Method and apparatus for training neural networks for determining molecular retrosynthetic pathways, computer device, and computer program

The neural network-based method for molecular retrosynthesis routes addresses time and efficiency issues by decomposing molecules into layers, using cost dictionaries to optimize pathways, enhancing search efficiency and accuracy.

JP7590064B2Active Publication Date: 2024-11-26TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
JP2023518017
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2020-11-04
Filing Date
2021-10-09
Publication Date
2024-11-26
Estimated Expiration
2041-10-09

AI Technical Summary

Technical Problem

Existing methods for designing molecular retrosynthesis routes using artificial intelligence are time-consuming, with small maximum search heights leading to incomplete constructions of complex molecules and large search heights causing exponentially increasing times, resulting in decreased efficiency and accuracy.

Method used

A method for training a neural network to determine molecular retrosynthesis pathways by decomposing molecules into multiple layers, using cost dictionaries to identify optimal decomposition routes, and employing a layered reinforcement learning approach to reduce computational overhead and enhance search efficiency.

Benefits of technology

The method significantly reduces computational time and increases accuracy in determining molecular retrosynthesis routes by dividing the search into multiple layers, ensuring efficient and precise synthesis pathway determination.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007590064000008
    Figure 0007590064000008
  • Figure 0007590064000009
    Figure 0007590064000009
  • Figure 0007590064000010
    Figure 0007590064000010
Patent Text Reader

Abstract

The present application discloses a method and apparatus for training neural networks to determine molecular retrosynthetic pathways. The neural network training method for determining molecular retrosynthetic pathways uses the concept of hierarchical learning to decompose the training process of molecular retrosynthetic pathways with large search depths into multiple layers for training, and instead of the complete retrosynthetic reaction process, uses multiple layers of molecular retrosynthetic pathways. After completing the training of one layer of molecular retrosynthetic pathways, a representative molecule is selected as the starting molecule for the molecular retrosynthetic pathway of the next layer using a molecular screening method, effectively improving the efficiency of molecular retrosynthetic pathway search and more effectively extracting accurate molecular cost value information. The hierarchical approach significantly reduces the large amount of computational overhead involved in determining molecular retrosynthetic pathways, ensuring the accuracy of the molecular retrosynthetic pathways while reducing the time required for determining molecular retrosynthetic pathways.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] This application claims priority to a Chinese patent application filed with the China Patent Office on November 4, 2020, bearing application number 2020112189912 and title "Method and apparatus for training neural networks for determining molecular retrosynthetic pathways," the entire contents of which are incorporated herein by reference.

[0002] The present application relates to the technical field of artificial intelligence, and in particular to a method, apparatus, device and readable storage medium for training a neural network for determining molecular retrosynthetic pathways. [Background technology]

[0003] In recent years, artificial intelligence technology has been developed rapidly and gradually introduced into various scientific fields, and has played an important role. In the field of chemistry, due to the infinite changes in chemical reactions under different conditions, in the past, when producing compound molecules, researchers only had to spend a large amount of time and energy to design a reasonable organic synthesis route. Assisting researchers in designing organic synthesis routes based on artificial intelligence technology can greatly improve the efficiency of researchers in researching and developing chemical molecules and other compounds.

[0004] Currently, methods for designing molecular retrosynthetic pathways based on artificial intelligence include the following: first, performing random search steps based on the Monte Carlo Tree Search (MCTS) algorithm until a solution is found or the maximum depth is reached, and introducing symbolic artificial intelligence to complete the design of a molecular retrosynthetic pathway; second, determining a template selection policy for each stage of the molecular retrosynthetic reaction based on deep reinforcement learning technology, and finally obtaining a molecular retrosynthetic pathway; third, based on a distributed training architecture, combining deep reinforcement learning technology to accelerate the construction of an optimal molecular retrosynthetic pathway and the network fitting of the cost function, and completing the design of a training set molecular retrosynthetic pathway by the network. Summary of the Invention [Problem to be solved by the invention]

[0005] However, the design of molecular retrosynthetic pathways using the above methods is time-consuming. In the above methods, the maximum search height must be determined at the beginning of constructing the molecular retrosynthetic tree. If the maximum search height is too small, it will be difficult for some complex molecules to complete the construction of the molecular retrosynthetic tree within a finite height. Conversely, if the maximum search height is too large, the time required will increase exponentially, and the efficiency and accuracy of molecular retrosynthetic pathway design will be reduced.

[0006] Embodiments of the present application provide methods for training neural networks for determining molecular retrosynthetic pathways, methods for determining molecular retrosynthetic pathways, devices, apparatus, and readable storage media. [Means for solving the problem]

[0007] According to one embodiment, there is provided a method for training a neural network for determining a molecular retrosynthetic pathway executed by a computing device, the method comprising: determining a first degradation pathway for each of a plurality of first molecules based on molecular representation information of the first molecules; obtaining a first cost dictionary based on a first degradation pathway of each of the first molecules, the first cost dictionary including molecular representation information and corresponding cost value information of each of the first molecules, the cost value information of the first molecules indicating a cost for separating the first molecules according to the corresponding first degradation pathway; determining molecular representation information of at least one second molecule based on a first degradation pathway of each of the first molecules, wherein each of the second molecules is a molecule that can be decomposed to obtain a molecule among molecules obtained by decomposing a corresponding first molecule based on the first degradation pathway; determining a plurality of third molecules from each of the second molecules, each of the third molecules representing a class of the second molecules; obtaining a second cost dictionary based on a second decomposition pathway of each of the third molecules, the second cost dictionary including molecular representation information and corresponding cost value information of each of the third molecules, the cost value information of the third molecules indicating a cost for separating the third molecules according to the corresponding second decomposition pathway; The method includes: training based on the first cost dictionary and the second cost dictionary to obtain a target neural network, the target neural network outputs corresponding cost value information of the target molecule based on the input molecular representation information of the target molecule; and synthesizing a retrosynthetic pathway of the target molecule using the corresponding cost value information of the target molecule.

[0008] According to another aspect, there is provided a computational device implemented method for determining a molecular retrosynthetic pathway, said method comprising: receiving a molecular representation of a target molecule, the molecular representation indicating a three-dimensional chemical structure of the target molecule; inputting the molecular representation of the target molecule into a neural network for determining a molecular retrosynthetic pathway; determining a target degradation pathway of the target molecule based on the neural network that determines the molecular retrosynthetic pathway, the target degradation pathway being a degradation pathway with the lowest cost value among at least one degradation pathway; and obtaining molecular retrosynthetic pathway information of the target molecule based on the target decomposition pathway.

[0009] According to another aspect, there is provided an apparatus for training a neural network for determining a molecular retrosynthetic pathway, the apparatus being provided on a computing device, the apparatus comprising: A first determination module that determines a first degradation pathway for each of a plurality of first molecules based on molecular expression information of the first molecules, wherein a path depth of the first degradation pathway is equal to or less than a target depth; a first cost dictionary generation module for obtaining a first cost dictionary based on a first decomposition pathway of each of the first molecules, the first cost dictionary including molecular representation information and corresponding cost value information of each of the first molecules, the cost value information of the first molecules indicating a cost for separating the first molecules according to the corresponding first decomposition pathway; A second determination module that determines molecular expression information of at least one second molecule based on a first degradation pathway of each of the first molecules, wherein each of the second molecules is a molecule that can be decomposed to obtain a molecule among molecules obtained by decomposing a corresponding first molecule based on the first degradation pathway; a third determination module for determining a plurality of third molecules from each of the second molecules, each of the third molecules representing a class of the second molecules; a second cost dictionary generation module for obtaining a second cost dictionary based on a second decomposition pathway of each of the third molecules, the second cost dictionary including molecular representation information and corresponding cost value information of each of the third molecules, the cost value information of the third molecules indicating a cost for separating the third molecules according to the corresponding second decomposition pathway; a training module for training based on the first cost dictionary and the second cost dictionary to obtain a target neural network, the target neural network outputs corresponding cost value information of the target molecule based on input molecular expression information of the target molecule, and the corresponding cost value information of the target molecule synthesizes a retrosynthetic pathway of the target molecule.

[0010] 1. A molecular retrosynthetic path determination apparatus implemented in a computing device, the apparatus comprising: a receiving module for receiving molecular representation information of a target molecule, the molecular representation information indicating a three-dimensional chemical structure of the target molecule; an input module for inputting the molecular representation of the target molecule into a neural network for determining a molecular retrosynthetic pathway; a decomposition path determination module that determines a target decomposition path of the target molecule based on the neural network that determines the molecular retrosynthesis path, the target decomposition path being a decomposition path with the lowest decomposition cost value among at least one decomposition path; and a route determination module for obtaining molecular retrosynthetic route information of the target molecule based on the target decomposition route.

[0011] According to another aspect, a computing device is provided, the computing device including one or more processors and a memory, the memory storing at least one computer-readable instruction, the at least one computer-readable instruction being loaded and executed by the one or more processors to realize operations performed by the methods in the embodiments of the present application.

[0012] According to another aspect, one or more computer-readable storage media are provided, the computer-readable storage media storing at least one computer-readable instruction, the at least one computer-readable instruction being read and executed by the one or more processors to realize operations performed by the methods in the embodiments of the present application.

[0013] According to another aspect, a computer program product is provided, the computer program product including computer readable instructions stored in a computer readable storage medium, the computer readable instructions being read by one or more processors of a computing device from the computer readable storage medium, and the one or more processors executing the computer readable instructions causing the computing device to perform the steps of the methods in any of the preferred implementations described above.

