an embodiment of the present invention, and each unit in the training apparatus 500 is configured to perform the corresponding flow in method 300. For simplicity, this will not be repeated here.

[0239] Furthermore, each unit in the prediction device 400 or training device 500 according to an embodiment of the present invention may be integrated into one or more other units, or some of the units therein may be further divided into a plurality of functionally smaller units, thereby enabling the same operation without impairing the realization of the technical effects of the embodiment of the present invention. The above units are divided based on logical function, and in actual application, the function of one unit may be realized by multiple units, or the function of multiple units may be realized by one unit. In another embodiment of the present invention, the prediction device 400 or training device 500 may include other units, and in actual application, these functions may be realized by the cooperation of other units, or by the cooperation of multiple units. According to another embodiment of the present invention, a prediction device 400 or training device 500 according to an embodiment of the present invention can be configured by executing a computer program (including program code) capable of executing each step of the corresponding method on a general-purpose computing device of a general-purpose computer including processing elements and storage elements such as a central processing unit (CPU), a random access storage medium (RAM), and a read-only storage medium (ROM), thereby realizing the method of the embodiment of the present invention. The computer program is written on, for example, a computer-readable storage medium, and the corresponding method of the embodiment of the present invention is realized by loading the computer-readable storage medium into an electronic device and executing it therein.

[0240] In other words, the above units can be implemented in hardware form, in software form as instructions, or in the form of a combination of software and hardware. Specifically, each step of the method embodiment in the embodiment of the present invention may be implemented by hardware integrated logic circuits and / or software form instructions within a processor, and the steps of the method disclosed in the embodiment of the present invention may be embodied in a hardware coding processor and executed directly, or may be implemented as a combination of hardware and software execution within a coding processor. Exemplarily, the software may be stored in a storage medium familiar in the art, such as random memory, flash memory, read-only memory, programmable read-only memory, electrically rewritable programmable memory, registers, etc. The storage medium is placed in memory, and the processor reads the information in memory and combines it with its hardware to implement the steps in the method embodiment described above.

[0241] Figure 8 is a schematic diagram of the electronic device 600 provided by an embodiment of the present invention.

[0242] As shown in Figure 8, the electronic device 600 comprises at least a processor 610 and a computer-readable storage medium 620. Here, the processor 610 and the computer-readable storage medium 620 may be connected by a bus or in other forms. The computer-readable storage medium 620 is configured to store a computer program 621, the computer program 621 comprises computer instructions, and the processor 610 is configured to execute the computer instructions stored in the computer-readable storage medium 620. The processor 610 is the computing core and control core of the electronic device 600 and is configured to implement one or more computer instructions, specifically, to implement a corresponding method flow or corresponding function by loading and executing one or more computer instructions.

[0243] For example, processor 610 can be called a central processing unit (CPU). Processor 610 includes, but is not limited to, general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.

[0244] For example, the computer-readable storage medium 620 may be high-speed RAM memory, non-volatile memory such as at least one disk memory, or, exemplary, at least one computer-readable storage medium located away from the aforementioned processor 610. Specifically, the computer-readable storage medium 620 includes, but is not limited to, volatile memory and / or non-volatile memory. Here, non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may be random access memory (RAM) used as an external cache. From illustrative but not limited explanations, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synch-linked dynamic random access memory (SLDRAM), and direct memory bus random access memory (DR RAM).

[0245] As shown in Figure 8, the electronic device 600 may further include a transceiver 630.

[0246] Here, the processor 610 can control the transceiver 630 to communicate with other devices, specifically, to transmit information or data to other devices or to receive information or data transmitted by other devices. The transceiver 630 may include a transmitter and a receiver. The transceiver 630 may further include an antenna, which may be one or more.

[0247] Furthermore, each component of the electronic device 600 is connected via a bus system, where the bus system further includes a power bus, a control bus, and a status signal bus in addition to the data bus.

[0248] In one implementation, the electronic device 600 may be an electronic device having data processing capabilities, and a first computer instruction is stored in the computer-readable storage medium 620. The first computer instruction stored in the computer-readable storage medium 620 is loaded and executed by the processor 610 to realize the corresponding steps in the method embodiment shown in Figure 1. In a specific implementation, the first computer instruction in the computer-readable storage medium 620 is loaded by the processor 610 and the corresponding steps are executed, and to avoid duplication, the explanation is omitted here.

