Method for predicting the retrosynthesis of compound molecules and related device
The method employs a graph neural network and reactant generation network to enhance the accuracy of retrosynthesis predictions by recursively expanding leaf nodes and screening reactants, addressing the inaccuracies of template-based algorithms in predicting compound molecular pathways.
Patent Information
- Application Number
- JP2023528742
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2021-01-27
- Filing Date
- 2022-01-21
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-01-21
AI Technical Summary
Template-based retrosynthesis algorithms require large amounts of reaction templates and can be unpredictable or inaccurately predict compound molecular retrosynthesis processes, affecting the accuracy of retrosynthesis predictions.
A method and device using a graph neural network and reactant generation network to predict retrosynthesis by expanding leaf nodes recursively, determining bond cleavage positions, and screening reactants based on predetermined conditions to ensure accurate multi-step reaction retrosynthesis.
The method improves the accuracy of retrosynthesis predictions by ensuring reliable reactant determination through recursive expansion and path tracing, enhancing the reliability and precision of compound molecule synthesis pathways.
Smart Images

Figure 0007580851000016 
Figure 0007580851000017 
Figure 0007580851000018
Abstract
Description
[Technical field]
[0001] This application was filed with the China Patent Office on January 27, 2021, bearing application number 2021101122078, and claims priority to a Chinese patent application entitled "Method for predicting the retrosynthesis of compound molecules and related apparatus," 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 technology for predicting the retrosynthesis of compound molecules. [Background technology]
[0003] Due to the rapid development of artificial intelligence technology, it has been gradually introduced into various scientific fields and plays an important role. In the field of chemistry, chemical reactions change infinitely under different conditions, so when producing compound molecules, researchers generally have to spend a lot of time designing a reasonable organic synthesis route. By using artificial intelligence technology to help researchers design organic synthesis routes, the efficiency of researchers' research and development of chemical drug molecules and other compounds can be greatly improved.
[0004] At present, artificial intelligence retrosynthesis algorithms mainly include template-based retrosynthesis algorithms. In template-based retrosynthesis algorithms, first, a template or rule for describing a chemical reaction conversion rule is obtained, which can be manually labeled or extracted from an existing chemical reaction library. Then, a predicted chemical reaction is matched to a target molecule based on the obtained template or rule. Summary of the Invention [Problem to be solved by the invention]
[0005] However, template-based retrosynthetic algorithms generally require a large number of reaction templates, and in the case of a retrosynthetic process without a reaction template, it may be impossible to predict or may make an incorrect prediction, which may further affect the accuracy of compound molecule retrosynthesis prediction.
[0006] In view of this, the present application provides a compound molecule retrosynthesis prediction method and related device, which can effectively improve the accuracy of compound molecule retrosynthesis prediction. [Means for solving the problem]
[0007] A first aspect of the present application provides a method for predicting the retrosynthesis of a compound molecule executed by a computer device, specifically comprising: obtaining a target molecule and setting the target molecule as a root node of a tree structure, the root node being associated with a first leaf node of the tree structure, the tree structure including a retrosynthetic pathway of the target molecule; A step of obtaining a plurality of second leaf nodes by performing an expansion process on the first leaf node using a target retrosynthesis model, the target retrosynthesis model including a graph neural network and a reactant generation network, the graph neural network predicting a bond cleavage position of a compound molecule corresponding to the first leaf node, and the reactant generation network determining a predicted molecule set based on the bond cleavage position; determining a terminal node that satisfies a predetermined condition by performing a recursive process on the predicted molecule set corresponding to the second leaf node; and traversing the pathway information corresponding to the terminal nodes to determine a retrosynthetic pathway for the target molecule.
[0008] A second aspect of the present application provides a device for predicting the retrosynthesis of a compound molecule, a obtaining unit for obtaining a target molecule and making the target molecule a root node of a tree structure, the root node being associated with a first leaf node of the tree structure, the tree structure including a retrosynthetic pathway of the target molecule; an expansion unit for performing an expansion process on the first leaf node using a target retrosynthesis model to obtain a plurality of second leaf nodes, the target retrosynthesis model including a graph neural network and a reactant generation network, the graph neural network predicting a bond cleavage position of a compound molecule corresponding to the first leaf node, and the reactant generation network determining a predicted molecule set based on the bond cleavage position; a processing unit for performing a recursive process on the predicted molecule set corresponding to the second leaf node to determine a terminal node that satisfies a predetermined condition; and a prediction unit for traversing the pathway information corresponding to the terminal nodes to determine a retrosynthetic pathway of the target molecule.
[0009] A third aspect of the present application provides a computer device, comprising a memory, a processor and a bus system, the memory storing a program code, and the processor executing the method for predicting the retrosynthesis of a compound molecule according to the first aspect above, according to the command of the program code.
[0010] A fourth aspect of the present application provides a computer-readable storage medium having stored thereon instructions that, when executed on a computer, cause the computer to carry out the method for predicting the retrosynthesis of a compound molecule as described in the first aspect above.
[0011] A fifth aspect of the present application provides a computer program product or a computer program, the computer program product or the computer program comprising computer instructions, the computer instructions being stored in a computer readable storage medium, a processor of a computer device reading the computer instructions from the computer readable storage medium, and the processor executing the computer instructions to cause the computer device to perform the method for predicting the retrosynthesis of a compound molecule according to the first aspect above. Effect of the Invention
[0012] As can be seen from the above technical solutions, the embodiments of the present application have the following advantages: A target molecule is obtained, and the target molecule is a root node of a tree structure, and the root node is associated with a first leaf node of the tree structure, and the tree structure includes a retrosynthetic pathway of the target molecule. Next, a target retrosynthetic model performs an expansion process on the first leaf node to obtain a plurality of second leaf nodes, and the target retrosynthetic model includes a graph neural network and a reactant generation network, and the graph neural network predicts a bond cleavage position of a compound molecule at the first leaf node, and the reactant generation network determines a predicted molecule set based on the bond cleavage position, and further performs a recursive process on the predicted molecule set corresponding to the second leaf node to determine a terminal node that satisfies a predetermined condition, and further traverses the path information corresponding to the terminal node to determine a retrosynthetic pathway of the target molecule. In this way, a multi-step reaction retrosynthetic prediction process is realized, and the leaf nodes are gradually expanded and screened recursively, and path tracing is performed on the terminal node when the predetermined condition is satisfied, thereby ensuring the reliability of the reactants determined by the multi-step reaction retrosynthetic prediction process and improving the accuracy of the retrosynthetic prediction of the compound molecule. [Brief description of the drawings]
[0013] [Figure 1] FIG. 1 is a diagram showing a network architecture implemented by a system for predicting the retrosynthesis of compound molecules. [Diagram 2] FIG. 1 is an architecture diagram of the retrosynthetic prediction of compound molecules provided in the examples of the present application. [Diagram 3] 1 is a flow chart of a method for predicting the retrosynthesis of a compound molecule provided in the examples of the present application. [Figure 4] FIG. 2 is a schematic diagram of the operation process of a search tree provided in an embodiment of the present application. [Diagram 5] FIG. 1 is a schematic diagram of a process for predicting intermediate reactants provided in the examples of the present application. [Figure 6] FIG. 1 is a schematic diagram of the operation process of the target retrosynthesis model provided in the examples of the present application. [Figure 7] FIG. 1 is a schematic diagram of a retrosynthetic reaction process provided in the examples of the present application. [Figure 8] FIG. 2 is a schematic diagram of the working process of the graph neural network provided in the embodiments of the present application. [Figure 9] FIG. 2 is a schematic diagram of a string conversion provided in an embodiment of the present application. [Figure 10] 1 is a flow chart of a method for predicting the retrosynthesis of another compound molecule provided in the examples of the present application. [Figure 11] FIG. 1 is a schematic diagram of the retrosynthetic prediction process for compound molecules provided in the examples of the present application. [Figure 12] FIG. 1 is a schematic diagram of the retrosynthetic prediction process for another compound molecule provided in the examples of the present application. [Figure 13] FIG. 1 is a schematic diagram of the retrosynthetic prediction process for another compound molecule provided in the examples of the present application. [Figure 14] FIG. 1 is a schematic diagram of the retrosynthetic prediction process for another compound molecule provided in the examples of the present application. [Figure 15] FIG. 1 is a structural schematic diagram of a retrosynthetic prediction device for compound molecules provided in the examples of the present application. [Figure 16] FIG. 2 is a structural schematic diagram of a terminal device provided in an embodiment of the present application; [Figure 17] FIG. 2 is a structural schematic diagram of a server provided in an embodiment of the present application; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0014] First, we will interpret some terms that may appear in the examples of this application.