[0014] The details of one or more embodiments of the present application are presented in the drawings and description that follow. Based on the description, drawings and claims of the present application, other features, objects and advantages of the present application will become more apparent. [Brief description of the drawings]

[0015] In order to more clearly explain the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application, and those skilled in the art can obtain other drawings based on these drawings on the premise that they do not make efforts worthy of inventive step. [Figure 1] FIG. 1 is a schematic diagram of an implementation environment for a method for training a neural network for determining molecular retrosynthetic pathways provided by an embodiment of the present application. [Diagram 2] FIG. 2 is an architecture diagram of a molecular retrosynthesis tree constructed based on a hierarchical scheme, according to an embodiment of the present application. [Diagram 3] 1 is a flow chart of a method for training a neural network for determining molecular retrosynthetic pathways provided by an embodiment of the present application. [Figure 4] 1 is a flow chart of a method for training a neural network to determine another molecular retrosynthetic pathway provided by an embodiment of the present application. [Diagram 5] 1 is a flowchart of a method for obtaining a second cost dictionary provided by an embodiment of the present application; [Figure 6] FIG. 1 is an architecture diagram of a method for training a neural network to determine molecular retrosynthetic pathways provided by an embodiment of the present application. [Figure 7] 1 is a flow chart of a method for determining a molecular retrosynthetic pathway provided by an embodiment of the present application. [Figure 8] FIG. 1 is a block diagram of an apparatus for training a neural network for determining molecular retrosynthetic pathways provided by an embodiment of the present application. [Figure 9] FIG. 2 is a structural schematic diagram of a server provided by an embodiment of the present application; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] In order to make the objectives, technical solutions and advantages of the present application clearer, the embodiments of the present application will be described in more detail below in conjunction with the drawings. Hereinafter, exemplary embodiments will be described in detail, examples of which are shown in the drawings. When describing the drawings below, the same numerals in different drawings indicate the same or similar elements unless otherwise specified. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as described in the appended claims. Below, we will briefly introduce the techniques and terms that may be used in the embodiments of the present application.

[0017] Artificial Intelligence (AI) is the theory, method, technology, and operating system that simulates, extends, and expands human intelligence using digital computers or devices controlled by digital computers to sense the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, AI is a comprehensive technology of computer science that aims to create new intelligent machines that understand the essence of intelligence and can respond in a manner similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines so that machines can sense, reason, and make decisions.

[0018] Machine Learning (ML): is a multidisciplinary field that includes probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It is the study of how computers can mimic or realize human learning behavior, acquire new knowledge and skills, restructure existing knowledge structures, and continuously improve their own performance. Machine learning is the key to artificial intelligence and the fundamental means of making computers intelligent, and its applications span various fields of artificial intelligence. Machine learning and deep learning generally include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0019] Markov Decision Process (MDP) is a mathematical model of sequential decision, which is used to simulate random policies and rewards that an intelligent agent can realize in an environment with Markovian system states. MDP is constructed based on a set of interactive objects, i.e., an intelligent agent and an environment, and its elements include state, action, policy, and reward. In simulating MDP, an intelligent agent senses the current system state and performs actions on the environment according to the policy to change the state of the environment and obtain rewards, and the accumulation of rewards over time is called reward.

[0020] The molecular representation information is information that indicates the three-dimensional chemical structure of a molecule. For example, the molecular representation information is the SMILES notation (Simplified molecular input line entry specification, SMILES) of the molecule, that is, the chemical structure of the molecule is expressed by a character string. For example, the molecular representation information is a molecular graph that indicates the molecular structure, as shown in Table 1, which is the normal properties of nodes and edges in a molecular graph. [Table 1]

[0021] The Jaccard similarity coefficient is a coefficient for comparing the similarity and difference between finite sample sets. The larger the value of the Jaccard similarity coefficient, the more similar the samples are. The Tanimoto coefficient is obtained by extending the Jaccard coefficient and is also called the broad Jaccard similarity coefficient.

[0022] The Synthetic Accessibility Score (SA Score) is a method for quickly evaluating the difficulty of decomposing a large number of compounds based on the complexity of the molecule. This method uses the assumption that molecular structures that occur frequently are easy to synthesize, and calculates the decomposability of a molecule by weighting the frequency of Extended-Connectivity Fingerprints (ECFPs) with a diameter of 4 in 1 million compounds obtained from the Organic Small Molecule Bioactivity Database (PubChem). The frequency of occurrence and molecular complexity are used as evaluation indices. The synthetic accessibility score is normalized to a value of 1 (easy) to 10 (complex).

[0023] The implementation environment of the training method provided in the embodiments of the present application will be described below. FIG. 1 is a schematic diagram of the implementation environment of the training method for a neural network for determining a molecular retrosynthetic pathway provided in the embodiments of the present application. The implementation environment includes a terminal 101 and a server 102.

[0024] The terminal 101 and the server 102 may be directly or indirectly connected by wire or wireless communication, and the present application is not limited thereto. In one embodiment, the terminal 101 may be, but is not limited to, a smartphone, a tablet, a laptop, a desktop computer, etc. The terminal 101 provides the server 102 with basic information required for a method of determining a molecular retrosynthetic route, such as molecular expression information (including a molecular graph structure, a SMILES notation, etc.), molecular cost value reference information (e.g., a molecular synthetic accessibility score), and molecular retrosynthesis reference information (e.g., a molecular decomposability solution), etc.

[0025] The server 102 may be an independent physical server, a server cluster or a distributed system consisting of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, big data and artificial intelligence platforms. The server 102 executes the method for determining a molecular retrosynthetic pathway provided by the embodiments of the present application, and trains a neural network based on the basic information provided by the terminal 101. In one embodiment, the server 102 is equipped with a Linux operating system and GPU computing resources.

[0026] In one embodiment, during the process of training a neural network to determine a molecular retrosynthetic pathway, the server 102 is responsible for the main computing operations and the terminal 101 is responsible for the secondary computing operations, or the server 102 is responsible for the secondary computing operations and the terminal 101 is responsible for the main computing operations, or the server 102 or the terminal 101 may be responsible for the computing operations individually.

[0027] In one embodiment, the above training process adopts distributed training, for example, multiple computing nodes are used for training respectively. The server 102 includes a training server, which is a server cluster and includes multiple servers as computing nodes, and each computing node executes a part of the training task respectively, and transmits the neural network model through the training to the target server to provide the corresponding function to the user.

[0028] In one embodiment, the terminal 101 generally refers to one of multiple terminals, and this embodiment describes only the terminal 101 as an example. As can be understood by those skilled in the art, the number of the above terminals 101 may be greater. For example, the above terminals 101 may be tens or hundreds, or even more, in which case the implementation environment of the above method for determining a molecular retrosynthetic pathway further includes other terminals. The embodiment of the present application does not limit the number of terminals and the type of equipment.

[0029] In one embodiment, the wireless or wired network employs standard communication technologies and / or protocols. The network may generally be the Internet, but may be any network, including, but not limited to, any combination of Local Area Network (LAN), Metropolitan Area Network (MAN), Wide Area Network (WAN), mobile, wired or wireless networks, private networks or virtual private networks. In some embodiments, technologies and / or formats including HyperText Mark-up Language (HTML), Extensible Markup Language (XML), and the like, are used to represent data exchanged over the network. Also, all or some links may be encrypted using conventional encryption technologies, such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), Internet Protocol Security (IPsec), and the like. In some other embodiments, custom and / or proprietary data communication technologies may be used to replace or supplement the above data communication technologies.

[0030] In the embodiment of the present application, a method for determining a molecular retrosynthetic route is provided, and a hierarchical reinforcement learning method is adopted. In the field of chemistry, retrosynthetic analysis is an important method for solving molecular synthetic routes, and is also the simplest and most basic method for molecular synthetic route design. Its essence is the separation of a target molecule, and by analyzing the structure of the target molecule, it is decomposed into intermediates and raw materials that are simpler and easier to synthesize, and the design of a molecular synthetic route is completed. The intermediates are precursor compounds required for the synthesis of the target molecule, that is, organic compounds that are difficult to obtain on the market and require self-synthesis, and the raw materials are organic compounds that are easily available on the market and are easy to synthesize the target molecule. Searching for a molecular retrosynthetic route is a process of constructing a molecular retrosynthetic tree. In the related art, when training the construction of a molecular retrosynthetic tree, it is generally performed based on a preset maximum search height. If such a method is adopted, if the maximum search height is too small, it will be difficult for some complex molecules to complete the construction of a molecular retrosynthetic tree within a finite height, and conversely, if the maximum search height is too large, the time required for construction will increase exponentially. In the embodiment of the present application, a hierarchical construction method is adopted in the training process of constructing a molecular retrosynthesis tree, which greatly reduces the large amount of calculation and time overhead involved in the process of constructing a molecular retrosynthesis tree. For more detailed description, please refer to the following embodiment.

[0031] Hereinafter, the following embodiments will be described by taking the example of dividing the molecular retrosynthesis tree into two layers, but the embodiments of the present invention are not limited to the two-layered method, and the molecular retrosynthesis tree construction process may be divided into multiple layers, and the embodiments of the present application are not limited thereto.

[0032] For example, refer to Fig. 2, which is an architecture diagram of constructing a molecular retrosynthesis tree based on the hierarchical method provided by the embodiment of the present application. As shown in Fig. 2, the maximum depth of the molecular retrosynthesis tree is 10 as an example, and the entire construction of the molecular retrosynthesis tree is divided into two layers, upper and lower, and two small molecular retrosynthesis trees are used instead of the complete retrosynthesis reaction process. After completing the construction of the upper layer molecular retrosynthesis tree, a molecule with representativeness is selected as the starting molecule of the lower layer molecular retrosynthesis tree through a molecular clustering screening method, which effectively improves the search efficiency of the molecular retrosynthesis route, and more effectively extracts accurate molecular cost value information.

[0033] Hereinafter, the method for determining a molecular retrosynthetic pathway provided by the examples of the present application will be specifically described.

[0034] 3 is a flow chart of a method for determining a molecular retrosynthetic pathway provided in an embodiment of the present application, which is taken as an example for application on a server in an embodiment of the present application, as shown in FIG. 3. The method includes the following steps: 301: The server determines a first degradation pathway for each of a plurality of first molecules based on molecular representation information of the first molecules, and a pathway depth of the first degradation pathway is less than or equal to a target depth.