[0249] In another embodiment of the present invention, embodiments of the present invention further provide a computer-readable memory medium (Memory) which is a storage device within an electronic device 600 for storing programs and data. For example, a computer-readable memory medium 620. The computer-readable memory medium 620 hereof may comprise an internal storage medium within the electronic device 600, and of course may comprise an extended storage medium supported by the electronic device 600. The computer-readable memory medium provides a memory space for storing the operating system of the electronic device 600. Furthermore, the memory space further stores one or more computer instructions that are loaded and executed by a processor 610, and these computer instructions may comprise one or more computer programs 621 (including program code).

[0250] In another embodiment of the present invention, embodiments of the present invention further provide a computer program product or computer program including computer instructions, the computer instructions being stored in a computer-readable storage medium, for example, computer program 621. In this case, the electronic device 600 may be a computer, and a processor 610 reads the computer instructions from the computer-readable storage medium 620, and the processor 610 executes the computer instructions to cause the computer to perform the methods provided in the various optional manner described above.

[0251] In other words, when implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded into a computer and executed, the flow of the embodiment of the present invention is executed in whole or in part, or the functions of the embodiment of the present invention are realized. The computer may be a general-purpose computer, a dedicated computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions are transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, radio, or microwave) method.

[0252] Those skilled in the art will be able to implement the units and flow steps of each example described in the embodiments disclosed herein by means of electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application of the technical solution and the design constraints. While experts in the art may use different methods for each specific application to implement the described functions, such implementations should not be considered beyond the scope of the invention.

[0253] Finally, although the above is merely a specific embodiment of the present invention, the scope of protection of the present invention is not limited thereto. Modifications or substitutions that are easily conceivable to those skilled in the art within the technical scope of the disclosure of the present invention are included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be the same as the scope of protection of the claims.

Claims

1. A method for predicting reactant molecules, performed by an electronic device, A step of extracting features from a product molecule to obtain the characteristics of the product molecule, wherein the characteristics of the product molecule include the characteristics of atoms contained in the product molecule and the characteristics of chemical bonds between adjacent nodes of the atoms. A step of predicting a conversion pathway from the product molecule to multiple reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule, wherein the conversion pathway includes an editing sequence and a synthesizer completion sequence, and the reverse reaction prediction model is a machine learning model that has learned the relationship between the tasks of predicting synthesizers contained in the product molecule and completing those synthesizers. A step of obtaining a plurality of synthesizers corresponding to the product molecule by editing the target to be edited according to the edited state indicated by each editing operation in the editing sequence, wherein the target to be edited is an atom or chemical bond in the product molecule. A step of obtaining a plurality of reactant molecules corresponding to the plurality of synthesizers by adding a basic graph indicated by each synthesizer, according to the interface atoms indicated by each synthesizer, based on at least one synthesizer completion operation corresponding to each synthesizer in the synthesizer completion sequence, wherein the basic graph includes a plurality of atoms or atomic edges for connecting the plurality of atoms. A method for predicting reactant molecules containing the above.

2. The step of predicting the conversion pathway from the product molecule to multiple reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule is: A step of obtaining the input characteristics of the t-th action based on the (t-1)-th action obtained by the prediction by the inverse reaction prediction model, wherein t is an integer greater than 1. The process includes the steps of: predicting the t-th operation and obtaining the conversion path, based on the input features corresponding to the t-th operation and the hidden features corresponding to the t-th operation, that the operation predicted by the inverse reaction prediction model is a synthesizer complementation operation, and that the process has been traversed through all the attached atoms on the plurality of synthesizers and all the attached atoms on the basic graph added to the plurality of synthesizers; The hidden features of the t-th action are associated with the action predicted by the inverse response prediction model prior to the t-th action. The method for predicting reactant molecules according to claim 1.

3. The step of obtaining the input characteristics of the t-th action based on the (t-1)-th action obtained by the prediction using the inverse reaction prediction model is: The steps include obtaining the k first prediction results with the highest scores from the prediction results of the (t-1)th operation, according to a beam search method for hyperparameter k, The step of determining k first input features corresponding to the t-th operation based on the k first prediction results, The step of obtaining the aforementioned conversion path is: The steps include predicting the t-th operation based on each of the k first input features and the hidden features corresponding to the t-th operation, and obtaining the k second prediction results with the highest scores from the prediction results obtained by the prediction, according to a beam search method of hyperparameter k. The steps include determining k second input features corresponding to the (t+1)th operation based on the k second prediction results, Based on the k second input features and the hidden features corresponding to the operation of (t+1) in the k second input features, the operation predicted by the reverse reaction prediction model is a synthesizer complementation operation, and the step of predicting the operation of (t+1) until the traverse of all attached atoms on the plurality of synthesizers and all attached atoms on the basic graph added to the plurality of synthesizers is obtained, thereby obtaining the conversion path. The method for predicting reactant molecules according to claim 2.