[0015] SMILES (Simplified molecular input line entry specification): A specification for unambiguously describing molecular structures using ASCII strings.
[0016] Transformer: A neural network model that performs incremental learning based on the attention mechanism.
[0017] Graph Neural Networks (GNNs): Neural networks that process graph data, for example a set of neural network algorithms that operate on molecular maps (also called "molecular graphs").
[0018] Reaction Center: A subgraph consisting of the key vertices to be cut in retrosynthesis and their first-level adjacent edges.
[0019] Monte Carlo Tree Search (MCTS): A heuristic search algorithm for decision processes.
[0020] Atom-mapping: A method of describing chemical reactions that involves one-to-one matching of atoms between reactants and products. Commonly used in template-based retrosynthetic algorithms.
[0021] Basis molecule set: a compound library that determines the predicted end point of a reaction in a multi-step retrosynthesis, i.e., when a predicted reactant is in the basis molecule set, the prediction of the reactant can be stopped without being further decomposed to predict a synthetic pathway. Based on this, the amount of compounds included in the basis molecule set can determine the number of steps in the retrosynthetic pathway.
[0022] Root node: A search start node in a tree structure, which indicates a target synthetic compound in the retrosynthesis prediction of a compound.
[0023] Leaf node: A node below the root node, which indicates an intermediate compound predicted by a single-step or multi-step retrosynthesis in the prediction of the retrosynthesis of a compound.
[0024] Terminal node: a node in a tree structure that satisfies a search termination condition, for example, a node in a Monte Carlo tree for retrosynthetic prediction that satisfies a search termination condition, which may be that the number of expansions of the terminal node reaches a predetermined value, or that the terminal node is a molecule in the base molecule set.
[0025] Retrosynthetic pathway: A route leading from a target synthetic compound to compounds that participate in a reaction.
[0026] Here, the compound molecule retrosynthesis prediction method provided in the present application is applied to a computer device, specifically, it can be executed by a system or program having a compound molecule retrosynthesis prediction function, such as a drug synthesis assistant, executed on a computer device. Specifically, the compound molecule retrosynthesis prediction system can be executed in the network architecture of FIG. 1. As shown in FIG. 1, the compound molecule retrosynthesis prediction system can provide compound molecule retrosynthesis prediction services for multiple information sources, for example, by triggering an operation on the terminal side, sending the target molecule to a server, and the server predicts reactants and feeds back the prediction result to the terminal. Various terminal devices are shown in FIG. 1, and the terminal devices may be computer devices, and in an actual scenario, more or less types of terminal devices may participate in the compound molecule retrosynthesis prediction process, and the specific number and type are determined according to the actual scenario, and are not limited here. Also, FIG. 1 shows one server, and in an actual scenario, multiple servers may participate, and specifically, the number of servers participating in the compound molecule retrosynthesis prediction process is determined according to the actual scenario.
[0027] In this embodiment, the server may be an independent physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing basic cloud computing services. The terminal may be, but is not limited to, a smartphone, a tablet, a laptop, a desktop computer, a smart speaker, a smart watch, etc. The terminal and the server may be directly or indirectly connected by wired or wireless communication, and the terminal and the server may be connected to form a blockchain network, but this is not limited to the present disclosure.
[0028] Here, the above-mentioned compound molecule retrosynthetic prediction system can be executed on a personal mobile terminal, for example, as an application program similar to a drug synthesis assistant, and may be executed on a server, or may be executed on a third-party device to perform retrosynthetic prediction of compound molecules and obtain the retrosynthetic prediction processing result of the compound molecules from the information source. Specifically, the compound molecule retrosynthetic prediction system may be executed on the above-mentioned device as a program, or may be executed as a system component in the above-mentioned device, or may even be one of the cloud service programs. The specific execution mode is determined according to the actual scenario and is not limited here.
[0029] Currently, template-based algorithms are mainly used to realize retrosynthetic prediction of compound molecules. Template-based algorithms require a large number of reaction templates. In the case of a retrosynthetic process without a reaction template, prediction may not be possible or may be erroneous, which affects the accuracy of compound molecule retrosynthetic prediction.
[0030] In order to solve the above problems, the present application presents a method for predicting the retrosynthesis of a compound molecule, which is applied to the architecture of the retrosynthesis prediction of a compound molecule shown in FIG. 2. As shown in FIG. 2, the user performs a target operation of inputting a target molecule through a terminal, and the terminal sends the target molecule to a server. The server then combines single-step reaction prediction and multi-step reaction prediction to predict a reactant corresponding to the target molecule. Specifically, the molecular graph structure information and SMILES string information of the existing product (target molecule) are used to supplement the synthon after bond cleavage to obtain the reactant. In the prediction process, the molecular graph structure and SMILES string information are fully used, and the entire flow of the retrosynthesis is first given the bond cleavage position and supplemented with the synthon after bond cleavage, so that the entire flow is easy to visualize and has interpretability. Then, the single-step results are sorted and selected by the method of Monte Carlo tree search, and multi-step reaction prediction is realized to obtain the reactant.
[0031] Here, the method provided by the present application may be a program written, or may be realized as a processing logic in a hardware system, or as a compound molecule retrosynthesis prediction device, in an integrated or external form, by implementing the above processing logic. In the implementation method, the compound molecule retrosynthesis prediction device obtains a target molecule, and the target molecule is set as a root node of a tree structure, and the root node is associated with a first leaf node of the tree structure, and the tree structure includes a retrosynthesis route of the target molecule, and the first leaf node is expanded based on a target retrosynthesis model to obtain a plurality of second leaf nodes, and the target retrosynthesis model includes a graph neural network and a reactant generation network, and the graph neural network determines a bond cleavage position of the compound molecule in the first leaf node, and the reactant generation network obtains a predicted molecule set based on the bond cleavage position, and further performs a recursive process on the predicted molecule set in the second leaf node to determine a terminal node when a predetermined condition is reached, and further traverses the route information corresponding to the terminal node to determine the retrosynthesis route of the target molecule. In this way, a multi-step reaction retrosynthetic prediction process is realized, and in the retrosynthetic prediction process, a step-by-step recursive expansion is performed to gradually recursively expand and screen the leaf nodes, and path tracing is performed for the terminal nodes that satisfy the specified conditions, thereby ensuring the reliability of the reactants predicted by the multi-step reaction retrosynthetic prediction process and improving the accuracy of compound molecule retrosynthetic prediction.
[0032] The solutions provided in the embodiments of the present application relate to artificial intelligence machine learning technology, and are specifically illustrated by the following embodiments.
[0033] The following describes the method for predicting the retrosynthesis of a compound molecule provided in the embodiments of the present application. Refer to FIG. 3. FIG. 3 is a flow chart of the method for predicting the retrosynthesis of a compound molecule provided in the embodiments of the present application. The prediction method is executed by a computer device, specifically, may be executed by a terminal, may be executed by a server, or may be executed by a terminal and a server together. The embodiments of the present application include at least the following steps: 301: A target molecule is obtained and the target molecule is set as a root node in a tree structure.
[0034] In this embodiment, the target molecule may be a drug molecule or a compound molecule having other uses, and the specific molecular form is determined according to the actual scenario.
[0035] In addition, the root node is associated with a first leaf node of the tree structure, the tree structure includes a retrosynthetic pathway of the target molecule, and the tree structure may be a Monte Carlo tree, which is a heuristic search algorithm tree for a decision process, in which the root node is an upper node of the leaf nodes, and the synthetic pathway of the target molecule can be determined by tracing the pathway from the leaf nodes to the root node.