[0035] In the embodiment of the present application, the first molecule is an existing molecule. The molecular representation information represents a three-dimensional chemical structure of the molecule. For example, the molecular representation information is a SMILES notation of the molecule, that is, the chemical structure of the molecule is represented by a character string. Also, for example, the molecular representation information is a molecular graph representing the molecular structure, but the embodiment of the present application is not limited thereto. The first decomposition pathway refers to a pathway with the lowest cost for decomposing the first molecule until the target decomposition condition is satisfied after the decomposition.

[0036] 302: The server obtains a first cost dictionary based on a first decomposition pathway of each first molecule, the first cost dictionary including molecular representation information and corresponding cost value information of each first molecule, and the cost value information of the first molecule indicates a cost for separating the first molecule according to the corresponding first decomposition pathway.

[0037] In the embodiment of the present application, the molecular expression information and the corresponding cost value information in the first cost dictionary exist in a one-to-one manner.

[0038] 303: The server determines molecular representation information of at least one second molecule based on a first decomposition pathway of each said first molecule, each said second molecule being a molecule that can be decomposed to obtain a molecule.

[0039] In the embodiment of the present application, after each first molecule is decomposed according to the corresponding first decomposition pathway, a plurality of molecules are obtained, among which there are some molecules whose decomposition is not completed and which have a decomposable pathway, and these molecules are determined as second molecules, that is, each second molecule is a molecule that can be decomposed to obtain a molecule among the obtained molecules after the corresponding first molecule is decomposed according to the first decomposition pathway.

[0040] 304: The server determines a plurality of third molecules from each of the second molecules, each of the third molecules representing a class of the second molecules.

[0041] In an embodiment of the present application, the second molecules are divided into a number of sets based on the structural similarity between the molecules, the second molecules in each set have similar molecular structures, and one third molecule is determined from each set, where the third molecule is a representative molecule in the set to which it belongs.

[0042] 305: The server obtains a second cost dictionary based on a second decomposition pathway of each of the third molecules, the second cost dictionary including molecular representation information and corresponding cost value information of each of the third molecules, and the cost value information of the third molecules indicates a cost for separating the third molecules according to the corresponding second decomposition pathway.

[0043] In the embodiment of the present application, the second decomposition pathway refers to a pathway with the lowest cost for decomposing a third molecule until the target decomposition condition is met after the decomposition. The molecular representation information and the corresponding cost value information in the second cost dictionary exist in a one-to-one manner.

[0044] 306: The server performs training based on the first cost dictionary and the second cost dictionary to obtain a target neural network, and the target neural network outputs corresponding cost value information of the target molecule based on the input molecular expression information of the target molecule.

[0045] Here, the decomposition pathway corresponding to the corresponding cost value information of the target molecule is the decomposition pathway with the lowest decomposition cost value among all possible decomposition pathways of the target molecule. Therefore, a retrosynthetic pathway of the target molecule is synthesized based on the corresponding cost value information of the target molecule.

[0046] In one embodiment, the target molecule is separated according to the decomposition pathway corresponding to the cost value information, and molecular retrosynthetic pathway information is obtained according to the separation result.

[0047] In an embodiment of the present application, the server performs training based on a first cost dictionary to obtain a first neural network, performs training based on a second cost dictionary to obtain a second neural network, and finally combines the two neural networks to obtain a target neural network.

[0048] Here, in the embodiment of the present application, according to the above steps 301-306, the server divides the search for retrosynthetic pathways of multiple molecules into multiple layers, and then obtains cost dictionaries corresponding to each layer based on the obtained Lth layer and L+1th layer, where L is greater than or equal to 1, and a corresponding cost dictionary can be obtained for each layer divided in the search process. Correspondingly, when performing training, the server trains a neural network based on the first cost dictionary, the second cost dictionary and the cost dictionaries of the previous L-1 layers, and obtains the trained neural network to obtain a multi-layer molecular retrosynthetic pathway of the molecule. In the embodiment of the present application, a method for training a neural network for determining a molecular retrosynthetic pathway is provided. For a plurality of molecules, when determining the retrosynthetic pathway of each molecule, the concept of hierarchical learning is adopted, and the training process of the molecular retrosynthetic pathway with large search depth is decomposed into multiple layers for training, so as to increase the training speed, and instead of the complete retrosynthetic reaction process, multiple layers of molecular retrosynthetic pathways are used. After completing the training of one layer of molecular retrosynthetic pathway, a representative molecule is selected as the starting molecule of the molecular retrosynthetic pathway of the next layer in a molecular screening manner, so as to effectively improve the search efficiency of the molecular retrosynthetic pathway, and more effectively extract the cost value information of the accurate molecule. The hierarchical manner greatly reduces the large amount of calculation overhead caused by the determination of the molecular retrosynthetic pathway, ensures the accuracy of the molecular retrosynthetic pathway, and reduces the time required for determining the molecular retrosynthetic pathway.

[0049] 4 is a flowchart of another method for training a neural network for determining a molecular retrosynthetic pathway provided in an embodiment of the present application, in which, as shown in FIG. 4, the method is applied to a server in an embodiment of the present application, and takes as an example a method for dividing a molecular retrosynthetic pathway into two layers. The method includes the following steps:

[0050] 401: The server partitions a decomposition task of each first molecule into a plurality of first subtasks based on molecular expression information of each first molecule, and the decomposition task is to separate each first molecule according to a decomposition pathway.

[0051] In an embodiment of the present application, the decomposition task is to separate a molecule into at least one molecule that is simpler and easier to synthesize based on the molecular structure analysis of the molecule, according to a decomposition pathway based on the molecular representation information of the molecule.

[0052] In a preferred embodiment, the server is associated with a molecular database, which stores molecular representation information of existing molecules, and the server extracts molecular representation information of each first molecule from the molecular database, generates a decomposition task for each first molecule according to the molecular representation information of each first molecule, and then partitions the decomposition task for each first molecule into a plurality of first subtasks, each first subtask including at least one decomposition task for the first molecule.

[0053] On the premise that one first molecule corresponds to one decomposition task, when the server divides the decomposition tasks of multiple first molecules, the server includes any one of the following methods:

[0054] In a first preferred embodiment, the server is evenly partitioned according to the number of current first molecules and the number of current computing nodes, for example, the number of current computing nodes is divided by the number of current first molecules, and the obtained number determines the number of first subtasks in each computing node, and the number of decomposition tasks of the first molecules in each first subtask is similar.

[0055] In a second preferred embodiment, the server partitions the molecules based on the molecular representation information of each first molecule according to the complexity of the chemical three-dimensional structure of the molecule, for example, partitioning a decomposition task of a first molecule having a single benzene ring into one first subtask, partitioning a decomposition task of a first molecule having one or more benzene rings into another first subtask, etc., such that the complexity of the chemical three-dimensional structure of each first molecule corresponding to one first subtask is similar.

[0056] In a third preferred embodiment, the server partitions based on the number of current computing nodes and the computing capabilities of the computing nodes. For example, if the number of subtasks currently being executed by a computing node is greater than a threshold, the server partitions 10 first molecule decomposition tasks into one first subtask; if the number of subtasks currently being executed by a computing node is less than the threshold, the server partitions 100 first molecule decomposition tasks into another first subtask, and so on, with the number of first molecule decomposition tasks in each first subtask being different.

[0057] Here, the embodiments of the present application do not specifically limit the partitioning method of the decomposition task of multiple first molecules, and may adopt any one of the preferred methods provided above, or may combine any preferred methods to obtain a more complex partitioning method.

[0058] 402: The server assigns each first subtask to a plurality of computing nodes, and each computing node calculates a first initial cost value function of each first molecule and returns the first initial cost value function, where the first initial cost value function is obtained by the corresponding computing node calculating based on molecule cost value reference information.

[0059] In an embodiment of the present application, the server processes data based on multiple computing nodes at the same time, and the server assigns multiple first subtasks to multiple computing nodes, and each computing node is responsible for the decomposition task of at least one first molecule. Using multiple computing nodes to perform parallel calculations accelerates the training process and increases the training speed. The first initial cost value function is an initial cost value function of each first molecule obtained by the computing node using the molecular cost value reference information based on the molecular expression information of each first molecule. The molecular cost value reference information indicates the decomposability of the molecule. The cost value function determines a decomposition path with the minimum cost for decomposing the molecule based on the molecular expression information of the molecule. At the beginning of the decomposition of each first molecule, the cost value function of each first molecule is initialized by the molecular cost value reference information, and the cost value function randomly initialized at the beginning of training avoids the situation where the policy is prone to instability, and improves the stability of the policy for determining the molecular decomposition path based on the cost value function.

[0060] In a preferred implementation, the numerator cost value reference information is a composite accessibility score, and the server uses the composite accessibility score to initialize the numerator cost value function, and the a priori dominance of the composite accessibility score, which represents the numerator difficulty, enhances the convergence speed of the numerator cost value function to some extent.

[0061] In a preferred embodiment, when the server allocates each first sub-task to a plurality of computing nodes, the server includes one of the following methods:

[0062] In the first kind of preferred implementation, if the number of the first subtasks is similar to the number of the current computing nodes, the server allocates according to a one-to-one manner.

[0063] In a second preferred embodiment, when the number of the first subtasks is greater than the number of the current computing nodes, the server allocates them based on the current computing power of each computing node, for example, allocates more than one first subtask to a computing node with strong computing power, and allocates only one first subtask to a computing node with weak computing power.

[0064] In a third preferred embodiment, the server assigns the first subtask based on the number of decomposition tasks of the first molecule in the first subtask and the computing power of the computing node, for example, assigning a first subtask including 100 decomposition tasks of the first molecule to a computing node with strong computing power, and assigning a first subtask including 10 decomposition tasks of the first molecule to a computing node with weak computing power.

