Efficient system for multi-step inverse synthesis planning of chemical organic molecules

Abstract

The invention provides an efficient system for multi-step inverse synthesis planning of chemical organic molecules, and the system comprises a one-step inverse synthesis prediction module which takes to-be-synthesized molecules as input, and generates chemical reactions and reactants; the molecular synthesis cost estimation module is used for estimating the input molecular synthesis cost; the search module takes a multi-step inverse synthesis planning problem as a search problem in the one-step inverse synthesis prediction module, combines exploration attributes of Monte Carlo tree search with A * search, and performs inverse synthesis planning; and the training module is used for taking the path consistency as a training constraint of the molecular synthesis cost estimation module and improving the generalization ability of the model. According to the method, Monte Carlo tree search is used as pre-search of A * search through the search module, exploratory performance is introduced for the A * search, so that the search process can effectively escape from a local optimal solution, extra consumption caused by deviation of a guide function is reduced, and the success rate of finding an inverse synthesis path by a molecular inverse synthesis planning task within specified time is increased.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical organic molecule retrosynthetic planning, and in particular, to an efficient system for performing multi-step retrosynthetic planning of chemical organic molecules. Background Art

[0002] For a given target molecule, chemical organic molecule retrosynthetic planning aims to find a feasible and economically viable synthetic route from a vast chemical reaction space, using known or commercially available basic molecular combinations. This is a fundamental problem in synthetic chemistry and plays a crucial role in a wide range of applications such as drug design and materials science. For example, the expenditure required to bring a drug to market exceeds $2 billion to $3 billion, mainly for the early discovery stage and clinical trials. In recent years, using data-driven retrosynthetic planning tools to accelerate and reduce failures in new drug synthesis has attracted widespread attention.

[0003] Considering that the synthesis of most molecules usually requires multiple chemical reactions and the number of possible chemical reactions is huge, even for experienced chemists, retrosynthetic planning is a challenging task.

[0004] The difficulty of synthesizing a molecule is greatly affected by its structural complexity and the available starting molecules. In some cases, finding a feasible synthetic path may require intensive work by human experts for several weeks. Computer-aided retrosynthetic planning can help chemists find higher-quality synthetic paths, usually achieved by a search guided by a one-step retrosynthetic prediction model.

[0005] With the development of artificial intelligence technology, machine learning-based methods have shown extraordinary performance and can automatically learn to generate recommendations. These methods usually regard one-step retrosynthetic prediction as a classification task based on reaction templates, or use a sequence-to-sequence translation model in natural language processing to regard the one-step retrosynthetic prediction task as a translation problem between products and reactants. A one-step retrosynthetic prediction model can predict the most likely chemical reaction that can directly synthesize the target molecule, which can be extended to a complete route design by searching.

[0006] Since there may be a large number of possible precursors at each step, a one-step retrosynthetic prediction model is used as a guiding search strategy to screen out a few of the most likely chemical reactions to avoid combinatorial explosion. The chemical synthesis path can be transformed into a search tree. The nodes in the search tree represent the current synthesis state, which contains all the molecules required to synthesize the current target molecule. If all the molecules in a state are available starting molecules, then the target molecule can be synthesized. The edges in the search tree correspond to the chemical reactions that guide the state transition between two connected nodes.

[0007] Excellent search can efficiently find reasonable retrosynthetic planning routes from a vast chemical space. Traditional heuristic search methods, such as depth-first search, etc., have poor performance, and the quality of the paths they discover cannot be guaranteed. Inspired by the excellent performance of deep reinforcement learning in solving games, using neural network-guided Monte Carlo tree search (MCTS) for retrosynthetic planning has almost reached the level of literature routes and achieved good results in double-blind AB tests. The traversal tree strategy of MCTS achieves a balance between exploitation and exploration, and the selected leaf nodes evaluate their expected subsequent synthesis costs and propagate them back to all nodes on the traversal path, affecting the subsequent search process. In addition, the search based on A * has achieved better results in molecular retrosynthetic planning. Although the A * search can theoretically guarantee finding the optimal solution, it is still a challenging problem to obtain the optimal solution more efficiently in practice. Compared with MCTS, the search process of the A * search lacks exploratory behavior, and once the estimated value is determined, the estimated value of the synthesis cost will not be updated even if more information is available, limiting the efficiency of the search. Both the A * search and MCTS used in the retrosynthetic planning system rely heavily on the quality of the heuristic guiding function, including the one-step retrosynthetic prediction model and the cost estimation model. Under the guidance of a biased heuristic function, the A * search may spend a lot of time expanding states on non-optimal branches, which will significantly reduce the efficiency of the search process due to the lack of exploratory behavior. The search process of MCTS depends on the balance between exploitation and exploration. If the exploration is insufficient, MCTS may exhibit behavior similar to that of the A * search. If too much emphasis is placed on exploration, a large amount of time will be spent forcibly exploring unnecessary branches. Pruning can be used during the search process to reduce these meaningless explorations, but this may also lead to the exclusion of the optimal solution due to insufficient simulation in the early search stage.