[0036] For ease of understanding, we refer to the schematic diagram of the operation process of the search tree provided in the embodiment of the present application shown in Figure 4. Each node corresponds to a molecule set (e.g., a node corresponding to a reactant set includes multiple reactant molecules), and the node corresponding to the terminal molecule (reactant) is defined as the terminal node. Starting from the root node (target molecule), the search tree can iteratively perform four steps of selection, expansion, rollout and update, and then return to selection again after the update is completed, and repeat the above process.
[0037] Specifically, the selection process means that the search tree traverses from the root node (target molecule) to the leaf nodes, and in this process, selects a subnode A with a high score. The score is generated by a scoring function, which is a weighted sum of the rewards fed back by all the terminal nodes expanded from the root node and the inverse of the access count of the root node.
[0038] The development process develops the node A with a high score to obtain the node B and node C, that is, adds a leaf node according to the target retrosynthesis model. For example, first, 10 reaction types are traversed, and three results are predicted for each reaction type, to obtain a total of 30 results. Then, the 30 results can be filtered to reserve some with high probability using some rules, retrosynthesis probability model, forward synthesis prediction model probability, and other methods. Some rules include: (1) The generated result molecule is a molecule that exists in nature; (2) The generated result molecule is an organic substance; (3) The number of rings of the generated result molecule does not increase. The retrosynthesis probability model and the forward synthesis prediction model are empirical models obtained collectively based on synthesis experience, and both can reflect the feasibility of the developed compound.
[0039] An exemplary implementation of the rollout process is as follows: first, traverse 10 reaction types, predict one result for each reaction type, and obtain a total of 10 results. Then, filter the 10 results and use the reverse synthesis probability, forward synthesis prediction probability, etc. to select the most likely reaction for the next development. The above process is repeated until a stopping condition is met. The specific stopping condition is that all molecules corresponding to the leaf nodes are in the basic molecule set, or a predetermined number of times is reached.
[0040] The update process means calculating the reward score of the terminal node and gradually updating it to the nodes in the upper layer. Regarding the calculation of the reward, a synthetic route capable of generating the terminal node is obtained by rollout, and for the molecules corresponding to all the leaf nodes in the synthetic route, the ratio of the number in the raw material library to the number of all the leaf node molecules is calculated as the reward, thereby updating the scoring of the tree structure, and the subsequent selection-expansion-rollout-update cycle can be easily performed, and appropriate reactants can be accurately extracted based on the tree structure.
[0041] The scoring in the tree structure of this embodiment is set as a cyclic process, and the specific scoring function may use the scoring function of any round in the cyclic process, and it should be understood that the numerical explanations in the above step interpretation are only examples, and the specific numerical values are determined according to the actual scenario.
[0042] In 302, a target reverse synthesis model is used to perform an unfolding process on the first leaf node to obtain a plurality of second leaf nodes.
[0043] In this embodiment, the target retrosynthesis model includes a graph neural network and a reactant generation network, in which the graph neural network predicts the bond cleavage positions of the compound molecule corresponding to the first leaf node, and the reactant generation network determines a predicted molecule set based on the bond cleavage positions; that is, the process of determining intermediate reactants includes a step of predicting the potential bond cleavage positions of the molecule and generating a synthon after bond cleavage, and a step of supplementing the synthon with reactant information to obtain intermediate reactants.
[0044] Optionally, in the process of obtaining the intermediate reactant (i.e., the process of determining the second leaf node), since there may be multiple candidate results, it is necessary to screen to obtain the intermediate reactant. FIG. 5 is a schematic diagram of the process of predicting the intermediate reactant provided in the embodiment of the present application. As shown in FIG. 5, for compound molecule A, after determining its bond cleavage position, first, expansion processing is performed based on the reaction type, and then, a preliminary selection of the reaction type is performed to obtain candidate sequence 1 (R1-Rn). For example, 10 types with the highest scores among each reaction type are selected, and validity screening is performed to obtain candidate sequence 2 (R1-Rk, k < n). For example, it is determined whether each molecule in candidate sequence 1 is a molecule stable in terms of effect. Furthermore, atom conservation screening is performed to obtain candidate sequence 3 (R1-Rm, m < k). By calculating the synthesis probability in candidate sequence 3, screening is performed to obtain possible reactant sets B and C. The synthesis probability can be obtained based on the statistics of the success rate or usage rate of different types of reactions in the database.
[0045] Also, regarding the determination of the intermediate reactant, the bond cleavage probability may be determined directly based on the set in which the reaction type is expanded, or it may be obtained by determining the reactant prediction probability after the preliminary selection of the reaction type. The reactant set at the second leaf node may be determined based on the overlap of both.
[0046] For one complex molecule, generally, it is necessary to perform multiple single-step reaction predictions. And based on the sorting results of a single step, for the obtained reactants or intermediate products, the next single-step reaction prediction is performed until reaching the known or commercially available raw material library in the literature (here defined as the basic molecule set for determining the end point of retrosynthetic analysis). Therefore, the accuracy of the single-step reaction prediction plays an important role in the prediction of the entire multi-step retrosynthesis.
[0047] The process of single-step prediction, that is, the process of using the target retrosynthesis model, is described below. Figure 6 is a schematic diagram of the operation process of the target retrosynthesis model provided in the embodiment of the present application, and the process of single-step prediction is shown in the figure, which firstly determines the first string of the compound molecule corresponding to the first leaf node, then converts the first string into a first molecular graph, and then inputs a graph based on the first molecular graph through a graph neural network to determine the bond cleavage position of the compound molecule corresponding to the first leaf node (bond cleavage position determination), and determines at least one synthon based on the input of the bond cleavage position through a reactant generation network, and finally filters at least one synthon according to a predetermined rule to determine a predicted molecular set (reactant set).
[0048] After obtaining the reactant set, a specific reaction process can be obtained. FIG. 7 is a schematic diagram of the retrosynthetic reaction process provided in the examples of the present application. The diagram shows synthon 1 and synthon 2 determined based on the target molecule, reactant 1 and reactant 2 corresponding to synthon 1 and synthon 2, and reactant 1 is determined by performing a base molecule set search for synthon 1, that is, searching for molecules associated with the synthon in the base molecule set.
[0049] It should be understood that the problem of processing bond breaking prediction is a graph-to-graph conversion problem. In the above steps, the molecular graph determination process is specifically as follows: first, convert the SMILES string of a given product into a molecular graph Gp having a corresponding Np number of nodes, then input the molecular graph Gp into the graph neural network, and then predict the bond breaking in the molecular graph Gp.
[0050] Specifically, the bond breaking prediction process is carried out based on the target key features, specifically, firstly, obtain bond breaking information based on the first molecular graph and the limit of the number of bond breaking by the graph neural network, and determine the target key features based on the target molecule, such as the types of atoms and the types of bonds contained in the target molecule, and further extract the bond breaking information based on the target key features and determine the bond breaking position.
[0051] The target key features may include atom key features and bond key features, and the atom (node) and bond (edge) features used by the target key features are automatically extracted using the open source RDKit. Specifically, it should be understood that the atom key features may include at least one of atom type, bond number, formal charge, chirality, hydrogen atom number, atom hybridization state, aromaticity, atomic weight, high frequency reaction center feature, and reaction type, and the bond key features may include at least one of bond type, conjugation, ring bond, and molecular stereochemistry feature.
[0052] Specifically, the atom type is the atomic number of the atom (referring to the number by which elements are arranged in the periodic table), the number of bonds is the number of different chemical bonds to which the atom belongs, the formal charge is the charge of the atom assigned to the product molecule, the chirality refers to the non-superimposability of a molecule on its mirror image, such as the non-superimposability of a human hand between a left hand and its mirror image, the right hand, on each hand, the number of hydrogen atoms is the number of hydrogen atoms connected to the atom, the atomic hybridization state includes sp, sp2, sp3, sp3d or sp3d2, the aromaticity characterizes whether the atom is in an aromatic ring system, the atomic weight is the weight of the atom, the high frequency reaction center feature characterizes whether an atom has a high frequency reaction center feature and depends on whether the molecular sub-map containing the atom is a high frequency reaction center, the high frequency reaction center is the center of a frequent reaction extracted from the products of the retrosynthesis training set, and the reaction type is the chemical reaction type of the retrosynthetic reaction, which may be a feature of the atom.