[0065] Here, the embodiments of the present application do not specifically limit the manner of allocating each first sub-task to multiple computing nodes, and may adopt any one of the preferred manners provided above, or may combine any preferred manners to obtain a more complex allocation manner.

[0066] 403: The server receives a first initial cost value function fed back from each computing node.

[0067] In an embodiment of the present application, based on the first initial cost value function obtained by each computing node calculating in step 402 above, a first initial cost value function corresponding to each first molecule is received, which is returned from each computing node.

[0068] Here, the above steps 401 to 403 are an embodiment of performing a molecular retrosynthetic pathway searching process in parallel based on the distributed training framework provided in the embodiments of the present application, and adopting a distributed computing method maintains the original computing effect while accelerating the computing process and increasing the training speed. In another preferred embodiment, the server independently performs the decomposition task of each first molecule, and does not need to perform it based on the distributed training framework, for example, the server obtains a first initial cost value function of each first molecule based on the molecular expression information and molecular cost value reference information of each first molecule, and the embodiments of the present application are not specifically limited thereto.

[0069] 404: When any one of the decomposition levels of each first molecule is completed, the server updates a first initial cost value function of each first molecule based on the decomposition cost value corresponding to the decomposition path of any one layer of each first molecule to obtain a first target cost value function of each first molecule, and the first target cost value function determines a decomposition path of the first molecule having the minimum decomposition cost value.

[0070] In an embodiment of the present application, when performing a decomposition task for each first molecule, a complete decomposition pathway of one first molecule consists of at least one decomposition level, and one decomposition level performs one-step decomposition for one first molecule, i.e., the decomposition pathway of one first molecule consists of at least one-step decomposition.

[0071] There are various decomposition methods for each decomposition stage. Each decomposition method corresponds to a decomposition cost value, and the decomposition cost value indicates that the first molecule is decomposed to a cost required for the current level according to the current decomposition method. When any one of the decomposition levels of each first molecule is completed, the server obtains at least one decomposition cost value according to at least one decomposition method existing in the any one of the decomposition levels. According to the at least one decomposition cost value, the server updates a first initial value function of each first molecule, and obtains a first target cost value function of each first molecule, and the first target cost value function determines a decomposition path of the first molecule having a minimum decomposition cost value.

[0072] For example, when decomposing a first molecule, the server determines a decomposition cost value a for one first molecule at a first decomposition level A, i.e., when performing a first stage of decomposition on the first molecule. 1 , a 2 , a 3 A corresponding to 1 , A 2 , A 3 There are three decomposition methods, and the first initial cost value function of the first molecule is updated based on the three decomposition cost values, and the decomposition method A is determined based on the updated cost value function. 1 When decomposed by, the corresponding decomposition cost value a 1 is the smallest, and therefore the server selects decomposition method A at the first decomposition level A. 1 Similarly, a second decomposition level B is obtained based on the decomposition cost value b 1 , b 2 , b 3 B 1 , B 2 , B 3 There are three decomposition methods, and the decomposition method with the smallest decomposition cost value is determined in the decomposition level B by continuously updating the updated cost value function in the decomposition level A according to the three decomposition cost values. The server performs the above method to the next decomposition level in order, and when any one of the decomposition levels is completed, iteratively updates the first initial cost value function to finally obtain the first target cost value function.

[0073] In a preferred embodiment, the constraint condition for determining the decomposition path with the minimum decomposition cost value is called a policy, and when performing a decomposition task for each first molecule, the decomposition cost value required for the decomposition path obtained after the first molecule completes at least one decomposition level according to the policy is obtained by calculating according to formula (1), in which case, different decomposition paths, i.e., different decomposition cost values, are obtained according to different decomposition methods. When any one decomposition level of the first molecule completes decomposition, the decomposition cost value of at least one decomposition path can be obtained, and the first initial cost value function of the first molecule is updated according to the at least one decomposition cost value calculated, to obtain the first target cost value function v * (m). In one embodiment, after obtaining the first target cost value function, iteratively update the policy according to formula (3), and π(r|m) is the policy for decomposing the molecule, that is, the decomposition path of the first molecule is generated according to the policy, and the policy is iteratively updated to obtain an updated policy π'(r|m), and the cost required when decomposing the first molecule according to the updated policy π'(r|m) is minimum. Formulas (1) to (3) are as follows:

number

number

number

[0074] In a preferred embodiment, when the server executes the decomposition task for each first molecule, there are K decomposition levels, K is 1 or more, and each time a decomposition level is completed, the corresponding decomposition cost value is 1, and if the molecule obtained by decomposing according to at least one decomposition method present in any one of the decomposition levels is a molecule present in the molecule database, the decomposition cost value of the current decomposition method is 0, and if the obtained molecule is a molecule that does not exist, for example, if the molecule does not exist in the molecule database, the decomposition cost value of the current decomposition method is 100. The server calculates according to the above formula (1) to obtain the decomposition cost value obtained by each first molecule based on any one of the decomposition levels. Finally, the server updates the first initial cost value function of the first molecule based on the decomposition cost value according to any one of the decomposition levels to obtain a first target cost value function.

[0075] In a preferred embodiment, the server performs a decomposition task for each first molecule based on the molecular retrosynthesis reference information, which provides at least one decomposition method according to any one decomposition level of each first molecule. For example, the molecular retrosynthesis reference information is a computer-assisted retrosynthesis method based on molecular similarity, which provides a suitable decomposition solution for the molecule. Also, for example, the molecular retrosynthesis reference information is a template neural network, which provides at least one reaction template, i.e., a decomposition method, for the decomposition of the molecule, and the reaction template describes the process of a type of chemical reaction including breaking old chemical bonds and creating new chemical bonds.

[0076] In a preferred embodiment, in the decomposition process of each first molecule, the decomposition process of each first molecule is modeled into an unfolding game, and π(r|m) is a policy for decomposing a molecule, where m is a product and r is a selectable reaction, so that a cost function of the current molecule can be obtained, and in the unfolding process, a first target cost value function v *By continuously updating (m), we determine the disassembly path with the smallest disassembly cost, and for each round of deployment, we iteratively update the policy.

[0077] 405: When the decomposition task of any one of the first molecules satisfies the target decomposition condition, the server determines a first decomposition path for each of the first molecules based on a first target cost value function for each of the first molecules.

[0078] In the embodiment of the present application, the target decomposition condition refers to the depth of the decomposition path obtained by each first molecule based on at least one decomposition level being smaller than the target depth, and the decomposition method does not exist in the obtained molecule, or the depth of the decomposition path obtained by each first molecule based on at least one decomposition level being equal to the target depth. The target depth is a depth threshold preset by the server. For example, the depth threshold is set to 5.

[0079] The following is an example of a target depth of 5. After a first molecule has gone through P decomposition levels (P is greater than or equal to 1 and less than 5), if the depth of the obtained decomposition path is less than 5 and there are no further decomposition methods for the obtained molecule, i.e., all of the obtained molecules are molecules that can be obtained from the molecular database, then the decomposition task of the first molecule satisfies the target decomposition condition, and the server obtains a first target cost value function for the first molecule based on the current decomposition level, and determines a decomposition path having the smallest decomposition cost value based on the function, and the decomposition path is the first decomposition path of the first molecule.

[0080] After a first molecule goes through Q decomposition levels (in this case, Q is equal to 5), if the depth of the obtained decomposition path is 5, the decomposition task of the first molecule satisfies the target decomposition condition, and the server obtains a first target cost value function of the first molecule based on the current decomposition level, and determines a decomposition path having the smallest decomposition cost value based on the function, where the decomposition path is the first decomposition path of the first molecule.

[0081] 406: The server determines corresponding cost value information for each first molecule based on the first degradation pathway of each first molecule.

[0082] In an embodiment of the present application, the corresponding cost value information of each first molecule indicates the cost of separating the first molecule according to the corresponding first decomposition pathway. In a preferred embodiment, the server calculates a first target cost value function for each first molecule, i.e., v * (m) is used to determine corresponding cost value information of each first molecule. In another preferred embodiment, when decomposing each first molecule, each time any one decomposition level is completed, a decomposition path having a minimum decomposition cost value based on the current decomposition level, i.e., corresponding cost value information of each first molecule based on the current decomposition level, can be obtained.

[0083] 407: The server obtains a first cost dictionary based on the molecular representation information and corresponding cost value information of each first molecule, where the first cost dictionary includes the molecular representation information and corresponding cost value information of each first molecule.

[0084] In the embodiment of the present application, the molecular representation information and the corresponding cost value information in the first cost dictionary are in a one-to-one manner. For example, in the first cost dictionary, the molecular representation information of each molecule and its corresponding cost value information are represented in the form of key-value pairs, that is, there are N key-value pairs in the first cost dictionary, where N is 1 or more, and {Smile 1 :cost 1};{Smile 2 :cost 2}……{Smile N :cost N} respectively indicate numerator 1 to numerator N and their corresponding cost value information.

[0085] Through the above steps 401 to 407, the server completes the search for upper-layer retrosynthetic routes of a plurality of first molecules, and obtains a first cost dictionary corresponding to the upper-layer retrosynthetic routes.

[0086] 408: The server determines molecular representation information of at least one second molecule based on the first decomposition pathway of each first molecule, each second molecule being a molecule decomposable to obtain the molecule.

[0087] In the embodiments of the present application, after each first molecule is decomposed based on the corresponding first decomposition pathway, a plurality of molecules can be obtained, among which there are some molecules whose decomposition is not completed and which have a decomposable pathway, and these molecules are determined to be second molecules.

[0088] In a preferred embodiment, when the server performs a decomposition task based on a first molecule, the first molecule is decomposed based on its corresponding first decomposition pathway, after which at least one molecule is obtained, and then molecular representation information of the at least one acquired molecule is determined, based on which the server determines whether the acquired molecule can be further decomposed, and if there is a molecule that can be further decomposed, determines whether a decomposition method exists for the molecule, and if there is a decomposition method, determines the molecule to be a second molecule.