[0008] A more reliable guiding function is of great benefit to the search. From a chemical perspective, the synthesis difficulty of a molecule is related to its structural complexity. The high structural diversity of chemical molecules and the sparsity of training data make it challenging to accurately estimate the synthesis cost of chemical molecules. In addition, the synthesis cost is also affected by the available starting materials. If the precursors are easily available, complex molecules can be rapidly synthesized in a few steps. Due to the different available precursors, the synthesis costs of molecules with similar structures may vary greatly. Therefore, the generalization ability of the cost estimator is crucial for the success of retrosynthetic planning.

[0009] No description or report of similar technologies to the present invention has been found, and no similar materials at home and abroad have been collected. Summary of the Invention

[0010] Aiming at the defects in the prior art, the purpose of the present invention is to provide an efficient system for multi-step retrosynthetic planning of chemical organic molecules.

[0011] According to one aspect of the present invention, there is provided an efficient system for multi-step retrosynthetic planning of chemical organic molecules, including:

[0012] One-step retrosynthetic prediction module: taking the molecule to be synthesized as input, and generating chemical reactions and reactants for synthesizing the input molecule to be synthesized;

[0013] Molecular synthesis cost estimation module: estimating the cost required for synthesizing with raw material molecules for a single input molecule; estimating the overall synthesis cost for multiple input molecules simultaneously;

[0014] Search module: regarding the multi-step retrosynthetic planning problem as a search problem in the one-step retrosynthetic prediction module, combining the exploration attribute of Monte Carlo tree search with A* * search, and performing retrosynthetic planning based on the molecular synthesis cost estimation module;

[0015] Training module: taking path consistency as a constraint for training the molecular synthesis cost estimation module to improve the generalization ability of the model.

[0016] Preferably, the one-step retrosynthetic prediction module is a neural network including two fully connected layers, and performs a classification task based on chemical reaction templates;

[0017] The molecule to be synthesized is used as the input molecule of the one-step retrosynthetic prediction module. According to the chemical reaction template, the probability of being able to synthesize the input molecule is obtained, and the N chemical reaction templates with the highest probability are selected and applied to the input molecule to obtain several optional chemical reactions and reactants for synthesizing the input molecule as the output.

[0018] Preferably, the molecular synthesis cost estimation module estimates the cost required for synthesizing with raw material molecules for a single input molecule, including:

[0019] According to a single input molecule, obtain the synthesis path from the USPTO dataset;

[0020] For each chemical reaction r in the synthesis path i , use the one-step retrosynthetic prediction model to calculate the probability Pr(r i ) of this chemical reaction occurring, and obtain the cost c of this chemical reaction occurring i =-logPr(r i );

[0021] Add up the costs of all chemical reactions in the synthesis path to obtain the synthesis cost of this single input molecule.

[0022] Preferably, the overall synthesis cost estimated by the molecular synthesis cost estimation module for multiple input molecules simultaneously includes:

[0023] Multiple molecules are used as inputs simultaneously;

[0024] Each molecule passes through a fully connected network with a scale of 128 to obtain the feature expression of each molecule;

[0025] The elements at the corresponding positions of the feature expressions of all molecules are summed to obtain the global expression of multiple molecules;

[0026] This global expression is used as the estimation of the synthesis cost output through a neural network layer.