[0053] The bond type indicates the type to which the chemical bond belongs, for example, single bond, double bond, triple bond, aromatic bond, etc.; the conjugation characteristics indicate whether the chemical bond is conjugated, the ring bond characteristics indicate whether the chemical bond is part of a ring bond, the molecular stereochemistry characteristics are achirality element, any of the chirality elements, or stereochemistry of a double bond, etc.
[0054] Optionally, since a given target molecule may have a combination of various bond breaks, the total number of bond breaks can be limited by increasing auxiliary tasks. Figure 8 is a schematic diagram of the operation process of the graph neural network provided in the embodiment of the present application. The figure shows the bond break prediction process by atom features (a), and the bond break prediction process by atom features and bond features (b), and the accuracy of the determined bond break position can be improved by setting a limit on the number of bond breaks, bond break type, or bond break energy value based on the bond break prediction process of atom features.
[0055] The training process of the graph neural network of this embodiment is described below, and the model training data for the training process may be from the USPTO_50k and USPTO_480k data, which are subsets of the 1.8 million datasets in the US patent database USPTO. Reaction data containing chiral molecules in the reaction are removed, and the whole is processed by atom-mapping. The distribution of reaction types, as shown in Table 1, mainly includes 10 chemical reaction types. [Table 1]
[0056] Specifically, in the training process of the graph neural network, firstly obtain a first training molecule and a corresponding first training synthon, then determine the node features (atom features) and edge features (bond features) of the first training molecule and the first training synthon, the node features indicate the atomic relationship between the first training molecule and the first training synthon, and the edge features indicate the chemical bond relationship between the first training molecule and the first training synthon, and then determine a first loss function according to the node features and edge features, determine the bond cleavage probability of the chemical bond between the first training molecule and the first training synthon, determine a second loss function according to the bond cleavage probability, and further update the model parameters of the graph neural network according to the first loss function and the second loss function.
[0057] Specifically, the first loss function determination process, i.e., for a given input, Gp = {Ap,Ep,Xp} Gp is the molecular graph of the first training molecule, Ap is the atomic features of the first training molecule, Ep is the binding features of the first training molecule, and Xp are the parameters of the GAT model.
[0058]
number
[0059] First, the vector representation (embedding) of the node is multiplied by a weight.
number
number
number
number
number
number
[0060]
number
[0061] Furthermore, the goal of optimizing bond cleavage prediction is to i,j and real y i,j The goal is to minimize all likelihoods ∈{0,1}, and calculate the negative logarithm of a likelihood using a binary cross-entropy loss function, which is:
number
[0062] join b i,j is the corresponding adjacent element a i,j exists when it is not zero. For ground truth, y i,j It should be understood that if = 1, it means that the chemical bond between the i-th atom and the j-th atom has been broken.
[0063] Optionally, the model training step can use a multistep method, but can be optimized to use a cosine annealing training method to increase the learning rate, i.e., the learning rate can be changed cyclically.
[0064] Specifically, cosine annealing can reduce the learning rate by a cosine function. In the cosine function, as x increases, the cosine value decreases slowly, then decreases at an accelerated rate, and then decreases slowly again. This change mode, combined with the learning rate, can provide a very effective calculation method with good results. The specific calculation formula is as follows:
number
[0065] Through the above calculation process, we can break out of the current local optimum and find a new local optimum. After each cycle of calculation is completed, the model parameters of different local optimum points are saved and the learning rate is smoothly reduced.
[0066] The above embodiment describes the use and training process of a graph neural network. Hereinafter, the use and training process of a reactant production network will be described.
[0067] Specifically, after predicting the bond cleavage sites, the target molecule can be split to obtain synthons. The split synthon subgraph can be converted from the RDKit to a SMILES representation. Note that a synthon is not necessarily a real reactant, but can be a substructure of a reactant. Then, a neural network for reactant prediction based on Transformer (reactant generation network) can be constructed.
[0068] Specifically, the application process of the reactant generation network is as follows: first, a splitting process is performed on the target molecule according to the bond cleavage position determined by the graph neural network to obtain at least one synthon molecular graph; then, the synthon molecular graph is converted into a second string; and the second string is updated according to a predetermined reaction type to obtain a third string, where the predetermined reaction type may be any one of the reaction types shown in Table 1; and then, at least one synthon is determined according to the third string by the reactant generation network.
[0069] In a possible scenario, FIG. 9 is a schematic diagram of the string conversion provided in the embodiment of the present application, where each reaction type can be represented as one string RXN, for example, the k-th reaction type can be represented as RXN_k, each product molecule can be represented as one SMILES string and denoted as Product, each synthon can also be represented as one SMILES string and denoted as Synthon, and the molecule corresponding to each synthon can also be represented as one SMILES string and denoted as Reactant. Here, the source sequence data information is obtained by concatenating the reaction type information (if any), the standard product SMILES string, and the corresponding synthon. The target sequence corresponds to the SMILES information of the reactant for each synthon. Furthermore, an attention mechanism can be used to allow both the product and the synthon to be observed, thereby enhancing the interpretability of the algorithm.
[0070] In addition, in the training process of the reactant generation network, the association relationship between the training molecule, the training synthon and the training reactant needs to be expressed in the character dimension. Specifically, firstly, a second training molecule, a second training synthon and a training reactant are obtained, and then a first training string is determined according to the string corresponding to the second training molecule, the string corresponding to the second training synthon and a predetermined reaction type, and a second training string corresponding to the training reactant is determined, and then the first training string and the second training string are associated with each other to determine a first training sample pair, and then the reactant generation network can be trained according to the first training sample pair.
[0071] In a possible scenario, assuming there are two second training synthons and the SMILES strings of the two training synthons are Synthon1 and Synthon2, respectively, a long string (first training string) can be obtained by concatenating the strings corresponding to the reaction type, product molecule, and synthon, as follows: U=<RXN_i> Product <link> Sython1.Synthon2 Furthermore, the corresponding correct target molecule strings Reactant1 and Reactant2 (assuming here that there are only two reactants) are concatenated into a long target string (a second training string). V = Reactant1.Reactant2 Then, construct a training sample (first training sample pair) from U and V: (U,V)
[0072] Optionally, in order to improve the robustness of the reactant generation network, the SMILES string of the synthon predicted in the first step (bond breaking division) can be used as training data, and the prediction result may be inaccurate. Specifically, first obtain the candidate string predicted by the graph neural network (prediction result), then add the candidate string to the string corresponding to the second training synthon to update the first training string to a third training string, and then associate the third training string with the second training string to determine a second training sample pair, and finally train the reactant generation network based on the second training sample pair.
[0073] In a possible scenario, suppose the first staircase model predicted three synthons (candidate strings).
number
number
number
[0074] Specifically, in the training process, the training objective is to minimize the difference between the predicted string and the real reactant string, and the specific formula is as follows:
number
[0075] Optionally, in order to reduce the difficulty of learning the sequence model, the edit distance between the learning objective, i.e., the expression of the SMILES of the reactants, and the synthon can be minimized, i.e., a canonical process can be performed on the SMILES of each synthon. Specifically, a target character format is first determined, and processed using, for example, an open source chemical information tool RDKit, and then the string corresponding to the second training synthon and the first training string determined based on the predetermined reaction type are updated based on the target character format, thereby reducing the distance between the string corresponding to the second training synthon and the first training string determined based on the predetermined reaction type, thereby reducing the difficulty of learning the sequence model.
[0076] At 303, a terminal node that satisfies a predetermined condition is determined by performing a recursive process on the predicted molecule set corresponding to the second leaf node.