4. The step of obtaining the aforementioned conversion path is: If the (t-1)th operation is the editing operation, the steps include determining the input features of the tth operation based on the features of the subgraph obtained by the editing performed by the (t-1)th operation, The process includes the steps of using the inverse response prediction model to predict the editing target and the edited state indicated by the t-th operation, and obtaining the editing sequence, based on the input characteristics of the t-th operation and the hidden characteristics of the t-th operation, until the operation predicted by the inverse response prediction model is the last editing operation in the editing sequence. The method for predicting reactant molecules according to claim 2.

5. The step of obtaining the aforementioned conversion path is: If the (t-1)th operation is the last editing operation or the synthesizer completion operation, the steps include determining the input characteristics of the tth operation based on the characteristics of the subgraph obtained by the editing by the (t-1)th operation and the characteristics of the attached atoms corresponding to the (t-1)th operation, Based on the input features of the t-th operation and the hidden features of the t-th operation, the operation predicted by the reverse reaction prediction model is a synthesizer completion operation, and the steps include predicting the fundamental graph and interface atoms indicated by the t-th operation until traversing all attached atoms on the plurality of synthesizers and all attached atoms on the fundamental graph added to the plurality of synthesizers, thereby obtaining the synthesizer completion sequence. The method for predicting reactant molecules according to claim 2.

6. The aforementioned editing operation is represented by an action tag for instructing editing, a tag for indicating the target of editing, and a tag for indicating the state after editing; the aforementioned synthesizer completion operation is represented by an action tag for instructing synthesizer completion, a tag for indicating the basic graph, and a tag for indicating the interface atom. The method for predicting reactant molecules according to claim 1.

7. If the object to be edited by the editing operation is an atom, the edited state indicated by the editing operation is a change in the number of charges on the atom or a change in the number of hydrogen atoms on the atom; if the object to be edited by the editing operation is a chemical bond, the edited state indicated by the editing operation is one of the following: addition of a chemical bond, deletion of a chemical bond, or change in the type of chemical bond. The method for predicting reactant molecules according to claim 1.

8. A method for training an inverse response prediction model performed by an electronic device, A step of extracting features from a product molecule to obtain the characteristics of the product molecule, wherein the characteristics of the product molecule include the characteristics of atoms contained in the product molecule and the characteristics of chemical bonds between adjacent nodes of the atoms. A step of predicting a conversion pathway from the product molecule to a plurality of reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule, and obtaining the conversion pathway, The transformation path includes an editing sequence and a synthesizer completion sequence, where each editing operation in the editing sequence is used to indicate the object to be edited and the edited state, where the object to be edited is an atom or chemical bond in the product molecule, and for a plurality of synthesizers of the product molecule obtained by the editing sequence, the synthesizer completion sequence includes at least one synthesizer completion operation corresponding to each of the plurality of synthesizers, where each synthesizer completion operation in the at least one synthesizer completion operation is used to indicate a base graph and interface atoms, where the base graph includes a plurality of atoms or atomic edges for connecting the plurality of atoms, and A method for training a reverse reaction prediction model, comprising the step of training the reverse reaction prediction model based on the loss of the conversion path and the training path, wherein the trained reverse reaction prediction model is a machine learning model that has learned the relationship between the tasks of predicting and complementing the synthetic atoms contained in the product molecule.

9. Before the step of obtaining the conversion path, the training method is The steps include obtaining candidate reactant molecules corresponding to the aforementioned product molecules, The steps include obtaining a basic graph dictionary by comparing the molecular structures of the product molecule and the candidate reactant molecule, The steps include obtaining the training path based on the aforementioned basic graph dictionary and A method for training an inverse response prediction model according to claim 8, further comprising:

10. The step of obtaining the training path based on the aforementioned basic graph dictionary is: A step of constructing a linked tree based on the aforementioned basic graph dictionary, wherein the linked tree includes a tree structure in which the plurality of compositers are root nodes and the basic graphs in the aforementioned basic graph dictionary are subnodes; The step includes determining the shortest path as the training path by traversing the aforementioned linked tree. A method for training an inverse response prediction model according to claim 9.