[0027] Preferably, in the search module, the root node of the search tree is the target molecule, the nodes of the search tree are molecule sets, the edges are chemical reactions, the first molecule of the parent node is decomposed into several reactants through this chemical reaction to obtain the molecule set of the child node, the nodes in the search tree that can be further expanded are OPEN nodes, and the nodes that have been expanded are CLOSED nodes.

[0028] Preferably, in the search module, combining the exploration attribute of Monte Carlo tree search with A * search, and performing retrosynthesis planning based on the molecular synthesis cost estimation module, includes:

[0029] Simulation: Starting from the root node, use the exploratory tree traversal strategy of Monte Carlo tree search to traverse to an OPEN node without expanding new nodes, and use it as a candidate node. Repeat this several times to obtain a set of candidate nodes;

[0030] Selection: According to the selection method of A * search, select the node with the lowest synthesis cost in the set of candidate nodes as the next expansion node;

[0031] Expansion: Use the first non - raw material molecule in the molecule set included in the state of the selected node as the input of a one - step retrosynthesis prediction model to obtain several chemical reactions that can be used to synthesize this molecule, generating several child nodes of the expansion node; if all the chemical molecules in a certain child node are existing raw material molecules, the synthesis path is successfully found, otherwise, estimate the synthesis costs of all several child nodes using the synthesis cost estimation module and add all several child nodes to the search tree;

[0032] Loop: Continuously repeat the simulation, selection, and expansion steps. If the running time exceeds the specified time, or there are no nodes in the search tree that can be expanded, the search fails and the synthesis path cannot be found.

[0033] Preferably, the simulation process includes:

[0034] The tree traversal strategy selects the pUCT tree traversal strategy to obtain OPEN nodes. The pUCT formula includes a balance between exploitation and exploration, and the obtained OPEN nodes are exploratory;

[0035] Update the synthesis cost evaluation backup of the current OPEN node to the synthesis costs of all nodes on the path from the root node to the current OPEN node.

[0036] Preferably, the path consistency means that in the same optimal path, the synthesis cost estimates of all node states are the same; using the path consistency as the learning constraint of the molecular synthesis cost estimation module.

[0037] Preferably, the loss function of the molecular synthesis cost estimation module includes two parts:

[0038] One part is the minimum error loss function of the prediction cost,

[0039]

[0040] where f i is the prediction of the molecular synthesis cost estimation template, and z i is the molecular synthesis cost label;

[0041] The other part is the constraint of the path consistency, represented by the squared error between the cost estimate value of any node on the path and the mean of the cost estimates of all nodes on the path,

[0042]

[0043] where is the mean of the cost estimates of all nodes on the synthesis path with length n where f i is located.

[0044] Preferably, the policy function of the one-step retrosynthesis prediction module uses cross-entropy as the loss function,

[0045]

[0046] where y i represents the i-th template as the training label, and p i is the probability that the one-step retrosynthesis prediction module predicts that the i-th template is selected, and N is the number of training samples;

[0047] By minimizing the loss function L = L p + αL c + βL pc, the parameters of the model are updated using Adam gradient descent, where α and β are hyperparameters for adjusting the weights of the loss function.

[0048] Compared with the prior art, the embodiments of the present invention have at least one of the following beneficial effects:

[0049] In the high-efficiency system for multi-step retrosynthetic planning of chemical organic molecules in the embodiments of the present invention, the Monte Carlo tree search is used as a pre-search for A * search, introducing exploration for A * search, enabling the search process to effectively escape from local optimal solutions, reducing the additional consumption caused by the deviation of the guiding function, and improving the success rate of finding the retrosynthetic path for the molecular retrosynthetic planning task within a specified time.

[0050] In the high-efficiency system for multi-step retrosynthetic planning of chemical organic molecules in the embodiments of the present invention, by using path consistency as the training constraint of the synthesis cost model, the generalization ability of the prediction model is improved, thereby better guiding the search and further improving the success rate and efficiency of the search. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] By reading the following detailed description of the non-limiting embodiments with reference to the accompanying drawings, other features, objects, and advantages of the present invention will become more apparent:

[0052] Figure 1 is a schematic structural diagram of a high-efficiency system for multi-step retrosynthetic planning of chemical organic molecules in an embodiment of the present invention;

[0053] Figure 2 is a schematic structural diagram of a search tree in a preferred embodiment of the present invention;