[0077] In this embodiment, the recursive process is a method of solving a problem by decomposing the problem into similar sub-problems by repeating an operation, i.e., for a compound in the predicted molecule set corresponding to the second leaf node, the process of obtaining synthons by division in the above step 302 is repeatedly executed until a predetermined condition is satisfied.
[0078] Specifically, the recursive process is as follows: first, a first candidate molecule corresponding to the predicted molecule set corresponding to the second leaf node under a predetermined reaction type is determined; then, a third leaf node is obtained by performing an unfolding process on the first candidate molecule based on a target retrosynthesis model; a second candidate molecule is determined by performing a recursive process on the predicted molecule set corresponding to the third leaf node; if the second candidate molecule satisfies a predetermined condition, the leaf node corresponding to the second candidate molecule is determined as a terminal node, thus realizing multi-step retrosynthesis prediction.
[0079] Specifically, for the setting of the predetermined condition, the synthon after the above-mentioned recursive processing may be a molecule in the basic molecule set; that is, first, a plurality of pathway molecules are obtained by traversing based on the second candidate molecule; if the pathway molecule is a molecule in the basic molecule set, it is determined that the second candidate molecule satisfies the predetermined condition, and the leaf node corresponding to the second candidate molecule is determined to be a terminal node, thus ensuring that all compound molecules in the predicted and obtained synthetic pathway are simple basic molecules, and ensuring the feasibility of this proposal in a practical scenario.
[0080] In addition, the selection of the base molecule set should consider the rationality of the base molecule set for the predicted results of the multi-step retrosynthesis, where rationality means that in the actual synthesis, the chemist will consider the price and commercial availability of the intermediates at each stage of the synthetic route and decide whether to continue the retrosynthesis according to these factors. Therefore, in the selection of the base molecule set, it is necessary to ensure as much as possible that the molecules are all commercially available chemical intermediates, not drug molecules used for virtual screening of drugs. The difference between the two is as follows: chemical intermediates generally have large packaging specifications (g units), while drug molecules for screening generally have small packaging specifications such as milligram units and are relatively expensive. Based on this principle, it is possible to select building block libraries from some compound reagent companies with public reagent catalogs, and other companies with other molecular libraries. Building block libraries have cost advantages in actual use scenarios because their molecules are small and their structures are simple.
[0081] It should be understood that the selection of the base molecule set mainly takes into consideration the commercial availability of the compounds and whether they are known, and therefore the common compound raw material supplier eMolecules (Plus library and SC library, more than 20 million), and the common compound library companies Enamine (97,000) and ChemDiv (70,000) that provide building blocks may be selected. After merging and deduplication of several libraries, the finally determined base molecule set comprises about 23 million molecules.
[0082] Optionally, the predetermined condition may be the number of recursive processes, i.e., the number of expansions of the second candidate molecule. Specifically, first, in the tree structure, the number of expansions corresponding to the second candidate molecule is determined, and when the number of expansions reaches a predetermined value, it is determined that the second candidate molecule satisfies the predetermined condition, and the leaf node corresponding to the second candidate molecule is determined as the terminal node. For example, if the number of expansions of a given second candidate molecule is 2, the second candidate molecule may be divided to obtain a third candidate molecule, and then the third candidate molecule may be further divided to obtain a fourth candidate molecule, and the fourth candidate molecule may be determined as the terminal node.
[0083] At 304, the pathway information corresponding to the terminal nodes is traversed to determine a retrosynthetic pathway for the target molecule.
[0084] In this embodiment, the pathway information corresponding to the terminal nodes is traversed, i.e., a reverse search is performed based on the direction from the terminal node to the root node, and the related compounds in the pathway are associated to determine a retrosynthetic pathway for the target molecule.
[0085] In a possible scenario, the above multi-step retrosynthesis prediction method has a prediction accuracy rate of 64% for the top-10 of the disclosed ChEMBL 100 molecule test set, i.e., it can predict 64 out of 100 molecules, and the Top-1 prediction accuracy rate is 43%, which is in the top position compared with the expression effect of some of the disclosed platforms.
[0086] As can be seen from the above embodiment, a target molecule is obtained, the target molecule is the root node of a tree structure, the root node is associated with a first leaf node of the tree structure, the tree structure includes a retrosynthetic pathway of the target molecule, and then a target retrosynthetic model performs an expansion process on the first leaf node to obtain a plurality of second leaf nodes, the target retrosynthetic model includes a graph neural network and a reactant generation network, the graph neural network predicts the bond cleavage position of the compound molecule corresponding to the first leaf node, the reactant generation network determines a predicted molecule set based on the bond cleavage position, and further performs a recursive process on the predicted molecule set corresponding to the second leaf node to determine a terminal node that satisfies a predetermined condition, and further traverses the path information corresponding to the terminal node to determine the retrosynthetic pathway of the target molecule. In this way, a multi-step reaction retrosynthetic prediction process is realized, and the leaf nodes are gradually expanded and screened recursively, and path tracing is performed on the terminal nodes that meet the predetermined conditions, thereby ensuring the reliability of the reactants obtained in the multi-step reaction retrosynthetic prediction process and improving the accuracy of compound molecule retrosynthesis prediction.
[0087] In the above embodiment, the retrosynthesis prediction process of a multi-step reaction is described. In the development process, different rules (e.g., the generated result molecule is a molecule that exists in nature, the generated result molecule is an organic substance, the number of rings of the generated result molecule does not increase, etc.) and models (e.g., a retrosynthesis probability model, a forward synthesis prediction model, etc.) are used to screen the nodes, and the rules and models are specifically set, which may be automatically called or may be set by the relevant person as necessary. The following describes this scenario. Referring to FIG. 10, which is a flowchart of a retrosynthesis prediction method for another compound molecule provided in the embodiment of the present application, the embodiment of the present application includes at least the following steps: In 1001, a target molecule is obtained and the target molecule is made the root node in a tree structure. In 1002, a target reverse synthesis model is used to perform an unfolding process on the first leaf node to obtain a plurality of second leaf nodes.
[0088] In this embodiment, the process of implementing steps 1001-1002 is similar to the process of implementing steps 301-302 in the embodiment of FIG. 3, so for an explanation of the related features, refer to the explanation of FIG. 3 and no detailed explanation will be given here.
[0089] At 1003, reaction screening requirements are obtained in response to the target operation.
[0090] In this embodiment, the reaction screening requirements are to screen the nodes that will subsequently participate in the rollout, for example, screening compounds with competing reaction risks, selecting compounds with protecting groups, screening compounds based on the distribution characteristics of C-H bonds, selecting compounds related to named reactions, etc., and the specific reaction screening requirements are determined according to the actual scenario.
[0091] The target operation may be an input operation triggered by a person in the background, or may be an operation of setting a screening rule between nodes of a tree structure. It should be understood that the specific operation method is determined according to the actual scenario.
[0092] Optionally, the above reaction screening requirements can also be applied to screening in the expansion process. For example, the expansion process first traverses 10 reaction types and predicts 3 outcomes for each reaction type, up to a total of 30. Then, when filtering the 30 results, the above reaction screening requirements are used to screen and retain some with high likelihood.
[0093] At 1004, the compound molecules corresponding to the second leaf nodes are screened based on reaction screening requirements.
[0094] In this example, the above reaction screening requirements will be explained based on a specific reaction. (1) Operation of competitive reactions. In this embodiment, the reaction pathway is rationally designed by the operation of competitive reactions to avoid functional group interference. Figure 11 is a schematic diagram of the retrosynthesis prediction process of the compound molecule provided in the embodiment of this application. The key step of the synthetic pathway in the figure is to convert C=O to C=S double bond. When predicting the pathway with compound 3, the thio reaction will cause multiple competitive reactions of C=O double bonds, namely, the competitive reactions of aldehyde carbonyl and amide carbonyl, and therefore it is necessary to remove the nodes containing the corresponding reactants. This allows the given pathway to avoid predictions involving competitive reactions, i.e., the thio reaction will occur when compound 5 is selected, and then the aldehyde group is introduced to obtain compound 3, improving the accuracy of retrosynthesis prediction. (2) Use of protecting groups. The rational use of protecting groups has been a challenge in computer-aided retrosynthetic prediction, because the computer must analyze groups in the reaction substrate that may cause competing reactions, rationally protect groups that are not desired to participate in the reaction, and then remove the protecting groups.