[0089] Here, in an embodiment of the present application, the server determines cost value information of each first molecule based on the first decomposition pathway of each first molecule in a front-to-back order according to the above steps 406 to 408, obtains a first cost dictionary, and then determines at least one second molecule. In another preferred embodiment, the server first determines at least one second molecule based on the first decomposition pathway of each first molecule according to step 408, and then performs steps 406 and 407 to obtain the first cost dictionary. The embodiment of the present application is not specifically limited thereto.

[0090] 409: The server clusters each second molecule to obtain a plurality of sets, each set including at least one second molecule having a similar molecular structure.

[0091] In an embodiment of the present application, clustering is a process of dividing each second molecule into a plurality of sets based on the structural similarity between each molecule. The second molecules in each set have similar molecular structures. In one embodiment, the server may classify at least one second molecule based on other classification methods, such as a naive Bayes classification algorithm. The embodiment of the present application is not specifically limited thereto.

[0092] In a preferred embodiment, the server clusters each second molecule based on the Taylor Butina (TB) algorithm and determines the structural similarity between two second molecules based on the Tanimoto coefficient, as shown in Equation (4).

number

[0093] If the Tanimoto coefficient between two molecules is smaller than a predetermined threshold, the two molecules are considered to belong to the same class of molecules, and the two molecules are classified into the same set. For example, the predetermined threshold is 0.4, and the embodiment of the present application does not specifically limit the setting of the predetermined threshold, and it can be adjusted according to the actual situation.

[0094] Here, in the embodiment of the present application, the TB algorithm is adopted to cluster at least one second molecule, and in another preferred embodiment, the server adopts other clustering algorithms, such as k-means clustering algorithm, mean shift clustering algorithm, etc. to cluster the second molecule, and the embodiment of the present application is not specifically limited thereto.

[0095] 410: The server determines cluster centers for each set to a plurality of third molecules, where the third molecules are representative molecules in the set to which they belong.

[0096] In an embodiment of the present application, the cluster center is a center in a plurality of sets generated after clustering each second molecule, and the center is a representative second molecule in the set to which it belongs, and the representative second molecule is determined to be a third molecule.

[0097] After searching for upper-layer retrosynthetic pathways for each first molecule through the above steps 408 to 410, the server screens a plurality of molecules for which the obtained decomposition is incomplete and decomposition methods exist to obtain a plurality of representative third molecules, and these third molecules are used as starting molecules for lower-layer retrosynthetic pathways, thereby avoiding the situation in which the decomposition must be continued until the maximum depth is reached or the entire pathway is searched for, which takes a long time, when searching for the retrosynthetic pathway of each molecule. In the hierarchical process, a representative molecule is selected as the third molecule, which greatly reduces the amount of calculation in the subsequent training process, shortens the time required to determine the molecular retrosynthetic pathway, and further increases the training speed.

[0098] 411: The server obtains a second cost dictionary based on a second decomposition pathway of each third molecule, the second cost dictionary including molecular representation information and corresponding cost value information of each third molecule, and the cost value information of the third molecule indicates a cost for separating the third molecule according to the corresponding second decomposition pathway.

[0099] In the embodiment of the present application, the process of obtaining the second cost dictionary based on the second decomposition pathway of each third molecule can be referred to FIG. 5. FIG. 5 is a flowchart of the method for obtaining the second cost dictionary provided in the embodiment of the present application, which is specifically divided into the following steps 501 to 507. Here, steps 501 to 507 are similar to the execution process of steps 401 to 407 above, so steps 501 to 507 are briefly described below, and specific preferred implementation modes will not be described in detail.

[0100] 501: The server partitions a decomposition task of each third molecule into a plurality of second subtasks based on the molecular expression information of each third molecule, and the decomposition task is to separate each third molecule according to a decomposition pathway.

[0101] In an embodiment of the present application, the third molecules are molecules obtained by performing the above step 410. The server partitions the decomposition task of the multiple third molecules into multiple subtasks according to the molecular expression information of each third molecule.

[0102] 502: The server assigns each second subtask to a plurality of computing nodes, and each computing node calculates a second initial cost value function of each third molecule and returns the second initial cost value function, which is obtained by the corresponding computing node calculating based on molecule cost value reference information.

[0103] In an embodiment of the present application, the server can process data based on multiple computing nodes at the same time, and the server assigns multiple second subtasks to multiple computing nodes, and one second computing node is responsible for the decomposition task of at least one third molecule. The second initial cost value function is an initial cost value function of each third molecule obtained by the computing node calculating using the molecular cost value reference information based on the molecular expression information of each third molecule. The cost value function determines a decomposition path with the minimum cost for decomposing the molecule based on the molecular expression information of the molecule. At the beginning of the decomposition of each third molecule, the cost value function of each third molecule is initialized using the molecular cost value reference information, thereby avoiding the situation in which the cost value function due to random initialization at the beginning of training is prone to policy instability, and improving the stability of the policy for determining the molecular decomposition path based on the cost value function.

[0104] 503: The server receives the second initial cost value function fed back from each computing node.

[0105] In an embodiment of the present application, in the above step 502, based on the second initial cost value function obtained by calculation by each computing node, the server can receive a second initial cost value function corresponding to each third molecule returned from each computing node.

[0106] 504: When any one of the decomposition levels of each of the third molecules is completed, based on the decomposition cost value corresponding to the decomposition path of any one layer of each of the third molecules, the server updates a second initial cost value function of each of the third molecules to obtain a second target cost value function of each of the third molecules, and the second target cost value function determines a decomposition path of the third molecule having the minimum decomposition cost value.

[0107] In an embodiment of the present application, when performing a decomposition task for each third molecule, the complete decomposition pathway of one third molecule consists of at least one decomposition level, and one decomposition level performs one stage of decomposition for one third molecule, i.e., the decomposition pathway of one third molecule consists of at least one stage of decomposition.

[0108] There are various decomposition methods for each stage of decomposition. Each decomposition method corresponds to a decomposition cost value, and the decomposition cost value indicates the cost of decomposing the third molecule to the current level according to the current decomposition method. When any one of the decomposition levels of each third molecule is completed, the server obtains at least one decomposition cost value according to at least one decomposition method existing in the any one of the decomposition levels. According to the at least one decomposition cost value, the server updates a second initial value function of each third molecule, and obtains a second target cost value function of each third molecule, and the second target cost value function determines a decomposition path of the third molecule having a minimum decomposition cost value.

[0109] 505: When the decomposition task of any one of the third molecules satisfies the target decomposition condition, the server determines a second decomposition path for each of the third molecules based on a second target cost value function for each of the third molecules.

[0110] In the embodiment of the present application, the target decomposition condition is that the depth of the decomposition path obtained by each third molecule based on at least one decomposition level is smaller than the target depth, and no decomposition method exists for the obtained molecule, or the depth of the decomposition path obtained by each third molecule based on at least one decomposition level is equal to the target depth. The target depth is a depth threshold preset by the server. For example, the depth threshold is set to 5. The second decomposition path refers to a path with the smallest cost for decomposing the third molecule after decomposition until the target decomposition condition is met.

[0111] 506: The server determines corresponding cost value information for each third molecule based on the second degradation pathway of each third molecule.

[0112] In an embodiment of the present application, the corresponding cost value information for each third molecule indicates the cost of separating that third molecule according to the corresponding second decomposition pathway.

[0113] 507: The server obtains a second cost dictionary according to the molecular representation information and corresponding cost value information of each third molecule, where the second cost dictionary includes the molecular representation information and corresponding cost value information of each third molecule.

[0114] In the embodiment of the present application, the molecular representation information and the corresponding cost value information in the second cost dictionary are in a one-to-one manner. For example, in the second cost dictionary, the molecular representation information of each molecule and its corresponding cost value information are represented in the form of key-value pairs, that is, there are M key-value pairs in the second cost dictionary, where M is 1 or more, and {Smile 1 :cost 1}, {Smile 2 :cost 2}……{Smile M :cost M} respectively indicate molecule 1 to molecule M and their corresponding cost value information.

[0115] The server completes the search of the lower-layer retrosynthetic route through the above steps 501 to 507, and obtains a second cost dictionary corresponding to the lower-layer retrosynthetic route. Based on this hierarchical method, the search of the entire retrosynthetic route of a plurality of first molecules is hierarchicalized, the search of the upper-layer retrosynthetic route is completed, and after screening for representative molecules, the search of the lower-layer retrosynthetic route is completed, thereby greatly reducing the search time for molecular retrosynthetic route.

[0116] 412: The server trains a second neural network based on the molecular representation information and corresponding cost value information of each molecule in the second cost dictionary.

[0117] In an embodiment of the present application, the neural network includes a first neural network and a second neural network, the first neural network being obtained by training based on information in a first cost dictionary, and the second neural network being obtained by training based on information in a second cost dictionary. The above step 412 includes the following steps: the server inputs molecular representation information of a molecule in the second cost dictionary into the second neural network, and obtains corresponding predicted cost value information of the molecule by performing calculation according to network parameters of the second neural network; the server determines a model loss of the second neural network according to the corresponding predicted cost value information of the molecule and the corresponding cost value information of the molecule in the second cost dictionary; the server adjusts network parameters in the second neural network according to the model loss of the second neural network, and again inputs molecular representation information of a new molecule according to the adjusted second neural network, and iteratively adjusts network parameters in the second neural network according to the predicted cost value information of each input and the corresponding cost value information of the input molecule until the model loss of the second neural network meets the target condition, and then the server determines the current second neural network as the second neural network after training.

[0118] Here, in the embodiment of the present application, the training process of the second neural network may further include other steps, which are not mentioned here.

[0119] 413: The server updates the first cost dictionary based on the second cost dictionary to obtain an updated first cost dictionary.