[0054] Figure 3 is a framework diagram of the MEEA * search module in a preferred embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several modifications and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0056] The present invention provides an embodiment, providing a high-efficiency system for multi-step retrosynthetic planning of chemical organic molecules, as Figure 1 shown, which is a schematic structural diagram of the system, including:

[0057] One-step retrosynthesis prediction module: Taking the molecule to be synthesized as input, generating chemical reactions and reactants for synthesizing the said molecule to be synthesized as input;

[0058] Molecular synthesis cost estimation module: Estimating the cost required for synthesizing using raw material molecules for a single input molecule; Estimating the overall synthesis cost for multiple input molecules simultaneously; where the existing materials available for synthesizing the target molecule are raw material molecules, and if the reactants obtained from one-step retrosynthesis are not raw material molecules, then this molecule needs to be further synthesized;

[0059] Search module: Regarding the multi-step retrosynthesis planning problem as a search problem in the said one-step retrosynthesis prediction module, combining the exploration attribute of Monte Carlo tree search with A * search, and performing retrosynthesis planning based on the said molecular synthesis cost estimation module;

[0060] Training module: Taking path consistency as the constraint for training the said molecular synthesis cost estimation module to improve the generalization ability of the model.

[0061] In the present invention, chemical molecules are represented by Simplified Molecular-Input Line-Entry System (SMILES), and are converted into 2048-dimensional Morgan molecular fingerprint vectors as the input for the one-step retrosynthesis prediction module and the molecular synthesis cost estimation module.

[0062] In the above embodiment, the Monte Carlo tree search is used as the pre-search for A * search by the search module, introducing exploration for A * search, enabling the search process to effectively escape from local optimal solutions, reducing the additional consumption caused by the deviation of the guiding function, and improving the success rate of finding the retrosynthesis path for the molecular retrosynthesis planning task within the specified time.

[0063] In a preferred embodiment of the present invention, a preferred structure of the one-step retrosynthesis prediction module is provided. The one-step retrosynthesis prediction model uses a multi-classification problem for chemical reaction templates. Chemical reaction templates are a form of representation that describes the general or typical transformation patterns in chemical reactions. They are a structured expression for describing the reactants, products, and their transformation relationships in chemical reactions. Given an input product, the corresponding reactants can be inferred according to the reaction template. The chemical reaction templates are extracted from the chemical reactions collected in the United States Patent and Trademark Office (USPTO) dataset using the rdkit library. The products of the chemical reactions are used as the input of the one-step retrosynthesis prediction model, which is converted into a 2048-dimensional Morgan molecular fingerprint representation, and the corresponding chemical reaction templates are represented in one-hot form as the output labels, and used as a classification task to learn to select the templates that can be used to synthesize the input reactants from the existing chemical reaction templates. Since the reactants can be synthesized by multiple chemical reaction templates, this task is a multi-classification problem.

[0064] The one-step retrosynthesis prediction model is a neural network that includes two fully connected layers. The input is the molecule to be synthesized, and the output is a number of chemical reactions and reactants corresponding to the synthesis of the molecule to be synthesized. Specifically, the target molecule is used as the input, and the selection probability of each chemical reaction template is output. The 50 chemical reaction templates with the highest probabilities are selected and applied to the input molecule to obtain a number of chemical reactions that may generate the target molecule. The set of reactants for all chemical reactions is used as the output of the one-step retrosynthesis prediction model.

[0065] Furthermore, the neural network of the one-step retrosynthesis prediction model also includes a hidden layer with a scale of 1024 dimensions. During training, the cross-entropy is used as the loss function to update the network parameters. Specifically,

[0066]

[0067] where y i represents the i-th template as the training label, and p i is the probability predicted by the one-step retrosynthesis prediction module that the i-th template is selected, and N is the number of training samples.

[0068] In a preferred embodiment of the present invention, a preferred structure of the molecular synthesis cost estimation module is provided. This module takes a molecule as the input and estimates the cost of synthesizing the molecule. Specifically, for the input molecule, the USPTO dataset contains the corresponding synthesis path. For each chemical reaction r i in the path, the one-step retrosynthesis prediction model is used to calculate the probability Pr(r i ) of this chemical reaction occurring. The cost of this chemical reaction occurring is represented by c i =-logPr(r i ), and the synthesis cost of the input molecule is the sum of the costs of all chemical reactions occurring in the synthesis path.