[0095] In a possible scenario, FIG. 12 is a schematic diagram of the retrosynthetic prediction process of the compound molecule provided in the examples of the present application, for the target synthetic compound 9, the free NH2 group will undergo a competitive reaction, so compound 10 obtained based on NHBoc is protected, and then compound 10 can be easily obtained from compounds 11 and 12; otherwise, when the free NH2 is in 12, an intramolecular reaction will occur, which will affect the reaction result.
[0096] In another possible scenario, hydroxyl groups often need to be protected during oxidation or in certain reactions under alkaline conditions. Specific protection methods include: (1) etherification reactions can be used to prevent hydroxyl groups from being affected by bases, e.g., ROH +ROH→H2O+ROR. (2) esterification reactions can be used to prevent hydroxyl groups from being oxidized.
[0097] Also, it may be necessary to protect the carboxyl group under high temperature or alkaline conditions. The most commonly used method for protecting the carboxyl group is esterification reaction. Or, unsaturated carbon-carbon bonds are protected, and the protection is mainly conjugated by addition reaction, since carbon-carbon double bonds are easily oxidized. Or, for the protection of carbonyl groups (often appearing in the form of information problems), in particular, aldehyde groups need to be protected during oxidation reaction or in the presence of base, and the protection of carbonyl groups generally uses a method of generating acetal or ketal, and the generated acetal or ketal is hydrolyzed to the original aldehyde or ketone. The determination of the protective group is specifically determined by the actual scenario, and is not limited here.
[0098] (3) Use of regioselectivity. In this embodiment, the regioselectivity problem of CH substitution reaction on aromatic ring is solved. Figure 13 is a schematic diagram of the retrosynthetic prediction process of compound molecules provided in the embodiment of this application. When predicting molecule 15, select NBS (16) and the bromination reaction of compound 17. The CH (pyridine ring and furan ring) of multiple aromatic rings of compound 17 can be brominated by compound 16. Therefore, the rational use of features such as the distribution of partial charges of each CH is the key to accurately predicting regioselectivity, and therefore, in the process of setting reaction screening requirements, specific CH distribution rules can be defined to select the corresponding reactants.
[0099] (4) Use of classical name responses. In this embodiment, the organic named reaction is the most reliable chemical reaction determined by the research and use of many organic chemists, so the substrate universality is also high, and the reaction path is relatively easy to realize. Therefore, in retrosynthesis analysis, the use of classical named reactions is a relatively reliable path, which is relatively easy to be patronized by users, and can predict chemical reactions with rationality and robustness. Figure 14 is a schematic diagram of the retrosynthesis prediction process of compound molecules provided in the embodiment of this application, in which the Mitsunobu reaction is shown, that is, in the reactant screening process, compounds 21 and 22 are preferentially selected.
[0100] Specifically, the operation of classical name responses includes, but is not limited to, the following examples: Beckman Beckmann rearrangement: reaction of ketoximes to amides under acidic conditions (caprolactam) Cannizzarro disproportionation: A reaction in which an α-hydroxylated aldehyde is converted to an alcohol and a carboxylic acid (benzaldehyde) in the presence of a strong base. Claisen condensation: A reaction in which an ester forms a carbanion in the presence of a strong base and undergoes nucleophilic addition or elimination to another ester. Clemmensen Clemmensen reduction: A reaction in which aldehydes and ketones are reduced to hydrocarbons (carbonyl groups to methylene groups) in the presence of zinc amalgam and concentrated hydrochloric acid. Cope elimination reaction: elimination reaction of tertiary amines treated with hydrogen peroxide and then heated (Hoffmann rule) Corey-House Corey-House-Posner-Whitesides reaction: Coupling of halogenated hydrocarbons with dialkylcopper lithium reagents (an important reaction for linking carbon chains) cram Cram's rule: nucleophiles attack preferentially from the less hindered side Dickerman Dieckmann condensation: the reaction resembles ester condensation and ring formation Diels-Alder reaction: Generally, the reaction of a derivative of 1,3-butadiene with a derivative of ethylene (synergistic reaction) Fehling's solution: freshly prepared copper hydroxide oxidizes aldehydes to acids. Friedel-Crafts reaction: A reaction in which a hydrocarbon group or an acyl group is introduced into a benzene ring.
[0101] It should be understood that in a practical scenario, one or more of the above reaction screening requirements can be used, and the specific number and order of determination will be determined by the practical scenario.
[0102] In 1005, a terminal node is determined when a predetermined condition is satisfied by performing a recursive process on the predicted molecule set corresponding to the second leaf node after screening.
[0103] Specifically, in the recursive process, for each simulated object, 10 reaction types are traversed and one result is predicted for each reaction type, for a total of 10. Then, the 10 results are filtered, and the above reaction screening requirements are used to combine the reverse synthesis probability, the forward synthesis prediction probability, etc. to select the most likely reaction for the next expansion. Then, the above process is repeated until the stopping condition is met. The stopping condition is that the compound molecules corresponding to all leaf nodes are in the basic molecule set, or a predetermined number of recursions is reached.
[0104] At 1006, the pathway information corresponding to the terminal nodes is traversed to determine a retrosynthetic pathway for the target molecule.
[0105] In this embodiment, the process of implementing steps 1005 and 1006 is similar to the process of implementing steps 303 and 304 in the embodiment of FIG. 3, so reference may be made to the description of related features, and no detailed description will be given here.
[0106] As can be understood, the above reaction screening requirement settings may also be applied to a single-step retrosynthesis prediction scenario, i.e., the present application also discloses a single-step retrosynthesis prediction method, specifically refer to step 302 in the embodiment of FIG. 3, which will not be described here in detail.
[0107] For the single-step retrosynthesis prediction process, the prediction accuracy rate reaches 62.4% even for data without reaction type labels, which makes the method more universal and practical, because in actual use scenarios, such reaction type labels may not be available. To some extent, it is verified that large-scale data sets contribute to improving the accuracy of single-step prediction. Specifically, the prediction accuracy possible with a 50,000 data set is 58%, and when the data set is increased to 480,000, the accuracy increases to 62.4%.
[0108] In this embodiment, the prediction accuracy of multi-step retrosynthesis is currently in the leading position compared with the expression efficiency of some of the disclosed platforms. In the disclosed ChEMBL 100 molecule test set, the top-10 prediction accuracy is 64%, that is, 64 out of 100 molecules can be predicted, and the top-1 prediction accuracy is 43%, which has excellent feasibility.
[0109] The embodiment of the present application further provides a related device for implementing the above solution. Referring to Figure 15, Figure 15 is a structural schematic diagram of a compound molecule retrosynthesis prediction device provided in the embodiment of the present application, the prediction device 1500 includes: an acquisition unit 1501 for acquiring a target molecule and making the target molecule a root node of a tree structure, the root node being associated with a first leaf node of the tree structure, the tree structure including a retrosynthetic pathway of the target molecule; an expansion unit 1502 for performing an expansion process on the first leaf node by a target retrosynthesis model to obtain a plurality of second leaf nodes, the target retrosynthesis model including a graph neural network and a reactant generation network, the graph neural network predicting a bond cleavage position of a compound molecule corresponding to the first leaf node, and the reactant generation network determining a predicted molecule set based on the bond cleavage position; a processing unit 1503 for determining a terminal node that satisfies a predetermined condition by performing a recursive process on the predicted molecule set corresponding to the second leaf node; and a prediction unit 1504 for traversing the pathway information corresponding to the terminal nodes to determine a retrosynthetic pathway of the target molecule.
[0110] Preferably, in some possible implementations of the present application, the unfolding unit 1502 specifically includes: determining a first string of a compound molecule corresponding to the first leaf node; converting the first string into a first molecular graph; determining, by the graph neural network, a bond cleavage position of a compound molecule corresponding to the first leaf node based on the first molecular graph; determining at least one synthon based on the bond cleavage location by the reactant generation network; The predicted molecule set is determined by filtering the at least one synthon based on a predetermined rule.