[0120] In an embodiment of the present application, each third molecule in the second cost dictionary is determined by clustering a plurality of molecules obtained by decomposing the first molecule in the first cost dictionary. The second cost dictionary includes molecular expression information and corresponding cost value information of each third molecule, and the server updates the cost value information of each first molecule according to the molecular expression information and corresponding cost value information of each third molecule to obtain the updated first cost dictionary. The server improves the accuracy of determining the molecular retrosynthesis cost by updating the cost value information of each first molecule from bottom to top.

[0121] In a preferred embodiment, the specific process of the server updating the first cost dictionary according to the second cost dictionary is as follows: based on the molecular expression information of each third molecule, determine the molecular expression information of each first molecule corresponding to each third molecule, and then, based on the cost value information of each third molecule in the second cost dictionary and the cost value information of each corresponding first molecule in the first cost dictionary, add them correspondingly to obtain the cost value information of each first molecule after updating, that is, the first cost dictionary after updating. For example, the minimum decomposition cost value obtained by the server according to the cost value information of a certain third molecule for decomposing the third molecule is 10, and the minimum decomposition cost value of the corresponding first molecule is 20, so that the minimum decomposition cost value of the first molecule after updating is 30, and the embodiment of the present application does not specifically limit the updating method of the first cost dictionary.

[0122] 414: The server trains a first neural network based on the molecular representation information and corresponding cost value information of each molecule in the updated first cost dictionary.

[0123] In an embodiment of the present application, step 414 includes the following steps: the server inputs molecular representation information of a molecule in the first cost dictionary into the first neural network, and calculates according to network parameters of the first neural network to obtain corresponding predicted cost value information of the molecule; the server determines a model loss of the first neural network according to the corresponding predicted cost value information of the molecule and the corresponding cost value information of the molecule in the first cost dictionary; the server adjusts network parameters in the first neural network according to the model loss of the first neural network, and again inputs molecular representation information of a new molecule according to the adjusted first neural network, and continues to iteratively adjust the network parameters in the first neural network according to the predicted cost value information of each input and the corresponding cost value information of the input molecule until the model loss of the first neural network meets a target condition, and then the server determines the current first neural network to be the first neural network after training.

[0124] Here, in the embodiment of the present application, the training process of the first neural network may further include other steps, which are not mentioned here.

[0125] 415: The server combines the trained second neural network with the first neural network to obtain a target neural network.

[0126] In the embodiment of the present application, the target neural network outputs the corresponding cost value information of the target molecule according to the input molecular expression information of the target molecule. The server combines the second neural network after training with the first neural network to obtain the target neural network. For example, the server obtains the target neural network according to a serial manner, and for a target molecule, inputs the molecular expression information of the target molecule into the first neural network, and outputs the cost value information of the target molecule according to the upper layer retrosynthesis route, and then the server determines the molecule that the target molecule has obtained through the upper layer retrosynthesis route and has further decomposition methods, inputs the molecule that has further decomposition methods into the second neural network, and outputs the cost value information of the target molecule according to the lower layer retrosynthesis route, and finally obtains the complete retrosynthesis route of the target molecule. The embodiment of the present application does not specifically limit the acquisition manner of the target neural network.

[0127] Through the above steps 412 to 415, the server obtains a target neural network by training based on the second cost dictionary and the updated first cost dictionary, and the generalization ability of the target neural network of the present invention to molecules that have not participated in the training process can be maximized, and a decomposition pathway for any molecule can be obtained based on the target neural network.

[0128] Here, in the embodiment of the present application, the server trains the second neural network in a front-to-back order according to the above steps 412-415, then updates the first cost dictionary, and trains the first neural network. In another preferred embodiment, the server first performs step 413 to update the first cost dictionary, and then performs steps 412, 414, and 415 to obtain the target neural network. The embodiment of the present application is not specifically limited thereto.

[0129] Here, in the embodiment of the present application, the above steps 412 to 415 are one embodiment of the process of obtaining a target neural network by training provided by the embodiment of the present application. In one embodiment, the server may obtain the target neural network through other training methods, which are not specifically limited in the embodiment of the present application.

[0130] Also, referring to Table 2, Table 2 is a comparison of the experimental effects between the related method and the examples of the present application, and the experiment was conducted based on a standard data set to obtain the time required to successfully decompose the same number of molecules. As can be seen from the experimental results, compared with the related method, if the method provided in the examples of the present application is adopted, when the same number of molecules is decomposed, the time required is obviously shorter than some other methods. [Table 2]

[0131] Refer to Table 3, which is a comparison of the experimental effect between the related method and the embodiment of the present application, and the experiment is based on a standard data set to obtain the decomposition results of decomposing the same number of molecules. As can be seen from the experimental results, compared with the related method, if the method provided by the embodiment of the present application is adopted, when decomposing the same number of molecules, the number of decomposition successes is the largest, because the number of failures due to inability to decompose is the smallest, and the number of failures due to the number of decomposition layers being too large is also the smallest. [Table 3]

[0132] Of course, the above different embodiments may be combined with each other to form different implementation solutions, to which the embodiments of the present invention are not limited.

[0133] In the embodiment of the present application, a method for training a neural network for determining a molecular retrosynthetic pathway is provided. For a plurality of molecules, when determining the retrosynthetic pathway of each molecule, the concept of hierarchical learning is adopted to decompose the training process of the molecular retrosynthetic pathway with large search depth into multiple layers for training, thereby increasing the training speed, and instead of the complete retrosynthetic reaction process, multiple layers of molecular retrosynthetic pathways are used. After completing the training of one layer of molecular retrosynthetic pathway, a representative molecule is selected as the starting molecule of the molecular retrosynthetic pathway of the next layer in a molecular screening manner, effectively improving the search efficiency of molecular retrosynthetic pathways and more effectively extracting accurate molecular cost value information. The hierarchical manner greatly reduces the large amount of computation overhead caused by the determination of molecular retrosynthetic pathways, ensures the accuracy of molecular retrosynthetic pathways, and reduces the time required for determining molecular retrosynthetic pathways.

[0134] In the following, in combination with a practical situation, an example will be described of the neural network training method for determining a molecular retrosynthetic pathway provided in the embodiment of the present application. As shown in FIG. 6, FIG. 6 is a training architecture diagram of a neural network for determining a molecular retrosynthetic pathway provided in the embodiment of the present application, which is divided into two layers.

[0135] First, for the molecules in the molecular training set, a distributed training framework is adopted to explore upper-layer retrosynthetic pathways in parallel, and all molecular decomposition tasks are assigned to N computational nodes, where N is greater than or equal to 1. Each computational node executes the molecular decomposition task respectively, and at the beginning of the decomposition, each molecule adopts the synthetic accessibility score of the molecule to initialize its cost value function, and the a priori dominance representing the difficulty of molecular decomposition is used to increase the convergence speed of the molecular cost value function. Then, each molecule is expanded based on the policy π(r|m), and finally the cost of each molecule in the corresponding policy is obtained. When the calculation of each computational node is completed, the molecular representation information and its cost value information of all molecules are compiled to obtain the upper-layer cost dictionary.

[0136] After completing the upper-layer retrosynthetic route search, the molecules that have not been decomposed and have decomposition methods are collected from among the molecules obtained by decomposition, and the molecules are clustered using the TB algorithm, and the representative molecules are screened as the initial training set for the lower-layer molecule retrosynthetic route search, i.e., the lower-layer training data. A molecular expansion process similar to that of the upper layer is adopted to expand the molecules in the lower-layer molecule training set, and a lower-layer cost dictionary is obtained.

[0137] In addition, the upper-layer cost dictionary is updated based on the lower-layer cost dictionary, the corresponding neural networks are trained based on the two cost dictionaries, and the two neural networks are supervised trained using deep learning technology to obtain a target neural network that can perform retrosynthetic pathway search for a new target molecule. The process of supervised training of the neural network is as follows: input the molecular expression information of the molecule, perform batch normalization on the data through the first fully connected layer (Dense), then perform batch normalization on the data through the second fully connected layer, and repeat the process of batch normalization on the data through the second fully connected layer five times, and finally perform 500-500e through the third fully connected layer. ―|x| where x is a variable.

[0138] 7 is a flow chart of a method for determining a molecular retrosynthetic pathway provided in an embodiment of the present application, which is illustrated in FIG 7 and is applied to a server in an embodiment of the present application. The method includes the following steps:

[0139] 701: A server receives molecular representation information of a target molecule, the molecular representation information indicating a three-dimensional chemical structure of the target molecule.

[0140] In an embodiment of the present application, the target molecule is a molecule that cannot be obtained from a molecular database, and the server receives molecular representation information of the target molecule, for example, the server receives a SMILES notation of the target molecule, i.e., a string of characters, representing the three-dimensional chemical structure of the molecule.

[0141] 702: The server inputs the molecular representation information of the target molecule into a neural network that determines a molecular retrosynthetic pathway.

[0142] In an embodiment of the present application, the server receives the molecular representation information of the target molecule, and then inputs it into a neural network that determines the molecular retrosynthetic pathway provided by an embodiment of the present invention.

[0143] 703: Based on the neural network that determines the molecular retrosynthetic pathway, the server determines a target degradation pathway for the target molecule, where the target degradation pathway is a degradation pathway with the lowest degradation cost value among at least one degradation pathway.

[0144] In the embodiment of the present application, based on the neural network for determining a molecular retrosynthetic pathway provided in the embodiment of the present application, corresponding cost value information of a target molecule is output, and the decomposition pathway corresponding to the cost value information is the decomposition pathway with the lowest decomposition cost value among all possible decomposition pathways of the target molecule.