[0069] The above is the estimation of the synthesis cost of a single molecule. For multiple molecules, the molecular synthesis cost estimation module takes multiple molecules as the input at the same time. Each molecule passes through a fully connected network with a scale of 128 to obtain the feature expression of each molecule. The elements at the corresponding positions of the feature expressions of all molecules are summed to obtain the global expression of multiple molecules. This global expression is used as the output of a neural network to estimate the synthesis cost.

[0070] In a preferred embodiment of the present invention, a preferred structure of the search module is provided, which is a search module guided by the one-step retrosynthesis prediction module and the molecular synthesis cost estimation module. First, the composition of the search tree in this module is introduced, as Figure 2As shown, the root node of the search tree is the target molecule. The node state s in the search tree is the set of molecules required to synthesize the target molecule currently. The edges are the chemical reactions a that connect the parent node and the child node. Each edge has two attributes, namely the traversal count N(s, a) and the synthesis cost Q(s, a) of this synthesis path. The nodes in the search tree are divided into OPEN nodes and CLOSED nodes. Among them, the leaf nodes that can be further expanded are OPEN nodes, and the nodes that have been expanded are CLOSED nodes.

[0071] Furthermore, the execution process of this search module mainly includes the following steps:

[0072] (1) Simulation: By performing K Monte Carlo tree search simulations, without expanding nodes, a set of candidate nodes is collected.

[0073] (2) Selection: Select the node with the smallest estimated synthesis cost from the set of candidate nodes as the next node to be expanded in the search tree.

[0074] (3) Expansion: Use the first non - raw material molecule in the molecule set included in the node to be expanded that is selected (the user pre - sets a feasible set of raw material molecules according to their own existing synthesis raw materials) as the input of a one - step retrosynthesis prediction model, and obtain several chemical reactions that can be used to synthesize this molecule, thereby generating several child nodes of the expanded node; directly add these child nodes to the search tree. The attributes N(s, a) and Q(s, a) of the newly added edges are initialized to 0, representing the access count and the average cost respectively. If there is a newly added node state that only contains raw material molecules, a synthesis path of the target molecule is successfully found.

[0075] Iteratively execute the above three steps until the time limit is exceeded, or all nodes in the search tree become CLOSED nodes and there are no OPEN nodes that can be further expanded.

[0076] In a preferred embodiment, two preferred steps for the simulation process are provided:

[0077] The first step is to traverse from the root node to the leaf node according to the tree policy: To further promote exploration, a variant of pUCT is adopted, and a uniform distribution is used as the prior policy. The next action is selected in the following way:

[0078]

[0079] where c puct is a hyperparameter used to control the exploration degree of the simulation, is the set of all possible chemical reactions that can be used as legal actions given the current state s. a and b are chemical reactions, N(s, a) and N(s, b) are the number of traversals, and Q(s, a) is the synthesis cost. Starting from the root node, the above tree policy is applied until a leaf node that can be used as a candidate is found.

[0080] The second step is to update the synthesis cost f of the candidate node selected by the simulation to all nodes on the path from the root node to this candidate node:

[0081]

[0082] In a preferred embodiment of the present invention, path consistency is proposed. Path Consistency is derived from the path optimality principle in classical A* search, that is, the estimated value of the cost spent on an optimal path should be the same. CNneim-A relies on this optimality for forward search and names this condition path consistency. Subsequently, path consistency was proposed to be combined with deep reinforcement learning to improve the learning efficiency by adding a weighted penalty term to the loss function as follows: * where λ is a hyperparameter.

[0083]

[0084] is the loss function in reinforcement learning, is the path consistency loss.

[0085] In a preferred embodiment of the present invention, the one-step inverse synthesis prediction module and the molecular synthesis cost estimation module are trained by defining an update mechanism; among them, the policy function of the one-step inverse synthesis prediction module uses cross-entropy as the loss function,

[0086]

[0087] where y i represents the i-th template as the training label, and p i is the probability that the i-th template is predicted to be selected by the one-step inverse synthesis prediction module, and N is the number of training samples.