[0111] Preferably, in some possible implementations of the present application, the unfolding unit 1502 specifically includes: determining bond break information based on the first molecular graph and a limit on the number of bond breaks by the graph neural network; and determining target key features based on the target molecule, the target key features including atom key features and bond key features, the atom key features including at least one of atom type, bond number, formal charge, chirality, hydrogen atom number, atom hybridization state, aromaticity, atomic weight, high frequency reaction center feature, and reaction type, and the bond key features including at least one of bond type, conjugation, ring bond, and molecular stereochemistry feature; and extracting the bond cleavage information based on the target key features to determine the bond cleavage locations.
[0112] Preferably, in some possible implementations of the present application, the unfolding unit 1502 specifically includes: obtaining a first training molecule and a corresponding first training synthon; determining node features and edge features of the first training molecule and the first training synthon, the node features indicating atomic relationships between the first training molecule and the first training synthon, and the edge features indicating chemical bond relationships between the first training molecule and the first training synthon; determining a first loss function based on the node features and the edge features; determining a bond cleavage probability of a chemical bond between the first training molecule and the first training synthon; determining a second loss function based on the bond breaking probabilities; and updating model parameters of the graph neural network based on the first loss function and the second loss function.
[0113] Preferably, in some possible implementations of the present application, the unfolding unit 1502 specifically includes: performing a division process on the target molecule based on the bond cleavage positions to obtain at least one synthon molecule graph; converting said synthon molecular graph into a second string; updating the second string based on a predetermined reaction type to obtain a third string; and determining, by a reactant generation network, the at least one synthon based on the third string.
[0114] Preferably, in some possible implementations of the present application, the unfolding unit 1502 specifically includes: obtaining a second training molecule, a second training synthon, and a training reactant; determining a first training string based on a string corresponding to the second training molecule, a string corresponding to the second training synthon, and the predetermined reaction type; determining a second training string corresponding to the training reactant; determining a first training sample pair by associating the first training string with the second training string; used to train the reactant generation network based on the first training sample pair.
[0115] Preferably, in some possible implementations of the present application, the unfolding unit 1502 specifically includes: Obtaining candidate character strings predicted by the graph neural network; adding the candidate string to a string corresponding to the second training synthon to update the first training string to a third training string; determining a second training sample pair by associating the third training string with the second training string; is used to train the reactant production network based on the second training sample pair.
[0116] Preferably, in some possible implementations of the present application, the unfolding unit 1502 specifically includes: Determine the target character format; and updating the first training string based on the target character format.
[0117] Preferably, in some possible implementations of the present application, the processing unit 1503 specifically comprises: determining a first candidate molecule corresponding to the set of predicted molecules corresponding to the second leaf node under a given reaction type; performing an unfolding process on the first candidate molecule using the target retrosynthesis model to obtain a third leaf node; determining a second candidate molecule by performing a recursive process on the set of predicted molecules corresponding to the third leaf node; If the second candidate molecule satisfies a predetermined condition, the leaf node corresponding to the second candidate molecule is used to determine that the leaf node is the terminal node.
[0118] Preferably, in some possible implementations of the present application, the processing unit 1503 specifically comprises: traversing based on the second candidate molecule to obtain a plurality of pathway molecules; If the pathway molecule is a molecule in a base molecule set, then the second candidate molecule is determined to satisfy the predetermined condition, and the leaf node corresponding to the second candidate molecule is determined to be the terminal node.
[0119] Preferably, in some possible implementations of the present application, the processing unit 1503 specifically comprises: determining a number of expansions in the tree that correspond to the second candidate molecule; If the number of expansions reaches a predetermined value, the second candidate molecule is determined to satisfy the predetermined condition, and the leaf node corresponding to the second candidate molecule is determined to be the terminal node.
[0120] According to the above-mentioned device, a retrosynthetic prediction process for a multi-step reaction is realized, and leaf nodes are gradually recursively expanded and screened, and path tracing is performed for terminal nodes that satisfy specified conditions, thereby ensuring the reliability of reactants determined by the retrosynthetic prediction process for a multi-step reaction and improving the accuracy of retrosynthetic prediction of compound molecules.
[0121] The embodiments of the present application further provide a terminal device, and Figure 16 is a structural schematic diagram of another terminal device provided in the embodiments of the present application, for ease of explanation, only the parts related to the embodiments of the present application are shown, and for the specific technical details not disclosed, please refer to the embodiment method part of the present application. The terminal may be any terminal device, including a mobile phone, a tablet, a personal digital assistant (PDA), a point of sales (POS), a car-mounted computer, etc., and the terminal is taken as an example of a computer.
[0122] Figure 16 shows a block diagram of a part of the structure of a computer related to a terminal provided in an embodiment of the present application. Referring to Figure 16, the computer includes components such as a radio frequency (RF) circuit 1610, a memory 1620, an input unit 1630 (including a touch panel 1631 and other input devices 1632), a display unit 1640 (including a display panel 1641), a sensor 1650, an audio circuit 1660 (connected to a speaker 1661 and a microphone 1662), a wireless fidelity (WiFi) module 1670, a processor 1680, and a power supply 1690. As can be understood by those skilled in the art, the computer structure of Figure 16 does not limit the computer, and may include more or fewer components than those shown, or may combine some components, or may use a different component arrangement.
[0123] The mobile phone further includes a power source 1690 (e.g., a battery) for powering each component, and preferably, the power source is logically connected to the processor 1680 via a power management system, thereby enabling functions such as charging, discharging management, and power consumption management to be realized via the power management system.
[0124] Although not shown, the mobile phone may include a camera, a Bluetooth module, etc., which are not mentioned here in detail.
[0125] In an embodiment of the present application, the processor 1680 included in the terminal is used to execute the functions of each step of the above-mentioned method for predicting the retrosynthesis of compound molecules.
[0126] An embodiment of the present application further provides a server, and reference is made to FIG. 17, which is a structural schematic diagram of a server provided in an embodiment of the present application, in which the server 1700 can vary widely in configuration or performance, and includes one or more central processing units (CPUs) 1722 (e.g., one or more processors) and a memory 1732, one or more storage media 1730 (e.g., one or more mass storage devices) for storing application programs 1742 or data 1744. The memory 1732 and the storage media 1730 may be temporary or permanent. The program stored in the storage media 1730 includes one or more modules (not shown), each of which includes a series of command operations in the server. Furthermore, the central processing unit 1722 is configured to communicate with the storage media 1730 and execute the series of command operations in the storage media 1730 in the server 1700.
[0127] The server 1700 may include one or more power sources 1726, one or more wired or wireless network interfaces 1750, one or more input / output interfaces 1758, and / or one or more operating systems 1741, such as Windows Server 2003 (Windows Server 2003 R2), such as Windows 8.1 (Windows Server 2003 R2), or any combination thereof. TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM It may further include the following.
[0128] Specifically, the central processing unit 1722 of the server 1700 is specifically used to execute the functions of each step of the above-mentioned method for predicting the retrosynthesis of compound molecules.
[0129] An embodiment of the present application further provides a computer-readable storage medium, which stores a compound molecule retrosynthesis prediction command, and when executed on a computer, causes the computer to execute the steps executed by the compound molecule retrosynthesis prediction device in the method described in the embodiment of Figures 3 to 14 above.
[0130] An embodiment of the present application further provides a computer program product including a compound molecule retrosynthesis prediction command, which, when executed on a computer, causes the computer to execute the steps executed by the compound molecule retrosynthesis prediction device in the method described in the embodiment of Figures 3 to 14 above.
[0131] An embodiment of the present application further provides a compound molecule retrosynthetic prediction system, which may include a compound molecule retrosynthetic prediction device in the embodiment described in FIG. 15, or a terminal device in the embodiment described in FIG. 16, or a server described in FIG. 17.
[0132] As can be clearly understood by those skilled in the art, for the sake of convenient and concise description, the specific operation processes of the above-described systems, devices and units can be referred to the corresponding processes in the above method embodiments, and will not be described in detail here.