[0145] In a preferred embodiment, step 703 above comprises: The neural network for determining the molecular retrosynthetic pathway includes a first neural network and a second neural network, and includes the steps of: outputting cost value information of the upper layer retrosynthetic pathway of the target molecule based on the first neural network; determining a first decomposition pathway of the target molecule, where the pathway depth of the first decomposition pathway is equal to or less than the target depth; determining a molecule for which a further decomposition method exists from molecules obtained by decomposing the target molecule based on the first decomposition pathway; inputting molecular representation information of the molecule for which a further decomposition method exists into the second neural network; outputting cost value information of the lower layer retrosynthetic pathway of the target molecule based on the second neural network; determining a second decomposition pathway of the target molecule; and determining a target decomposition pathway of the target molecule based on the first decomposition pathway and the second decomposition pathway.

[0146] 704: The server obtains molecular retrosynthetic pathway information of the target molecule based on the target decomposition pathway.

[0147] In the embodiment of the present application, based on the target decomposition pathway in step 703, the decomposition reaction of the target molecule at each stage of the decomposition pathway is obtained, and finally the complete retrosynthetic pathway of the target molecule is determined.

[0148] In the embodiments of the present application, a method for determining a molecular retrosynthetic pathway is provided, in which a molecular retrosynthetic pathway determining neural network is used to obtain a retrosynthetic pathway of a target molecule, which requires a short time and has a high accuracy rate.

[0149] It should be noted that although the steps of the flowcharts of the above embodiments are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly described herein, the execution of these steps is not strictly limited to the order, and they may be executed in other orders. Furthermore, at least some of the steps of the flowcharts of the above embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily have to be executed simultaneously, and may be executed at different times and orders. In addition, the order of execution is not necessarily executed consecutively, and may be executed alternately with at least some of other steps or sub-steps or stages of other steps.

[0150] 8 is a block diagram of a neural network training apparatus for determining a molecular retrosynthetic pathway provided in an embodiment of the present application, which performs the steps of performing the above-mentioned neural network training method for determining a molecular retrosynthetic pathway. Referring to FIG. 8, the apparatus includes: A first determination module 801 that determines a first decomposition pathway for each of a plurality of first molecules based on molecular expression information of the first molecules, the first determination module 801 having a path depth equal to or less than a target depth; a first cost dictionary generation module 802 for obtaining a first cost dictionary based on a first decomposition pathway of each first molecule, the first cost dictionary including molecular representation information and corresponding cost value information of each first molecule, the cost value information of the first molecule indicating a cost for separating the first molecule according to the corresponding first decomposition pathway; A second determination module 803 for determining molecular representation information of at least one second molecule based on a first degradation pathway of each first molecule, where each second molecule is a molecule that can be decomposed to obtain a molecule; a third determination module 804 for determining a plurality of third molecules from each of the second molecules, each of the third molecules representing a class of second molecules; a second cost dictionary generation module 805 for obtaining a second cost dictionary based on a second decomposition pathway of each third molecule, the second cost dictionary including molecular representation information and corresponding cost value information of each third molecule, the cost value information of the third molecule indicating a cost for separating the third molecule according to the corresponding second decomposition pathway; a training module 806 for performing training based on the first cost dictionary and the second cost dictionary to obtain a target neural network, the target neural network including: a training module 806 for outputting corresponding cost value information of the target molecule based on input molecular expression information of the target molecule.

[0151] In a preferred embodiment, the first determination module 801: an obtaining unit for obtaining a first initial cost value function of each first molecule according to molecular representation information and cost value reference information of each first molecule; an update unit for updating a first initial cost value function of each first molecule according to a decomposition cost value corresponding to a decomposition path of any one layer of each first molecule when the decomposition of any one decomposition level of each first molecule is completed, to obtain a first target cost value function of each first molecule, the first target cost value function being for determining a decomposition path of the first molecule having a minimum decomposition cost value; a determining unit for determining a first decomposition path for each of the first molecules based on a first target cost value function for each of the first molecules when the decomposition task for any one of the first molecules satisfies the target decomposition condition.

[0152] In a preferred embodiment, the acquisition unit comprises: partitioning a decomposition task of each of the first molecules into a plurality of first subtasks according to the molecular representation information of each of the first molecules, the decomposition task being to separate each of the first molecules according to a decomposition pathway; Assign each first subtask to a plurality of computing nodes, and each computing node calculates a first initial cost value function for each first molecule and returns the first initial cost value function, the first initial cost value function being obtained by the corresponding computing node calculating based on molecule cost value reference information, and the molecule cost value reference information indicating the decomposability of the molecule; A first initial cost value function fed back from each computing node is received.

[0153] In a preferred embodiment, the first cost dictionary generation module 802 determining corresponding cost value information for each first molecule based on a first degradation pathway for each first molecule; A first cost dictionary is obtained based on the molecular representation information and the corresponding cost value information of each first molecule.

[0154] In a preferred embodiment, the third determination module 804 is clustering each of the second molecules to obtain a plurality of sets, each set including at least one second molecule having a similar molecular structure; The cluster centers of each set are determined to be a plurality of third molecules, each of which is a representative molecule in the set to which it belongs.

[0155] In a preferred implementation, the training module 806 includes: a first training unit for training a second neural network according to the molecular representation information and corresponding cost value information of each molecule in the second cost dictionary; a second cost dictionary updating unit for updating the first cost dictionary based on the second cost dictionary to obtain an updated first cost dictionary; a second training unit for training the first neural network according to the molecular expression information and corresponding cost value information of each molecule in the updated first cost dictionary; and a combination unit for combining the second neural network after training with the first neural network to obtain a target neural network.

[0156] In a preferred embodiment, the first training unit comprises: Input the molecular representation information of each molecule in the second cost dictionary into a second neural network to obtain corresponding predicted cost value information of each molecule; Determine a model loss of the second neural network based on the corresponding predicted cost value information of each molecule and the corresponding cost value information of each molecule in the second cost dictionary; A network parameter in the second neural network is adjusted based on the model loss of the second neural network.

[0157] In the embodiment of the present application, a neural network training device for determining molecular retrosynthetic pathways is provided. For a plurality of molecules, when determining the retrosynthetic pathways of each molecule, the idea of ​​hierarchical learning is adopted to decompose the training process of the molecular retrosynthetic pathway with large search depth into multiple layers for training, thereby increasing the training speed, and using multiple molecular retrosynthetic pathways instead of the complete retrosynthetic reaction process, after completing the training of one layer of molecular retrosynthetic pathways, a representative molecule is selected as the starting molecule of the molecular retrosynthetic pathway of the next layer in a molecular screening manner, effectively improving the search efficiency of molecular retrosynthetic pathways, and more effectively extracting accurate molecular cost value information. The hierarchical manner greatly reduces the large amount of calculation overhead caused by the determination of molecular retrosynthetic pathways, ensures the accuracy of molecular retrosynthetic pathways, and reduces the time required for determining molecular retrosynthetic pathways.

[0158] In one embodiment, a molecular retrosynthetic routing apparatus is provided, the apparatus comprising: a receiving module for receiving a molecular representation of a target molecule, the molecular representation indicating a three-dimensional chemical structure of the target molecule; an input module for inputting the molecular representation information of the target molecule into a neural network for determining a molecular retrosynthetic pathway; A decomposition path determination module that determines a target decomposition path of the target molecule based on a neural network for determining the molecular retrosynthesis path, the target decomposition path being a decomposition path with the lowest decomposition cost value among at least one decomposition path; and a route determination module for obtaining molecular retrosynthetic route information of the target molecule based on the target decomposition route.

[0159] In one embodiment, the neural network for determining the molecular retrosynthetic pathway includes a first neural network and a second neural network. The decomposition pathway determination module further outputs cost value information of the upper retrosynthetic pathway of the target molecule based on the first neural network, determines a first decomposition pathway of the target molecule based on the cost value information of the upper retrosynthetic pathway of the target molecule, determines molecular representation information of a molecule having an additional decomposition device from a molecule obtained by decomposing the target molecule based on the first decomposition pathway, inputs the molecular representation information of the molecule having an additional decomposition device into the second neural network, outputs cost value information of the lower retrosynthetic pathway of the target molecule based on the second neural network, determines a second decomposition pathway of the target molecule based on the cost value information of the lower retrosynthetic pathway of the target molecule, and determines a target decomposition pathway of the target molecule based on the first decomposition pathway and the second decomposition pathway.

[0160] Here, the above-mentioned embodiment provides an apparatus, and only the partition of each of the above-mentioned functional modules is described as an example, and in actual application, the above-mentioned functions can be allocated to be completed by different functional modules according to needs, that is, the internal structure of the apparatus can be partitioned into different functional modules to complete all or part of the above-mentioned functions. In addition, the above-mentioned embodiment provides an apparatus and the corresponding method embodiment belong to the same concept, and the specific implementation process can be referred to the method embodiment, and will not be described here.

[0161] When the computer device is configured as a server, Fig. 9 is a structural schematic diagram of a server provided by the embodiments of the present application, and the server 900 may have a large difference due to different configurations or performances, and may include one or more processors (Central Processing Units, CPU) 901 and one or more memories 902, in which at least one computer-readable instruction is stored, and the at least one computer-readable instruction is read and executed by the one or more processors 901 to realize the method for training a neural network for determining a molecular retrosynthetic pathway or the method for determining a molecular retrosynthetic pathway provided by each of the above method embodiments. Of course, the server 900 may further include components such as a wired or wireless network interface, a keyboard, and an input / output interface to perform input and output, and the server 900 may include other components for realizing the functions of the device, which will not be described here.

[0162] An embodiment of the present application further provides one or more computer-readable storage media, the one or more computer-readable storage media being applied to a computer device, the computer-readable storage media storing at least one computer-readable instruction, the at least one computer-readable instruction being read and executed by one or more processors to realize the operations performed by the computer device in each of the methods of the above embodiments.

[0163] An embodiment of the present application further provides a computer program product, which includes computer-readable instructions, the computer-readable instructions being stored in a computer-readable storage medium, one or more processors of a computing device reading the computer-readable instructions from the computer-readable storage medium, and the one or more processors executing the computer-readable instructions, thereby causing the computing device to perform the steps of each method provided by the various preferred implementations above.