[0088] The loss function of the molecular synthesis cost estimation module consists of two parts. First is the minimum error loss function of the predicted cost,

[0089]

[0090] where f i is the prediction of the molecular synthesis cost estimation template, and z i is the molecular synthesis cost label.

[0091] Next is the constraint of path consistency, which is represented by the mean square error between the cost estimate value of any node on the path and the mean cost estimate of all nodes on the path. The learning objective is to synthesize the average value of cost evaluation along the optimal path

[0092]

[0093] where is the length of f i and the mean cost estimate of all nodes on the synthesis path. By minimizing the loss function,

[0094] L = L p + αL c + βL pc ,

[0095] where α and β are hyperparameters for adjusting the weights of the loss function, and the parameters of the model are updated using Adam gradient descent.

[0096] The above embodiments improve the generalization ability of the prediction model by using path consistency as the training constraint of the synthesis cost model, thus better guiding the search and further improving the success rate and efficiency of the search.

[0097] Taking the synthesis of CC(=O)NC[C@H]1CN(c2ccc3c(c2)CCCc2c(C(C)C)n[nH]c2-3)C(=O)O1 molecule as an example, in the synthesis path sought by this method, the molecule can be synthesized by amide condensation reaction CC(=O)NC[C@H]1CN(c2ccc3c(c2)CCCC(C(=O)C(C)C)C3=O)C(=O)O1+NN>>CC(=O)NC[C@H]1CN(c2ccc3c(c2)CCCc2c(C(C)C)n[nH]c2-3) C(=O)O1 can be synthesized by amide condensation reaction and acyl chloride substitution reaction to synthesize CC(=O)NC[C@H]1CN(c2ccc3c(c2)CCCC(C(=O)C(C)C)C3=O)C(=O)O1+CC(C)C(=O)Cl>>CC(=O)NC[C@H]1CN(c2ccc3c(c2)CCCC(C(=O)C(C)C)C3=O)C(=O)O1. Among them, the molecule CC(=O)NC[C@H]1CN(c2ccc3c(c2)CCCCC3=O)C(=O)Ol needs to be further synthesized by amide condensation reaction CC(=O)NC[C@H]1CO1+CCOC(=O)Nc1ccc2c(c1)CCCCC2=O>>CC(=O)NC[C@H]1CN(c2ccc3c(c2)CCCCC3=O)C(=O)O1. At this point, all reactants can be synthesized from raw material molecules, and the synthesis path of the target molecule is successfully found.

[0098] It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention. The above preferred features may be used in any combination without conflicting with each other.

Claims

1. An efficient system for multi-step retrosynthetic planning of chemical organic molecules, characterized in that, Including: One-step retrosynthesis prediction module: taking the molecule to be synthesized as input, and generating chemical reactions and reactants for synthesizing the molecule to be synthesized as the input; Molecular synthesis cost estimation module: estimating the cost required for synthesizing using raw material molecules for a single input molecule; estimating the overall synthesis cost for multiple input molecules simultaneously; Search module: Regarding the multi-step retrosynthesis planning problem as a search problem in the one-step retrosynthesis prediction module, combining the exploration property of Monte Carlo tree search with A * search, and performing retrosynthesis planning based on the molecular synthesis cost estimation module; Training module: using path consistency as a constraint for training the molecular synthesis cost estimation module to improve the generalization ability of the model.

2. The efficient system for multi-step retrosynthetic planning of chemical organic molecules according to claim 1, characterized in that, The one-step retrosynthesis prediction module is a neural network including two fully connected layers, and performs a classification task based on chemical reaction templates; The molecule to be synthesized is used as the input molecule of the one-step retrosynthesis prediction module. According to the chemical reaction templates, the probabilities of being able to synthesize the input molecule are obtained, and the N chemical reaction templates with the highest probabilities are selected and applied to the input molecule to obtain several optional chemical reactions and reactants for synthesizing the input molecule as the output.

3. The efficient system for multi-step retrosynthetic planning of chemical organic molecules according to claim 1, characterized in that, The molecular synthesis cost estimation module estimates the cost required for synthesizing using raw material molecules for a single input molecule, including: Obtaining a synthesis path from the USPTO dataset according to a single input molecule; For each chemical reaction r in the synthesis pathway i , use a one-step retrosynthesis prediction model to calculate the probability Pr(r i ) of the occurrence of this chemical reaction, and obtain the cost c i = -logPr(r i ); Adding up the costs of all chemical reactions in the synthesis path to obtain the synthesis cost of the single input molecule.