11. The step of obtaining a basic graph dictionary by comparing the molecular structures of the product molecule and the candidate reactant molecule is as follows: The steps include determining multiple molecular fragments other than multiple synths in the candidate reactant molecule as multiple candidate subgraphs, If, among the plurality of candidate subgraphs, the first candidate subgraph contains a first atom and a second atom, and the first atom and the second atom belong to different rings, the chemical bond between the first atom and the second atom is broken to obtain a plurality of first subgraphs; If a second candidate subgraph among the plurality of candidate subgraphs includes a linked third atom and a fourth atom, and one of the third atom and the fourth atom belongs to a ring, and the degree of the other atom of the third atom and the fourth atom is greater than or equal to a predetermined value, the chemical bond between the third atom and the fourth atom is broken to obtain a plurality of second subgraphs; The process includes the step of determining the candidate subgraphs excluding the first candidate subgraph and the second candidate subgraph from the plurality of candidate subgraphs, the plurality of first subgraphs and the plurality of second subgraphs as basic graphs in the basic graph dictionary. A method for training an inverse response prediction model according to claim 9.

12. In the step of training the inverse response prediction model based on the losses of the conversion path and the training path, The loss between the transformation path and the training path is, The determination is made based on the following information: the difference between the tags of the predicted actions in the conversion path and the tags of the training actions in the training path; the difference between the score of the edited target corresponding to the predicted action and the score of the edited target corresponding to the training action; the difference between the edited state corresponding to the predicted action and the edited state corresponding to the training action; the difference between the basic graph indicated by the predicted action and the basic graph indicated by the training action; and the difference between the interface atoms indicated by the predicted action and the interface atoms indicated by the training action. A method for training the inverse response prediction model according to claim 8.

13. A device for predicting reactant molecules, An extraction unit configured to perform feature extraction on a product molecule and obtain the characteristics of the product molecule, wherein the characteristics of the product molecule include the characteristics of atoms contained in the product molecule and the characteristics of chemical bonds between adjacent nodes of the atoms. A prediction unit configured to predict a conversion pathway from the product molecule to a plurality of reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule, wherein the conversion pathway includes an editing sequence and a synthesizer completion sequence, and the reverse reaction prediction model is a machine learning model that has learned the relationship between the tasks of predicting synthesizers contained in the product molecule and completing the synthesizers, and An editing unit configured to edit the target to be edited according to the edited state indicated by each editing operation in the editing sequence, in order to obtain a plurality of synthesizers corresponding to the product molecule, wherein the target to be edited is an atom or chemical bond in the product molecule; A predictor of reactant molecules comprising an additional unit configured to obtain a plurality of reactant molecules corresponding to the plurality of synthesizers by adding a basic graph indicated by each synthesizer complementation operation according to the interface atoms indicated by each synthesizer complementation operation in the at least one synthesizer complementation operation, based on at least one synthesizer complementation operation corresponding to each synthesizer in the synthesizer complementation sequence, wherein the basic graph includes a plurality of atoms or atomic edges for connecting the plurality of atoms.

14. A training device for an inverse response prediction model, An extraction unit configured to perform feature extraction on a product molecule and obtain the characteristics of the product molecule, wherein the characteristics of the product molecule include the characteristics of atoms contained in the product molecule and the characteristics of chemical bonds between adjacent nodes of the atoms. A prediction unit configured to predict the conversion pathway from the product molecule to a plurality of reactant molecules using a reverse reaction prediction model based on the characteristics of the product molecule, The transformation path includes an editing sequence and a synthesizer completion sequence, where each editing operation in the editing sequence is used to indicate the object to be edited and the edited state, where the object to be edited is an atom or chemical bond in the product molecule, and for a plurality of synthesizers of the product molecule obtained by the editing sequence, the synthesizer completion sequence includes at least one synthesizer completion operation corresponding to each of the plurality of synthesizers, where each synthesizer completion operation in the at least one synthesizer completion operation is used to indicate a base graph and interface atoms, where the base graph includes a prediction unit, which includes a plurality of atoms or atomic edges for connecting the plurality of atoms, A training apparatus for a reverse reaction prediction model, comprising a training unit configured to train the reverse reaction prediction model based on the loss between the conversion path and the training path, wherein the trained reverse reaction prediction model is a machine learning model that has learned the relationship between the tasks of predicting and complementing the synthetic atoms contained in the product molecule.

15. A processor that runs computer programs, An electronic device comprising a computer-readable storage medium that, when executed by the processor, causes the processor to execute a computer program which is a method for predicting reactant molecules according to any one of claims 1 to 7, or a method for training a reverse reaction prediction model according to any one of claims 8 to 12.

16. A computer program that causes a computer to execute the method for predicting reactant molecules according to any one of claims 1 to 7, or the method for training a reverse reaction prediction model according to any one of claims 8 to 12.

Images