[0164] As can be understood by those skilled in the art, the realization of all or part of the steps in the above embodiments may be completed by hardware, or may be completed by computer-readable instructions instructing related hardware, the computer-readable instructions being stored in one or more computer-readable storage media, and the above-mentioned storage media may be read-only memory, magnetic disks or optical disks, etc.

[0165] The above does not limit the present application, but is a preferred embodiment of the present application, and any modifications, equivalent replacements, improvements, etc. completed within the spirit and principle of the present application fall within the protection scope of the present application.

Claims

1. 1. A method for training a neural network for determining a molecular retrosynthetic pathway executed by a computing device, the method comprising: determining a first degradation pathway for each of a plurality of first molecules based on molecular representation information of the first molecules, obtaining a value of a first initial cost function of each of the first molecules from molecular representation information and cost value reference information of each of the first molecules based on a first initial cost function; When any one of the decomposition levels of each of the first molecules has completed decomposition, a step of updating a value of a first initial cost value function of each of the first molecules based on a decomposition cost value corresponding to a decomposition path of any one layer of each of the first molecules to obtain a value of a first target cost value function of each of the first molecules, wherein the first target cost value function is a function for determining a decomposition path of the first molecule having a minimum decomposition cost value; determining a first decomposition path for each of the first molecules based on a first target cost value function for each of the first molecules when the decomposition task for any one of the first molecules satisfies a target decomposition condition; obtaining a first cost dictionary based on a first decomposition pathway of each of the first molecules, the first cost dictionary including molecular representation information and corresponding cost value information of each of the first molecules, the cost value information of the first molecules indicating a cost for separating the first molecules according to the corresponding first decomposition pathway; determining molecular representation information of at least one second molecule based on a first decomposition pathway of each of the first molecules, wherein each of the second molecules is a molecule that can be decomposed to obtain a molecule among molecules obtained by decomposing a corresponding first molecule based on the first decomposition pathway; determining a plurality of third molecules from each of the second molecules, each of the third molecules representing a class of the second molecules; obtaining a second cost dictionary based on a second decomposition pathway of each of the third molecules, the second cost dictionary including molecular representation information and corresponding cost value information of each of the third molecules, the cost value information of the third molecules indicating a cost for separating the third molecules according to the corresponding second decomposition pathway; and a step of obtaining a target neural network by training based on the first cost dictionary and the second cost dictionary, wherein the target neural network outputs corresponding cost value information of the target molecule based on input molecular representation information of the target molecule, and the corresponding cost value information of the target molecule is information for synthesizing a retrosynthetic pathway of the target molecule.

2. The step of acquiring a value of a first initial cost value function of each of the first molecules from molecular representation information and cost value reference information of each of the first molecules based on a first initial cost value function includes: partitioning a decomposition task of each of the first molecules into a plurality of first subtasks based on molecular representation information of each of the first molecules, the decomposition task separating each of the first molecules according to a decomposition pathway; assigning each of the first subtasks to a plurality of computing nodes, each computing node calculating a value of a first initial cost value function for each of the first molecules, and returning a value of the first initial cost value function, the value of the first initial cost value function being obtained by a corresponding computing node calculating based on the first initial cost value function from molecule cost value reference information, the molecule cost value reference information indicating decomposability of the molecule; and receiving a first initial cost function value fed back from each of the computational nodes.

3. The step of obtaining a first cost dictionary based on a first decomposition pathway of each of the first molecules includes: determining corresponding cost value information for each of the first molecules based on a first degradation pathway for each of the first molecules; and obtaining a first cost dictionary based on the molecular representation information and corresponding cost value information of each of the first molecules.

4. The step of determining a plurality of third molecules indicative of a class of the second molecules from each of the second molecules comprises: clustering each of the second molecules to obtain a plurality of sets, each of the sets including at least one second molecule having a similar molecular structure; 2. The method of claim 1, further comprising: determining a cluster center of each of the sets to a third molecule corresponding to the set, thereby obtaining a plurality of third molecules, each of the third molecules being a representative molecule in the set to which it belongs.

5. The step of obtaining a target neural network by training based on the first cost dictionary and the second cost dictionary includes: training a second neural network based on the molecular representation information and corresponding cost value information of each molecule in the second cost dictionary; updating the first cost dictionary based on the second cost dictionary to obtain an updated first cost dictionary; training a first neural network based on the molecular representation information and corresponding cost value information of each molecule in the updated first cost dictionary; and combining the second neural network after training with the first neural network to obtain a target neural network.

6. training a second neural network based on the molecular representation information and corresponding cost value information of each molecule in the second cost dictionary, inputting the molecular representation information of each molecule in the second cost dictionary into the second neural network to obtain corresponding predicted cost value information of each molecule; determining a model loss of the second neural network based on the corresponding predicted cost value information of each molecule and the corresponding cost value information of each molecule in the second cost dictionary; and adjusting network parameters in the second neural network based on a model loss of the second neural network.

7. 1. A method for determining a molecular retrosynthetic pathway implemented by a computational device, the method comprising: receiving a molecular representation of a target molecule, the molecular representation indicating a three-dimensional chemical structure of the target molecule; inputting the molecular representation information of the target molecule into a neural network for determining a molecular retrosynthetic pathway, the neural network being trained by the method of any one of claims 1 to 6; determining a target degradation pathway of the target molecule based on a neural network for determining the molecular retrosynthetic pathway; and obtaining molecular retrosynthetic pathway information for the target molecule based on the target decomposition pathway.

8. The step of determining a target degradation pathway of the target molecule based on a neural network for determining the molecular retrosynthetic pathway comprises: The neural network for determining a molecular retrosynthetic pathway includes a first neural network and a second neural network, outputting cost value information of the upper layer retrosynthetic pathway of the target molecule based on the first neural network; determining a first decomposition pathway of the target molecule based on cost value information of an upper layer retrosynthetic pathway of the target molecule; determining molecular representation information of a molecule for which a further decomposition method exists from molecules obtained by decomposing the target molecule based on the first decomposition pathway; inputting molecular representation information of the molecule for which further decomposition methods exist into the second neural network; outputting cost value information of a lower-level retrosynthetic pathway of the target molecule based on the second neural network; determining a second decomposition pathway of the target molecule based on cost value information of a lower-layer retrosynthetic pathway of the target molecule; and determining a target degradation pathway for the target molecule based on the first degradation pathway and the second degradation pathway.

9. 1. An apparatus for training a neural network for determining a molecular retrosynthetic pathway, the apparatus being provided on a computing device, the apparatus comprising: A first determination module that determines a first degradation pathway for each of a plurality of first molecules based on molecular representation information of the first molecules, an acquiring unit for acquiring a value of a first initial cost function of each of the first molecules from the molecular representation information and the cost value reference information of each of the first molecules according to the first initial cost function; an update unit for updating a value of a first initial cost value function of each first molecule according to a decomposition cost value corresponding to a decomposition path of any one layer of each first molecule when the decomposition of any one decomposition level of each first molecule is completed, to obtain a value of a first target cost value function of each first molecule, the first target cost value function being used to determine a decomposition path of the first molecule having a minimum decomposition cost value; A first determination module including: a determination unit for determining a first decomposition path of each of the first molecules based on a first target cost value function of each of the first molecules when the decomposition task of any one of the first molecules satisfies a target decomposition condition; a first cost dictionary generation module for obtaining a first cost dictionary based on a first decomposition pathway of each of the first molecules, the first cost dictionary including molecular representation information and corresponding cost value information of each of the first molecules, the cost value information of the first molecules indicating a cost for separating the first molecules according to the corresponding first decomposition pathway; A second determination module that determines molecular expression information of at least one second molecule based on a first decomposition pathway of each of the first molecules, wherein each of the second molecules is a molecule that can be decomposed to obtain a molecule among molecules obtained by decomposing a corresponding first molecule based on the first decomposition pathway; a third determination module for determining a plurality of third molecules from each of the second molecules, each of the third molecules representing a class of the second molecules; a second cost dictionary generation module for obtaining a second cost dictionary based on a second decomposition pathway of each of the third molecules, the second cost dictionary including molecular representation information and corresponding cost value information of each of the third molecules, the cost value information of the third molecules indicating a cost for separating the third molecules according to the corresponding second decomposition pathway; a training module that performs training based on the first cost dictionary and the second cost dictionary to obtain a target neural network, the target neural network outputting corresponding cost value information of the target molecule based on input molecular expression information of the target molecule, the corresponding cost value information of the target molecule being information for synthesizing a retrosynthetic pathway of the target molecule.

10. 1. A molecular retrosynthetic pathway determination apparatus provided on a computing device, the apparatus comprising: a receiving module for receiving a molecular representation of a target molecule, the molecular representation indicating a three-dimensional chemical structure of the target molecule; an input module for inputting molecular representation information of the target molecule into a neural network for determining a molecular retrosynthetic pathway, the neural network being trained by the method according to any one of claims 1 to 6; a decomposition pathway determination module that determines a target decomposition pathway of the target molecule based on a neural network that determines the molecular retrosynthetic pathway; and a route determination module that obtains molecular retrosynthetic route information for the target molecule based on the target decomposition route.

11. A computing device comprising one or more processors and a memory, the memory storing at least one computer readable instruction, the at least one computer readable instruction being read by the one or more processors to perform a method according to any one of claims 1 to 8.

12. A computer program comprising at least one computer readable instruction, said at least one computer readable instruction for carrying out a method according to any one of claims 1 to 8.

Citation Information

Patent Citations

  • Method for evaluating synthesization of organic small-molecule compounds based on reverse synthesis

    CN101789047A

  • Method and device for determining chemical synthesis route

    CN109872780A

  • Method, equipment and device for determining reactant of target compound and medium

    CN111223532A

  • Artificial intelligence-based retrosynthesis prediction method and device, equipment and storage medium

    CN111524557A