4. The efficient system for multi-step retrosynthetic planning of chemical organic molecules according to claim 3, characterized in that, The molecular synthesis cost estimation module estimates the overall synthesis cost for multiple input molecules simultaneously, including: Multiple molecules are used as inputs simultaneously; Each molecule passes through a fully connected network with a scale of 128 to obtain the feature expression of each molecule; The elements at the corresponding positions of the feature expressions of all molecules are summed to obtain the global expression of multiple molecules; This global expression is used as an estimate of the synthesis cost output through a neural network.

5. The efficient system for multi-step retrosynthetic planning of chemical organic molecules according to claim 1, characterized in that, In the search module, the root node of the search tree is the target molecule, the nodes of the search tree are sets of molecules, the edges are chemical reactions, the first molecule of the parent node is decomposed into several reactants through this chemical reaction to obtain the set of molecules of the child node, the nodes in the search tree that can be further expanded are OPEN nodes, and the nodes that have been expanded are CLOSED nodes.

6. The efficient system for multi-step retrosynthetic planning of chemical organic molecules according to claim 5, characterized in that, In the search module, combining the exploration attribute of Monte Carlo tree search with A * search, and performing retrosynthesis planning based on the molecular synthesis cost estimation module, including: Simulation: Starting from the root node, using Monte Carlo tree search with an exploratory tree traversal strategy to traverse to OPEN nodes without expanding new nodes, and taking them as candidate nodes, repeating several times to obtain a set of candidate nodes; Selection: According to A * Search for the selection method, and select the node with the lowest synthesis cost in the candidate node set as the next expansion node; Expansion: Using the first non-raw material molecule in the set of molecules included in the state of the selected node as the input of the one-step retrosynthesis prediction model to obtain several chemical reactions that can be used to synthesize this molecule, and generating several child nodes of the expanded node; if all the chemical molecules in a certain child node are existing raw material molecules, a synthesis path is successfully found, otherwise, the synthesis costs of all several child nodes are speculated using the synthesis cost estimation module, and all several child nodes are added to the search tree; Loop: Continuously repeat the simulation, selection, and expansion steps. If the running time exceeds the specified time, or there are no nodes in the search tree that can be expanded, the search fails and no synthesis path is found.

7. The efficient system for multi-step retrosynthetic planning of chemical organic molecules according to claim 6, characterized in that, The simulation process includes: The tree traversal strategy selects the pUCT tree traversal strategy to obtain OPEN nodes. The pUCT formula includes a balance between exploitation and exploration, and the obtained OPEN nodes are exploratory; Update the synthesis cost evaluation backup of the current OPEN node into the synthesis costs of all nodes on the path from the root node to the current OPEN node.

8. The efficient system for multi-step retrosynthetic planning of chemical organic molecules according to claim 1, characterized in that, The path consistency means that in the same optimal path, the synthesis cost estimates of all node states are the same; use the path consistency as the learning constraint of the molecular synthesis cost estimation module.

9. An efficient system for multi-step retrosynthetic planning of chemical organic molecules according to claim 8, characterized in that, The loss function of the molecular synthesis cost estimation module includes two parts: One part is the minimum error loss function of the predicted cost, where f i is the prediction of the molecular synthesis cost estimation template, and z i is the molecular synthesis cost label; The other part is the constraint of the path consistency, which is represented by the squared error between the cost estimate value of any node on the path and the mean of the cost estimates of all nodes on the path. Among them is f i The average value of the cost estimates of all nodes on the synthesis path with length n where it is located.

10. An efficient system for multi-step retrosynthetic planning of chemical organic molecules according to claim 9, characterized in that, The policy function of the one-step retrosynthesis prediction module uses cross-entropy as the loss function. Among them, y i represents the i-th template as the training label, and p i is the probability predicted by the one-step retrosynthesis prediction module that the i-th template is selected, and N is the number of training samples; By minimizing the loss function \(L = L p +\alpha L c +\beta L pc , the parameters of the model are updated using Adam gradient descent, where \(\alpha\) and \(\beta\) are hyperparameters that adjust the weights of the loss function.

Images