[0133] As described above, the above embodiments are only used to explain the present invention, and are not limited thereto. Although the present application has been described in detail with reference to the above embodiments, it should be understood that a person skilled in the art may still modify the technical solutions described in the above embodiments or equally replace some technical features therein, and such modifications or replacements do not deviate from the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application. [Explanation of symbols]
[0134] 1500 Prediction Device 1501 Acquired Units 1502 Deployment Unit 1503 Processing Unit 1504 prediction units 1610 Radio Frequency Circuits 1620 Memory 1630 Input Unit 1631 Touch Panel 1632 Input Device 1640 display unit 1641 Display Panel 1650 Sensor 1660 Audio Circuit 1661 Speaker 1662 Microphone 1670 WiFi Module 1680 processor 1690 Power supply 1700 Server 1722 Central Processing Unit 1726 Power supply 1730 Storage medium 1732 Memory 1741 Operating Systems 1742 Application Program 1744 Data 1750 Wireless Network Interface 1758 Input / Output Interface
Claims
1. A method for predicting the retrosynthesis of a compound molecule executed by a computer device, comprising: obtaining a target molecule and setting the target molecule as a root node of a tree structure, the root node being associated with a first leaf node of the tree structure, the tree structure including a retrosynthetic pathway of the target molecule; a step of acquiring a plurality of second leaf nodes by performing an expansion process on the first leaf node using a target retrosynthesis model, the target retrosynthesis model including a graph neural network and a reactant generation network, the graph neural network predicting a bond cleavage position of a compound molecule corresponding to the first leaf node, and the reactant generation network determining a predicted molecule set based on the bond cleavage position; determining a terminal node that satisfies a predetermined condition by performing a recursive process on the predicted molecule set corresponding to the second leaf node; traversing pathway information corresponding to the terminal nodes to determine a retrosynthetic pathway for the target molecule; The graph neural network comprises: obtaining a first training molecule and a corresponding first training synthon; determining node features and edge features of the first training molecule and the first training synthon, the node features indicating atomic relationships between the first training molecule and the first training synthon, and the edge features indicating chemical bond relationships between the first training molecule and the first training synthon; determining a first loss function based on the node features and the edge features; determining a bond cleavage probability of a chemical bond between the first training molecule and the first training synthon; determining a second loss function based on the bond breaking probabilities; updating a model parameter of the graph neural network based on the first loss function and the second loss function; To be trained method.
2. The step of acquiring a plurality of second leaf nodes by performing an expansion process on the first leaf node using the target reverse synthesis model includes: determining a first string of compound molecules corresponding to the first leaf node; converting the first string into a first molecular graph; determining, by the graph neural network, a bond cleavage position of a compound molecule corresponding to the first leaf node based on the first molecular graph; determining, by the reactant generation network, at least one synthon based on the bond cleavage location; and determining the predicted molecule set by filtering the at least one synthon based on a predetermined rule. The method of claim 1.
3. The step of determining, by the graph neural network, a bond cleavage position of a compound molecule corresponding to the first leaf node based on the first molecular graph, includes: determining bond break information based on the first molecular graph and a limit on the number of bond breaks by the graph neural network, and determining target key features based on the target molecule, the target key features including atom key features and bond key features, the atom key features including at least one of atom type, bond number, formal charge, chirality, hydrogen atom number, atom hybridization state, aromaticity, atomic weight, high frequency reaction center feature, and reaction type, and the bond key features including at least one of bond type, conjugation, ring bond, and molecular stereochemistry feature; and extracting the bond break information based on the target key features to determine the bond break locations. The method of claim 2.
4. determining, by the reactant generation network, at least one synthon based on the bond cleavage location, performing a segmentation process on the target molecule based on the bond cleavage positions to obtain at least one synthon molecule graph; converting the synthon molecular graph into a second string; updating the second string based on a predetermined reaction type to obtain a third string; determining, by the reactant generating network, the at least one synthon based on the third string. The method of claim 2.
5. The reactant production network comprises: obtaining a second training molecule, a second training synthon, and a training reactant; determining a first training string based on the strings corresponding to the second training molecules, the strings corresponding to the second training synthons, and the predetermined reaction type, and determining a second training string corresponding to the training reactants; determining a first training sample pair by associating the first training string with the second training string; training the reactant generation network based on the first training sample pair; To be trained The method according to claim 4.
6. obtaining candidate strings predicted by the graph neural network; adding the candidate string to a string corresponding to the second training synthon to update the first training string to a third training string; determining a second training sample pair by associating the third training string with the second training string; training the reactant generation network based on the second training sample pair. The method according to claim 5.
7. determining a target character format; and updating the first training string based on the target character format. The method according to claim 5.
8. The step of determining a terminal node that satisfies a predetermined condition by performing a recursive process on the predicted molecule sets corresponding to the second leaf nodes includes: determining a first candidate molecule corresponding to the set of predicted molecules corresponding to the second leaf node under a given reaction type; performing an unfolding process on the first candidate molecule using the target retrosynthesis model to obtain a third leaf node; determining a second candidate molecule by performing a recursive process on the set of predicted molecules corresponding to the third leaf node; and if the second candidate molecule satisfies the predetermined condition, determining that the leaf node corresponding to the second candidate molecule is the terminal node. The method of claim 1.
9. determining that a leaf node corresponding to the second candidate molecule is the terminal node if the second candidate molecule satisfies the predetermined condition, traversing based on the second candidate molecule to obtain a plurality of pathway molecules; if the pathway molecule is a molecule in a base molecule set, determining that the second candidate molecule satisfies the predetermined condition, and determining that a leaf node corresponding to the second candidate molecule is the terminal node. The method according to claim 8.
10. determining that a leaf node corresponding to the second candidate molecule is the terminal node if the second candidate molecule satisfies the predetermined condition, determining a number of expansions in the tree structure that correspond to the second candidate molecule; if the number of expansions reaches a predetermined value, determining that the second candidate molecule satisfies the predetermined condition and determining that the leaf node corresponding to the second candidate molecule is the terminal node. The method according to claim 8.
11. The method of claim 1, wherein the target molecule is a drug molecule, the tree structure is a Monte Carlo tree, and the compound molecule corresponding to the second leaf node is obtained by screening based on at least one of competitive reactions of groups, operation of protecting groups, reaction characteristics of characteristic groups, and named reaction rules.
12. A device for predicting the retrosynthesis of a compound molecule, comprising: a obtaining unit for obtaining a target molecule and making the target molecule a root node of a tree structure, the root node being associated with a first leaf node of the tree structure, the tree structure including a retrosynthetic pathway of the target molecule; an expansion unit for performing an expansion process on the first leaf node using a target retrosynthesis model to obtain a plurality of second leaf nodes, the target retrosynthesis model including a graph neural network and a reactant generation network, the graph neural network predicting a bond cleavage position of a compound molecule corresponding to the first leaf node, and the reactant generation network determining a predicted molecule set based on the bond cleavage position; a processing unit for performing a recursive process on the predicted molecule set corresponding to the second leaf node to determine a terminal node that satisfies a predetermined condition; a prediction unit for traversing pathway information corresponding to the terminal nodes to determine a retrosynthetic pathway of the target molecule; The graph neural network comprises: obtaining a first training molecule and a corresponding first training synthon; determining node features and edge features of the first training molecule and the first training synthon, the node features indicating atomic relationships between the first training molecule and the first training synthon, and the edge features indicating chemical bond relationships between the first training molecule and the first training synthon; determining a first loss function based on the node features and the edge features; determining a bond cleavage probability of a chemical bond between the first training molecule and the first training synthon; determining a second loss function based on the bond breaking probabilities; updating a model parameter of the graph neural network based on the first loss function and the second loss function; To be trained Device.
13. a memory for storing program code; A processor for executing the method for predicting the retrosynthesis of a compound molecule according to any one of claims 1 to 11 in accordance with the commands of the program code. Computer equipment.
14. A computer program that, when executed on a computer, causes the computer to execute the method for predicting the retrosynthesis of a compound molecule according to any one of claims 1 to 11.
